An Adaptive Fault-Tolerant Control Method for a Three-Level ANPC Inverter

Abstract

This invention discloses an adaptive fault-tolerant control method for a three-level ANPC inverter. Specifically, based on the influence of a faulty switching transistor on the output voltage vector, the inverter's fault-tolerant modes are classified into zero-level loss and zero-level retention fault-tolerant modes. The maximum modulation index of the inverter's output voltage vector is calculated under both fault-tolerant modes to determine whether the inverter system needs derating. For the zero-level retention fault-tolerant mode, a model predictive current control strategy with duty cycle modulation is adopted, utilizing a virtual zero-voltage vector synthesis method to solve the inverter's midpoint voltage imbalance problem. For the zero-level loss fault-tolerant mode, a cost function representing the fault-tolerant control performance index is designed, and a neural network is used to optimize the weight factors of the cost function. This invention achieves adaptive fault-tolerant control for open-circuit faults in inverters by adopting corresponding model predictive control optimization methods for different fault-tolerant modes, thereby improving the reliability of the inverter system.

Description

Technical Field

[0001] This invention belongs to the field of inverter fault-tolerant control, specifically relating to an adaptive fault-tolerant control method for a three-level ANPC inverter. Background Technology

[0002] With the rapid development of renewable energy, inverters, as one of the core components in the field of power electronics, have become a key factor in energy conversion and distribution. Compared with traditional two-level inverters, three-level inverters have advantages such as low switching losses, high current quality, and low electromagnetic interference. In particular, three-level ANPC inverters, with their high efficiency and low harmonic characteristics, are widely used in solar power generation, electric vehicle drive, and energy storage. However, as the number of active power devices in ANPC inverters increases, the probability of failure also increases. To ensure continuous power supply from the inverter and minimize the impact of failures on the system, it is necessary to develop advanced fault-tolerant control technologies for various inverter failure scenarios to improve the reliability and safety of the inverter system. Most existing fault-tolerant control methods are based on carrier modulation and space vector pulse width modulation techniques, achieving fault tolerance through effective switching state combinations or voltage vector reconstruction after a fault. However, in fault-tolerant operation mode, it is difficult to simultaneously achieve multiple control objectives, such as capacitor voltage balance, common-mode voltage suppression, and current harmonic optimization. Model predictive control (MPC), which has emerged in recent years, can effectively solve multi-objective optimization problems compared to traditional vector control, and has been gradually applied in industrial fields such as wind power generation, ship propulsion, and traction transmission. MPC typically uses a specific cost function to achieve multi-objective optimization, where weight factors are assigned to each control objective. Therefore, some research has introduced MPC into inverter fault-tolerant control to achieve optimal control under system fault conditions. Traditional finite set MPC can only select one voltage vector at a time. Under inverter switching faults, the effective output voltage vector decreases, resulting in the system's steady-state performance not reaching the ideal state under fault-tolerant control mode. Therefore, some scholars have attempted to introduce duty cycle modulation into traditional MPC, designing duty cycle-based MPC strategies. Meanwhile, for the control weight optimization problem in model predictive control, existing research has utilized neural networks to optimize the weight factors of the MPC cost function. Summary of the Invention

[0003] To address the shortcomings and gaps in existing technologies, this invention provides an adaptive fault-tolerant control method for three-level ANPC inverters. For the fault-tolerant mode where the inverter retains zero voltage, duty cycle modulation is introduced into finite set model predictive control. Simultaneously, a virtual zero-voltage vector synthesis method is employed to achieve reference current tracking and inverter midpoint voltage balance. For the fault-tolerant mode where the inverter loses zero voltage, a cost function representing the performance index of fault-tolerant control is designed, and a neural network is used to optimize the weighting factors of the cost function, achieving comprehensive optimization of multiple fault-tolerant control objectives.

[0004] The flowchart of the adaptive fault-tolerant control method for a three-level ANPC inverter of the present invention is as follows: Figure 1 As shown, the specific steps include:

[0005] S1: Based on the influence of the faulty switching transistor on the output voltage vector, the inverter fault tolerance mode is classified into two types: zero-level loss fault tolerance mode and zero-level retention fault tolerance mode.

[0006] S2: Calculate the maximum modulation index of the inverter output voltage vector under the two fault-tolerant modes to determine whether the inverter system needs to be derated.

[0007] S3: For the fault-tolerant mode that retains zero level, a model predictive current control strategy based on duty cycle modulation is adopted, and the problem of voltage imbalance at the inverter midpoint is solved by using the virtual zero voltage vector synthesis method.

[0008] S4: For the fault-tolerant mode that loses zero level, design a cost function that can represent the performance index of fault-tolerant control, and use a neural network to optimize the weight factor of the cost function.

[0009] Furthermore, step S1 specifically includes:

[0010] S1.1: The effective positive level P, zero level OU, zero level OL, and negative level N of the inverter output under switching transistor fault conditions are calculated based on Boolean operations and expressed as follows:

[0011]

[0012] Among them, the switching transistor T xn (x = a, b, c; n = 1, 2, 3, 4, 5, 6) is in a normal state, and the corresponding logical value is S. xn (x = a, b, c; n = 1, 2, 3, 4, 5, 6) is 1, and the switching transistor T xn When in a fault state, the corresponding logic value S xn =0; P x (x = a, b, c), OU x OL x and N x A logic value of 0 indicates that the inverter cannot output a certain level. x OU x OL x and N x A logic value of 1 indicates that the inverter can output an effective level.

[0013] S1.2: Analyze the impact of a faulty switching transistor on the inverter output positive level P. When the switching transistor T... x1 or T x2 Fault, corresponding logic value S x1or S x2 P is 0 x =0 indicates that the switching transistor T x1 or T x2 Under fault conditions, the inverter cannot output a positive level P.

