A sensorless motor control method based on neural network sliding mode observer

Abstract

The present invention relates to the technical field of motor sliding mode control, and specifically to a motor position sensorless control method based on a neural network sliding mode observer, including: using a sliding mode observer to track the stator current of a permanent magnet synchronous motor in an α-β axis stationary coordinate system to obtain an estimated current value, an actual current value, and a current error value; using a convolutional neural network, taking the actual current value and the current error value as input, and taking the back electromotive force estimate value as output, and tracking the stator current estimate through training optimization; entering a sliding mode surface and sliding on the sliding mode surface, the sliding mode control converges to an equilibrium point through a switching function, and the back electromotive force is obtained by using a back electromotive force expansion equation; extracting rotor position and speed information from the back electromotive force based on a phase-locked loop. The present invention effectively improves the system jitter phenomenon, improves the observation accuracy of the rotor position, and the system's speed regulation capability of sudden loading and unloading can be significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of motor sliding mode control, and in particular to a motor position sensorless control method. Background Art

[0002] Permanent magnet synchronous motor (PMSM) is widely used in the industrial field due to its advantages such as low loss, low temperature rise, high power factor, high efficiency, short start-up time, and high overload capacity. In order to achieve high-precision and high-efficiency control, the permanent magnet synchronous motor needs to calculate the position information of the rotor. At present, most traditional detection methods use position sensors to measure the position of the motor rotor. This method is simple to detect, but the sensor is difficult to install, the cost is high, and it is not suitable for large electromagnetic interference and harsh environments. Therefore, sensorless control methods are currently the focus of permanent magnet synchronous motor research. The current research methods include Kalman filter observer, sliding mode observer and model reference adaptive method. Since the sliding mode observer is independent of the control parameters during its movement, it has strong robustness and is easy to design. Therefore, most applications use this method for sensorless control research.

[0003] The principle of the sliding mode observer is mainly to reconstruct the back electromotive force by observing the error between the current and the actual current, and finally extract the rotor position information through the PLL phase-locked loop. However, the traditional sliding mode observer has difficulty in fully tracking the sampled current when the motor is running, which makes it difficult to estimate the rotor position information. At the same time, the back electromotive force obtained by the traditional sliding mode observer contains high-frequency harmonic components, which contains a large number of discontinuous signals. Therefore, a low-pass filter is used for filtering in actual application. When a low-pass filter is used, the signal phase will lag, causing the system to vibrate. Summary of the invention

[0004] The object of the present invention is to provide a motor position sensorless control method based on a neural network sliding mode observer to solve the above technical problems; the technical problems solved by the present invention can be achieved by the following technical solutions: a motor position sensorless control method based on a neural network sliding mode observer, characterized in that it includes: step S1, using a sliding mode observer to track the stator current of a permanent magnet synchronous motor in an α-β axis stationary coordinate system to obtain an estimated current value, an actual current value, and a current error value between the estimated current and the actual current; step S2, using a convolutional neural network, taking the actual current value and the current error value as inputs of the convolutional neural network, taking the back electromotive force estimation value as output of the convolutional neural network, and tracking the stator current estimation through training optimization; step S3, using the current error value as a state variable to construct a sliding surface, enter the sliding surface and slide on the sliding surface, the sliding mode control converges to an equilibrium point through a switching function, and using the back electromotive force expansion equation to obtain the back electromotive force; step S4, extracting the rotor position and speed information in the back electromotive force based on a phase-locked loop.

[0005] Preferably, in step S1, the mathematical model of the permanent magnet synchronous motor in the α-β axis stationary coordinate system is:

[0006]

[0007] in, , represents the stator current, , represents the stator voltage, , represents the stator inductance, is the stator resistance, represents the electrical angular velocity, , represents the back electromotive force;

[0008] The back EMF expanded equation is,

[0009]

[0010] Where: represents the permanent magnet flux, , represents the components of the stator current on the d-axis and q-axis of the dq rotating coordinate system, Indicates the rotor electrical angle.

