A three-phase permanent magnet synchronous motor full-speed range drive control system and control method thereof

Abstract

The present invention discloses a full-speed domain drive control system for a three-phase permanent magnet synchronous motor and a control method thereof. The system is composed of a torque limitation module, a full-speed domain control neural network module, a current control module, a coordinate transformation module-Ⅰ, a coordinate transformation module-Ⅱ, a coordinate inverse transformation module, an SVPWM module, a SiC MOSFET drive module, a full-speed domain position and speed observation module, a weak magnetic feedback module and a three-phase permanent magnet synchronous motor. The control method is used to drive the control system. The present invention realizes accurate tracking of the optimal dq axis reference current operating point based on deep learning, and realizes precise control of three operating modes of MTPA, weak magnetic control and MTPV; adopts an improved moving horizon estimation algorithm to accurately estimate the rotor position and rotor speed within the full speed domain; uses SiC MOSFET as an inverter switch tube, and designs a new drive circuit to adapt to the operating conditions of electric vehicles.

Description

Technical Field

[0001] The present invention relates to the field of motor control technology, and in particular to a three-phase permanent magnet synchronous motor full-speed drive control system and a control method thereof. Background Art

[0002] At present, the motor drive control of electric vehicles has achieved certain research results, but there are still many problems. In order to adapt to the operating conditions of electric vehicles, when the motor runs below the base speed, MTPA control is usually used to reduce motor losses; when the speed gradually increases, the motor back electromotive force also increases, and the inverter output voltage is close to saturation, and weak magnetic control is used to further expand the speed range of the motor; and as the inverter output voltage reaches saturation, MTPV control is used to maintain a certain torque output capacity at ultra-high speeds. Based on this, from MTPA control to weak magnetic control to MTPV control, it is currently a research hotspot in the field of electric vehicle motor drive control.

[0003] The traditional offline calculation method based on lookup table (LUT) is currently the most popular and widely used method in the industry, in which the MTPA trajectory, MTPV trajectory and all weak magnetic working points are pre-calculated offline; in the online control implementation of the motor, the d-axis and q-axis current reference values ​​are obtained from the LUT according to the input torque command and flux limit, and loaded into the on-board computing system of the electric vehicle. However, this method requires a lot of storage space to save the lookup table; in addition, the size of the lookup table and the interpolation method will affect the control accuracy, and the fixed lookup table cannot cope with the uncertainty of motor parameters caused by differences in motor components and time changes. As for the online calculation method, due to the need to consider the motor operation between modes, it usually has a very complex structure, and its poor convergence and computational burden will affect the control accuracy and system stability.

[0004] In this regard, the control method based on machine learning can well solve the shortcomings of the above conventional methods. The use of neural networks to achieve offline training and online implementation can perfectly overcome the shortcomings of offline calculation and online calculation, while inheriting the advantages of both. However, there are still many defects in the current research on the application of neural networks in permanent magnet synchronous motors. "Adjustable flux permanent magnet synchronous motor full-speed efficiency optimal control magnetization state selection method and online control method" (patent publication number CN112468035A) provides an adjustable flux permanent magnet synchronous motor full-speed efficiency optimal control method, which integrates a neural network model in the control, and adjusts the magnetization state of the motor in real time, so that the motor works in the magnetization state with the best efficiency at each operating point in the full speed domain, and realizes the optimal control of the adjustable flux motor full-speed efficiency. However, the patent focuses on the adjustable flux permanent magnet synchronous motor, which changes the magnetization state of the motor by applying a pulse current in the armature winding to achieve air gap magnetic field regulation, and selects the magnetization state according to the principle of optimal efficiency under different magnetization states. This control method is not reproducible for conventional permanent magnet synchronous motors. "Current trajectory search method and online control method for optimal control of full-speed efficiency of permanent magnet synchronous motor" (patent publication number CN112468036A) provides a load flux model that takes into account the nonlinearity of the motor, and adopts a current trajectory search method for optimal control of full-speed efficiency based on a double golden section iteration method and an online control algorithm for optimal full-speed efficiency based on a neural network model, which can quickly and accurately achieve optimal control of full-speed efficiency of permanent magnet synchronous motors; however, the current angle iteration and current amplitude iteration methods used in this patent may have problems of slow convergence speed and poor convergence, affecting the control accuracy; and the specific implementation method of its neural network needs to be discussed. "A permanent magnet synchronous motor control method based on fuzzy neural network" (patent publication number CN115065285A) reduces overshoot and avoids oscillation by adjusting membership functions, learning rates, and control force parameters. At the same time, the use of detection modules is reduced through an adaptive sliding mode control algorithm, thereby enhancing the adaptability of the system. However, the model constructed in the research of this patent is relatively idealized, and although the adaptive sliding mode control algorithm used can well observe the rotation speed in the high-speed domain, it is difficult to ensure prediction accuracy in the low-speed domain.

[0005] In addition, in order to adapt to the speed regulation of the electric drive system of electric vehicles within the full speed range of the drive motor, a higher switching frequency is usually required, which results in higher switching losses. For this reason, the traditional IGBT-based inverter drive method needs to be changed. Summary of the invention

[0006] In view of the deficiencies of the existing technology in the above-mentioned background technology, the present invention proposes a three-phase permanent magnet synchronous motor full-speed range drive control system and a control method thereof, which can accurately track the optimal dq-axis reference current operating point based on a neural network, and realize precise control of the three operating modes of MTPA, weak magnetic control, and MTPV.

[0007] The technical solution to achieve the purpose of the present invention is:

[0008] A three-phase permanent magnet synchronous motor full-speed drive control system, comprising a torque limit module, a full-speed control neural network module, a current control module, a coordinate transformation module-Ⅰ, a coordinate transformation module-Ⅱ, a coordinate inverse transformation module, an SVPWM module, a SiC MOSFET drive module, a full-speed position and speed observation module, a weak magnetic feedback module and a three-phase permanent magnet synchronous motor;

[0009] The torque limit module adopts an LSTM neural network to ensure that the expected torque provided to the full-speed domain control neural network module is lower than the maximum allowable torque command under the flux limit. If the expected torque exceeds the limit range, it is maintained at the maximum allowable torque under the flux limit; the LSTM neural network is a long short-term memory neural network;

[0010] The full-speed control neural network module is based on a deep learning algorithm and uses a neural network that combines a variational autoencoder and a multi-head attention mechanism to ensure that the control system can accurately track the optimal dq axis reference current operating point and achieve smooth switching and precise control of the three operating modes of MTPA, weak magnetic control, and MTPV within the full speed range; the MTPA is the maximum torque current ratio control, and the MTPV is the maximum torque voltage ratio control;

[0011] The SiC MOSFET drive module includes a three-phase full-bridge control module and a power tube drive module; the three-phase full-bridge control module is a permanent magnet synchronous motor three-phase full-bridge inverter circuit; the power tube drive module includes six variable gate resistor gate drive circuits; the variable gate resistor gate drive circuit includes: a totem pole driver, switches S1 and S2, an on resistor , off resistance , Gate resistance , 4 window comparators and two control blocks; the SiC MOSFET is a silicon carbide metal oxide semiconductor field effect transistor;

[0012] The coordinate transformation module-I has a built-in coordinate transformation matrix , used to implement Clark transform and Park transform;

[0013] The coordinate transformation module-II has a built-in coordinate transformation matrix , used to implement Clark transformation;

[0014] The coordinate inverse transformation module has a built-in coordinate transformation matrix , used to implement Park's inverse transform;

[0015] The full-speed position and speed observation module consists of an information collection and preprocessing module, an initial prediction module, a loss function calculation and construction module, an optimization solution module, and an estimation module. It uses sensorless technology and estimates the rotor position and speed based on an improved moving horizon estimation algorithm.

[0016] Furthermore, the torque limit module adopts an LSTM neural network model, which is trained according to the maximum allowable torque data at the maximum current circle and the MTPV trajectory under different flux linkage limits; the maximum current circle is in the dq axis current plane. Under constraints, The trajectory formed is a circular curve.