[0014] S1.3: Analyze the impact of a faulty switching transistor on the inverter output zero level OU. When the switching transistor T... x2 or T x5 Fault, corresponding logic value S x2 or S x5 0, OU x =0 indicates that the switching transistor T x2 or T x5 Under fault conditions, the inverter cannot output a zero level (OU).

[0015] S1.4: Analyze the impact of a faulty switching transistor on the inverter output zero level OL. When the switching transistor T... x3 or T x6 Fault, corresponding logic value S x3 or S x6 0, OL x =0 indicates that the switching transistor T x3 or T x6 Under fault conditions, the inverter cannot output a zero level (OL).

[0016] S1.5: Analyze the impact of a faulty switching transistor on the inverter output negative level N. When the switching transistor T... x3 or T x4 Fault, corresponding logic value S x3 or S x4 =0, N x =0 indicates that the switching transistor T x3 or T x4 Under fault conditions, the inverter cannot output a negative level N.

[0017] S1.6: Based on the impact of different faulty switching transistors on the inverter output level, the inverter fault-tolerant modes are classified, and the switching transistor T... x1 -T x4 Under fault conditions, the inverter cannot output a positive level P or a negative level N. The inverter operates in a zero-level fault-tolerant mode, and the switching transistor T... x5 -T x6 In the event of a fault, the inverter cannot output a zero level (OU or OL), and the inverter operates in a zero-level loss fault-tolerant mode.

[0018] Furthermore, step S2 specifically involves:

[0019] S2.1: Analyze the voltage vector output by the inverter in fault-tolerant mode. When the fault switch is located in phase a of the inverter, in the zero-level fault-tolerant mode, the inverter outputs 6 small vectors (ONN, OON, ONO, OPP, OOP, OPO), 2 medium vectors (ONP, OPN), and 1 zero vector (OOO); in the zero-level fault-tolerant mode, the inverter outputs 6 large vectors (PNN, PPN, NPN, NPP, NNP, PNP), 4 medium vectors (PON, NPO, NOP, PNO), 6 small vectors (POO, PPO, NON, NOO, NNO, POP), and 2 zero vectors (PPP, NNN).

[0020] S2.2: Calculate the maximum stator voltage vector amplitude of the inverter output under both fault-tolerant modes, retaining the maximum stator voltage vector amplitude u of the inverter output under the zero-level fault-tolerant mode. r1 for In zero-level loss fault-tolerant mode, the maximum stator voltage vector amplitude u of the inverter output r2 for

[0021] S2.3: Calculate the maximum modulation index of the inverter output voltage vector under the two fault-tolerant modes, expressed as:

[0022]

[0023] Among them, u dc This is the output voltage on the DC bus side of the inverter.

[0024] S2.4: In the zero-level fault-tolerant mode, the modulation index k1 is 0.5, and the inverter system needs to operate at a reduced rate. In the zero-level fault-tolerant mode, the modulation index k2 is 1, and the inverter system does not need to operate at a reduced rate.

[0025] Furthermore, step S3 specifically includes:

[0026] S3.1: Define the cost function J for model predictive current control in zero-level fault-tolerant mode as follows:

[0027]

[0028]

[0029] Among them, i d (k+1),i q (k+1) represent the predicted currents along the d-axis and q-axis at the next time step, respectively. λ represents the reference current along the d-axis and q-axis, respectively. dc u represents the weighting factor for the voltage balance at the midpoint of the model's predicted current control. c1(k+1), u c2 (k+1) represent the predicted inverter upper and lower capacitor voltages at the next time step, respectively. max This represents the maximum allowable output current amplitude in fault-tolerant mode.

[0030] S3.2: A model predictive current control strategy using duty cycle modulation is employed. Among the six small vectors (ONN, OON, ONO, OPP, OOP, OPO), an optimal small vector V is selected through a defined cost function J. opt1 The other voltage vector is fixed and chosen as the zero vector V0 (OOO).

[0031] S3.3: Calculate the small vector V using the deadbeat principle of q-axis current. opt1 duty cycle γ opt1 The duty cycle of the zero vector V0 is calculated to be 1-γ. opt1 .

[0032] S3.4: Use a pair of complementary small vectors (V n V p To synthesize a virtual zero voltage vector U0, thereby reducing inverter midpoint voltage fluctuations in fault-tolerant mode.

[0033] S3.5: Three pairs of complementary small vectors (ONN, OPP), (OON, OOP), and (ONO, OPO) are used as candidate vectors to synthesize the virtual zero voltage vector U0. The midpoint current i under the action of the three pairs of small vectors... o i a -i b and -i c By judging the deviation u of the upper and lower capacitor voltages of the inverter er Midpoint current i under the action of complementary small vectors o Let's choose a pair of complementary small vectors, represented as:

[0034] u er =i o <0

[0035] S3.6: Select a suitable pair of complementary small vectors (V n V p To synthesize a virtual zero vector U0 to replace the zero vector V0, it is represented as:

[0036] U0t0=V0t0=(V n t0+V p t0) / 2

[0037] S3.7: Calculate the duty cycle of the corresponding small vector, and complete the voltage vector duty cycle allocation according to the action sequence and action time of the small vector through the PWM principle, and control the switching transistor to complete the fault-tolerant control.

[0038] Furthermore, step S4 specifically involves:

[0039] S4.1: Determine the control objectives in the zero-level fault-tolerant mode. The fault-tolerant control objectives include reference current tracking, midpoint voltage balance, and switching frequency reduction.