[0011] Preferably, in step S1, the current state equation is:

[0012]

[0013] in, , represents the stator current, , represents the stator voltage, , represents the stator inductance, is the stator resistance, represents the electrical angular velocity, , represents the back electromotive force;

[0014] Based on the mathematical model of permanent magnet synchronous motor, the equation of the sliding mode observer is,

[0015]

[0016] in, , represents the estimated value of the current, represents the sliding mode control law;

[0017] Subtract Formula 3 from Formula 4 to obtain the current error equation:

[0018]

[0019] in, , represents the current error value, , .

[0020] Preferably, the switching function in step S3 is a sigmoid function, expressed as follows:

[0021]

[0022] Zhongzai The interval is approximately linear control; when hour Approximately a symbolic function ;

[0023] The expression of the sliding mode reaching law is:

[0024]

[0025] in, represents the error on the sliding surface, represents the design parameter, K represents the sliding mode switching gain, , represents the estimated value of the current, , represents the stator current;

[0026] The expression of sliding mode control law is:

[0027]

[0028] in , represents the current error value, , Indicates a control signal.

[0029] Preferably, a state differential equation is obtained based on the sliding mode control law:

[0030]

[0031] in , represents the current error value, represents the state matrix, represents the design parameters, represents the sliding mode switching gain, represents the stator inductance, , represents the back electromotive force;

[0032] Based on the equivalent control principle,

[0033] .

[0034] Preferably, step S3 further includes filtering the high frequency component of the control signal using a low pass filter, which is expressed as:

[0035]

[0036] in, is the cutoff frequency of the low-pass filter, represents a complex frequency variable, , represents the estimated value of the back electromotive force.

[0037] Preferably, step S3 further includes using a Lyapunov function stability criterion to verify stability and construct a positive definite Lyapunov function.

[0038]

[0039] in , represents the current error value;

[0040] By taking the derivative of formula 12, we can obtain

[0041]

[0042] in, represents the sliding mode switching gain, is the stator resistance, represents the stator inductance, , represents the back electromotive force;

[0043] According to Lyapunov's stability criterion, for the equilibrium state , if there exists a continuous function V, if it satisfies:

[0044]

[0045] Then the system is uniformly asymptotically stable over a large range at S=0; according to formulas 12 and 13, the stability condition is:

[0046]

[0047] in represents the sign function, this condition indicates that the sliding mode switching gain is greater than the back EMF , The sliding mode observer is stable when .

[0048] Preferably, in step S2, the convolution layer of the convolutional neural network uses a one-dimensional convolution kernel to extract data features, and the formula is:

[0049]

[0050] in, represents the input signal, represents the hybrid filter matrix, For the The weights of the convolution kernels; For the The bias of the convolution kernel.

[0051] Preferably, in step S2, the convolutional layer of the convolutional neural network uses a linear rectification function as an activation function, and the calculation formula of the linear rectification function mapping feature is:

[0052]

[0053] in, represents the mapping feature, Represents the activation function.

[0054] Preferably, in step S2, the convolution layer of the convolutional neural network includes a first convolutional layer for inputting a feature vector for feature extraction and outputting a feature map and a second convolutional layer for encoding the feature, and the convolutional neural network also includes an input layer connected to the first convolutional layer for inputting a feature vector; a pooling layer connected to the second convolutional layer for selecting features and filtering information; a fully connected layer connected to the pooling layer for integrating feature information; an output layer connected to the fully connected layer for outputting a prediction result; the training process of the convolutional neural network includes forward propagation and error back propagation, the forward propagation is the predicted value obtained after inputting the convolutional neural network; the error back propagation is to compare the predicted value obtained by the forward propagation with the true value, and use the cross entropy function as the loss function to calculate the gradient of the loss error for each weight parameter, update the weight parameter through the Adam optimizer, repeat the forward propagation and the error back propagation to continuously update the weight parameter, and make the loss function reach the optimal value as the gradient decreases.

