Xây dựng giải pháp chẩn đoán sự cố trong máy biến áp 3 pha sử dụng các phương pháp xử lý tín hiệu thông minh TT TIENG ANH

i TABLE OF CONTENTS PREAMBLE 1 THE URGENCY OF THE SUBJECT RESEARCH PURPOSES RESEARCH METHODS RESEARCH SUBJECTS RESEARCH SCOPE SCIENTIFIC AND PRACTICAL MEANINGS OF THE TOPIC .1 CHAPTER 1: OVERVIEW OF DIAGNOSTIC METHODS IN TRANSFORMERS 1.1 THE IMPORTANCE OF TRANSFORMERS’ FAULT DIAGNOSIS 1.2 REVIEW OF METHODS FOR TRANSFORMERS’ FAULT DIAGNOSIS 1.2.1 International works .2 1.2.2 Domestic works 1.2.3 The limitations of reviewed diagnosis methods 1.2.4 Thesis proposal 1.3 CONCLUSION of CHAPTER CHAPTER 2: THEORETICAL BASIS OF THE THESIS PROPOSALS 2.1 VIBRATION IN THE TRANSFORMER 2.1.1 Vibration of winding 2.1.2 Vibration of steel core .2 2.2 THE NEED TO ANALYZE THE TRANSFORMER VIBRATIONS 2.3 FREQUENCY DOMAIN ANALYSIS OF THE VIBRATIONS 2.3.1 The fundamentals of frequency response analysis .3 2.3.2 Application scope of the method 2.3.3 Review on the method of vibration analysis in the frequency domain 2.4 THE FINITE ELEMENTs METHOD 2.4.1 A general introduction to the finite elements method 2.4.2 Operation flow chart using the finite elements method 2.4.3 Generalized system of Maxwell equations for the electromagnetic field 2.5 APPLICATION OF FINITE ELEMENTS METHOD IN ANSYS MAXWELL SOFTWARE TO IMPLEMENT A MODEL OF TRANSFORMER 2.5.1 Equation of electromagnetic field 2.5.2 System of mechanical equations 2.6 NEUTRAL NETWORK MLP .7 2.6.1 The architecture of the MLP Neural Network MLP 2.6.2 Learning process for MLP 2.6.3 The steepest descend gradient algorithm .8 2.6.4 Levenberg–Marquardt algorithm for MLP 2.7 CONCLUSION OF CHAPTER ii CHAPTER 3: IMPLEMENTATION OF A MODELS IN ANSYS SOFTWARE FOR DISTRIBUTED TRANSFORMER IN SELECTED FAULTY STATES 3.1 GENERAL INTRODUCTION TO ANSYS SOFTWARE .9 3.1.1 Main modules of ANSYS software .9 3.1.2 ANSYS Maxwell as electromagnetic simulation function block 3.1.3 Structural simulation function block using ANSYS Structure .9 3.1.4 ANSYS design modeler and ANSYS meshing 3.1.5 ANSYS mechanical workbench 3.1.6 ANSYS mechanical 3.2 IMPLEMENTING A 400KVA 22-0.4KV Y-Y0 DISTRIBUTION TRANSFORMER MODEL IN ANSYS 10 3.2.1 Working principle of a transformer 10 3.2.2 Implementing a 400kVA 22-0.4kV Y-Y0 distribution transformer model 10 3.3 MODELS FOR SIMULATION OF NORMAL AND FAULTY STATES OF THE DISTRIBUTION TRANSFORMER 11 3.3.1 The mesh for the model of transformer in normal state 11 3.3.2 The mesh for the model of transformer with a loosen coil 11 3.3.3 The meshes for the model of transformer with shortages: 2-turn, 5% of total rounds, 10% of total rounds of the high-voltage winding, phase B 11 3.3.4 The mesh for the model of transformer with one coil fixing bolt loosen 12 3.4 CONCLUSION OF CHAPTER 12 CHAPTER 4: NUMERICAL RESULTS OF SIMULATION AND EXPERIMENT 12 4.1 DATA SETS FROM SIMULATION IN ANSYS SOFTWARE 12 4.1.1 Normal operation of MBA, 50% load (case A-1)` 12 4.1.2 Case of short-circuit of two high-voltage rounds 14 4.1.3 Case of wire loop loosening problem 14 4.1.4 Case of loosening of coil fixing bolt 15 4.1.7 Review of the simulation results 15 4.2 THE RESULTS OF TRAINING OF THE MLP NETWORK 16 4.2.1 The features extracted from the simulation data 16 4.2.2 Results training network MLP 17 4.3 PRACTICAL EXPERIMENT ON ACTUAL DISTRIBUTION TRANSFORMER 20 4.4 CONCLUSION OF CHAPTER 23 CONCLUSIONS AND RECOMMENDATIONS 23 PUBLICATIONS 24 PREAMBLE THE URGENCY OF THE SUBJECT During operation, transformers may encounter various problems such as insulation failure between turns of wire, short circuit, broken wire, earth fault, equipment malfunction or user's fault, overload condition and aging of equipment, When a fault occurs in a transformer, relay protection will act to separate the faulty element from the electrical system and eliminate the influence of the fault elements Diagnosing the fault type in a 3-phase transformer is an urgent problem to help to detect and troubleshoot a very important device in the power system The successful development of a solution to diagnose potential problems in transformers in general and 22/0.4kV distribution transformers in particular will have good practical significance, if put into application, it will help operators to recognize early transformer failures thereby avoiding economic losses due to repair or replacement of new transformers, as well as improving power supply continuity RESEARCH PURPOSES The thesis researches and provides solutions for fault diagnosis in 22/0.4kV 3-phase distribution transformers ANSYS software is used to implement a 22/0.4kV distributed transformers model The signals features are then processed by an MLP neural network trained with Levenberg - Marquadrt learning algorithm to diagnose potential fault types RESEARCH METHODS Studying research works on ANSYS software and signal processing to build a 22/0.4kV distribution transformer model in normal and fault working states Simulate the transformer in normal working state and fault cases in ANSYS to generate samples of electrical signals and mechanical vibrations These signals will be analyzed and feature parameters extracted to train the recognition models using MLP neural networks to detect the types of potential failure in the transformers The training algorithm was implemented with the Levenberg - Marquadrt algorithm and the Neural Network Toolbox library in Matlab Verify transformers model built on ANSYS software by experiment with device using accelerometer to measure vibration signal of transformer in normal working mode when