[0040] S4.2: To achieve inverter midpoint voltage balance in zero-level loss fault-tolerant mode, the predicted upper and lower capacitor voltages are used to define the cost function J. dc for:

[0041] J dc =(u c1 (k+1)-u c2 (k+1)) 2

[0042] S4.3: Quantification of the ability to balance the midpoint voltage; the voltage deviation between the upper and lower capacitors on the DC bus is defined as:

[0043] u er (k)=u c1 (k)-u c2 (k)

[0044]

[0045] Among them, u er (k) represents the deviation of the upper and lower capacitor voltages on the DC bus at time k, T s V represents a sampling period, N represents the total number of sampling points within a sampling period, and V represents the total number of sampling points within a sampling period. ag This represents the average deviation of the midpoint voltage.

[0046] S4.4: To reduce the inverter switching frequency, the cost function J is defined by the number of switching state transitions. sw for:

[0047] J sw =|S a (k+1)-S a (k)|+|S b (k+1)-S b (k)|+|S c (k+1)-S c (k)|

[0048] Among them, S x (k)(x=a,b,c) represents the current switching state at time k, S x (k+1) represents the on / off state at time k+1 in the future.

[0049] S4.5: The switching frequency of the inverter over a period of time is expressed as the average switching frequency, as follows:

[0050]

[0051] Among them, f sw (k) represents the switching frequency at time k.

[0052] S4.6: Define the cost function J in the loss-zero level fault-tolerant mode as follows:

[0053]

[0054] Where, λ dc , λ sw These represent the weighting factors for midpoint voltage balance and switching frequency adjustment in the zero-level fault-tolerant mode, respectively.

[0055] S4.7: Determine the weighting factor λ dc and λ sw The initial range is determined, and the weighting factor λ is further determined based on the simulation results. dc and λ sw The selection range is determined by specifying the motor speed N. p and torque T e Operating conditions.

[0056] S4.8: Generate the data needed for neural network training, and use a fitness function to evaluate the motor control performance index (THD) in fault-tolerant mode. i 、V ag f ag Fitness function f res Represented as:

[0057]

[0058]

[0059] in, and These represent the normalized total harmonic distortion (THD) of the current, respectively. i Midpoint voltage balance V ag and switching frequency f ag The indicators max(M) and min(M) are the maximum and minimum values ​​of the motor's control performance indicators under specific operating conditions.

[0060] S4.9: Design the structure of the neural network, and assign weight factors λ dc , λ sw and motor operating conditions N p 、T e As the input layer of the neural network, the fault-tolerant control performance index (THD) will be used. i 、Vag f ag As the output layer, select the number and structure of the hidden layers in the network, determine the number of neurons in each hidden layer, and define the objective function f of the BP neural network. BP for:

[0061]

[0062] S4.10: Backpropagation training of the neural network is performed using gradient descent. Hyperparameters such as batch size, number of iterations, and learning rate are designed. The trained network is evaluated based on RMSE (Real-Time Sequence), and the hyperparameters of the network structure are adjusted based on the training results to select the optimal network. This is expressed as:

[0063]

[0064] RMSE is a quantitative indicator. y is the predicted value of the neural network. i represents the true value in the dataset, and m is the number of samples in the dataset.

[0065] S4.11: Use a weight factor with a shorter step size as the input layer of the optimal network, and use the trained neural network prediction model to predict and control the optimal weight factor.

[0066] The beneficial technical effects of this invention are as follows:

[0067] (1) Based on the influence of the fault switch tube on the output voltage vector, the present invention classifies the fault tolerance mode of the three-level ANPC inverter into two fault tolerance modes: zero level loss and zero level retention. When the system detects a fault in the inverter, it directly switches the inverter control system to the corresponding fault tolerance mode, thus realizing adaptive fault tolerance control from normal mode to fault mode.

[0068] (2) The present invention optimizes fault-tolerant control. For the fault-tolerant mode that retains zero level, the model prediction current control strategy of duty cycle modulation is introduced into the fault-tolerant control of the inverter, thereby achieving reference current tracking and neutral point voltage balance of the inverter system under fault-tolerant operation mode. For the fault-tolerant mode that loses zero level, the neural network optimization model is used to predict the weight factor of the fault-tolerant control cost function, thereby achieving the optimal comprehensive control performance of the motor drive system fed by the three-level ANPC inverter under fault-tolerant operation mode. Attached Figure Description

[0069] Figure 1 This is a flowchart of the adaptive fault-tolerant control method for the three-level ANPC inverter of the present invention.

[0070] Figure 2Circuit diagram of a permanent magnet synchronous motor drive system powered by a three-level ANPC inverter.

[0071] Figure 3 Reserve the zero-level fault-tolerant mode space voltage vector diagram for the three-level ANPC inverter.

[0072] Figure 4 This is a block diagram illustrating the principle of the model predictive current control strategy based on duty cycle modulation in an embodiment of the present invention.

[0073] Figure 5 This is a waveform diagram of fault-tolerant control under the zero-level fault-tolerant mode in an embodiment of the present invention.

[0074] Figure 6 This is a space voltage vector diagram for a three-level ANPC inverter in zero-level fault-tolerant mode.

[0075] Figure 7 This is a block diagram illustrating the principle of the model prediction adaptive fault-tolerant control strategy based on neural network optimization, according to an embodiment of the present invention.

[0076] Figure 8 This is the training result of the BP neural network in an embodiment of the present invention.

[0077] Figure 9 This is a waveform diagram of fault-tolerant control under the lost zero-level fault-tolerant mode according to an embodiment of the present invention. Detailed Implementation

[0078] The present invention will be further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0079] This invention takes a permanent magnet synchronous motor drive system powered by a three-level active neutral point clamped (3L-ANPC) inverter as an example, and provides an adaptive fault-tolerant control method for the three-level ANPC inverter based on model predictive control. The main circuit schematic of the permanent magnet synchronous motor drive system powered by the three-level ANPC inverter is shown below. Figure 2 As shown, the system mainly includes a three-phase three-level ANPC inverter and a permanent magnet synchronous motor.