[0055] Beneficial effects of the present invention: Due to the adoption of the above technical scheme, the present invention uses a continuous function for switching at the zero point to obtain a continuous output, uses a convolutional neural network to take the actual current value and the current error value as input for network training, and achieves system convergence. Finally, a phase-locked loop is used to extract the rotor position information of the back electromotive force, which effectively improves the system jitter phenomenon, improves the observation accuracy of the rotor position, and the system's speed regulation capability of sudden loading and unloading can be significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A schematic diagram of the steps of a motor position sensorless control method according to an embodiment of the present invention;

[0057] Figure 2 Schematic diagram of a motor position sensorless control method according to an embodiment of the present invention;

[0058] Figure 3 It is the topology diagram of convolutional neural network;

[0059] Figure 4 It is a topological diagram after the convolutional neural network is optimized in an embodiment of the present invention;

[0060] Figure 5 A system simulation model of the motor position sensorless control method constructed in an embodiment of the present invention.

[0061] Figure 6 It is a schematic diagram of iterative training results of a convolutional neural network in an embodiment of the present invention;

[0062] Figure 7a This is the speed waveform of the traditional observer when the motor is running at 50HZ in a no-load state;

[0063] Figure 7b This is a speed waveform diagram of the sliding mode observer of the present invention when the motor is running at 50HZ in a no-load state;

[0064] Figure 8a This is the speed waveform of the traditional observer when the motor is running at 50HZ in a steady state with no load and a sudden addition of a load of 233NM;

[0065] Figure 8b The speed waveform diagram of the sliding mode observer of the present invention when the motor is running at 50HZ in a steady state with no load and a sudden addition of a load of 233NM;

[0066] Figure 9a This is the speed waveform of the motor controlled by the traditional sliding mode when the load of 233NM is suddenly unloaded during steady-state operation with load;

[0067] Figure 9b This is a speed waveform diagram of the motor optimized by the convolutional neural network in the present invention when the motor is suddenly unloaded with a load of 233NM during steady-state operation with load;

[0068] Fig.10 This is the schematic diagram of the inverter of the permanent magnet synchronous motor experimental platform;

[0069] Fig.11a Schematic diagram of the average phase error of the rotor position observed by the traditional observer;

[0070] Fig.11b The average phase error of the rotor position observed by the observer optimized by the convolutional neural network of the present invention;

[0071] Fig.12a Schematic diagram of current fluctuation amplitude of traditional observer;

[0072] Figure 12b Schematic diagram of current fluctuation amplitude of the sliding mode observer of the present invention. DETAILED DESCRIPTION

[0073] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0074] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0075] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0076] The motor position sensorless control method based on neural network sliding mode observer, such as Figure 1 As shown, it includes step S1, using a sliding mode observer to track the stator current of the permanent magnet synchronous motor in the α-β axis stationary coordinate system to obtain an estimated current value, an actual current value, and a current error value between the estimated current and the actual current; step S2, using a convolutional neural network, taking the actual current value and the current error value as the input of the convolutional neural network, taking the back electromotive force estimate value as the output of the convolutional neural network, and tracking the stator current estimate through training optimization; step S3, using the current error value as a state variable to construct a sliding surface, enter the sliding surface and slide on the sliding surface, the sliding mode control converges to the equilibrium point through the switching function, and the back electromotive force is obtained by using the back electromotive force expansion equation; step S4, extracting the rotor position and speed information in the back electromotive force based on the phase-locked loop.

[0077] Specifically, the present invention aims to solve the problem of system jitter caused by phase lag caused by filtering in traditional sliding mode observers and the difficulty in sampling the estimated current, which leads to errors in the estimation of the rotor position information and thus causes system vibration. A PMSM position sensorless control method based on a neural network sliding mode observer is now provided.