[0017] Furthermore, the full-speed domain control neural network module is used to estimate the optimal dq axis reference current to achieve accurate tracking of the dq axis current operating point; the full-speed domain control neural network module includes a variational autoencoder module, a multi-head attention mechanism module, and a prediction module;

[0018] The full-speed control neural network module has two input features and two output features: the input features are the expected torque and flux linkage limit ,in, According to the inverter DC voltage and motor rotor speed The calculation formula is: ; The output characteristics are d-axis reference current and q-axis reference current ;

[0019] The preprocessing before input is as follows:

[0020] Step A1: normalize the input features as follows:

[0021] ,

[0022] in, is the normalized expected torque, is the normalized flux linkage limit, and are the mean and standard deviation of the expected torque, respectively; and are the mean and standard deviation of the flux linkage limit, respectively;

[0023] Step A2: Construct input feature vector , ;

[0024] The variational autoencoder module uses an encoder-decoder architecture to learn the latent space representation of the input features:

[0025] Step B1: The encoder inputs the feature vector Gaussian distribution mapped to latent space, output mean and standard deviation as follows:

[0026] ,

[0027] in, is the mean of the Gaussian distribution in the latent space, is the standard deviation of the Gaussian distribution in the latent space, and is the weight matrix, and is the bias vector;

[0028] Step B2: Use the reparameterization technique to sample latent variables from the Gaussian distribution as follows:

[0029] ,

[0030] in, is a hidden variable, From the standard normal distribution Random noise sampled in ; is the identity matrix, represents the covariance matrix of the standard normal distribution;

[0031] Step B3: The decoder converts the hidden variable Mapping back to the reconstructed feature space:

[0032] ,

[0033] in, To reconstruct the feature vector, and are the weight matrix and bias of the decoder respectively, is the activation function;

[0034] The feature quantity input by the multi-head attention mechanism module includes an input feature vector , hidden variables , reconstruct the feature vector , which is used to model global dependencies between input features and latent variables:

[0035] Step C1: The attention input is expressed as follows:

[0036] ,

[0037] Step C2: Calculate the attention weight using the query vector, key vector, and value vector:

[0038] ,

[0039] in, Used to represent the attention weight, , , are query vector, key vector and value vector respectively, is the dimension of the key vector;

[0040] Step C3: Concatenate the results of multiple attention heads:

[0041] ,

[0042] in, Used to represent the concatenation of attention heads, Represents the output of each attention head in the multi-head attention mechanism. For the output of the i-th attention head, ; Represents a linear transformation weight matrix used to project the concatenated vector into the output feature space;

[0043] Step C4, the output is normalized and residual connected:

[0044] ,

[0045] Said is the attention output representation;

[0046] The prediction module uses a fully connected network to map the attention-optimized features to the optimal dq-axis reference current:

[0047] Step D1, generate output through two layers of fully connected network:

[0048] ,

[0049] ,

[0050] in, is the output of the first fully connected layer, and are the weights and biases of the first layer of the fully connected network, and are the weights and biases of the second layer of the fully connected network, and are the normalized d-axis reference current and q-axis reference current respectively;

[0051] Step D2: Denormalize the normalized dq axis reference current to the actual value:

[0052] ,

[0053] in, and are the standard deviations of the d-axis and q-axis reference current training data, respectively; and are the means of the d-axis and q-axis reference current training data respectively; and is calculated for each sample separately, representing the i The d-axis reference current and q-axis reference current of samples.

[0054] Furthermore, the loss function in the variational autoencoder module consists of two parts: reconstruction loss and KL divergence:

[0055] A. The reconstruction loss is ;

[0056] B. KL divergence is ; d is a hidden variable The dimension size of The encoder generates i The mean of the distribution of the latent variables, The encoder generates i The standard deviation of the distribution of the latent variable;

[0057] Total loss function .

[0058] Further, in the SiC MOSFET drive module, the three-phase full-bridge control module includes first to sixth power tubes, wherein one end of the first power tube, the third power tube, and the fifth power tube are respectively connected to the three-phase windings A, B, and C, and the other ends are connected to the positive pole of the DC bus; one end of the second power tube, the fourth power tube, and the sixth power tube are respectively connected to the three-phase windings A, B, and C, and the other ends are connected to the negative pole of the DC bus; the first to sixth power tubes are SiCMOSFETs;

[0059] In the power tube driving module, Q1 of the totem pole driver is connected to the left side of switch S1, and Q2 is connected to the left side of switch S2; the switch S1 is connected to the turn-on resistor In parallel, switch S2 and the off resistor In parallel; the turn-on resistor With off resistance Right side and gate resistor The window comparators are arranged in pairs and are respectively connected to corresponding control blocks, and the control blocks are used to control the on and off of switches S1 and S2.

[0060] Furthermore, the full-speed range position and speed observation module includes an information collection and preprocessing module, an initial prediction module, a loss function calculation and construction module, an optimization solution module, and an estimation module, and uses an improved moving horizon estimation algorithm to estimate the rotor position and rotor speed;

[0061] The functions of the information acquisition and preprocessing module are: filtering the input voltage and current signals to remove high-frequency noise interference and improve the quality and stability of the signals; performing sample-and-hold operations to ensure that accurate voltage and current values ​​are obtained at the sampling moment; and t , construct the information vector according to the definition The information vector Integrate the time window from the past N The measured current from the sampling moment to the current moment And the past N Stator voltage from the sampling moment to the previous sampling moment ; ; The measurement current is taken into account the measurement noise n After the measurement current, , for time t, we have ; is the state vector at time t , ; , are the α-axis and β-axis components of the stator current in the stationary coordinate system; is the speed position coupled sinusoidal term, , is the speed position coupling cosine term, ; is the rotor speed, is the rotor position; is the measurement matrix, ;

[0062] The initial prediction module obtains the discrete state equation Compute the initial forecast for the current moving time window ; is the input voltage vector, subscript t is the sampling time, , are the α-axis and β-axis components of the stator voltage in the stationary coordinate system; is the discretized system matrix, is the discretized input matrix, is the discretized system noise term; the specific calculation formula is as follows:

[0063] ,

[0064] in, is the state at the previous moment, is the stator voltage at the previous moment, is the system noise at the previous moment;

[0065] The functions of the loss function calculation and construction module are: The measured current obtained from the state estimation and the predicted current based on the state estimation calculated using the state estimation value and the measurement matrix , calculate the values ​​of each item in the loss function; construct the matrix and vector , establish a quadratic optimization model with loss function , construct the constraint matrix and , the constraints are ; , is the state time vector, and the elements in the matrix are all the state vectors collected in the time window at the current sampling time; the loss function is ; , is the regularization parameter; the first term is the regularization term, Indicates the initial value of the moving time window With the initial forecast The square of the error; the second term is the residual term of the least squares estimate, Indicates the measured current, is the predicted current based on state estimation; the third term Indicates noise;

[0066] The optimization solution module uses the equivalent KKT condition to solve the quadratic optimization model constructed by the loss function calculation and construction module. Specifically, first, the constraint solution is used as the initial prediction matrix ; Secondly, calculate the Lagrange multiplier vector and the state variable correction vector ; Finally, the estimated value of the state variable is calculated ; The estimated value of the state variable ;

[0067] The estimation module estimates the value of the state variable from extract and , calculate the rotor position estimate and the estimated rotor speed :

[0068] ,

[0069] .