[0080] A three-level ANPC inverter can output three switching states: positive level (P), zero level (O), and negative level (N), corresponding to the three phase voltages u output by the inverter. xo (x = a, b, c): u dc / 2, 0, -u dc / 2, the inverter DC bus voltage is u dc .

[0081] The three-level ANPC inverter has redundant switching states for output zero level. The switching state O has two redundant states, OU and OL. The phase voltage output by the ANPC inverter and the corresponding switching state are shown in Table 1:

[0082] Table 1 Switching states and output phase voltages of a three-level ANPC inverter

[0083]

[0084] In the table, S x (x = a, b, c) represents the corresponding output phase voltage u xo The switching state function is expressed as:

[0085]

[0086] Analyzing the impact of a faulty switching transistor on the output voltage level of the ANPC inverter, taking phase a as an example, the effective positive voltage level P, zero voltage level OU, zero voltage level OL, and negative voltage level N of the inverter output under a switching transistor fault are calculated based on Boolean operations, and expressed as follows:

[0087]

[0088] Among them, S an (n = 1, 2, 3, 4, 5, 6) represents a logic value of 1 when the corresponding switch T an (n=1,2,3,4,5,6) is in a normal state, S an The logic value represented is 0 when the corresponding switch T an In a faulty state, P a OU a OL a and N a A logic value of 0 indicates that the inverter cannot output a valid level. a OU a OL a and N a A logic value of 1 indicates that the inverter can output an effective level.

[0089] Switching transistor T a1 or T a2 Fault, its corresponding logical value S a1 or S a2 P is 0 a =0 indicates that the switching transistor T a1 or T a2 Under fault conditions, the ANPC inverter cannot output a positive level P.

[0090] Switching transistor T a2 or T a5 Fault, its corresponding logical value Sa2 or S a5 0, OU a =0 indicates that the switching transistor T a2 or T a5 Under fault conditions, the ANPC inverter cannot output a zero level (OU).

[0091] Switching transistor T a3 or T a6 Fault, its corresponding logical value S a3 or S a6 0, OL a =0 indicates that the switching transistor T a3 or T a6 Under fault conditions, the ANPC inverter cannot output a zero level (OL).

[0092] Switching transistor T a3 or T a4 Fault, its corresponding logical value S a3 or S a4 =0, N a =0 indicates that the switching transistor T a3 or T a4 Under fault conditions, the ANPC inverter cannot output a negative level N.

[0093] The impact of different switching transistor faults on the inverter output voltage vector was analyzed, and the three-level ANPC inverter can operate in two fault-tolerant modes.

[0094] Fault-tolerant mode 1: Three-level ANPC inverter switching transistor T a1 -T a4 In the event of an open-circuit fault, the faulty phase cannot output a P or N switching state. To ensure the balance of the inverter's switching losses, only the faulty phase is allowed to output an O switching state. The ANPC inverter operates in a zero-level fault-tolerant mode. In the zero-level fault-tolerant mode, the inverter outputs 6 small vectors (ONN, OON, ONO, OPP, OOP, OPO), 2 medium vectors (ONP, OPN), and 1 zero vector (OOO).

[0095] Fault-tolerant mode 2: Three-level ANPC inverter switching transistor T a5 or T a6 In the event of an open-circuit fault, the faulty phase cannot output OU or OL switching status. To ensure the balance of the inverter's switching losses, the ANPC inverter operates in a zero-level loss fault-tolerant mode. In the zero-level loss fault-tolerant mode, the inverter outputs 6 large vectors (PNN, PPN, NPN, NPP, NNP, PNP), 4 medium vectors (PON, NPO, NOP, PNO), 6 small vectors (POO, PPO, NON, NOO, NNO, POP), and 2 zero vectors (PPP, NNN).

[0096] Three-level ANPC inverter switching transistor T a1 -T a6 The fault tolerance modes under open-circuit faults are shown in Table 2.

[0097] Table 2 Classification of Fault Tolerance Modes for Three-Level ANPC Inverters

[0098]

[0099] This analysis examines the voltage and current constraints of an inverter-fed motor drive system in fault-tolerant mode. In a permanent magnet synchronous motor (PMSM) drive system, an open-circuit fault occurs in the ANPC inverter, but the motor's topology remains unchanged, thus not affecting the current output capability. The motor can operate at low speeds while still meeting certain torque requirements. During steady-state operation of the PMSM, neglecting the voltage drop across the stator resistance, the voltage and current constraints are expressed as follows:

[0100]

[0101] Among them, u max i represents the maximum stator voltage vector magnitude output by the inverter. max To determine the maximum vector magnitude of the dq-axis current that is allowed to flow, u d u q For the d-axis and q-axis stator voltage components; i d 、i q ω represents the stator current components along the d-axis and q-axis; e Let ψ be the electric angular velocity. f It is a permanent magnet flux linkage.

[0102] Analyze the radius u of the maximum voltage limit circle of the ANPC inverter output voltage vector in fault-tolerant mode. r , Figure 3 In the zero-level fault-tolerant mode, the radius u of the maximum voltage limit circle r1 for Figure 6 In the zero-level loss fault-tolerant mode, the radius u of the maximum voltage limit circle r2 for Calculate the maximum voltage vector modulation of the inverter output under both fault-tolerant modes:

[0103]

[0104] In the zero-level fault-tolerant mode, the modulation index k1 is 0.5, and the inverter system needs to operate at a reduced rate. In the zero-level fault-tolerant mode, the modulation index k2 is 1, and the inverter system does not need to operate at a reduced rate.

[0105] In practical implementation, the switching transistor T a1Under open-circuit fault conditions, the three-level ANPC inverter operates in a zero-level fault-tolerant mode. The inverter outputs nine switching states, including six small vectors (ONN, OON, ONO, OPP, OOP, OPO), two medium vectors (ONP, OPN), and one zero vector (OOO). The space voltage vector of the inverter output under fault conditions is as follows: Figure 3 As shown.