[0078] Specifically, the present invention uses a switching function sigmoid(x) that is continuous at zero point to replace a switching function sign(x) that is discontinuous at zero point on the basis of a traditional sliding mode observer, and uses the autonomous learning ability of a convolutional neural network to optimize a method for controlling rotor position parameters without a position sensor, comprising the following steps: Step 1, using a sliding mode observer to track the stator current of a permanent magnet synchronous motor in an α-β axis stationary coordinate system to obtain a deviation between an estimated current and an actual current; Step 2, using a convolutional neural network to convert the actual current value , And the current error value, , As the input of the convolutional neural network, the back EMF estimate , As the output of the convolutional neural network, the neural network is used to continuously learn and train to obtain the best approximation law, making the stator current estimation and tracking more accurate.

[0079] Step 3: The system reaches the sliding surface and slides on the sliding surface. The stator current deviation is switched by the switching function sigmoid(x) to obtain a continuous output vector. The back electromotive force is obtained using the back electromotive force formula of the permanent magnet synchronous motor.

[0080] Step 4, using the PLL phase-locked loop to extract the rotor position and speed information from the back electromotive force in step 3 to achieve sensorless control of the PMSM.

[0081] The technical advantage of the present invention is mainly reflected in the proposal of a PMSM position sensorless control method based on a neural network sliding mode observer, using the continuous function sigmoid (x) to switch at the zero point to obtain a continuous output, using a convolutional neural network to take the α-β axis current and current error value as input for network training, and using the gradient method to achieve system convergence, and finally using a PLL phase-locked loop to extract the rotor position information of the back electromotive force. The stability of the controller is proved by the Lyapunov stability criterion, and simulation and physical platforms are built for experiments. The experimental results show that this method effectively improves the system jitter phenomenon, improves the observation accuracy of the rotor position, and the system's speed regulation capability of sudden loading and unloading can be significantly improved.

[0082] In a preferred embodiment, in step S1, the mathematical model of the permanent magnet synchronous motor in the α-β axis stationary coordinate system is:

[0083]

[0084] in, , represents the stator current, , represents the stator voltage, , represents the stator inductance, is the stator resistance, represents the electrical angular velocity, , represents back electromotive force;

[0085] The expanded equation for back EMF is,

[0086]

[0087] Where: represents the permanent magnet flux, , represents the components of the stator current on the d-axis and q-axis of the dq rotating coordinate system, Indicates the rotor electrical angle.

[0088] In a preferred embodiment, in step S1, the current state equation is:

[0089]

[0090] in, , represents the stator current, , represents the stator voltage, , represents the stator inductance, is the stator resistance, represents the electrical angular velocity, , represents back electromotive force;

[0091] Based on the mathematical model of permanent magnet synchronous motor, the equation of sliding mode observer is,

[0092]

[0093] in, , represents the estimated value of current, represents the sliding mode control law;

[0094] Subtract Formula 3 from Formula 4 to obtain the current error equation:

[0095]

[0096] in, , Indicates the current error value, , .

[0097] Specifically, in permanent magnet synchronous motor applications, most sliding mode observer algorithms use the α-β axis stationary coordinate system. From formula 2, we can see that the back electromotive force contains the motor speed and rotor position information. From formula 4, we can see that the sliding mode observer can reconstruct the back electromotive force by observing the error between the current and the actual current, and finally extract the rotor position information through the PLL phase-locked loop. As the observation value of the sliding mode observer, For control input.

[0098] In a preferred embodiment, the switching function in step S3 is a sigmoid function, and the expression is:

[0099]

[0100] Zhongzai The interval is approximately linear control; when hour Approximately a symbolic function ;

[0101] The expression of the sliding mode reaching law is:

[0102]

[0103] in, represents the error on the sliding surface, represents the design parameter, which acts as an error suppression term to accelerate the convergence of the sliding mode error. K represents the sliding mode switching gain. , represents the estimated value of current, , represents the stator current;

[0104] The expression of sliding mode control law is:

[0105]

[0106] in , Indicates the current error value, , Indicates a control signal.

[0107] Specifically, the switching function is continuous at zero The function expression is as follows:

[0108]

[0109] ,exist When , the sliding mode control is approximately linear control.

[0110] when hour The function becomes , which has the same effect as the existing sliding mode observer.

[0111] In a preferred embodiment, the state differential equation is obtained based on the sliding mode control law.

