CNN-LSTM-based motor torque ripple suppression method

Abstract

The invention discloses a CNN-LSTM-based motor torque ripple suppression method. The method comprises the following steps: S1, establishing a mathematical model of a permanent magnet synchronous motor; s2, establishing an inverter model considering an inverter dead zone and tube voltage drop; s3, establishing a harmonic compensation voltage prediction model based on CNN-LSTM to optimize a DQ-axis stator voltage, reduce current harmonics and suppress motor vibration and torque fluctuation; and S4, the harmonic compensation voltage prediction model based on CNN-LSTM in the step S3 is embedded into a motor control system, and torque fluctuation of the motor is suppressed. According to the invention, the stator voltage of the DQ shaft is effectively optimized, fifth and seventh current harmonics which have great influence on the torque ripple are reduced, the torque ripple is suppressed, control parameters are reduced, the response speed of the system is improved, and the suppression effect of the motor on the torque ripple at a high rotating speed is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of permanent magnet synchronous motors, and in particular to a method for suppressing motor torque fluctuation based on CNN-LSTM. Background Art

[0002] Permanent magnet synchronous motor (PMSM) has been widely used in the electric drive system of electric vehicles due to its advantages such as high power density, high efficiency and wide speed regulation range. However, the torque fluctuation of the permanent magnet synchronous motor in the electric drive system is an important cause of electromagnetic noise and torsional vibration problems. On the one hand, the electromagnetic noise of the electric drive system is different from the noise of traditional automobile gasoline engines. Its frequency is higher and it is easy to make people feel uncomfortable. On the other hand, serious motor torque fluctuation will cause the hysteresis and eddy current loss of the motor to increase, increase the electrical stress of the motor material, and shorten the service life of the motor. The smoothness of torque output has become one of the important indicators to measure the dynamic performance and steady-state performance of motor control. How to suppress the torque fluctuation of permanent magnet synchronous motor has become one of the current research hotspots.

[0003] In recent years, scholars have devoted themselves to studying the impact of inverter dead time on output current harmonics. From the perspective of control, the current harmonic distortion and torque fluctuation problems caused by the inverter dead time effect are mainly suppressed by the following three methods. The first type is the iterative learning control method, which is a model-free control method that "self-learns" the periodic deviation signal in a memory manner; the second type is the model compensation method, which establishes a voltage distortion model caused by the dead time of the voltage source inverter (VSI) and uses feedforward control for compensation; the third type is the current harmonic injection method, which uses a control algorithm to suppress the distortion of current harmonics.

[0004] Each of the above methods has its limitations. The iterative learning control method is difficult to provide accurate compensation information when the speed is uncertain; the model compensation method relies on the judgment of the current direction, and misjudgment is prone to occur near the current zero point; the current harmonic injection method adds a PI controller, which has many adjustment parameters and is more difficult to design; and the above three methods have poor suppression effects on motor torque fluctuations at high speeds. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a method for suppressing motor torque fluctuations based on CNN-LSTM. By establishing a harmonic compensation voltage prediction model based on CNN-LSTM, the DQ axis stator voltage is optimized, the current harmonics are reduced, and the torque fluctuation is suppressed. At the same time, the control parameters are reduced, the system response speed is improved, and the motor's suppression effect on torque fluctuations at high speed is enhanced.

[0006] The technical solution adopted by the present invention to solve the technical problem is: a method for suppressing motor torque fluctuation based on CNN-LSTM, comprising the following steps:

[0007] S1: Establish the mathematical model of permanent magnet synchronous motor;

[0008] S2: Establish an inverter model that takes into account the inverter dead zone and tube voltage drop;

[0009] S3: Establish a harmonic compensation voltage prediction model based on CNN-LSTM to optimize the DQ axis stator voltage, reduce current harmonics, and suppress motor vibration and torque fluctuation; specifically, the following steps are included:

[0010] S31: Establishing current harmonic model;

[0011] S32: Collect training data;

[0012] S33: Establish a harmonic compensation voltage prediction model based on CNN-LSTM;

[0013] S4: Embed the CNN-LSTM-based harmonic compensation voltage prediction model of step S3 into the motor control system to suppress the torque fluctuation of the motor.