[0070] A control method for a three-phase permanent magnet synchronous motor full-speed drive control system, the steps are as follows:

[0071] Step S1, obtaining the current parameters of the three-phase permanent magnet synchronous motor, the current parameters including the three-phase current , , And three-phase voltage , , ; Input the current parameters into the coordinate transformation module-Ⅰ and the coordinate transformation module-Ⅱ;

[0072] Step S2, coordinate transformation module-II transforms the current parameters of the three-phase permanent magnet synchronous motor into a stationary coordinate system and outputs the αβ axis current signal , And voltage signal , ;

[0073] Step S3: The αβ axis current signal output from step S2 is , And voltage signal , Input full-speed position and speed observation module, output rotor position estimation value and the estimated rotor speed ; The estimated rotor position Input coordinate transformation module-Ⅰ and coordinate inverse transformation module;

[0074] Step S4: Input expected torque And the flux limit obtained in step S8 to the torque limit module, which Processing is performed and a processed expected torque signal is output;

[0075] Step S5: The expected torque signal processed in step S4 and the flux linkage limit obtained in step S8 are processed. Input to the full speed domain control neural network module, output d-axis reference current and q-axis reference current ;

[0076] Step S6, coordinate transformation module-I inputs three-phase current , , Converted into actual dq axis current value , ;

[0077] Step S7: The d-axis reference current obtained in step S5 is , q-axis reference current and the actual dq axis current value obtained in step S6 , Input current control module; the current control module , and , Compare and generate control signals through control algorithms to adjust the actual current to track the reference current; the current control module uses PI control to calculate the d-axis reference voltage , q-axis reference voltage , input weak magnetic feedback module and coordinate inverse transformation module; the PI control formula is: , ; is the proportionality coefficient, is the integral coefficient; , ;

[0078] Step S8: Obtaining DC bus voltage ,Will , the rotor speed estimation value output by the full-speed position and speed observation module , calculated by the current control module , Input weak magnetic feedback module, weak magnetic feedback module output flux limit ;

[0079] Step S9: The coordinate inverse transformation module transforms the coordinates processed in step S7 , Converted into αβ axis voltage signal , , input to SVPWM module;

[0080] Step S10: SVPWM module generates the voltage signal of the αβ axis according to the αβ axis voltage signal. , Generate pulse width modulation signal and input it to SiCMOSFET drive module;

[0081] Step S11, the power tube driving module in the SiC MOSFET driving module receives the pulse width modulation signal, and controls the on and off of the SiC MOSFET in the three-phase full-bridge control module according to the pulse width modulation signal, thereby controlling the three-phase winding current of the motor and driving the permanent magnet synchronous motor to operate.

[0082] Furthermore, the full speed control neural network module in step S5 needs to train the neural network model in the offline state before the motor starts. The training data of the model training consists of a series of data pairs including input and output. The training data is composed of samples, and the method for generating the training data is as follows:

[0083] Step S5-1, load the flux parameter data of the motor, obtain the nonlinear relationship between the d-axis and q-axis flux and the dq-axis current, and establish the corresponding table data , ;

[0084] Step S5-2: Generate a series of grid points on the 2D plane of the dq axis reference current , used to determine the flux linkage at different current values;

[0085] Step S5-3: For each Grid points, from and Get the dq axis flux in the lookup table and then calculate the expected torque and flux linkage limit ;

[0086] Step S5-4, determining the MTPA path, the MTPV path and the magnetic weakening path based on the data acquired in steps S5-1 to S5-3;

[0087] Step S5-5: Obtain all input and output data pairs in the MTPA path, MTPV path and weak magnetic path , saved in the data storage;

[0088] Step S5-6, expanding each training data sample along the MTPA path to the range of all flux linkage limits based on the expected torque;

[0089] Step S5-7: Expand each training data sample along the MTPV path to the range of all expected torques based on the flux limit.

[0090] Furthermore, in step S3, the full-speed range position and speed observation module is used to observe the rotor position and rotor speed to obtain the rotor position estimation value. and the estimated rotor speed The steps are as follows:

[0091] Step S3-1: Initialize the system when the motor starts and obtain the initial state of the motor at the initial moment and motor parameters, and set the sampling period , moving time window length N and the regularization parameter ; The initial state Including initial current signal , , initial voltage signal , ; The motor parameters include stator winding resistance , stator inductance , permanent magnet flux ; After initialization, enter the cyclic sampling moment from step S3-2;

[0092] Step S3-2, obtain the permanent magnet synchronous motor αβ axis current signal output by the coordinate transformation module-Ⅱ , And voltage signal , , input information collection and preprocessing module, output information vector ;

[0093] Step S3-3: Estimated value of the state variable obtained from the last sampling moment Get the previous state , stator voltage at the previous moment , input initial prediction module, output initial prediction ; In the first sampling period, the state of the previous moment , stator voltage at the previous moment Using the motor initial state obtained in step S3-1 ;

[0094] Step S3-4: The information vector obtained in step S3-2 is And the initial prediction obtained in step S3-3 Input loss function calculation and construction module, output matrix ,vector And the constraint matrix , The value of each element;

[0095] Step S3-5: The matrix output from step S3-4 ,vector And the constraint matrix , The values ​​of each element are input into the optimization solution module, and the estimated value of the state variable is output. ;

[0096] Step S3-6: The estimated value of the state variable output in step S3-5 Input estimation module, output rotor position estimation value , rotor speed estimate ;

[0097] Step S3-7, enter the next sampling moment, repeat the operations of step S3-2 to step S3-6, continuously update the state estimation, and realize real-time observation of the rotor position and speed.

[0098] Furthermore, in the SiC MOSFET driving module in step S11, the power tube driving module is responsible for controlling the on and off of the six power tubes in the three-phase full-bridge control module, and the three-phase full-bridge control module is responsible for controlling the three-phase winding of the permanent magnet synchronous motor;

[0099] The on and off of switches S1 and S2 in the power tube driver module are realized by comparing four voltage references with a window comparator. The window comparator continuously compares the gate-source voltage With four reference voltages: , , , Compare and apply two delay times and To control the on and off timing of switches S1 and S2;

[0100] A complete opening and closing process includes the following steps:

[0101] Step S11-1, when the input pulse width modulation signal is at a high level, Q1 and switch S1 in the totem pole driver are turned on at the same time;

[0102] Step S11-2: Reaching the threshold voltage When the drain current starts to rise, achieve When the switch S1 is turned off, the resistor Conducting to carry current;

[0103] Step S11-3: Exceed and reach Time, Delay ,at this time remain essentially unchanged;

[0104] Step S11-4: When the delay After the end, switch S1 is turned on again. Continue to rise until the power tube is fully turned on;

[0105] Step S11-5: When the input pulse width modulation signal is at a low level, the totem pole driver Q2 and the switch S2 are turned on at the same time. Start to descend;

[0106] Step S11-6: Start to descend to When the switch S2 is turned off, the current path is switched, and the turn-off resistor Current carrying, delay ,at this time remain essentially unchanged;

[0107] Step S11-7, when the delay After the end, Start to descend, and when it drops to When the power tube is turned off, the switch S2 is turned on again until the power tube shutdown transition is completed.

[0108] Compared with the prior art, the three-phase permanent magnet synchronous motor full-speed drive control system and control method described in the present invention have the following beneficial effects:

[0109] 1. A three-phase permanent magnet synchronous motor full-speed drive control system proposed in the present invention adopts technologies such as neural network and sensorless, which greatly reduces storage requirements at the software level, streamlines the electric drive system device architecture at the hardware level, effectively releases space, fully adapts to electric vehicle application requirements, and improves the overall performance and adaptability of the system; in addition, the close coordination between the modules can better meet the requirements of electric vehicles for motor control under different driving conditions.

[0110] 2. The full-speed domain control neural network module proposed in the present invention can well realize offline training and online implementation. It has the advantages of traditional lookup table methods such as good stability and high dynamic response, while requiring less storage space and can effectively deal with the problem of motor component parameters changing over time during the long-term operation of electric vehicles.

[0111] 3. The full-speed domain control neural network module proposed in the present invention adopts the optimal dq-axis reference current operating point tracking method based on neural network, based on the principles of MTPA control, weak magnetic field control, and MTPV control, with the expected torque and magnetic flux limit as input, and the dq-axis reference current as output. Compared with other control methods using neural networks, it can better realize the precise control of the three modes of MTPA control, weak magnetic field control, and MTPV control, and adapt to the actual operating conditions of electric vehicles; and this method can ensure that the optimal torque is generated in one sampling step. When the electric vehicle is running normally, the torque oscillation is smaller and the convergence time is shorter.

[0112] 4. The full-speed domain control neural network module proposed in the present invention is based on a deep learning algorithm and adopts a neural network that combines a variational autoencoder and a multi-head attention mechanism: the latent space representation of the variational autoencoder can capture the hidden variables of the complex state of the electric vehicle motor during operation; the multi-head attention mechanism can optimize the global dependency of features and enhance the accuracy of model prediction; finally, the fully connected layer is combined to realize real-time dq axis reference current prediction; this neural network architecture can adapt well to the operating conditions of the electric vehicle motor drive control system, has strong adaptability under high dynamic and complex conditions, and can well realize the full-speed domain speed regulation function.