[0112]

[0113] in , Indicates the current error value, represents the state matrix, represents the design parameters, represents the sliding mode switching gain, represents the stator inductance, , represents back electromotive force;

[0114] Based on the equivalent control principle,

[0115] .

[0116] Specifically, by substituting formula 8 into formula 5, we can obtain formula 9:

[0117]

[0118] According to the equivalent control principle in sliding mode theory, we can get formula 10:

[0119]

[0120] in , Indicates the current error value, represents the state matrix, represents the design parameters, represents the sliding mode switching gain, represents the stator inductance, , Represents back electromotive force.

[0121] Specifically, the present invention defines the sliding surface as , if the system is stable and meets the reachability condition, when the system reaches the sliding surface and slides on the sliding surface and finally reaches the equilibrium point, the current error is 0 and the current error change rate is 0, that is, According to the equivalent control principle, we can get:

[0122]

[0123] The designed sliding mode control law is:

[0124]

[0125] In the formula is the sliding mode switching gain, is a symbolic function.

[0126] In a preferred embodiment, step S3 further includes filtering the high frequency component of the control signal using a low pass filter, which is expressed as:

[0127]

[0128] in, is the cutoff frequency of the low-pass filter, represents a complex frequency variable, , Represents the estimated back-EMF value.

[0129] Specifically, referring to the sliding mode control law expression, it can be seen that the control quantity is a discontinuous high-frequency switching signal, which will aggravate the chattering phenomenon of the sliding mode, so it is necessary to use LPF to filter out the high-frequency component to obtain a continuous back-EMF estimate. , .

[0130]

[0131] In the formula is the cutoff frequency of the low-pass filter, represents a complex frequency variable, , Represents the estimated back-EMF value.

[0132] In a preferred embodiment, step S3 further includes verifying stability using a Lyapunov function stability criterion.

[0133] Construct a positive definite Lyapunov function,

[0134]

[0135] in , Indicates the current error value;

[0136] By taking the derivative of formula 12, we can obtain

[0137]

[0138] in, represents the sliding mode switching gain, is the stator resistance, represents the stator inductance, , represents back electromotive force;

[0139] According to Lyapunov's stability criterion, for the equilibrium state , if there exists a continuous function V, if it satisfies:

[0140]

[0141] Then the system is uniformly asymptotically stable over a large range at S=0; according to formulas 12 and 13, the stability condition is:

[0142]

[0143] in represents the sign function, this condition indicates that the sliding mode switching gain Greater than the back EMF , The sliding mode observer is stable.

[0144] Specifically, the stability of the sliding mode observer can be proved by using the Lyapunov function stability criterion. The system can reach the sliding mode surface in a short time under the action of the sliding mode control law. , and the system is is asymptotically stable.

[0145] The Lyapunov stability criterion simply states that if a properly chosen Lyapunov function is positive definite near an equilibrium point, and its derivative is always negative (that is, when the system is in a non-equilibrium state, the value of the Lyapunov function continues to decrease), then the system is asymptotically stable at this equilibrium point.

[0146] In the present invention, the Lyapunov function is constructed as,

[0147]

[0148] After taking the derivative of the Lyapunov function, we get formula 13, which represents the rate of change of the Lyapunov function over time. According to the Lyapunov stability criterion, the condition for system stability is the Lyapunov function Derivative with respect to time It is always negative, and the specific conditions for stability are obtained through formula 15.

[0149]

[0150] Through the above analysis, it can be concluded that as long as the sliding mode switching gain Specific back EMF , If it is large, the stability of the sliding mode observer can be guaranteed.

[0151] In a preferred embodiment, in step S2, the convolution layer of the convolutional neural network uses a one-dimensional convolution kernel to extract data features, and the formula is:

[0152]

[0153] in, represents the input signal, represents the hybrid filter matrix, For the The weights of the convolution kernels; For the The bias of the convolution kernel.