[0106] To prevent over- and under-regulation of the reference current in fault-tolerant mode, a model predictive current control strategy based on duty cycle modulation is adopted, which can effectively avoid over- and under-regulation of the reference current.

[0107] In a control cycle T s The internal action consists of two voltage vectors, including a small vector V. opt1 And a zero vector V0, a ​​small vector V opt1 Selected by a defined cost function. This is achieved by appropriately allocating small vectors V. opt1 The duty cycle of the zero vector V0 is used to track the reference current.

[0108] Define the cost function J in the zero-level fault-tolerant mode, and set a penalty term for midpoint voltage balance in the cost function, expressed as:

[0109]

[0110]

[0111] Among them, i d (k+1),i q (k+1) represent the predicted currents along the d-axis and q-axis at the next time step, respectively. λ represents the reference current along the d-axis and q-axis, respectively. dc u represents the weighting factor for the voltage balance at the midpoint of the model's predicted current control. c1 (k+1), u c2 (k+1) represent the predicted inverter upper and lower capacitor voltages at the next time step, respectively. max This represents the maximum allowable output current amplitude in fault-tolerant mode.

[0112] Among the six small vectors (ONN, OON, ONO, OPP, OOP, OPO), the optimal small vector V is selected using the defined cost function J. opt1 The other voltage vector is fixed and chosen as the zero vector V0 (OOO).

[0113] The small vector V is calculated using the deadbeat principle of q-axis current. opt1 The duty cycle of the zero vector V0 is expressed as:

[0114]

[0115] Among them, t opt1 For small vector V opt1 Duration of action, s opt1 It is a small vector V opt1 The slope of the q-axis current change when the zero vector V0 is applied; s0 is the slope of the q-axis current change when the zero vector V0 is applied.

[0116] s0 and s opt1 The calculation formula is:

[0117]

[0118]

[0119] Among them, u q_opt1 For small vector V opt1 The q-axis voltage component, L q R is the q-axis inductance of the permanent magnet synchronous motor. s is the stator resistance.

[0120] Small vector V opt1 duty cycle γ opt , represented as:

[0121]

[0122] The duty cycle of the zero vector V0 is calculated as 1 - γ. opt1 .

[0123] The model predictive current control strategy based on duty cycle modulation does not consider the midpoint voltage imbalance problem in the ANPC inverter, which will affect the fault-tolerant control performance. Since the zero vector V0 does not affect the midpoint voltage, a pair of complementary small vectors (V0) can be used based on the voltage deviation between the upper and lower capacitors of the ANPC inverter. n V p A virtual zero voltage vector U0 is synthesized to replace the zero vector V0, thereby reducing the voltage fluctuation at the inverter midpoint in fault-tolerant mode.

[0124] Analysis of the midpoint current i under the action of a small vector o As shown in Table 3.

[0125] Table 3 Midpoint current i under voltage vector action. o

[0126]

[0127] Three pairs of complementary small vectors (ONN, OPP), (OON, OOP), and (ONO, OPO) are used as candidate vectors to synthesize a virtual zero-voltage vector U0. This is achieved by judging the voltage deviation u between the upper and lower capacitors of the ANPC inverter. er Midpoint current i under the action of complementary small vectors o Let's choose a pair of complementary small vectors, represented as:

[0128] u er ×i o <0

[0129] Where u er The deviation of the voltage between the upper and lower capacitors of the inverter is calculated as follows:

[0130] u er =u c1 -u c2

[0131] By judging the voltage deviation u of the upper and lower capacitors er and three-phase current i a ,i b ,i c The symbols are selected for the synthesis of the virtual zero vector U0, and the rules for the synthesis of the virtual zero vector are shown in Table 4.

[0132] Table 4 Rules for the Synthesis of Virtual Zero Voltage Vector U0

[0133]

[0134] Choose a suitable pair of complementary small vectors (V) n V p To synthesize the virtual zero vector U0, we can use the following:

[0135] U0t0=V0t0=(V n t0+V p t0) / 2

[0136] In the formula, t0 is the duration of the zero vector V0 in the duty cycle model predictive current control.

[0137] The first optimal voltage vector V in this implementation case opt1 For ONO, a pair of complementary small vectors (OON, OOP) are used to synthesize a virtual zero voltage vector U0. The virtual zero voltage vector U0 replaces the V0 vector. In one control cycle T s Inside, the duty cycle of the small vector ONO is T. ONO =γ opt The duty cycle T of the voltage vector OON OON =(1-γ) opt ) / 2, the duty cycle T of the voltage vector OOP OOP =(1-γ)opt The voltage vector duty cycle is allocated according to the action sequence and action time of the small vectors using the PWM principle, and the switching transistors are controlled to achieve fault-tolerant control. The combined voltage vector and duty cycle of all virtual zero vectors U0 in the fault-tolerant mode of the ANPC inverter are shown in Table 5.

[0138] Table 5 Complementary small vectors and their duty cycles

[0139]

[0140] The block diagram of the proposed duty cycle modulation model predictive current control strategy is as follows: Figure 4 As shown, it mainly includes three parts: system sampling, cost function rolling, and virtual zero-voltage vector synthesis. The small vector V is selected based on the cost function. opt1 Calculate the small vector V opt1 duty cycle γ opt And using a pair of complementary small vectors (V n V p A virtual zero vector U0 is synthesized and then replaces the zero vector V0. For different complementary small vectors, the duty cycle of the corresponding small vector is calculated, and the switching transistor is controlled to complete the fault-tolerant control.