[0014] Furthermore, in step S31, the specific method of establishing the current harmonic model is as follows:

[0015] The three-phase voltage u of the permanent magnet synchronous motor a 、u b 、u c Transformed to the DQ rotation coordinate system, we get:

[0016]

[0017] In the formula, u d1 and u q1 are the d-axis and q-axis components of the fundamental voltage in the DQ rotating coordinate system, u5 and u7 are the amplitudes of the 5th and 7th harmonic voltages in the DQ rotating coordinate system, θ5 and θ7 are the initial phase angles of the 5th and 7th harmonic voltages, and ω is the rotor electrical angular velocity;

[0018] The mathematical equation of the motor current in the DQ rotating coordinate system is:

[0019]

[0020] In the formula, i d1 and i q1are the d-axis and q-axis components of the fundamental current in the DQ rotating coordinate system, i5 and i7 are the amplitudes of the 5th and 7th harmonic currents in the DQ rotating coordinate system, θ5 and θ7 are the initial phase angles of the 5th and 7th harmonic voltages, and ω is the rotor electrical angular velocity;

[0021] The stator voltage equation of the motor containing harmonic components is:

[0022]

[0023] In the formula, i d1 and i q1 are the d-axis and q-axis components of the fundamental current in the DQ rotating coordinate system, i5 and i7 are the amplitudes of the 5th and 7th harmonic currents in the DQ rotating coordinate system, R s is the stator resistance, L d and L q are the d-axis and q-axis inductances, θ5 and θ7 are the initial phase angles of the 5th and 7th harmonic voltages, ω is the rotor electrical angular velocity;

[0024] Since the fundamental wave is an AC component in the DQ rotating coordinate system of the 5th and 7th harmonics, the harmonic steady-state voltage equation in the 5th DQ rotating coordinate system is:

[0025]

[0026] In the formula, i d5 and i q5 are the d-axis and q-axis DC current components under the 5th DQ rotation coordinate axis, R s is the stator resistance, L d and L q are the d-axis and q-axis inductances respectively, ω is the rotor electrical angular velocity;

[0027] Similarly, the harmonic steady-state voltage equation in the 7th DQ rotating coordinate system is:

[0028]

[0029] In the formula, i d7 and i q7 are the d-axis and q-axis DC current components under the 7th DQ rotation coordinate axis, R s is the stator resistance, L d and L q are the d-axis and q-axis inductances respectively, ω is the rotor electrical angular velocity;

[0030] The three-phase current of the motor passes through the fifth and seventh rotation coordinate systems of the DQ axis, and then the fifth and seventh current harmonic components on the DQ axis are obtained by low-pass filtering;

[0031] By superimposing the voltages generated in the two lines of the harmonic steady-state voltage equation and the PI controller, the harmonic compensation voltage of each harmonic current in the rotating coordinate system is obtained;

[0032] Then sum the harmonic voltages of each current to obtain the harmonic compensation voltage in the DQ coordinate system. and

[0033] Right now:

[0034] Further, the specific method of collecting training data in step S32 is as follows: the initial speed of the motor is set to 50 rpm, and the motor torque is gradually increased from 0 Nm to the maximum torque, and the increment is set to 10 Nm; data is collected for 1 second in each working condition; the speed is increased to the maximum speed in increments of 50 rpm, and data is repeatedly recorded in each working condition; the d-axis current i under each working condition is recorded d , q-axis current i q , rotor electrical angular velocity θ, rotor angle ω and motor reference torque Data; then a large amount of experimental data is normalized and used to train CNN-LSTM.

[0035] Further, in step S33, the specific method of establishing the harmonic compensation voltage prediction model based on CNN-LSTM is:

[0036] Combine the convolutional neural network with the long short-term neural network to establish a CNN-LSTM network model; training the CNN-LSTM network includes two stages:

[0037] In the first stage, i d 、i q , θ, ω and The data is input into the convolutional neural network. In the second stage, the features extracted by the convolutional neural network are passed to the long-term short-term neural network, and the fully connected layer is used as the prediction result.