[0154] Specifically, the permanent magnet synchronous motor is a highly coupled time-varying nonlinear system. During its operation, its stator resistance and AC and DC axis inductance will change with load and temperature, and its model parameters will also change accordingly. At present, most nonlinear systems use neural network algorithms to improve the performance and robustness of the control system by using the nonlinear approximation ability of neural networks. Convolutional neural network (CNN) is a type of deep learning, which consists of input layer 1, convolution layer 2, activation layer 3, pooling layer 4, fully connected layer 5 and output layer 6. The basic structure is as follows: Figure 2 shown.

[0155] Convolutional neural networks are highly robust and can effectively reduce network complexity through pooling and weight sharing, making them easy to train. They can also adaptively learn parameter features and can process big data better than other traditional neural networks.

[0156] Specifically, convolutional neural networks are often used to process three-dimensional information, and in the present invention, the actual current value , Current error value They are all one-dimensional information, so the present invention adopts a one-dimensional convolution kernel to replace the three-dimensional convolution kernel to extract data features.

[0157] According to formula 2, the back electromotive force contains the rotor position and speed information, which can be extracted using a phase-locked loop (PLL). , And current error value As the input of the convolutional neural network, the back EMF estimate , As the output of the convolutional neural network, the neural network is used to continuously learn and train to obtain the best approximation law, so as to continuously correct the weights of the neural network. The structure diagram is as follows: Figure 3 shown.

[0158] Attached Figure 3 In the equation, v represents the control voltage applied to the permanent magnet synchronous motor (PMSM), and the state equation describing the change of the motor current is,

[0159]

[0160] in Indicates the estimated current value, is the stator resistance, represents the stator inductance, Indicates the input voltage, Represents the estimated value of back electromotive force. By tracking the stator current of the permanent magnet synchronous motor (PMSM) and combining the mathematical model of the permanent magnet synchronous motor, the actual current value and the estimated current value are obtained respectively, and the current error value is obtained. The continuous function sigmoid(x) is used to switch at the zero point to obtain a continuous output. The convolutional neural network is used to take the α-β axis current and the current error value as input for network training, and the gradient method is used to achieve system convergence to obtain the back electromotive force. Finally, the PLL phase-locked loop is used to detect the back electromotive force Extract rotor position information.

[0161] In a preferred embodiment, in step S2, the convolutional layer of the convolutional neural network uses a linear rectification function (RELU) as an activation function, and the calculation formula of the linear rectification function mapping feature is:

[0162]

[0163] in, represents the mapping feature, Represents the activation function.

[0164] In a preferred embodiment, in step S2, if Figure 4 As shown, the convolution layer of the convolutional neural network includes a first convolution layer 21 for inputting a feature vector for feature extraction and outputting a feature map and a second convolution layer 22 for encoding the feature. The convolutional neural network also includes an input layer 1, connected to the first convolution layer 21, for inputting a feature vector; a pooling layer 4, connected to the second convolution layer 22, for selecting features and filtering information; a fully connected layer 5, connected to the pooling layer 4, for integrating feature information; an output layer 6, connected to the fully connected layer 5, for outputting a prediction result; the training process of the convolutional neural network includes forward propagation and error back propagation, the forward propagation is the predicted value obtained after inputting the convolutional neural network; the error back propagation is to compare the predicted value obtained by the forward propagation with the true value, and use the cross entropy function as the loss function to calculate the gradient of the loss error for each weight parameter, and update the weight parameter through the Adam optimizer, repeat the forward propagation and error back propagation to continuously update the weight parameter, and make the loss function reach the optimal value as the gradient decreases.

[0165] The present invention uses 1 input channel and a feature vector of size 1 × 4. At the same time, in order to ensure the completeness of the extracted feature information, this paper uses two convolutional layers, and the convolutional layer parameters are shown in Table 1.

[0166]

[0167] Table 1 Convolutional layer parameters

[0168] The convolution kernels in the first convolution layer extract features from the input feature vector and output feature maps. The second convolution layer further encodes the features based on the above feature maps. Since the feature information extracted by the convolution layer may be large, the pooling layer is used to select features and filter information. The fully connected layer integrates the feature information and sends the results to the output layer. The optimized structure topology is shown in the figure below: Figure 4 shown.