[0141] Figure 5 To preserve the simulation results of the ANPC inverter in zero-level fault-tolerant mode, the a-phase switch state S a The state of phase b switch S remains at 0. b The output shows three switching states: 1, 0, and -1. The c-phase switching state is S. c It also outputs three switching states: 1, 0, and -1. In fault-tolerant operation mode, the three-phase current i... a 、i b 、i c It has good harmonic characteristics. Voltage u c1 and u c2 The voltage stabilized at 100V with fluctuations within ±3V, indicating that the proposed duty cycle modulation model predictive current control strategy achieved good fault-tolerant control performance.

[0142] In the zero-level fault-tolerant mode, the effective voltage vector of the inverter output decreases. At the same time, there are multiple fault-tolerant control objectives such as reference current tracking, midpoint voltage balance and switching frequency reduction. A BP neural network optimization model is used to predict the weight factors of the fault-tolerant control cost function.

[0143] To achieve voltage balance at the neutral point of the ANPC inverter in zero-level fault-tolerant mode, the predicted upper and lower capacitor voltages are used to define the cost function J. dc for:

[0144] J dc =(u c1(k+1)-u c2 (k+1)) 2

[0145] Among them, u c1 (k) represents the predicted upper capacitor voltage, u c2 (k) represents the predicted lower capacitor voltage.

[0146] Further quantification of the midpoint voltage balancing capability can better reveal the impact of different weighting factors. The voltage deviation between the upper and lower capacitors on the DC bus can be defined as:

[0147] u er (k)=u c1 (k)-u c2 (k)

[0148]

[0149] Among them, u er (k) is the deviation of the upper and lower capacitor voltages on the DC bus, N represents the total number of sampling points in one sampling period, V ag This represents the average deviation of the midpoint voltage.

[0150] The switching frequency of model predictive control is not fixed and varies greatly. To reduce the switching frequency of the ANPC inverter, the cost function J is defined by the number of switching state transitions. sw for:

[0151] J sw =|S a (k+1)-S a (k)|+|S b (k+1)-S b (k)|+|S c (k+1)-S c (k)|

[0152] Among them, S x (k)(x=a,b,c) represents the current switching state at time k, S x (k+1) represents the on / off state at time k+1 in the future.

[0153] Since the switching frequency of the ANPC inverter is not fixed under different operating conditions, the average switching frequency over a period of time is expressed as:

[0154]

[0155] Considering that the fault-tolerant control objectives of reference current tracking, midpoint voltage balance, and switching frequency adjustment need to be achieved simultaneously in the zero-level loss fault-tolerant mode, the comprehensive cost function J is defined as follows:

[0156]

[0157] Where, λ dc , λ sw These represent the weighting factors for midpoint voltage balance and switching frequency regulation in model predictive control, respectively.

[0158] The block diagram of the proposed neural network-optimized adaptive fault-tolerant control method for ANPC inverters is shown below. Figure 7 As shown, the proposed adaptive fault-tolerant control method mainly includes three research contents: training data generation, BP neural network design, network training and prediction. By optimizing the weight factor of the cost function through neural network, the multi-objective optimization of the inverter in fault-tolerant mode is realized.

[0159] The initial range of the weighting factor λ is determined through simulation. dc The initial range is 0.1 to 1, and the weighting factor λ sw The initial range is 0 to 1. Training data is generated using MATLAB / Simulink, and the initial increment of the weight factor is 0.1.

[0160] To obtain the optimal overall fault-tolerant control performance, the weighting factor λ is determined based on the simulation results. dc The selection range is 0.15 to 1, and the weighting factor λ sw The selection range is 0.05 to 0.25, which clarifies the operating conditions of the motor.

[0161] The generated training data includes the motor speed N. p and torque T e Weighting factor λ dc and λ sw A fitness function is used to evaluate the motor control performance index (THD) in fault-tolerant mode. i V ag ,f ag The fitness function is expressed as:

[0162]

[0163]

[0164] Among them, f res Represents the fitness function. and These represent the normalized total harmonic distortion (THD) of the current. i The parameters are: midpoint voltage balance and switching frequency. max(M) and min(M) are the maximum and minimum values ​​of the motor's control performance parameters under specific operating conditions.

[0165] The data used for neural network training is shown in Table 6. The motor speed setting range is 200–600 rpm / min, with a speed change step of 100 rpm / min; the motor torque setting range is 5–25 Nm, with a torque change step of 5 Nm; the weighting factor λ... dc The set range is 0.15 to 1, with a change step size of 0.05; weighting factor λ sw The range was set to 0.05–0.25, with a variation step size of 0.05. 2250 sets of simulation data were automatically generated using MATLAB / Simulink for network training and validation.

[0166] Table 6. Metric parameters for the generated dataset

[0167]

[0168] Design the structure of a BP neural network, and assign weight factors (λ) dc ,λ sw ) and permanent magnet synchronous motor operating conditions (N p ,T e ) is used as the input layer of a BP neural network, controlling the performance index (THD) i V ag ,f ag As the output layer, the number of hidden layers is selected as 3, and the number of neurons in the hidden layer is determined to be 10, 9, and 6 respectively. The hidden layers of the BP neural network used include fully connected layers (FClayer) and activation function layers (ReLU).

[0169] Define the objective function f of the BP neural network. BP for:

[0170]

[0171] The designed network operates according to the defined objective function f. BP This is used to optimize the network's hyperparameters. The training set is used to train the neural network, and the test set is used to evaluate the network's accuracy and generalization ability. The quantitative metrics for accuracy and generalization ability are expressed as:

[0172]

[0173] In the formula, RMSE is a quantitative indicator. y is the predicted value of the neural network. i represents the true value in the dataset, and m is the number of samples in the dataset.

[0174] Backpropagation training of a BP neural network is performed using gradient descent. Hyperparameters such as batch size, number of iterations, and learning rate are designed. The trained network is evaluated based on RMSE (Recovery Mean Squared Error), and the hyperparameters of the network structure are adjusted based on the training results to select the optimal network. The network training results are as follows: Figure 8 As shown.