load changes RESEARCH SUBJECTS The research object of the thesis is fault diagnosis of three-phase 400kVA 22-0.4kV Y-Y0 distribution transformer to improve the efficiency of the power system RESEARCH SCOPE Application of ANSYS software to implement fault models of three-phase 400kVA 22-0.4kV Y-Y0 distribution transformer (2-turn shortage in phase, shortage of 5% of total turns in phase, short of 10% of total turns in phase, loosen coils in one-phase winding, loosen winding mounting bolts) The generated signals from the simulations in ANSYS (electrical, force, and mechanical signals) will be used to generate sample signals for fault identification Selecting and building a recognition algorithm using MLP neural network to diagnose problems in distributed transformers Experiment using accelerometer to measure vibration on real transformer in normal working mode when load changes to verify the proposed approach SCIENTIFIC AND PRACTICAL MEANINGS OF THE TOPIC  Scientific significance : Propose a recognition algorithm using MLP neural network with simultaneous use of electrical and mechanical signals (vibrations) to diagnose potential problems in distributed transformers  The practical significance of the topic : - The thesis contributes to early prediction of potential problems that may occur for distribution transformers in order to improve the efficiency of power system operation - - The research results of the thesis are reference materials for students majoring in control and automation, master's students and graduate students interested in research on transformers fault diagnosis issues CHAPTER 1: OVERVIEW OF DIAGNOSTIC METHODS IN TRANSFORMERS 1.1 THE IMPORTANCE OF TRANSFORMERS’ FAULT DIAGNOSIS 1.2 REVIEW OF METHODS FOR TRANSFORMERS’ FAULT DIAGNOSIS 1.2.1 International works 1.2.2 Domestic works 1.2.3 The limitations of reviewed diagnosis methods 1.2.4 Thesis proposal 1.3 CONCLUSION of CHAPTER Chapter of the thesis has solved the following issues :  Synthesize domestic and international studies on potential fault diagnosis methods in transmission and distribution transformers  Discussed the limitations of the published methods on transformers fault diagnosis  Proposing a solution to diagnose distribution transformer faults by building an transformer model in ANSYS software to generate the electrical, mechanical (vibration) signals as a data set for identifying distributed transformer faults by MLP artificial neural network CHAPTER 2: THEORETICAL BASIS OF THE THESIS PROPOSALS 2.1 VIBRATION IN THE TRANSFORMER Vibration in transformers is caused by various forces present in the steel core and windings inside the transformer during operation 2.1.1 Vibration of winding The vibration of the windings caused by the electromagnetic forces when there is a current flowing in the coils 2.1.2 Vibration of steel core The vibration of the steel core is caused by a phenomenon called magnetostriction, which is the phenomenon when metal Figure 2.1: Magnetic circuit objects undergo a deformation in their shape when placed in a and transformer windings magnetic field Inside the transformer, the steel core, which is made in the form of laminated plates, also experiences expansion and contraction due to flux changes This expansion and contraction occurs twice in an alternating cycle 2.2 THE NEED TO ANALYZE THE TRANSFORMER VIBRATIONS 2.3 FREQUENCY DOMAIN ANALYSIS OF THE VIBRATIONS 2.3.1 The fundamentals of frequency response analysis The transformer is considered a complex network of RLC elements The contributions to this RLC complex network come from the resistance of the copper coil; the inductance of the windings and the capacitance coming from the insulating layers among the windings, between the winding and the winding, between the winding and the steel core, between the steel core and case, between the case and the winding Figure 2.2: Simplified isoval circuit However, we can use a simplified isotropic circuit with pooled RLC elements with the aggregated RLC elements (illustrated in Figure 2.2) to explain accurately the principle of frequency response technique The frequency response is carried out by applying a low voltage signal with variable frequencies into the windings of transformer and measuring both input and output signals The ratio of these two signals gives us the required response This ratio is called the transfer function of the transformer So we can obtain values of its magnitude and phase angle With different frequencies, the RLC network will give different impedance circuits Therefore, the transmission function at each frequency is a unit of measurement of the actual impedance of RLC network of the transformer 2.3.2 Application scope of the method Currently, in order to detect the displacement of the transformer windings, the maintenance units of the transformer use FRA measuring devices which are considered as a diagnostic tool to assist in the testing of damage assessment and fault investigation in MBAs The FRA technique has proven to be a powerful tool in terms of means to reliably and efficiently detect winding displacements and other failures that affect the impedance of the transformer 2.3.3 Review on the method of vibration analysis in the frequency domain 2.4 THE FINITE ELEMENTs METHOD 2.4.1 A general introduction to the finite elements method 2.4.2 Operation flow chart using the finite elements method 2.4.3 Generalized system of Maxwell equations for the electromagnetic field Table 2.1: System of Maxwell's equations Name Faraday's Law Ampere's Law Ampere's Gauss