[0169] Specifically, since the feature information extracted by the convolution layer may be large, the pooling layer 4 is used to select features and filter information. The fully connected layer 5 integrates the feature information and sends the result to the output layer 6.

[0170] Convolutional neural network training is divided into two processes: forward propagation and error back propagation. Forward propagation is to input the predicted value obtained through the above process. Back propagation is to compare the value obtained by forward propagation with the true value, and use the cross entropy function as the loss function to calculate the gradient of the loss error for each parameter, and update the parameters through the Adam optimizer. Repeat these two processes to continuously update the parameters, and as the gradient decreases, the loss function reaches the optimal value. The training parameters of the convolutional neural network are shown in Table 2.

[0171]

[0172] Table 2 CNN training parameters

[0173] In order to verify the effectiveness of the convolutional neural network-sliding mode observer, a simulation model was built using Matlab / Simulink. The system control block diagram is shown in the figure. Figure 5 The PMSM parameters are shown in Table 3.

[0174]

[0175] Table 3 Motor parameters

[0176] Figure 5 In the figure, the error between the target speed ω∗ and the actual speed ω will be input into the first PI controller, and the first PI controller outputs the d-axis current reference value id*. The error between id* and the actual current id enters the second PI controller. Similarly, the error between the q-axis current reference value iq* and the actual current iq enters the third PI controller. The second PI controller outputs the d-axis current reference value ud*, and the third PI controller outputs the q-axis voltage reference value uq*. After the α-β coordinate system transformation, the voltage signals uα* and uβ* are generated. The uα* and uβ* are input into the space vector pulse width modulation (SVPWM) module to generate the inverter input signal. The inverter output current iabc is sent to the permanent magnet synchronous motor (PMSM), and iα and iβ are obtained through the α-β coordinate system transformation. The iα and iβ are converted into the dq coordinate system to generate id and iq; and the iα and iβ are sent to the convolutional neural network sliding mode observer (CNNSMO) to identify the state quantities id and iq and output the actual speed ω.

[0177] according to Figure 6 It can be seen that when the network training iteration reaches 3000 times, the loss function reaches the minimum value.

[0178] Experimental verification: In order to verify the PMSM position sensorless control performance based on the neural network sliding mode observer, the method of the present invention is simulated and experimentally compared with the existing sliding mode observer control method. The parameters of the permanent magnet synchronous motor are: rated power 110KW, rated voltage 380V, rated current 176A, rated speed 1800r / min, stator resistance 22mΩ, d / q axis inductance 12.8mh. Load 233NM.

[0179] The simulation experiment was carried out under no-load and rated load conditions. Figure 7 shows the motor running at 50HZ steady state under no-load. Figure 8 shows the speed waveform of the motor when a 233NM load is suddenly added during steady state operation. Figure 9 shows the speed waveform of the motor when a 233NM load is suddenly removed during steady state operation with load.

[0180] Depend on Figure 7a It can be seen that when the motor is controlled by a traditional sliding mode device and the motor is running at 50HZ, the steady-state error is about 0.2%. Figure 7b It can be seen that using the improved observer using the convolutional neural network, the steady-state error of the motor is about 0.14% when it is running at 50HZ.

[0181] Depend on Figure 8a It can be seen that when the motor is controlled by the traditional sliding mode device, the motor is loaded in the steady-state operation at 50HZ, the overshoot is 3.84%, and the loading adjustment time is 2.63s. As shown in Figure 9a, when the motor is controlled by the traditional sliding mode device, the motor suddenly releases the load in the steady-state operation with load, the overshoot is 3.82%, and the loading adjustment time is 4.27s, and both have strong jitter.

[0182] Depend on Figure 8b It can be seen that using the observer improved by the convolutional neural network, the motor is loaded under 50HZ steady-state operation, the overshoot is 3.68%, and the loading adjustment time is 2.34s. As shown in Figure 9b, using the observer improved by the convolutional neural network, the motor suddenly releases the load under steady-state operation with load, the overshoot is 3.56%, the loading adjustment time is 4.51s, and the chattering phenomenon is effectively improved.