[0038] The convolutional neural network is used to perform initial feature recognition on the input data, and then the long-term and short-term neural network is used to iteratively train the selected feature values. The final output prediction result is the harmonic compensation voltage. d 、i q , θ, ω and is the input variable, and The mapping relationship for the output.

[0039] The present invention has the following beneficial effects compared with the prior art:

[0040] The present invention aims at the torque fluctuation problem of a permanent magnet synchronous motor for a vehicle caused by nonlinear factors of an inverter, and proposes a torque fluctuation suppression method based on CNN-LSTM; this method fully considers the current harmonics caused by the dead zone of the inverter and the tube voltage drop, and reduces the 5th and 7th current harmonics that have a greater impact on the torque fluctuation through a harmonic compensation voltage prediction model based on CNN-LSTM, thereby suppressing the torque fluctuation, reducing the control parameters, improving the system response speed, and enhancing the suppression effect of the motor on the torque fluctuation at high speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a flow chart of a motor torque fluctuation suppression method based on CNN-LSTM of the present invention.

[0042] Figure 2 It is a topological diagram of the inverter in the present invention.

[0043] Figure 3 It is the voltage waveform diagram of the A-phase bridge arm in the present invention.

[0044] Figure 4 It is a module for extracting the fifth and seventh current harmonics of permanent magnet synchronous motors.

[0045] Figure 5 It is the block diagram of the 5th and 7th current harmonic decoupling control of permanent magnet synchronous motor.

[0046] Figure 6 It is a CNN-LSTM network.

[0047] Figure 7 This is a flow chart of the training process of CNN-LSTM.

[0048] Figure 8 This is the schematic diagram of the motor torque fluctuation suppression method based on CNN-LSTM. DETAILED DESCRIPTION

[0049] The present invention is further described in detail below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments are only for illustrating the present invention, and are not intended to limit the scope of protection of the present invention. The accompanying drawings are only for illustrative purposes and cannot be understood as limiting the present invention; it is understandable to those skilled in the art that some well-known structures and their descriptions in the accompanying drawings may be omitted.

[0050] The invention discloses a motor torque fluctuation suppression method based on CNN-LSTM.

[0051] Reference Figure 1 , the motor torque fluctuation suppression method based on CNN-LSTM includes the following steps:

[0052] S1: Establish the mathematical model of permanent magnet synchronous motor.

[0053] In this embodiment, the basic parameters of the selected permanent magnet synchronous motor are maximum power 160kW, maximum torque 320Nm, maximum speed 16000Rpm, DC voltage 354V, pole pair number 4, stator flux 0.0255Wb and stator inductance 0.05mH.

[0054] In the synchronous rotating coordinate system, the voltage equation of the permanent magnet synchronous motor is:

[0055]

[0056] In the formula, u d and u q are the d-axis and q-axis voltages, i d and i q are the d-axis and q-axis currents, ψ d and ψ q are the components of the stator flux on the d-axis and q-axis respectively, ω is the rotor electrical angular velocity, R s is the stator resistance.

[0057] S2: Establish an inverter model that takes into account the inverter dead zone and tube voltage drop.

[0058] Inverter circuit for electric vehicles Figure 2 As shown. Define S a , S b , S c , S a ′、S b ′、S c ′ are the switch states of the inverter respectively. The working principle of the inverter is: when S a / S b / S c When it is in the 1 state, S a ′ / S b ′ / S c ′ is in the 0 state, at this time, the three switch devices in the upper bridge arm of the inverter circuit are turned on, and the three switch devices in the lower bridge arm of the inverter circuit are turned off. a ′ / S b ′ / S c When ' is 1, S a / S b / S c It is in the 0 state, at this time, the three switching devices of the upper bridge arm of the inverter circuit are turned off, and the three switching devices of the lower bridge arm of the inverter circuit are turned on.

[0059] Considering the inverter dead time and tube voltage drop, the actual output voltage waveform of phase A is: Figure 3 As shown in . When the A phase current i a When >0, the output voltage fluctuation of phase A bridge arm is Figure 3 (a). A phase current i a When <0, the output voltage fluctuation of phase A bridge arm is Figure 3 (b). In the figure, ν t is the conduction voltage drop of the switch tube, ν d is the conduction voltage drop of the freewheeling diode, T d is the dead time.