Ampere's Gauss (For magnetic field) Differential form    B .E  t     D .H  J  t   .D     .B  Integral form   d   Edl  BdA c dt  s     d   Hdl  JdA  DdA c s dt  s   Dd   A    dV s v   Bd   A  s The analysis and calculation of factors in the electric and magnetic fields can be based on the system of Maxwell's equations, The variable magnetic field generates an induced electric field and vice versa Electric and magnetic fields are closely related and transform each other The concept of the electromagnetic field was first stated by Maxwell (so they are now called Maxwell's equations):      D rot H  J  t     B rot E   t   divB    divD  (2.1)   In the magnetic material environment, the relationship between B and H according to the magnetic permeability coefficient of materials  are  as follows: B= H (2.2) In the SI system of units, the above quantities have the following units and dimensions:  Magnetic field strength vector A/m H  Magnetic induction vector T = kg/s2.A B  Current density vector A/m2 J  Magnetic permeability coefficient H/m of the material  Electromagnetic induction vector C/m2 D  Electric field strength vector V/m E Napla operator:        i j k x  y z (2.3) In the coordinate system Descartes:   A Ay A  A  div A  x   z x y z    i j k      rot A  x A  x y z Ax Ay Az   Equation rot H  J  (2.4) (2.5) D is equivalent to a system of three algebraic equations t  H z    y  H x    z  H y    x H y z  jx  Dx t Dy H z  jy  x t H x D  jz  z y t For ferromagnetic materials μ is a tensor (2.6)   xx  xy  xz        yx  yy  xz  (2.7)   zx  zy  zz    2.5 APPLICATION OF FINITE ELEMENTS METHOD IN ANSYS MAXWELL SOFTWARE TO IMPLEMENT A MODEL OF TRANSFORMER 2.5.1 Equation of electromagnetic field The problems of electromagnetic fields can be divided into different forms, each of which has a corresponding Maxwell equations system for its solving: two cases of steady-state systems (a stationary state, in which there is no variation of any quantity, and a steady state where all the physical properties are cyclical) and the transient state When the system is in the transiting state from one steady state to another, the temporal factor associated with time will be included as the basis for determining the instantaneous state of the system On that basis, in order to simplify the problem in implementing the finite element methods, in this thesis, we discussed the build of the characteristic equations for the elements according to the above three basic states The systems of equations at the element nodes apply to specific analytical models are later used in ANSYS software The static electromagnetic states are used only for models where only the static magnetic field formed by permanent magnets, electromagnets in different media in 3D space is present The electromagnetic equations are given by the formulas    H  J   B  (2.8)      B  0 ( H  M )  0  r  H  0  M p The 2D static magnetic state is used for models where only the static magnetic field formed by permanent magnets, electromagnets in different media is defined for 2D space This case applied to problems with circular symmetric geometric structure or when the size of one dimension is much larger than the other two dimensions, then the magnetic field derivative in one direction is zero The equation of the spatial variable is determined by:       J z ( x, y )     (  Az ( x, y ))  (2.9)  0  r  The 3D sinusoidal variable magnetic field state applies to the class of problems on electromagnetism in the magnetic field state generated by harmonic varying power sources, surface effects due to the combination of magnetic fields harmonic variation and harmonic variation current caused inside the conductor The equation of the spatial variable at the nodes is determined by:      H   j H    j  (2.10) 3D time-varying electromagnetic state applies to the class of problems with time-varying magnetic fields, currents in 3D space due to variable power sources or object movement Then the system of spatial equations at the nodes of the elements is determined by:     B   H  0  t   .B          .    .( )  t   (2.11) The 2D time-varying electromagnetic state is similar to the 3D electric field problem, but the derivative of the magnetic field and the current in a certain direction has zero value Then the spatial equation at the element node is zero determined by:           A   v  A  J s    V    H c  V    A t (2.12) State of fixed charges: Consider the class of problems about electric field distribution in 3D space without time variation The spatial equation of the system is established by the system: Apply to 2D model analysis:   .( r   )    v (2.13) Apply to 3D model analysis:   .( r   ( x, y ))    (2.14) Direct current: applied to a class of problems on analyzing conductive currents whose magnitude and direction not change with time The equation of the spatial variable at the element nodes is determined by: For 3D model analysis    J ( x, y )   E ( x, y )   ( x, y ) (2.15) For to 2D model analysis:   .( )  (2.16) Harmonic variable current: Applied to the class of problems of analyzing the amperage flow in the conductor of the harmonic variation system Equations of spatial variables at element nodes are determined by (applicable only to 2D problem class)    (2.17) .E  j  ( x, y )   Current varies with time: Applied to the problem model with time-varying amperage in 3D space, then the spatial equation at element nodes is determined by the formula         .    .