[0175] Using a weight factor with a shorter step size as the input layer of the optimal network, with a step size of 0.01, the trained BP neural network prediction model is used to predict and control the optimal weight factor.

[0176] Simulation results of zero-level fault-tolerant control are as follows Figure 9 As shown, the initial weighting factor λ for model predictive control dc =0.21, λ sw =0.15, before weight factor optimization, the phase current THD is 6.73%, and the electromagnetic torque T e The pulsation is relatively large, and the voltage deviation between the upper and lower capacitors is V. ag The value is 2.35V. The optimized weight factor λ of the BP neural network. dc =0.43, λ sw =0.13, when the optimized weighting factor is used, the phase current THD is 5.95%, and the electromagnetic torque T e The pulsation decreases, and the voltage deviation V between the upper and lower capacitors on the inverter DC bus decreases. ag The voltage is 1.27V. Comparing the simulation results before and after the optimization of the weight factor of the permanent magnet synchronous motor, the weight factor optimized based on the BP neural network can significantly improve the performance of model predictive control, thereby achieving the optimization of control performance in fault-tolerant mode.

Claims

1. An adaptive fault-tolerant control method for a three-level ANPC inverter, characterized in that, Includes the following steps: S1: Based on the impact of the faulty switching transistor on the output voltage vector, the inverter fault-tolerant modes are classified into zero-level loss fault-tolerant modes and zero-level retention fault-tolerant modes; switching transistor T x1 -T x4 Under fault conditions, the inverter cannot output a positive level P or a negative level N. The inverter operates in a zero-level fault-tolerant mode, and the switching transistor T... x5 -T x6 Under fault conditions, the inverter cannot output a zero level OU or OL, and the inverter operates in a zero-level loss fault-tolerant mode. S2: Calculate the maximum modulation index of the inverter output voltage vector under the two fault-tolerant modes to determine whether the inverter system needs to be derated. S3: For the fault-tolerant mode that retains zero level, a duty cycle modulation model predictive current control strategy is adopted, and the virtual zero voltage vector synthesis method is used to solve the problem of inverter midpoint voltage imbalance. S3.1: Define the cost function J for model predictive current control in zero-level fault-tolerant mode as follows: Among them, i d (k+1),i q (k+1) represent the predicted currents along the d-axis and q-axis at the next time step, respectively. λ represents the reference current along the d-axis and q-axis, respectively. dc u represents the weighting factor for the voltage balance at the midpoint of the model's predicted current control. c1 (k+1), u c2 (k+1) represent the predicted inverter upper and lower capacitor voltages at the next time step, respectively. max This represents the maximum allowable output current amplitude in fault-tolerant mode. S3.2: A model predictive current control strategy using duty cycle modulation is employed. Among the six small vectors, namely ONN, OON, ONO, OPP, OOP, and OPO, an optimal small vector V is selected through a defined cost function J. opt1 The other voltage vector is fixed and chosen as the zero vector V0, i.e., OOO; S3.3: Calculate the small vector V using the deadbeat principle of q-axis current. opt1 duty cycle γ opt1 The duty cycle of the zero vector V0 is calculated to be 1-γ. opt1 ; S3.4: Use a pair of complementary small vectors (V n V p To synthesize a virtual zero voltage vector U0, thereby reducing inverter midpoint voltage fluctuations in fault-tolerant mode; S3.5: Three pairs of complementary small vectors (ONN, OPP), (OON, OOP), and (ONO, OPO) are used as candidate vectors to synthesize the virtual zero voltage vector U0. The midpoint current i under the action of the three pairs of small vectors... o i a -i b and -i c By judging the deviation u of the upper and lower capacitor voltages of the inverter er Midpoint current i under the action of complementary small vectors o Let's choose a pair of complementary small vectors, represented as: in er ×and o <0 S3.6: Select a suitable pair of complementary small vectors (V n V p To synthesize a virtual zero vector U0 to replace the zero vector V0, it is represented as: U0t0=V0t0=(V n t0+V p t0) / 2 S3.7: Calculate the duty cycle of the corresponding small vector, and complete the voltage vector duty cycle allocation according to the action sequence and action time of the small vector through the PWM principle, and control the switching transistor to complete the fault-tolerant control; S4: For the fault-tolerant mode that loses zero level, design a cost function that can represent the performance index of fault-tolerant control, and use a neural network to optimize the weight factor of the cost function; S4.1: Determine the control objectives in the zero-level fault-tolerant mode. The fault-tolerant control objectives include reference current tracking, midpoint voltage balance, and switching frequency reduction. S4.2: To achieve inverter midpoint voltage balance in zero-level loss fault-tolerant mode, the predicted upper and lower capacitor voltages are used to define the cost function J. dc for: J dc =(you c1 (k+1)-u c2 (k+1)) 2 S4.3: Quantification of the ability to balance the midpoint voltage; the voltage deviation between the upper and lower capacitors on the DC bus is defined as: in er (k)=u c1 (k)-u c2 (k) Among them, u er (k) represents the deviation of the upper and lower capacitor voltages on the DC bus at time k, T s V represents a sampling period, N represents the total number of sampling points within a sampling period, and V represents the total number of sampling points within a sampling period. ag This represents the average deviation of the midpoint voltage; S4.4: To reduce the inverter switching frequency, the cost function J is defined by the number of switching state transitions. sw for: J sw =|S a (k+1)-S a (k)|+|S b (k+1)-S b (k)|+|S c (k+1)-S c (k)| Among them, S x (k), x = a, b, c represent the switch state at time k, S x (k+1) represents the on / off state at time k+1 in the future; S4.5: The switching frequency of the inverter over a period of time is expressed as the average switching frequency, as follows: Among them, f sw (k) represents the switching frequency at time k; S4.6: Define the cost function J in the loss-zero level fault-tolerant mode as follows: Where, λ dc , λ sw These represent the weighting factors for midpoint voltage balance and switching frequency adjustment in the zero-level fault-tolerant mode, respectively. S4.7: Determine the weighting factor λ dc and λ sw The initial range is determined, and the weighting factor λ is further determined based on the simulation results. dc and λ sw The selection range is determined by specifying the motor speed N. p and torque T e Operating conditions; S4.8: Generate the data needed for neural network training, and use a fitness function to evaluate the motor control performance indicators in fault-tolerant mode: THD. i 、V ag f ag fitness function f res Represented as: in, and These represent the normalized total harmonic distortion (THD) of the current, respectively. i Midpoint voltage balance V ag and switching frequency f ag The indicators, max(M) and min(M), are the maximum and minimum values ​​of the motor's control performance indicators under specific operating conditions. S4.9: Design the structure of the neural network, and assign weight factors λ dc , λ sw and motor operating conditions N p 、T e As the input layer of a neural network, the fault-tolerant control performance metric is: THD. i 、V ag f ag As the output layer, the number and structure of the hidden layers in the network are selected, and the number of neurons in each hidden layer is determined. The objective function f of the BP neural network is defined. BP for: S4.10: Backpropagation training of the neural network is performed using gradient descent. Hyperparameters such as batch size, number of iterations, and learning rate are designed. The trained network is evaluated based on RMSE (Real-Time Sequence of Errors). The hyperparameters of the network structure are adjusted based on the training results to select the optimal network, expressed as: RMSE is a quantitative indicator. y is the predicted value of the neural network. i represents the true value in the dataset, and m is the number of samples in the dataset; S4.11: Use a weight factor with a shorter step size as the input layer of the optimal network, and use the trained neural network prediction model to predict and control the optimal weight factor.