[0060] Assume that in one pulse width modulation cycle, T S1 、T S2 They are the upper and lower switch tubes S of inverter A phase respectively. a and S a The actual opening time of ′, T PWM is a modulation wave cycle, then T PWM =T d +T s1 +T s2 .

[0061] According to the area equivalence principle, the error voltage is averaged over the period of time, and the average error voltage Δu can be obtained as:

[0062]

[0063] In the formula, u dc is the DC voltage, T PWM is a modulation wave cycle, T d is the dead time, ν t is the conduction voltage drop of the switch tube, ν d is the conduction voltage drop of the freewheeling diode.

[0064] A square wave signal can be obtained by averaging the error voltage in a periodic time period. The mathematical expression of the average error voltage of the inverter can be obtained by Fourier decomposition of the square wave signal:

[0065]

[0066] Since most permanent magnet synchronous motors use three-phase star connection, the third harmonic components and integer multiples of three will be cancelled out. Therefore, there are only 5th, 7th, 11th and other harmonic components in the above formula. According to the harmonic amplitude, this method takes the 5th and 7th as the main research objects.

[0067] Therefore, the motor stator voltage equation in the three-phase stationary coordinate system is:

[0068]

[0069] Where: u1, u5 and u7 are the fundamental, 5th and 7th harmonic voltage amplitudes respectively; θ1, θ5 and θ7 are the initial phase angles of the fundamental, 5th and 7th harmonic voltages respectively.

[0070] S3: Establish a harmonic compensation voltage prediction model based on CNN-LSTM to optimize the DQ axis stator voltage, reduce current harmonics, and suppress vibration and torque fluctuations.

[0071] CNN, or convolutional neural network, is a feedforward neural network that is specifically designed to process data with a grid-like structure. It is mainly composed of an input layer, a convolution layer, a pooling layer, a fully connected layer, and an output layer. Since CNN uses convolution operations in its calculations, its computing speed is greatly improved compared to general matrix operations; the alternating use of convolution layers and pooling layers in CNN can effectively extract local features of data and reduce the dimensions of local features.

[0072] LSTM, or long short-term memory neural network, is a time-recurrent neural network, an improvement on the recurrent neural network (RNN). By adding a gate structure, it effectively solves the gradient vanishing and gradient exploding problems of RNN. The cell memory unit added to the LSTM network gives the LSTM network good memory capacity, and is widely used in time series prediction.

[0073] S31: Establishing current harmonic model

[0074] The three-phase voltage u of PMSM a 、u b 、u c Transformed to the DQ rotation coordinate system, we get:

[0075]

[0076] In the formula, u d1 and u q1 are the d-axis and q-axis components of the fundamental voltage in the DQ rotating coordinate system, u5 and u7 are the amplitudes of the 5th and 7th harmonic voltages in the DQ rotating coordinate system, θ5 and θ7 are the initial phase angles of the 5th and 7th harmonic voltages, and ω is the rotor electrical angular velocity.

[0077] The mathematical equation of the motor current in the DQ rotating coordinate system is:

[0078]

[0079] In the formula, i d1 and i q1 are the d-axis and q-axis components of the fundamental current in the DQ rotating coordinate system, i5 and i7 are the amplitudes of the 5th and 7th harmonic currents in the DQ rotating coordinate system, θ5 and θ7 are the initial phase angles of the 5th and 7th harmonic voltages, and ω is the rotor electrical angular velocity.

[0080] The stator voltage equation of the motor containing harmonic components is:

[0081]

[0082] In the formula, i d1 and i q1 are the d-axis and q-axis components of the fundamental current in the DQ rotating coordinate system, i5 and i7 are the amplitudes of the 5th and 7th harmonic currents in the DQ rotating coordinate system, R s is the stator resistance, L d and L q are the d-axis and q-axis inductances, θ5 and θ7 are the initial phase angles of the 5th and 7th harmonic voltages, and ω is the rotor electrical angular velocity.

[0083] Since the fundamental wave is an AC component in the DQ rotating coordinate system of the 5th and 7th harmonics, the steady-state voltage harmonic equation in the 5th DQ rotating coordinate system is:

[0084]

[0085] In the formula, i d5 and i q5 are the d-axis and q-axis DC current components under the 5th DQ rotation coordinate axis, R s is the stator resistance, L d and L q are the d-axis and q-axis inductances respectively, and ω is the rotor electrical angular velocity.