( )  t   (2.18) Method of calculating electromagnetic forces in Maxwell software: according to Lorentz's law of electromagnetic force, the software defines a quantity called Maxwell's force tensor by the formula:  B H  H x  By H x  Bz  H x  Bx     B H   (2.19)   H y  Bx H y  By  H y  Bz    B H   H z  Bx H z  By H z  Bz     Accordingly, the electromagnetic force is calculated according to the formula: dF    dA (2.20) 2.5.2 System of mechanical equations 2.5.3 Linking the electromagnetic field problem and mechanical problem 2.6 NEUTRAL NETWORK MLP 2.6.1 The architecture of the MLP Neural Network MLP MLP (MultiLayer Perceptron) network is a feedforward network built from the basic elements of McCulloch-Pitts neurons, in which neurons are arranged into layers consisting of a layer of input signal channels (input layer), a layer of output signal channels (output layer), and a number of intermediate layers known as hidden layers Figure 2.4 is a model of an MLP network with N inputs, one hidden layer with M neurons and K outputs We generally denote the concatenation weights between the input layer and the hidden layer as Wij ( i   M ; j   N ), denote the coupling weights between the hidden layer and the output layer as Vij ( i   K ; j   M ) The transfer functions of the hidden and output layer neurons are denoted f1 and f2, respectively In each model, the authors can choose different transfer functions Figure 2.4: MLP network model with according to experience and purpose The hidden layer commonly used functional forms are [37]: - Function logsig: logsig ( x)   e x  e x - Function tansig: tansig ( x)   e x - Linear function: linear ( x)  a  x  b In this thesis, the transfer function of hidden neurons is selected as the function f1 ( x )  tansig ( x) since this function has a range of values including both positive and negative values, it is more general than the function logsig(), also the nonlinear function will be more general than the linear function; The transfer function of the output neurons is the function f ( x)  linear ( x ) because this function can generate values greater than (because the thesis will use the status code of the transformer from d=0 to d=5) Then, with the input vector x   x1, x2 ,, xN    N (fixed bias input x0  ), the output is determined sequentially in the forward propagation direction as follows:  Total input excitation of the i-th hidden neuron i ( i   M ) equals: M ui   x j  Wij (2.21) j 0  Calculate the output of the i-th hidden neuron ( i   M ): vi  f1  ui   Total input excitations of the i-th output neuron i ( i   K ) (2.22) M gi   v jVij (2.23) j 0  And finally the i-th output of the network will be ( i   K ): yi  f2  gi  (2.24) In total, the transfer function of the MLP network is a nonlinear function given the following dependency formula: M  yi  f  gi   f   v jVij   j 0    M N   N       f  Vi   f1 u j Vij   f  Vi    f1  W j   xkW jk Vij        j 1 j 1  k 1       (2.25)   2.6.2 Learning process for MLP MLP networks are usually trained using supervised learning algorithms, i.e algorithms that train when there are samples that include both inputs and outputs, respectively With the sample data set being a set of p pairs of samples given in the form of input vectors – output vectors respectively xi , di  với i   p, xi  R N ; di  RK , we need to find an MLP network (including the determination of the structure parameters and the coupling weights corresponding to the selected structure) such that when given the vector xi into the MLP network, the output of the network will approximate the existing target value: i : MLP xi   di (2.26) Or the total error on the samples approaches some minimum value or is less than a pre-selected threshold   : E p  MLP  xi   d i i 1  (2.27) 2.6.3 The steepest descend gradient algorithm 2.6.4 Levenberg–Marquardt algorithm for MLP Algorithms that use gradients (first derivatives) have slow convergence When we need to improve the convergence speed, we can use the Levenberg–Marquardt (L–M) algorithm This algorithm is based on Taylor to quadratic expansion Considering the error according to formula (2.27) which is a function that depends on all the weights of the neurons, then we expand the function E in the neighborhood of the current weights W will be: T E (W  p)  E (W )   g (W )  p  with p – the increement amount, T p H (W ) p  O p3    E E E  g( W )  E   , , ,  Wn   W1 W2 (2.28) T is the gradient vector of the function E with regards to all the weights (grouped in the matrix W), and H is the symmetric square matrix of the second derivatives (also called the Hessian matrix) of E with respect to W with: 12 LabelID=IU_LA + 0.78ohm R22 LWinding_LA 0.78ohm R25 LWinding_LB 0 22000*sqrt(2) V LabelID=VV_HB LWinding_HB LabelID=IU_LC + 0.78ohm R28 5ohm R8 LabelID=IU_LB + 22000*sqrt(2) V LabelID=VV_HA LWinding_HA LWinding_LC 22000*sqrt(2) V LabelID=VV_HC LWinding_HC (5/2089*2) ohm R44 LWinding_T 5ohm R11 5ohm R14 Figure 3.4: Circuit diagram for transformer in case of short circuit of high voltage line phase B 3.3.4 The mesh for the model of transformer with one coil fixing bolt loosen 3.4 CONCLUSION OF CHAPTER Chapter of the thesis presented the following issues: - Simulate the transformer failures to generate the sample of electrical signals and mechanical vibrations (mechanical signals), - The thesis used ANSYS software to implement a distributed transformer model of 400kVA, 22-0.4kV, Y- Y0, - Boundary conditions, excitation conditions for the coils had been defined and set up for the simulation of the distributed transformer model 400kVA, 22-0.4kV, Y-Y0 and the original model was modified to support different scenarios: normal working