[0113] 5. The training data acquisition method of the neural network proposed in the present invention does not rely on the traditional formula to solve the relationship between the d-axis and q-axis magnetic flux and the dq-axis current, but incorporates the changes in motor parameters into it, which can better capture the nonlinear effects caused by the changes in motor parameters; and the training data can be collected and stored randomly and irregularly, without following the clearly defined structured method usually required by traditional data search and interpolation algorithms, which is more convenient and more operational; after collecting the training data, the data expansion along the MTPA path and the MTPV path effectively improves the adaptive ability of the neural network.

[0114] 6. The full-speed position and speed observation module proposed in the present invention adopts sensorless technology and estimates the rotor position and speed based on an improved moving horizon estimation algorithm. Compared with the adaptive sliding mode control, extended Kalman filtering and other methods used in existing research to adapt to the needs of electric vehicles to observe the rotor position and speed in the high-speed domain, it maintains the accurate estimation capability in the high-speed domain while still having high estimation accuracy and robustness in the low-speed domain, and has faster dynamic response; in addition, as an estimation method based on optimization, the improved moving horizon estimation algorithm is easier to implement and adjust, does not require complex signal processing, and is more suitable for electric drive systems of electric vehicles.

[0115] 7. The SiC MOSFET module proposed in the present invention adopts SiC MOSFET to replace the traditional IGBT to design the motor driver, which is more suitable for the high switching frequency electric vehicle electric drive system, and effectively reduces the switching loss, and improves the efficiency, power density and dynamic performance of the permanent magnet synchronous motor drive system.

[0116] 8. The SiC MOSFET module proposed in the present invention optimizes the power tube switching process by optimizing the reference voltage and delay parameters of the power tube driving module, which can reduce the voltage and current overshoot and oscillation during the switching process and improve the switching performance; in the power tube driving module, the reference voltage and delay parameters are set to control the switching action timing, thereby controlling the gate resistance value, current and voltage slope during the Miller platform, which can reduce the power tube switching loss, reduce the heat generation of the device, and improve the efficiency and stability of the drive control system. BRIEF DESCRIPTION OF THE DRAWINGS

[0117] Figure 1 This is a block diagram of a full-speed drive control system for a three-phase permanent magnet synchronous motor proposed by the present invention;

[0118] Figure 2 It is a schematic diagram of realizing three modes of MTPA control, weak magnetic control and MTPV control in the present invention;

[0119] Figure 3 This is a schematic diagram of the SiC MOSFET driving module proposed in the present invention;

[0120] Figure 4 The topology diagram of the variable gate resistance gate drive circuit proposed by the present invention;

[0121] Figure 5 This is a schematic diagram of the moving time window in the full-speed domain position and speed observation module proposed by the present invention;

[0122] Figure 6 To simulate complex working conditions, a simulation diagram is provided to test the speed estimation capability of the full-speed position speed observation module;

[0123] Figure 7 To simulate complex working conditions, a simulation diagram is provided to test the rotor position estimation capability of the full-speed position and speed observation module;

[0124] Figure 8 Under pure speed regulation, the speed increases from 2000rpm to 10000rpm, and the system speed simulation diagram;

[0125] Fig. 9 The simulation diagram of the dq axis reference current in different modes when the speed increases from 2000rpm to 10000rpm under pure speed regulation;

[0126] Fig.10 This is a simulation diagram of the actual dq axis current when the speed increases from 2000rpm to 10000rpm under pure speed regulation. DETAILED DESCRIPTION

[0127] The following is in conjunction with the instruction manual Figure 1-Figure 10, the present invention is described based on embodiments, but the present invention is not limited to these embodiments. In the detailed description of the present invention below, some specific details are described in detail. For those skilled in the art, the present invention can be fully understood without the description of these details.

[0128] like Figure 1 As shown, a full-speed domain drive control system of a three-phase permanent magnet synchronous motor is composed of a torque limitation module, a full-speed domain control neural network module, a current control module, a coordinate transformation module-Ⅰ, a coordinate transformation module-Ⅱ, a coordinate inverse transformation module, an SVPWM module, a SiC MOSFET drive module, a full-speed domain position and speed observation module, a weak magnetic feedback module and a three-phase permanent magnet synchronous motor;

[0129] The torque limit module adopts an LSTM neural network to ensure that the expected torque provided to the full-speed domain control neural network module is lower than the maximum allowable torque command under the flux limit. If the expected torque exceeds the limit range, it is maintained at the maximum allowable torque under the flux limit; the LSTM neural network is a long short-term memory neural network;

[0130] The full-speed control neural network module is based on a deep learning algorithm and uses a neural network that combines a variational autoencoder and a multi-head attention mechanism to ensure that the control system can accurately track the optimal dq axis reference current operating point and achieve smooth switching and precise control of the three operating modes of MTPA, weak magnetic control, and MTPV within the full speed range; the MTPA is the maximum torque current ratio control, and the MTPV is the maximum torque voltage ratio control;

[0131] like Figure 2 As shown, the OA segment is the dq current trajectory of MTPA; at point A, it switches to the weak magnetic field control mode; the AB segment is the weak magnetic field control; at point B, it switches to the MTPV control mode; the BC segment is the dq current trajectory of MTPV;

[0132] like Figure 3 As shown, the SiC MOSFET drive module includes a three-phase full-bridge control module and a power tube drive module; the three-phase full-bridge control module is a permanent magnet synchronous motor three-phase full-bridge inverter circuit; the power tube drive module includes six variable gate resistor gate drive circuits; the variable gate resistor gate drive circuit includes: a totem pole driver, switches S1 and S2, an on resistor , off resistance , Gate resistance , 4 window comparators and two control blocks; the SiC MOSFET is a silicon carbide metal oxide semiconductor field effect transistor;

[0133] The coordinate transformation module-I has a built-in coordinate transformation matrix , used to implement Clark transform and Park transform;

[0134] The coordinate transformation module-II has a built-in coordinate transformation matrix , used to implement Clark transformation;

[0135] The coordinate inverse transformation module has a built-in coordinate transformation matrix , used to implement Park's inverse transform;

[0136] The full-speed position and speed observation module consists of an information collection and preprocessing module, an initial prediction module, a loss function calculation and construction module, an optimization solution module, and an estimation module. It uses sensorless technology and estimates the rotor position and speed based on an improved moving horizon estimation algorithm.

[0137] Furthermore, the torque limit module adopts an LSTM neural network model, which is trained according to the maximum allowable torque data at the maximum current circle and the MTPV trajectory under different flux linkage limits; the maximum current circle is in the dq axis current plane. Under constraints, The trajectory formed is a circular curve.

[0138] Furthermore, the full-speed domain control neural network module is used to estimate the optimal dq axis reference current to achieve accurate tracking of the dq axis current operating point; the full-speed domain control neural network module includes a variational autoencoder module, a multi-head attention mechanism module, and a prediction module;

[0139] The full-speed control neural network module has two input features and two output features: the input features are the expected torque and flux linkage limit ,in, According to the inverter DC voltage and motor rotor speed The calculation formula is: ; The output characteristics are d-axis reference current and q-axis reference current ;

[0140] The preprocessing before input is as follows:

[0141] Step A1: normalize the input features as follows:

[0142] ,

[0143] in, is the normalized expected torque, is the normalized flux linkage limit, and are the mean and standard deviation of the expected torque, respectively; and are the mean and standard deviation of the flux linkage limit, respectively;

[0144] Step A2: Construct input feature vector , ;

[0145] The variational autoencoder module uses an encoder-decoder architecture to learn the latent space representation of the input features:

[0146] Step B1: The encoder inputs the feature vector Gaussian distribution mapped to latent space, output mean and standard deviation as follows:

[0147] ,

[0148] in, is the mean of the Gaussian distribution in the latent space, is the standard deviation of the Gaussian distribution in the latent space, and is the weight matrix, and is the bias vector;

[0149] Step B2: Use the reparameterization technique to sample latent variables from the Gaussian distribution as follows:

[0150] ,

[0151] in, is a hidden variable, From the standard normal distribution Random noise sampled in ; is the identity matrix, represents the covariance matrix of the standard normal distribution;