[0086] Similarly, the harmonic steady-state voltage equation in the 7th DQ rotating coordinate system is:

[0087]

[0088] In the formula, i d7 and i q7 are the d-axis and q-axis DC current components under the 7th DQ rotation coordinate axis, R s is the stator resistance, L d and L q are the d-axis and q-axis inductances respectively, and ω is the rotor electrical angular velocity.

[0089] After the three-phase current of the motor passes through the fifth and seventh rotation coordinate systems of the DQ axis, the fifth and seventh current harmonic components on the DQ axis are obtained by low-pass filtering, such as Figure 4 shown.

[0090] By superimposing the voltages generated in the two lines of the harmonic voltage steady-state equation and the PI controller, the harmonic compensation voltage of each harmonic current in the rotating coordinate system is obtained, as shown in Figure 5 shown.

[0091] Then sum the harmonic voltages of each current to obtain the harmonic compensation voltage in the DQ coordinate system. and

[0092] Right now:

[0093] S32: Collect training data

[0094] In order to ensure the accuracy of the harmonic compensation voltage prediction model, the data of the motor under each working condition is collected as much as possible, but there are countless working conditions of the motor. Therefore, in order to ensure the accuracy of the harmonic compensation voltage prediction model while improving the calculation efficiency, the initial speed of the motor is set to 50rpm, and the motor torque gradually increases from 0Nm to the maximum torque, and the increment is set to 10Nm. The data is collected for 1 second in each working condition. The speed is increased to the maximum speed in increments of 50rpm, and the data is recorded repeatedly under each working condition. Record the d-axis current i under each working condition d , q-axis current i q , rotor electrical angular velocity θ, rotor angle ω and motor reference torque Data; then a large amount of experimental data is normalized and used to train CNN-LSTM.

[0095] S33: Establishing a harmonic compensation voltage prediction model based on CNN-LSTM

[0096] Combine CNN (convolutional neural network) with LSTM (long short-term neural network) to establish a CNN-LSTM network model based on Figure 6 shown.

[0097] The training of CNN-LSTM network is divided into two stages: in the first stage, i d 、i q , θ, ω and The data is input into CNN, which consists of four layers: convolutional layer, Relu activation function layer, maximum pooling layer and flat layer. In the second stage, the features extracted by CNN are passed to LSTM, and the fully connected layer is used as the prediction result.

[0098] The CNN network is used to perform initial feature recognition on the input data, and then the LSTM network is used to iteratively train the selected feature values. The final output prediction result is the harmonic compensation voltage. d 、i q , θ, ω and is the input variable, and The training process of the harmonic compensation voltage prediction model is as follows: Figure 7 shown.

[0099] S4: Embed the harmonic compensation voltage prediction model into the motor control system to suppress the torque fluctuation of the motor.

[0100] The harmonic compensation voltage prediction model completed through the above training is embedded into the motor control system, and the motor i d 、i q , θ, ω and The data is input into the prediction model of harmonic compensation voltage, and the prediction results of CNN-LSTM are and Input into the control system to suppress the torque fluctuation of the motor, such as Figure 8 shown.

[0101] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A motor torque fluctuation suppression method based on CNN-LSTM, characterized by: The following steps are involved: S1: Establish the mathematical model of permanent magnet synchronous motor; S2: Establish an inverter model that takes into account the inverter dead zone and tube voltage drop; S3: Establish a harmonic compensation voltage prediction model based on CNN-LSTM to optimize the DQ axis stator voltage, reduce current harmonics, and suppress motor vibration and torque fluctuation; specifically, the following steps are included: S31: Establishing current harmonic model; S32: Collect training data; S33: Establish a harmonic compensation voltage prediction model based on CNN-LSTM; S4: Embed the CNN-LSTM-based harmonic compensation voltage prediction model of step S3 into the motor control system to suppress the torque fluctuation of the motor.