state and 05 faulty cases The data generated from the simulations are later used to build up a fault detection model CHAPTER 4: NUMERICAL RESULTS OF SIMULATION AND EXPERIMENT 4.1 DATA SETS FROM SIMULATION IN ANSYS SOFTWARE As described in Chapter 3, a model of 400kVA, 22-0.4kV, Y-Y0 distributed transformer model has been implemented in ANSYS software along with its modifications to simulate different faulty states For each scenario of faults, simulations were performed for different load levels of 50%, 80% and 100% of the norminal load Also the simulations were performed with 10 different initial phase values to enrich the training samples data set As results, a total of   (10  3)  234 simulations had been done 4.1.1 Normal operation of MBA, 50% load (case A-1)` 4.1.1.1 Calculation results over time of the force components of the coils and cores The results are given as the graph of Figure 4.1, Figure 4.2 and Figure 4.3 Figure 4.1: The two-end tensile force component in the radial direction of the coil HA, HB, HC 13 Figure 4.2: The two-end tensile force components in the radial direction of windings LA, LB, LC Figure 4.3: Graph of force in directions x, y, z acting on core 4.1.1.2 Displacement response analysis in frequency domain  Displacement analysis results in the x direction of the case: Figure 4.4: Displacement in the x-direction of the case  Displacement analysis results in the y direction of the case Figure 4.5: Displacement in the y-direction of the case 14  Displacement analysis results in the z direction of the case Figure 4.6: Displacement in the z-direction of the case 4.1.2 Case of short-circuit of two high-voltage rounds  Displacement analysis results in the x direction of the case Figure 4.7: Displacement in the x-direction of the case  Displacement analysis results in the y direction of the case Figure 4.8: Displacement in the y-direction of the case  Displacement analysis results in the z direction of the case Figure 4.9: Displacement in the z-direction of the case The maximum of displacement was 0.18406mm at the frequency of 50Hz 4.1.3 Case of wire loop loosening problem  Displacement analysis results in the x direction of the case 15 Figure 4.10: Displacement in the x-direction of the case  Displacement analysis results in the y direction of the case Figure 4.11: Displacement in the y-direction of the case  Displacement analysis results in the z direction of the case Figure 4.12: Displacement in the z-direction of the case 4.1.4 Case of loosening of coil fixing bolt 4.1.7 Review of the simulation results Through the simulation results, it is found that there is a difference between the electrical characteristics and mechanical vibration, the simulation results of the working modes of the transformer are as follows  Case of 50% nominal load Displacement (vibration amplitude) of Normal the transformer shell 6,4.10-5mm Displacement in at 115Hz the x-dir Displacement in 4,447 10-4 mm at 50Hz the y-dir Displacement in 6,14.10-3mm at 50Hz the z-dir Shorted two turns of high voltage wire Loosening of Shorted 5% of Loose problem high-voltage high voltage of coil fixing bolt coil windings wire loop 0,00146mm at 120Hz 0,0134mm at 50Hz 0,18406mm at 50Hz 6,4.10-5mm at 115Hz 4,447 10-4 mm at 50Hz 6,14 10-3mm at 50Hz 0,0036mm at 195Hz 0,00096mm at 195Hz 0,00054mm at 195Hz 0,0517 mm at 120Hz 0,4823mm at 50Hz 6,6003mm at 50hz Shorted 10% of high-voltage wire loop 0,1935mm at 120Hz 0,74037mm at 50Hz 10,183mm at 50Hz 16  Case of 80% nominal load Displacement (vibration amplitude) of the transformer shell Displacement in the x-dir Displacement in the y-dir Displacement in the z-dir Normal Shorted two turns of high voltage wire Loosening of high-voltage coil windings Loose problem of coil fixing bolt Shorted 5% of high voltage wire loop Shorted 10% of high-voltage wire loop 6,8 10-5 mm at 115Hz 4,53 10-4 mm at 120Hz 6,27 10-3mm at 50Hz 0,775mm at 125Hz 9,98mm at 50Hz 136mm at 50Hz 0,68mm at 115Hz 0,354 mm at 120Hz 60,27mm at 50Hz 89,6mm at 195Hz 21,0 mm at 80Hz 5,79mm at 80Hz 586 mm at 115Hz 325mm at 115Hz 627 mm at 40hz 409mm at 125Hz 316mm at 115Hz 654mm at 40Hz  Case of 100% nominal load Displacement (vibration amplitude) of the transformer shell Displacement in the x-dir Displacement in the y-dir Displacement in the z-dir Normal Shorted two turns of high voltage wire 7,2.10-5mm 2,02.10-3mm 7,14.10-5mm at 115Hz at 50Hz at 115Hz -4 Loosening of Loose problem high-voltage of coil fixing coil windings bolt -4 -4 Shorted 5% of high voltage wire loop Shorted 10% of high-voltage wire loop 1,35.10-2mm 2,68.10-2mm 7,5.10-2mm at 195Hz at 115Hz at 135Hz -3 4,61.10 mm 1,86.10 mm 4,61.10 mm 3,6.10 mm 0,19555mm 1,01mm at 50Hz at 125Hz at 50Hz at 195Hz at 50Hz at 50Hz 6,38.10-3mm 2,78.10-2mm 6,37.10-3mm at 2,02.10-3mm 2,6895mm 13,884mm at 50Hz at 50Hz 50Hz at 195Hz at 50Hz at 50Hz 4.2 THE RESULTS OF TRAINING OF THE MLP NETWORK 4.2.1 The features extracted from the simulation data In the thesis, it is proposed to use the characteristics of the signals to classify the status of the transformer extracted from the above measured signals as follows:  From the frequency spectrum of the displacement oscillations on the transformer housing in axes: use the maximum value of the spectrum on each axis, or we have: x1  max  25,30,,200  M x    ; x2  max   25,30,,200 max  M z     M y   ; x3  25,30, ,200  From the variable value of the force acting on the axes: using the maximum value of the force on each axis (for phase B, the high-voltage side, the phase used in the simulation is the phase where