[0152] Step B3: The decoder converts the hidden variable Mapping back to the reconstructed feature space:

[0153] ,

[0154] in, To reconstruct the feature vector, and are the weight matrix and bias of the decoder respectively, is the activation function;

[0155] The feature quantity input by the multi-head attention mechanism module includes an input feature vector , hidden variables , reconstruct the feature vector , which is used to model global dependencies between input features and latent variables:

[0156] Step C1: The attention input is expressed as follows:

[0157] ,

[0158] Step C2: Calculate the attention weight using the query vector, key vector, and value vector:

[0159] ,

[0160] in, Used to represent the attention weight, , , are query vector, key vector and value vector respectively, is the dimension of the key vector;

[0161] Step C3: Concatenate the results of multiple attention heads:

[0162] ,

[0163] in, Used to represent the concatenation of attention heads, Represents the output of each attention head in the multi-head attention mechanism. For the output of the i-th attention head, ; Represents a linear transformation weight matrix used to project the concatenated vector into the output feature space;

[0164] Step C4, the output is normalized and residual connected:

[0165] ,

[0166] Said is the attention output representation;

[0167] The prediction module uses a fully connected network to map the attention-optimized features to the optimal dq-axis reference current:

[0168] Step D1, generate output through two layers of fully connected network:

[0169] ,

[0170] ,

[0171] in, is the output of the first fully connected layer, and are the weights and biases of the first layer of the fully connected network, and are the weights and biases of the second layer of the fully connected network, and are the normalized d-axis reference current and q-axis reference current respectively;

[0172] Step D2: Denormalize the normalized dq axis reference current to the actual value:

[0173] ,

[0174] in, and are the standard deviations of the d-axis and q-axis reference current training data, respectively; and are the means of the d-axis and q-axis reference current training data respectively; and is calculated for each sample separately, representing the i The d-axis reference current and q-axis reference current of samples.

[0175] Furthermore, the loss function in the variational autoencoder module consists of two parts: reconstruction loss and KL divergence:

[0176] A. The reconstruction loss is ;

[0177] B. KL divergence is ; d is a hidden variable The dimension size of The encoder generates i The mean of the distribution of the latent variables, The encoder generates i The standard deviation of the distribution of the latent variable;

[0178] Total loss function .

[0179] Further, in the SiC MOSFET drive module, the three-phase full-bridge control module includes first to sixth power tubes, wherein one end of the first power tube, the third power tube, and the fifth power tube are respectively connected to the three-phase windings A, B, and C, and the other ends are connected to the positive pole of the DC bus; one end of the second power tube, the fourth power tube, and the sixth power tube are respectively connected to the three-phase windings A, B, and C, and the other ends are connected to the negative pole of the DC bus; the first to sixth power tubes are SiCMOSFETs;

[0180] In the power tube drive module, the variable gate resistor gate drive circuit is as follows: Figure 4 As shown, Q1 of the totem pole driver is connected to the left side of the switch S1, and Q2 is connected to the left side of the switch S2; the switch S1 is connected to the turn-on resistor In parallel, switch S2 and the off resistor In parallel; the turn-on resistor With off resistance Right side and gate resistor The window comparators are arranged in pairs and are respectively connected to corresponding control blocks, and the control blocks are used to control the on and off of switches S1 and S2.

[0181] Furthermore, the full-speed range position and speed observation module includes an information collection and preprocessing module, an initial prediction module, a loss function calculation and construction module, an optimization solution module, and an estimation module, and uses an improved moving horizon estimation algorithm to estimate the rotor position and rotor speed;

[0182] The functions of the information acquisition and preprocessing module are: filtering the input voltage and current signals to remove high-frequency noise interference and improve the quality and stability of the signals; performing sample-and-hold operations to ensure that accurate voltage and current values ​​are obtained at the sampling moment; and t , construct the information vector according to the definition The information vector Integrate the time window from the past N The measured current from the sampling moment to the current moment And the past N Stator voltage from the sampling moment to the previous sampling moment ; ; The measurement current is taken into account the measurement noise n After the measurement current, , for time t, we have ; is the state vector at time t , ; , are the α-axis and β-axis components of the stator current in the stationary coordinate system; is the speed position coupled sinusoidal term, , is the speed position coupling cosine term, ; is the rotor speed, is the rotor position; is the measurement matrix, ;

[0183] The initial prediction module obtains the discrete state equation Calculates the initial forecast for the current moving time window ,like Figure 5 shown; described is the input voltage vector, subscript t is the sampling time, , are the α-axis and β-axis components of the stator voltage in the stationary coordinate system; is the discretized system matrix, is the discretized input matrix, is the discretized system noise term; the specific calculation formula is as follows:

[0184] ,

[0185] in, is the state at the previous moment, is the stator voltage at the previous moment; is the system noise at the previous moment;

[0186] The functions of the loss function calculation and construction module are: The measured current obtained from the state estimation and the predicted current based on the state estimation calculated using the state estimation value and the measurement matrix , calculate the values ​​of each item in the loss function; construct the matrix and vector , establish a quadratic optimization model with loss function , construct the constraint matrix and , the constraints are ; , is the state time vector, and the elements in the matrix are all the state vectors collected in the time window at the current sampling time; the loss function is ; , is the regularization parameter; the first term is the regularization term, Indicates the initial value of the moving time window With the initial forecast The square of the error; the second term is the residual term of the least squares estimate, Indicates the measured current, is the predicted current based on state estimation; the third term Indicates noise;

[0187] The optimization solution module uses the equivalent KKT condition to solve the quadratic optimization model constructed by the loss function calculation and construction module. Specifically, first, the constraint solution is used as the initial prediction matrix ; Secondly, calculate the Lagrange multiplier vector and the state variable correction vector ; Finally, the estimated value of the state variable is calculated ; The estimated value of the state variable ;

[0188] The estimation module estimates the value of the state variable from extract and , calculate the rotor position estimate and the estimated rotor speed :

[0189] ,

[0190] .

[0191] A control method for a three-phase permanent magnet synchronous motor full-speed drive control system, the steps are as follows:

[0192] Step S1, obtaining the current parameters of the three-phase permanent magnet synchronous motor, the current parameters including the three-phase current , , And three-phase voltage , , ; Input the current parameters into the coordinate transformation module-Ⅰ and the coordinate transformation module-Ⅱ;

[0193] Step S2, coordinate transformation module-II transforms the current parameters of the three-phase permanent magnet synchronous motor into a stationary coordinate system and outputs the αβ axis current signal , And voltage signal , ;

[0194] Step S3: The αβ axis current signal output from step S2 is , And voltage signal , Input full-speed position and speed observation module, output rotor position estimation value and the estimated rotor speed ; The estimated rotor position Input coordinate transformation module-Ⅰ and coordinate inverse transformation module;

[0195] Step S4: Input expected torque And the flux limit obtained in step S8 to the torque limit module, which Processing is performed and a processed expected torque signal is output;

[0196] Step S5: The expected torque signal processed in step S4 and the flux linkage limit obtained in step S8 are processed. Input to the full speed domain control neural network module, output d-axis reference current and q-axis reference current ;

[0197] Step S6, coordinate transformation module-I inputs three-phase current , , Converted into actual dq axis current value , ;

[0198] Step S7: The d-axis reference current obtained in step S5 is , q-axis reference current and the actual dq axis current value obtained in step S6 , Input current control module; the current control module , and , Compare and generate control signals through control algorithms to adjust the actual current to track the reference current; the current control module uses PI control to calculate the d-axis reference voltage , q-axis reference voltage , input weak magnetic feedback module and coordinate inverse transformation module; the PI control formula is: , ; is the proportionality coefficient, is the integral coefficient; , ;

[0199] Step S8: Obtaining DC bus voltage , , the rotor speed estimation value output by the full-speed position and speed observation module , calculated by the current control module , Input weak magnetic feedback module, weak magnetic feedback module output flux limit ;

[0200] Step S9: The coordinate inverse transformation module transforms the coordinates processed in step S7 , Converted into αβ axis voltage signal , , input to SVPWM module;

[0201] Step S10: SVPWM module generates the voltage signal of the αβ axis according to the αβ axis voltage signal. , Generate pulse width modulation signal and input it to SiCMOSFET drive module;

[0202] Step S11, the power tube driving module in the SiC MOSFET driving module receives the pulse width modulation signal, and controls the on and off of the SiC MOSFET in the three-phase full-bridge control module according to the pulse width modulation signal, thereby controlling the three-phase winding current of the motor and driving the permanent magnet synchronous motor to operate.