2. The motor torque fluctuation suppression method based on CNN-LSTM according to claim 1 is characterized in that: The specific method of establishing the current harmonic model in step S31 is as follows: The three-phase voltage u of the permanent magnet synchronous motor a 、u b 、u c Transformed to the DQ rotation coordinate system, we get: In the formula, u d1 and u q1 are the d-axis and q-axis components of the fundamental voltage in the DQ rotating coordinate system, u5 and u7 are the amplitudes of the 5th and 7th harmonic voltages in the DQ rotating coordinate system, θ5 and θ7 are the initial phase angles of the 5th and 7th harmonic voltages, and ω is the rotor electrical angular velocity; The mathematical equation of the motor current in the DQ rotating coordinate system is: In the formula, i d1 and i q1 are the d-axis and q-axis components of the fundamental current in the DQ rotating coordinate system, i5 and i7 are the amplitudes of the 5th and 7th harmonic currents in the DQ rotating coordinate system, θ5 and θ7 are the initial phase angles of the 5th and 7th harmonic voltages, and ω is the rotor electrical angular velocity; The stator voltage equation of the motor containing harmonic components is: In the formula, i d1 and i q1 are the d-axis and q-axis components of the fundamental current in the DQ rotating coordinate system, i5 and i7 are the amplitudes of the 5th and 7th harmonic currents in the DQ rotating coordinate system, R s is the stator resistance, L d and L q are the d-axis and q-axis inductances, θ5 and θ7 are the initial phase angles of the 5th and 7th harmonic voltages, ω is the rotor electrical angular velocity; Since the fundamental wave is an AC component in the DQ rotating coordinate system of the 5th and 7th harmonics, the harmonic steady-state voltage equation in the 5th DQ rotating coordinate system is: In the formula, i d5 and i q5 are the d-axis and q-axis DC current components under the 5th DQ rotation coordinate axis, R s is the stator resistance, L d and L q are the d-axis and q-axis inductances respectively, ω is the rotor electrical angular velocity; Similarly, the harmonic steady-state voltage equation in the 7th DQ rotating coordinate system is: In the formula, i d7 and i q7 are the d-axis and q-axis DC current components under the 7th DQ rotation coordinate axis, R s is the stator resistance, L d and L q are the d-axis and q-axis inductances respectively, ω is the rotor electrical angular velocity; The three-phase current of the motor passes through the fifth and seventh rotation coordinate systems of the DQ axis, and then the fifth and seventh current harmonic components on the DQ axis are obtained by low-pass filtering; By superimposing the voltages generated in the two lines of the harmonic steady-state voltage equation and the PI controller, the harmonic compensation voltage of each harmonic current in the rotating coordinate system is obtained; Then sum the harmonic voltages of each current to obtain the harmonic compensation voltage in the DQ coordinate system. and 3. The motor torque fluctuation suppression method based on CNN-LSTM according to claim 2 is characterized in that: The specific method of collecting training data in step S32 is as follows: the initial motor speed is set to 50 rpm, the motor torque is gradually increased from 0 Nm to the maximum torque, and the increment is set to 10 Nm; data is collected for 1 second in each working condition; the speed is increased to the maximum speed in increments of 50 rpm, and data is repeatedly recorded in each working condition; the d-axis current i under each working condition is recorded d , q-axis current i q , rotor electrical angular velocity θ, rotor angle ω and motor reference torque Data; then a large amount of experimental data is normalized and used to train CNN-LSTM.

4. The motor torque fluctuation suppression method based on CNN-LSTM according to claim 3 is characterized in that: The specific method of step S33, establishing the harmonic compensation voltage prediction model based on CNN-LSTM is: Combine the convolutional neural network with the long short-term neural network to establish a CNN-LSTM network model; training the CNN-LSTM network includes two stages: In the first stage, i d 、i q , θ, ω and The data is input into the convolutional neural network; In the second stage, the features extracted by the convolutional neural network are passed to the long-term short-term neural network, and the fully connected layer is used as the prediction result; The convolutional neural network is used to perform initial feature recognition on the input data, and then the long-term and short-term neural network is used to iteratively train the selected feature values. The final output prediction result is the harmonic compensation voltage. d 、i q , θ, ω and is the input variable, and The mapping relationship for the output.

Images