the failure occurs), or we have: x4  max t 0,100 ms  F H x  t  ; x5  max t 0,100 ms  For the low voltage side we have: x7  max FxL  t  ; x8  max t 0,100 ms    t ,100 ms F F  H y L y  t   ; x6  t max  FzH  t  ; 0,100 ms   max  FzL  t   ;  t   ; x9  t 0,100 ms    From the variable value of the force acting on the radial: using the maximum value of the force on each axis, select the wire bundle of phase B, the high-voltage side (the phase used in the simulation is the phase where the fault occurs) ), or we have: 17 x10  max t 0,100 ms   Fcx  t   ; x11  max t 0,100 ms   Fcy  t   ; x12  max t 0,100 ms   Fcz  t   ;  From the variable value of phase B current amplitude on high voltage and low voltage side, phase B voltage on low voltage side, using the maximum value of the signals, or we have: x13  max  I H  t   ; x14  max  I L  t   , x15  max U L  t   ; t 0,100 ms  t 0,100 ms  t 0,100 ms  Thus, the input feature vector consists of 15 components The output of the identification system is the status code of the MBA, which consists of states considered as above:  MBA in normal mode  The transformer has a loosen coil fixing bolts  The transformer has a loosen coil loop around the shaft  The transformer has shorted consecutive turns of wire (at the B phase winding, high voltage side)  The transformer has a shorted of 5% of the total turns of wire (at phase B winding, high voltage side)  The transformer has a shorted of 10% of the total turns of wire (at phase B winding, high voltage side) 4.2.2 Results training network MLP The thesis uses the Levenberg - Marquadrt learning algorithm and the Neural Network Toolbox library in Matlab to perform the computations The test results with a feature vector of 15 inputs, output and an increasing number of hidden neurons are as follows:  With hidden neuron (a) Learning results with 180 samples (b) Test results with 54 samples Figure 4.13: Simulation results with hidden neuron Figure 4.13a and 4.13b presented the results of training the network with the learning samples set and then testing it with the test samples set The horizontal axis is the ordinal 18 number of the samples Because the training sample set consists of 180 samples, Fig 4.24a has a horizontal axis from to 180, the test sample set includes 54 samples, so Fig 4.24b has a horizontal axis from to 54 Value points '*' (blue) are the exact target values to be achieved for the learning process The 'o' value points (red) are the output values of the MLP network for each xi sample The solid purple line represents the absolute difference between the output of the MLP network and the target value The black dashed line is the y = 0.5 threshold value line for a quick assessment of the correlation between the absolute deviation and the 0.5 threshold value Since the target values have discrete values (d = 1, 2, …, 6), the output value of the network will be reduced to the nearest target value as the identification result For example, if the output value is 2.34 then the identification result will be the case 'd=2' (MBA is loose the coil fixing bolt) With this method, since the target values have a minimum distance of 1, if the difference between the output values is less than 0.5, there will be no identification error The learning results in Figure 4.13 show that the network has a simple structure (15 inputs, hidden neuron, output), so the patterns have not been successfully learned, so there are many error cases In which, many samples in cases and have failed to learn  With hidden neurons (a) Learning results with 180 samples (b) Test results with 54 samples Figure 4.14: Simulation results with hidden neurons Experimental results show that the network still has a too simple structure (15 inputs, hidden neurons, output) so it has not yet successfully learned the patterns, but the number of errors is less than the field merge hidden layer There are still a number of cases (group samples) being mistaken for form When checking with the 54 sample data set (as shown in Figure 4.14b), there are still cases of misclassification 19  With hidden neurons: (a) Learning results with 180 samples (b) Test results with 54 samples Figure 4.15: Simulation results with hidden neuron The learning results show that the network has successfully learned all the samples, all the cases of the learning samples and the test samples have small errors (less than 0.5 threshold)  With hidden neurons: (a) Learning results with 180 samples (b) Test results with 54 samples Figure 4.16: Simulation results with hidden neurons The learning results show that just like with the network with hidden neurons, the network with hidden neurons has successfully learned all the samples, all the learning and test samples have small errors (less than 0.5 threshold)  With hidden neurons: 20 (a) Learning results with 180 samples (b) Test results with 54 samples Figure 4.17: Simulation results with hidden neurons The learning results show that the network with hidden neurons has also successfully learned all the samples, all the learning samples and the test samples have small errors (less than 0.5 threshold) Therefore, it can be seen that the MLP network with 15 inputs, hidden layer with at least hidden neurons and output neuron can successfully learn to accurately identify the states of the MBA simulated in the thesis The smallest network model with hidden neurons is shown in Figure 4.18 below The structure of the neural network selected to recognize the feature vectors extracted from the signals of the transformer is shown in Figure 4.18 