[0203] Furthermore, the full speed control neural network module in step S5 needs to train the neural network model in the offline state before the motor starts. The training data of the model training consists of a series of data pairs including input and output. The training data is composed of samples, and the method for generating the training data is as follows:

[0204] Step S5-1, load the flux parameter data of the motor, obtain the nonlinear relationship between the d-axis and q-axis flux and the dq-axis current, and establish the corresponding table data , ;

[0205] Step S5-2: Generate a series of grid points on the 2D plane of the dq axis reference current , used to determine the flux linkage at different current values;

[0206] Step S5-3: For each Grid points, from and Get the dq axis flux in the lookup table and then calculate the expected torque and flux linkage limit ;

[0207] Step S5-4, determining the MTPA path, the MTPV path and the magnetic weakening path based on the data acquired in steps S5-1 to S5-3;

[0208] Step S5-5: Obtain all input and output data pairs in the MTPA path, MTPV path and weak magnetic path , saved in the data storage;

[0209] Step S5-6, expanding each training data sample along the MTPA path to the range of all flux linkage limits based on the expected torque;

[0210] Step S5-7: Expand each training data sample along the MTPV path to the range of all expected torques based on the flux limit.

[0211] Furthermore, in step S3, the full-speed range position and speed observation module is used to observe the rotor position and rotor speed to obtain the rotor position estimation value. and the estimated rotor speed The steps are as follows:

[0212] Step S3-1: Initialize the system when the motor starts and obtain the initial state of the motor at the initial moment and motor parameters, and set the sampling period , moving time window length N and the regularization parameter ; The initial state Including initial current signal , , initial voltage signal , ; The motor parameters include stator winding resistance , stator inductance , permanent magnet flux ; After initialization, enter the cyclic sampling moment from step S3-2;

[0213] Step S3-2, obtain the permanent magnet synchronous motor αβ axis current signal output by the coordinate transformation module-Ⅱ , And voltage signal , , input information collection and preprocessing module, output information vector ;

[0214] Step S3-3: Estimated value of the state variable obtained from the last sampling moment Get the previous state , stator voltage at the previous moment , input initial prediction module, output initial prediction ; In the first sampling period, the state of the previous moment , stator voltage at the previous moment Using the motor initial state obtained in step S3-1 ;

[0215] Step S3-4: The information vector obtained in step S3-2 is And the initial prediction obtained in step S3-3 Input loss function calculation and construction module, output matrix ,vector And the constraint matrix , The value of each element;

[0216] Step S3-5: The matrix output from step S3-4 ,vector And the constraint matrix , The values ​​of each element are input into the optimization solution module, and the estimated value of the state variable is output. ;

[0217] Step S3-6: The estimated value of the state variable output in step S3-5 Input estimation module, output rotor position estimation value , rotor speed estimate ;

[0218] Step S3-7, enter the next sampling moment, repeat the operations of step S3-2 to step S3-6, continuously update the state estimation, and realize real-time observation of the rotor position and speed.

[0219] Simulate complex working conditions to test the speed and rotor position estimation capabilities of the full-speed position and speed observation module. Figure 6 , Figure 7 shown.

[0220] Furthermore, in the SiC MOSFET driving module in step S11, the power tube driving module is responsible for controlling the on and off of the six power tubes in the three-phase full-bridge control module, and the three-phase full-bridge control module is responsible for controlling the three-phase winding of the permanent magnet synchronous motor;

[0221] The on and off of switches S1 and S2 in the power tube driver module are realized by comparing four voltage references with a window comparator. The window comparator continuously compares the gate-source voltage With four reference voltages: , , , Compare and apply two delay times and To control the on and off timing of switches S1 and S2;

[0222] A complete opening and closing process includes the following steps:

[0223] Step S11-1, when the input pulse width modulation signal is at a high level, Q1 and switch S1 in the totem pole driver are turned on at the same time;

[0224] Step S11-2: Reaching the threshold voltage When the drain current starts to rise, achieve When the switch S1 is turned off, the resistor Conducting to carry current;

[0225] Step S11-3: Exceed and reach Time, Delay ,at this time remain essentially unchanged;

[0226] Step S11-4: When the delay After the end, switch S1 is turned on again. Continue to rise until the power tube is fully turned on;

[0227] Step S11-5: When the input pulse width modulation signal is at a low level, the totem pole driver Q2 and the switch S2 are turned on at the same time. Start to descend;

[0228] Step S11-6: Start to descend to When the switch S2 is turned off, the current path is switched, and the turn-off resistor Current carrying, delay ,at this time remain essentially unchanged;

[0229] Step S11-7, when the delay After the end, Start to descend, and when it drops to When the power tube is turned off, the switch S2 is turned on again until the power tube shutdown transition is completed.

[0230] like Figure 8-Figure 10 As shown in the figure, given a torque of 20 N·m, the system performs a mode switching action under pure speed regulation, and successfully controls the speed from 2000 rpm to 10000 rpm in about 0.11 s. , The mode is switched. , Stable following, the speed regulation effect is very good.

[0231] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A three-phase permanent magnet synchronous motor full-speed drive control system, characterized in that: The system is composed of a torque limit module, a full-speed control neural network module, a current control module, a coordinate transformation module-I, a coordinate transformation module-II, a coordinate inverse transformation module, an SVPWM module, a SiC MOSFET drive module, a full-speed position and speed observation module, a weak magnetic feedback module and a three-phase permanent magnet synchronous motor; The torque limit module adopts an LSTM neural network to ensure that the expected torque provided to the full-speed domain control neural network module is lower than the maximum allowable torque command under the flux limit. If the expected torque exceeds the limit range, it is maintained at the maximum allowable torque under the flux limit; the LSTM neural network is a long short-term memory neural network; The full-speed control neural network module is based on a deep learning algorithm and uses a neural network that combines a variational autoencoder and a multi-head attention mechanism to ensure that the control system can accurately track the optimal dq axis reference current operating point and achieve smooth switching and precise control of the three operating modes of MTPA, weak magnetic control, and MTPV within the full speed range; the MTPA is the maximum torque current ratio control, and the MTPV is the maximum torque voltage ratio control; The SiC MOSFET drive module includes a three-phase full-bridge control module and a power tube drive module; the three-phase full-bridge control module is a permanent magnet synchronous motor three-phase full-bridge inverter circuit; the power tube drive module includes six variable gate resistor gate drive circuits; The variable gate resistor gate drive circuit includes: a totem pole driver, switches S1 and S2, an on resistor R on , turn-off resistance R off , Gate resistance R g_out , 4 window comparators and two control blocks; the SiC MOSFET is a silicon carbide metal oxide semiconductor field effect transistor; The coordinate transformation module-I has a built-in coordinate transformation matrix T Clark / Park , used to implement Clark transform and Park transform; The coordinate transformation module-II has a built-in coordinate transformation matrix T Clark , used to implement Clark transformation; The coordinate inverse transformation module has a built-in coordinate transformation matrix T 反Park , used to implement Park's inverse transform; The full-speed position and speed observation module consists of an information collection and preprocessing module, an initial prediction module, a loss function calculation and construction module, an optimization solution module, and an estimation module. It uses sensorless technology and estimates the rotor position and speed based on an improved moving horizon estimation algorithm.

2. A three-phase permanent magnet synchronous motor full-speed drive control system according to claim 1, characterized in that: The torque limit module adopts an LSTM neural network model, which is trained according to the maximum allowable torque data at the maximum current circle and the MTPV trajectory under different flux linkage limits; the maximum current circle is within the dq axis current plane. Under the constraint, i max The trajectory formed is a circular curve.