2. The adaptive fault-tolerant control method for a three-level ANPC inverter according to claim 1, characterized in that, Step S1 specifically involves: S1.1: The effective positive level P, zero level OU, zero level OL, and negative level N of the inverter output under switching transistor fault conditions are calculated based on Boolean operations and expressed as follows: Among them, the switching transistor T xn x = a, b, c; n = 1, 2, 3, 4, 5, 6 are in a normal state, corresponding to the logical value S. xn x = a, b, c; n = 1, 2, 3, 4, 5, 6 is 1, and the switching transistor T xn When in a fault state, the corresponding logic value S xn =0; P x x = a, b, c, OU x OL x and N x A logic value of 0 indicates that the inverter cannot output a certain level. x OU x OL x and N x A logic value of 1 indicates that the inverter can output an effective level. S1.2: Analyze the impact of a faulty switching transistor on the inverter output positive level P. When the switching transistor T... x1 or T x2 Fault, corresponding logic value S x1 or S x2 P is 0 x =0 indicates that the switching transistor T x1 or T x2 Under fault conditions, the inverter cannot output a positive level P; S1.3: Analyze the impact of a faulty switching transistor on the inverter output zero level OU. When the switching transistor T... x2 or T x5 Fault, corresponding logic value S x2 or S x5 0, OU x =0 indicates that the switching transistor T x2 or T x5 Under fault conditions, the inverter cannot output a zero level (OU). S1.4: Analyze the impact of a faulty switching transistor on the inverter output zero level OL. When the switching transistor T... x3 or T x6 Fault, corresponding logic value S x3 or S x6 0, OL x =0 indicates that the switching transistor T x3 or T x6 Under fault conditions, the inverter cannot output a zero-level (OL) signal. S1.5: Analyze the impact of a faulty switching transistor on the inverter output negative level N. When the switching transistor T... x3 or T x4 Fault, corresponding logic value S x3 or S x4 =0, N x =0 indicates that the switching transistor T x3 or T x4 Under fault conditions, the inverter cannot output a negative level N; S1.6: Based on the impact of different faulty switching transistors on the inverter output level, the inverter fault-tolerant modes are classified, and the switching transistor T... x1 -T x4 Under fault conditions, the inverter cannot output a positive level P or a negative level N. The inverter operates in a zero-level fault-tolerant mode, and the switching transistor T... x5 -T x6 In the event of a fault, the inverter cannot output a zero level (OU or OL), and the inverter operates in a zero-level loss fault-tolerant mode.

3. The adaptive fault-tolerant control method for a three-level ANPC inverter according to claim 2, characterized in that, Step S2 specifically involves: S2.1: Analyze the voltage vector output by the inverter in fault-tolerant mode. When the fault switch is located in phase a of the inverter, in the zero-level fault-tolerant mode, the inverter outputs 6 small vectors: ONN, OON, ONO, OPP, OOP, OPO, 2 medium vectors: ONP, OPN, and 1 zero vector: OOO; in the zero-level fault-tolerant mode, the inverter outputs 6 large vectors: PNN, PPN, NPN, NPP, NNP, PNP, 4 medium vectors: PON, NPO, NOP, PNO, 6 small vectors: POO, PPO, NON, NOO, NNO, POP, and 2 zero vectors: PPP, NNN. S2.2: Calculate the maximum stator voltage vector amplitude of the inverter output under both fault-tolerant modes, retaining the maximum stator voltage vector amplitude u of the inverter output under the zero-level fault-tolerant mode. r1 for In zero-level loss fault-tolerant mode, the maximum stator voltage vector amplitude u of the inverter output r2 for S2.3: Calculate the maximum modulation index of the inverter output voltage vector under the two fault-tolerant modes, expressed as: Among them, u dc This is the output voltage on the DC bus side of the inverter. S2.4: In the zero-level fault-tolerant mode, the modulation index k1 is 0.5, and the inverter system needs to operate at a reduced rate. In the zero-level fault-tolerant mode, the modulation index k2 is 1, and the inverter system does not need to operate at a reduced rate.

Images