Figure 4.18: Structure of neural network with 15 inputs, hidden 4.3 PRACTICAL EXPERIMENT ON ACTUAL DISTRIBUTION TRANSFORMER Vibration measuring equipment is designed, manufactured (details of design and fabrication are as shown in the appendix of the thesis) and installed for testing at the 3rd Industrial University Substation as shown in Figure 4.19 (a) (b) Figure 4.19: Transformer at Industrial University Station (a) and the measuring device mounted on the transformer case (b) 21 It can be seen that the vibration level of the transformer increased gradually between 10am and 12pm After that, it is relatively stable until about pm and increases rapidly to peak at pm Figure 4.20 shows the vibration measurement by accelerometer sensor at Industrial University Station from 9.00 to 18.30 on September 15, 2020 Figure 4.21 shows the results of data collected (remotely) by Thai Nguyen Electricity Company for the experimental transformer station in the same time period However, the current values (measured remotely) are quite limited in that they are only taken every 30 minutes to take the instantaneous value at that time It can be seen that the load variation graph has similar trends with the vibration chart It is an upward trend in the morning and fluctuates (around an average) during the afternoon hours and rapidly increases in the late afternoon To see more clearly the degree of correlation, in Fig 4.22 both lines of variation have been shown, in which the actual measured values of total P have been normalized (linearly) in the range [0,1] From Fig 4.22, we can see that there is a clear correlation between the vibration of the transformer and the transmission power of the transformer Conducted spectrum analysis of the vibration signals, the signal's spectrum was analyzed according to the windows with length of minutes, the number of calculation windows is 90, the total calculation time for spectral analysis is 450 minutes The analysis results show that the vibrational spectra have a relatively high similarity as can be seen in Fig 4.23 below Figure 4.20: Vibration measurement by accelerometer sensor on site from 9.00 to 18.30 on September 15, 2020 Figure 4.21: Values of total P measured by Thai Nguyen Electricity Company for the site from 9am to 18.30 on 15/9/2020 Figure 4.22: Graph showing simultaneous vibration and P sum signals of transformer normalized to range [0,1] during test measurement 22 (a) (c) (b) (d) Figure 4.23: Spectral analysis results for digital windows : (a), 10 (b), 20 (c) 30(d) Figure 4.23 shows the amplitude of the vibration spectrum using Fourrier analysis It can be noticed between the graphs that there is a difference in the pitch of the spectral peaks, but the position of the spectral peaks is quite stable The figures have marked some spectral peaks in the vicinity of fundamental frequencies such as 50Hz, 100Hz and 150Hz along with some non-harmonic spectral peaks of the fundamental frequency such as peaks at 75Hz and 107.5Hz To show more clearly the variation of the peak positions in the spectrum over time, Figure 4.24 shows the variation of the spectral peak in frequency ranges :  To monitor the variation of the neighboring spectral peak at 50Hz, we determine the extreme point of the frequency range [37.5Hz ; 62.5Hz],  To monitor the variation of the neighboring 75Hz spectral peak, we determine the extreme point of the frequency range [62.5Hz ; 87.5Hz],  To monitor the variation of the neighboring spectral peak 107.5Hz, we determine the extreme point of the frequency range [105Hz ; 110Hz] (the search band for this spectral peak is narrower to avoid falsely Figure 4.24: Variation of fundamental spectral peaks with sampling time 23 detecting the spectral peak in the vicinity of 100Hz),  To monitor the variation of the neighboring spectral peak of 150Hz, we determine the extreme point of the frequency range [137.5Hz ; 162.5Hz] It can be seen from Fig 4.24 that the basic spectral peaks have quite stable positions, among the observed positions, only the spectral peaks in the vicinity of 75Hz have clear fluctuations in the range from 70Hz to 80Hz Through this, it can be seen that in the case of normal operation, the frequency spectrum may vary greatly in amplitude when the instantaneous power of the transformer changes, but the main spectral peaks will have a relatively stable position Due to the short test measurement time, there are no conditions for testing with transformers that are having problems and not many data for comparison, but certain conclusions can be drawn as follows: The vibration of the transformer is greatly affected by the instantaneous transmission power of the transformer, so to detect abnormal vibrations, it is necessary to combine both the vibration signal and the electrical signal (measurement of the working point of the transformer) Accelerometer sensor can be used as a solution to measure the vibration of equipment when operating, thereby as a basis for building models to identify working modes of equipment In normal operating mode, when the load changes, the frequency spectrum of the actual measured vibration signals has fairly stable spectral peak positions This shows that when there are unusual vibrational components (due to different incidents), it is possible to detect these components as a basis for fault case identification 4.4 CONCLUSION OF CHAPTER Chapter of the thesis presented the following issues: - Using the transformer model built in Chapter by ANSYS software to simulate 234 results of working scenarios of the