3. A three-phase permanent magnet synchronous motor full-speed drive control system according to claim 1, characterized in that: The full-speed domain control neural network module is used to estimate the optimal dq axis reference current and realize accurate tracking of the dq axis current working point; the full-speed domain control neural network module includes a variational autoencoder module, a multi-head attention mechanism module, and a prediction module; The full-speed control neural network module has two input features and two output features: the input features are the expected torque T em * and flux linkage limit λ lim,ωe , where λ lim,ωe According to the inverter DC voltage V dc and the motor rotor speed ω e The calculation formula is: The output characteristics are respectively the d-axis reference current i d * and q-axis reference current i q * ; The preprocessing before input is as follows: Step A1: normalize the input features as follows: in, is the normalized expected torque, is the normalized flux limit, μ T and σ T are the mean and standard deviation of the expected torque respectively; μ λ and σ λ are the mean and standard deviation of the flux linkage limit, respectively; Step A2: construct the input feature vector x. The variational autoencoder module uses an encoder-decoder architecture to learn the latent space representation of the input features: Step B1: The encoder maps the input feature vector x to a Gaussian distribution in the latent space, and the output mean μ and standard deviation σ are as follows: μ=f μ (x)=W μ x+b μ ,log(σ 2 )=f σ (x)=W σ x+b σ Among them, μ is the mean of the Gaussian distribution of the latent space, σ is the standard deviation of the Gaussian distribution of the latent space, and W μ and W σ is the weight matrix, b μ and b σ is the bias vector; Step B2: Use the reparameterization technique to sample latent variables from the Gaussian distribution as follows: Among them, z is a hidden variable, ∈ is a standard normal distribution The random noise sampled in; I is the identity matrix, represents the covariance matrix of the standard normal distribution; Step B3: The decoder maps the latent variable z back to the reconstructed feature space: in, To reconstruct the eigenvector, W dec and b dec are the weight matrix and bias of the decoder respectively, and ReLU is the activation function; The feature quantities input by the multi-head attention mechanism module include input feature vector x, hidden variable z, reconstructed feature vector Used to model global dependencies between input features and latent variables: Step C1: The attention input is expressed as follows: Step C2: Calculate the attention weight using the query vector, key vector, and value vector: Among them, Attention is used to represent the attention weight, Q, K, and V are query vectors, key vectors, and value vectors respectively, and d k is the dimension of the key vector; Step C3: Concatenate the results of multiple attention heads: MultiHead(Q,K,V)=Concat(head1,…,head h )W O Among them, MultiHead is used to represent the splicing of each attention head, head represents the output of each attention head in the multi-head attention mechanism, and for the output of the i-th attention head, there is head i =Attention(Q i ,K i ,V i );W O Represents a linear transformation weight matrix used to project the concatenated vector into the output feature space; Step C4, the output is normalized and residual connected: h attn =LayerNorm(h input +MultiHead) The h attn is the attention output representation; The prediction module uses a fully connected network to map the attention-optimized features to the optimal dq-axis reference current: Step D1, generate output through two layers of fully connected network: h (1) =ReLU(W1h attn +b1) Among them, h (1) is the output of the first fully connected layer, W1 and b1 are the weight and bias of the first layer of the fully connected network, W2 and b2 are the weight and bias of the second layer of the fully connected network, and are the normalized d-axis reference current and q-axis reference current respectively; Step D2: Denormalize the normalized dq axis reference current to the actual value: in, and are the standard deviations of the d-axis and q-axis reference current training data, respectively; and are the means of the d-axis and q-axis reference current training data respectively; i * d,i and i * q,i are calculated separately for each sample and represent the d-axis reference current and q-axis reference current of the i-th sample respectively.

4. A three-phase permanent magnet synchronous motor full-speed drive control system according to claim 3, characterized in that: The loss function in the variational autoencoder module consists of two parts: reconstruction loss and KL divergence. composition: A. The reconstruction loss is B. KL divergence is The d is the dimension of the latent variable z, μ i is the distribution mean of the i-th latent variable generated by the encoder, σ i The standard deviation of the distribution of the i-th latent variable generated by the encoder; The total loss function L VAE =L rec +L KL .

5. A three-phase permanent magnet synchronous motor full-speed drive control system according to claim 1, characterized in that: In the SiC MOSFET drive module, the three-phase full-bridge control module includes first to sixth power tubes, wherein one end of the first power tube, the third power tube, and the fifth power tube are respectively connected to the three-phase windings A, B, and C, and the other ends are connected to the positive pole of the DC bus; one end of the second power tube, the fourth power tube, and the sixth power tube are respectively connected to the three-phase windings A, B, and C, and the other ends are connected to the negative pole of the DC bus; the first to sixth power tubes are SiC MOSFETs; In the power tube driving module, Q1 of the totem pole driver is connected to the left side of switch S1, and Q2 is connected to the left side of switch S2; the switch S1 is connected to the turn-on resistor R on In parallel, switch S2 and the turn-off resistor R off In parallel; the turn-on resistor R on With the off resistance R off The right side and the gate resistor R g_out The window comparators are arranged in pairs and are respectively connected to corresponding control blocks, and the control blocks are used to control the on and off of switches S1 and S2.

6. A three-phase permanent magnet synchronous motor full-speed drive control system according to claim 1, characterized in that: The full-speed domain position and speed observation module includes an information collection and preprocessing module, an initial prediction module, a loss function calculation and construction module, an optimization solution module, and an estimation module, and uses an improved moving horizon estimation algorithm to estimate the rotor position and rotor speed; The functions of the information acquisition and preprocessing module are: filtering the input voltage and current signals to remove high-frequency noise interference and improve the quality and stability of the signals; Perform a sample-and-hold operation to ensure accurate voltage and current values ​​at the sampling moment; construct an information vector according to the definition based on the current sampling moment t The information vector The measured current y in the time window from the Nth sampling moment in the past to the current moment is integrated t-N ,…,y t And the stator voltage v from the Nth sampling moment to the previous sampling moment t-N ,…,v t-1 ; The measurement current is the measurement current after taking into account the measurement noise n, y = [i α +ni β +n] T , for time t, there is y t =Cx t +n t ; the x t is the state vector x at time t, i α 、i β are the α-axis and β-axis components of the stator current in the stationary coordinate system; is the speed position coupled sinusoidal term, is the speed position coupling cosine term, The ω e is the rotor speed, θ e is the rotor position; C is the measurement matrix, The initial prediction module obtains the discrete state equation x t+1 =A d x t +B d v t +ξ t Calculates the initial forecast for the current moving time window The v t =[v α ,v β ] T is the input voltage vector, subscript t is the sampling time, v α 、v β are the α-axis and β-axis components of the stator voltage in the stationary coordinate system; d is the discretized system matrix, B d is the discretized input matrix, ξ t is the discretized system noise term; the specific calculation formula is as follows: Among them, x t-N-1,t-1 is the state at the previous moment, v t-N-1 is the stator voltage at the previous moment, ξ t-N-1 is the system noise at the previous moment; The functions of the loss function calculation and construction module are: The measured current obtained in the state estimation and the predicted current Cx calculated based on the state estimation using the state estimation value and the measurement matrix i,t , calculate the values ​​of each item in the loss function; construct the matrix Q and vector c T , establish a quadratic optimization model with loss function Construct constraint matrix A D and B D , the constraint condition is A D X t =B D ; is the state time vector, and the elements in the matrix are all the state vectors collected in the time window at the current sampling time; the loss function is The η>0 is a regularization parameter; the first term is the regularization term, Represents the initial value x of the moving time window t-N,t With the initial forecast The square of the error; the second term is the residual term of the least squares estimate, y i,t Indicates the measured current, Cx i,t is the predicted current based on state estimation; the third term Indicates noise; The optimization solution module uses the equivalent KKT condition to solve the quadratic optimization model constructed by the loss function calculation and construction module. Specifically, first, the constraint solution is used as the initial prediction matrix Next, calculate the Lagrange multiplier vector λ * and the state variable correction vector p; finally, the estimated value X of the state variable is calculated * ; The estimated value of the state variable The estimation module estimates the value of the state variable X * extract and Calculate rotor position estimate and the estimated rotor speed 7. A three-phase permanent magnet synchronous motor full-speed range drive control method, applicable to a three-phase permanent magnet synchronous motor full-speed range drive control system according to any one of claims 1 to 6, characterized in that: The method steps are as follows: Step S1, obtaining the current parameters of the three-phase permanent magnet synchronous motor, the current parameters including the three-phase current i a 、i b 、i c And the three-phase voltage v a 、v b 、v c ; Input the current parameters into the coordinate transformation module-Ⅰ and the coordinate transformation module-Ⅱ; Step S2, coordinate transformation module-II transforms the current parameters of the three-phase permanent magnet synchronous motor into a stationary coordinate system and outputs the αβ axis current signal i α 、i β And the voltage signal v α 、v β ; Step S3: The αβ axis current signal i outputted from step S2 is α 、i β And the voltage signal v α 、v β Input full-speed position and speed observation module, output rotor position estimation value and the estimated rotor speed The rotor position estimate Input coordinate transformation module-Ⅰ and coordinate inverse transformation module; Step S4: Input the expected torque T em * And the flux limit λ obtained in step S8 lim,ωe To the torque limit module, the torque limit module is T em * Processing is performed and a processed expected torque signal is output; Step S5: The expected torque signal processed in step S4 and the flux linkage limit λ obtained in step S8 are processed into lim,ωe Input to the full-speed domain control neural network module, output d-axis reference current i d * and q-axis reference current i q * ; Step S6, coordinate transformation module-I inputs three-phase current i a 、i b 、i c Converted into actual dq axis current value i d 、i q ; Step S7: The d-axis reference current i obtained in step S5 is d * , q-axis reference current i q * and the actual dq axis current value i obtained in step S6 d 、i q Input current control module; the current control module converts i d * 、i q * with i d 、i q Compare and generate control signals through control algorithms to adjust the actual current to track the reference current; The current control module uses PI control to calculate the d-axis reference voltage v d * , q-axis reference voltage v q * , input weak magnetic feedback module and coordinate inverse transformation module; the PI control formula is v d * =K p Δi d +K i ∫Δi d dt、v q * =K p Δi q +K i ∫Δi q dt; K p is the proportionality coefficient, the K i is the integral coefficient; the Δi d =i d * -i d , Δi q =i q * -i q ; Step S8: Obtain the DC bus voltage V dc ,Will The rotor speed estimate output by the full-speed position and speed observation module The current control module calculates v d * 、v q * Input weak magnetic feedback module, weak magnetic feedback module output flux limit λ lim,ωe ; Step S9: The coordinate inverse transformation module transforms the v processed in step S7 d * 、v q * Converted into αβ axis voltage signal v α 、v β , input to SVPWM module; Step S10: The SVPWM module generates a voltage signal v according to the αβ axis voltage signal v α 、v β Generate pulse width modulation signal and input it to SiC MOSFET driver module; Step S11, the power tube driving module in the SiC MOSFET driving module receives the pulse width modulation signal, and controls the on and off of the SiC MOSFET in the three-phase full-bridge control module according to the pulse width modulation signal, thereby controlling the three-phase winding current of the motor and driving the permanent magnet synchronous motor to operate.

8. A three-phase permanent magnet synchronous motor full-speed range drive control method according to claim 7, characterized in that: The full speed control neural network module in step S5 needs to train the neural network model in the offline state before the motor starts. The training data of the model training consists of a series of input and output data pairs [(T em * ,λ lim,ωe );(i d * ,i q * )], the training data is generated as follows: Step S5-1, load the flux parameter data of the motor, obtain the nonlinear relationship between the d-axis and q-axis flux and the dq-axis current, and establish the corresponding table data λ d (i d ,i q ), λ q (i d ,i q ); Step S5-2: Generate a series of grid points (i d * ,i q * ), used to determine the flux linkage at different current values; Step S5-3: for each (i d * ,i q * ) grid points, from λ d (i d ,i q ) and λ q (i d ,i q ) to obtain the dq axis flux in the lookup table, and then calculate the expected torque T em * and flux linkage limit λ lim,ωe ; Step S5-4, determining the MTPA path, the MTPV path and the magnetic weakening path based on the data acquired in steps S5-1 to S5-3; Step S5-5, obtain all input and output data pairs in the MTPA path, MTPV path and weak magnetic path [(T em * ,λ lim,ωe );(i d * ,i q * )], save it in the data storage; Step S5-6, expanding each training data sample along the MTPA path to the range of all flux linkage limits based on the expected torque; Step S5-7: Expand each training data sample along the MTPV path to the range of all expected torques based on the flux limit.

9. A three-phase permanent magnet synchronous motor full-speed range drive control method according to claim 7, characterized in that: In step S3, the full-speed range position and speed observation module is used to observe the rotor position and rotor speed to obtain the rotor position estimation value. and the estimated rotor speed The steps are as follows: Step S3-1: Initialize the system when the motor starts, obtain the initial state x0 and motor parameters of the motor at the initial moment, and set the sampling period T c , moving time window length N and regularization parameter η; The initial state x0 includes an initial current signal i α (0),i β (0), initial voltage signal v α (0), v β (0); The motor parameters include stator winding resistance R, stator inductance L, permanent magnet flux λ m ; After initialization, enter the cyclic sampling moment from step S3-2; Step S3-2, obtain the permanent magnet synchronous motor αβ axis current signal i output by the coordinate transformation module-II α 、i β And the voltage signal v α 、v β , input information collection and preprocessing module, output information vector Step S3-3: The estimated value X of the state variable obtained from the last sampling moment * Get the previous state x t-N-1,t-1 , the stator voltage v at the previous moment t-N-1 , input initial prediction module, output initial prediction In the first sampling period, the state x at the previous moment t-N-1,t-1 , the stator voltage v at the previous moment t-N-1 Using the motor initial state x0 obtained in step S3-1; Step S3-4: The information vector obtained in step S3-2 is And the initial prediction obtained in step S3-3 Input loss function calculation and construction module, output matrix Q, vector c T And the constraint matrix A D , B D The value of each element; Step S3-5: The matrix Q and vector c outputted from step S3-4 are T And the constraint matrix A D , B D The values ​​of each element are input into the optimization solution module, and the estimated value of the state variable X is output. * ; Step S3-6: The estimated value X of the state variable output in step S3-5 is * Input estimation module, output rotor position estimation value Estimated rotor speed Step S3-7, enter the next sampling moment, repeat the operations of step S3-2 to step S3-6, continuously update the state estimation, and realize real-time observation of the rotor position and speed.

10. The full-speed range drive control method of a three-phase permanent magnet synchronous motor according to claim 7, characterized in that: In the SiC MOSFET driving module in step S11, the power tube driving module is responsible for controlling the on and off of the six power tubes in the three-phase full-bridge control module, and the three-phase full-bridge control module is responsible for controlling the three-phase windings of the permanent magnet synchronous motor; The on and off of the switches S1 and S2 in the power tube drive module are realized by comparing the four voltage references of the window comparator. The window comparator continuously compares the gate-source voltage V gs With four reference voltages: V 1_High 、V 1_Low 、V 2_High 、V 2_Low comparison and apply two delay times t d_1 and t d_2 To control the on and off timing of switches S1 and S2; A complete opening and closing process includes the following steps: Step S11-1, when the input pulse width modulation signal is at a high level, Q1 and switch S1 in the totem pole driver are turned on at the same time; Step S11-2: When V gs Reaching the threshold voltage V gs_th When the drain current starts to rise, V gs Reach V 1_Low When the switch S1 is turned off, the resistor R on Conducting to carry current; Step S11-3: When V gs Exceed V 1_Low and reaches V 1_High Time, delay t d_1 , at this time V gs remain unchanged; Step S11-4, when the delay t d_1 After the end, switch S1 is turned on again, V gs Continue to rise until the power tube is fully turned on; Step S11-5: When the input pulse width modulation signal is at a low level, the totem pole driver Q2 and the switch S2 are turned on at the same time, V gs Start to descend; Step S11-6: When V gs Starts to drop to V 2_High When the switch S2 is turned off, the current path is switched, and the turn-off resistor R off Current carrying, delay t d_2 , at this time V gs remain unchanged; Step S11-7, when the delay t d_2 After the end, V gs starts to decrease, and when it drops to V 2_Low When the power tube is turned off, the switch S2 is turned on again until the power tube shutdown transition is completed.

Images