transformer, including: 01 normal working case and 05 incident cases For each case, the transformer is simulated with 50%, 80%, 100% nominal rated load - Conducted evaluation and analysis showed that the electrical signals and active force signals in the transformer can be used to detect the status of the transformer - Extracted the characteristic information of the received signals (from ANSYS simulation), trained the MLP network with 15 inputs, hidden neurons and output to identify the working state of the transformer The network was trained with 180 samples and tested with 54 samples out of a total of 234 samples collected Study results and test results are 100% accurate - Developed a vibration measuring device using accelerometer and applied actual test with a distribution transformer at the Industrial University Station The initial measurement results show that the measured vibration is consistent with the variation of the transformer's transmission power and the amplitude spectrum of the vibration signal has a set of fairly stable spectral peaks during normal-mode operation CONCLUSIONS AND RECOMMENDATIONS CONCLUSIONS When diagnosing transformer failures in general and distribution transformers in particular, it is very important to have complete information about transformers to help identify and accurately diagnose transformer problems The project has researched and successfully built a 24 distribution transformer model on Ansys software and used Levenberg - Marquardt algorithm combined with MLP neural network to identify and diagnose transformer problems New contributions of the thesis: - The thesis has built a 22/0.4kV Y-Y0 transformer model, in ANSYS software to serve as simulation results of electrical signals and mechanical signals (mechanical vibration) Conduct simulation of 234 working scenarios of transformer including normal working case and 05 incident cases For each case, the transformer is simulated with 50%, 80%, 100% of rated load - Proposing to extract 15 characteristic information of signals obtained from ANSYS simulation as a basis for building identification model - Building an identification model using MLP neural network with 15 inputs, hidden neurons and output to identify the working state of the MBA The network was trained with 180 samples and tested with 54 samples out of a total of 234 samples collected Study results and test results achieve high accuracy To verify the transformer model built on ANSYS software, the thesis conducted an experiment to measure the vibration of the transformer with a vibration measuring device using an accelerometer sensor for the distribution transformer at the Industrial University Station Results shows that the MBA model built is true to reality In addition, the thesis has studied the theoretical basis of the vibration phenomenon in the MBA and introduced a system of Maxwell's equations with boundary conditions for electromagnetic field problems Finite element theory is used to solve the system of Maxwell's equations and from the results of magnetic field calculations (electromagnetic force, deformation, displacement), the thesis has built a model of oscillation calculation by the partial method finite death RECOMMENDATIONS Although some results have been achieved as described above, the ideas and proposed solutions still have some problems that need to be further supplemented and researched, such as not taking into account the influence of the MBA's working process in the future overload mode or not taking into account the factors of natural climate conditions affecting the working mode of the MBA Therefore, the development direction of the thesis is to continue to study more new types of transformers and new fault cases of transformers as well as to expand potential fault diagnosis for other electrical equipment PUBLICATIONS Đào Duy Yên, Trần Xuân Minh, Trương Tuấn Anh (2020), “Chẩn Đoán Sự Cố Trong Máy Biến Áp P Sử Dụng Các Tín Hiệu Dịng, Áp Và Rung Động Cơ Khí”, Tạp chí Khoa học Công nghệ Đại học Thái nguyên, 225(09), trang 96 – 102 Truong Tuan Anh, Dao Duy Yen (2020), “Some solutions for online monitoring the vibration of transformers and the application to identify the state of the transformers”, IJRDO - Journal of Electrical and Electronics Engineering, Volume Issue Dao Duy Yen, Tran Xuan Minh, Tran Hoai Linh (2018), “A Solution for Fault Detection in Power Transformer using Vibration Signals and Mechanical Forces”, SSRG 25 International Journal of Electrical and Electronics Engineering (SSRG - IJEEE) – Volume Issue 10 – Oct 2018 Đào Duy Yên (2017), “Ứng dụng phần mềm ANSYS xây dựng mơ hình MBA ba pha phân tích chế độ làm việc MBA điều kiện làm việc bình thường cố”, Tạp chí Khoa học Cơng nghệ Đại học Thái Nguyên, Tập 176 - số 16, trang 185-189 Linh Tran Hoai, Yen Dao Duy (2017), “Transformer faults detection using electrical and mechanical vibration signals”, The 11th SEATUC Symposium, Vietnam ... electrical equipment PUBLICATIONS Đào Duy Yên, Trần Xuân Minh, Trương Tuấn Anh (2020), ? ?Chẩn Đoán Sự Cố Trong Máy Biến Áp P Sử Dụng Các Tín Hiệu Dịng, Áp Và Rung Động Cơ Khí”, Tạp chí Khoa học Công nghệ... of transformer in normal state Figure 3. 3: Meshing model and number of mesh elements MBA 3. 3.2 The mesh for the model of transformer with a loosen coil 3. 3 .3 The meshes for the model of transformer... function block 3. 1 .3 Structural simulation function block using ANSYS Structure 3. 1.4 ANSYS design modeler and ANSYS meshing 3. 1.5 ANSYS mechanical workbench 3. 1.6 ANSYS mechanical 10 3. 2 IMPLEMENTING