[0183] Experimental performance comparison: The present invention builds a permanent magnet synchronous motor towing platform based on TMS320F28377D, in which a battery pack supplies 600V DC power, and an inverter supplies 380V AC power. The inverter structure is as follows Fig.10 Figures 11 and 12 are comparisons of the experimental results of the sliding mode observer of the present application and the existing sliding mode observer. Figure 11 is a comparison of the actual rotor angle of the motor converted by the encoder under steady-state operation of the synchronous motor with the angle observed by the sliding mode observer and the angle observed by the observer after the convolutional neural network parameters are optimized.

[0184] Depend on Fig.11a It can be seen that the average phase error of the rotor position observed by the traditional sliding mode observer is 0.55 rad. Fig.11b It can be seen that the average phase error of the rotor position observed by the observer after convolutional neural network optimization is 0.12rad. It can be seen that the accuracy of the rotor position observation of the traditional sliding mode observer is improved by autonomous parameter learning through convolutional network.

[0185] As shown in Figure 12, after the physical platform motor IF is started, the control algorithm of this application enables the system to enter a steady state faster. After the system stabilizes, the load is suddenly increased at 0.2S. The current fluctuation amplitude of the improved control algorithm is significantly smaller than that of the traditional sliding mode observer, and the system can automatically adjust to steady-state operation in a shorter time. Experimental comparison shows that this application effectively improves the system jitter phenomenon, improves the observation accuracy of the rotor position, and the system's speed regulation capability of sudden loading and unloading can be significantly improved.

[0186] Aiming at the problem of system jitter caused by traditional sliding mode observer, this paper proposes a sensorless control method for optimizing the control of the system by using convolutional neural network. The α-β axis current and current error value are used as input for network training, and the gradient method is used to achieve system convergence. Finally, the PLL phase-locked loop is used to extract the rotor position information of the back electromotive force. The stability of the controller is proved by the Lyapunov stability criterion, and simulation and physical platforms are built for experiments. The experimental results show that this method effectively improves the system jitter phenomenon, improves the observation accuracy of the rotor position, and significantly improves the speed regulation capability of the system under sudden loading and unloading.

[0187] The above description is only a preferred embodiment of the present invention, and does not limit the implementation mode and protection scope of the present invention. For those skilled in the art, it should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the description and illustrations of the present invention should be included in the protection scope of the present invention.

Claims

1. A motor position sensorless control method based on a neural network sliding mode observer, characterized in that: The method comprises: step S1, using a sliding mode observer to track the stator current of the permanent magnet synchronous motor in an α-β axis stationary coordinate system to obtain an estimated current value, an actual current value, and a current error value between the estimated current and the actual current; Step S2, using a convolutional neural network, taking the actual current value and the current error value as inputs of the convolutional neural network, taking the back electromotive force estimation value as output of the convolutional neural network, and tracking the stator current estimation through training optimization; Step S3, constructing a sliding surface with the current error value as a state variable, entering the sliding surface and sliding on the sliding surface, the sliding mode control converges to an equilibrium point through a switching function, and the back electromotive force is obtained by using a back electromotive force expansion equation; Step S4, extracting rotor position and speed information from the back electromotive force based on a phase-locked loop; The switching function in step S3 is a sigmoid function, and the expression is: ； Zhongzai The interval is approximately linear control; when hour Approximately a symbolic function ; The expression of the sliding mode reaching law is: ； in, represents the error on the sliding surface, represents the design parameter, K represents the sliding mode switching gain, , represents the estimated current value, , represents the stator current; The expression of sliding mode control law is: ； in , represents the current error value, , Indicates a control signal.

2. The motor position sensorless control method based on neural network sliding mode observer according to claim 1 is characterized in that: In step S1, the mathematical model of the permanent magnet synchronous motor in the α-β axis stationary coordinate system is: ； in, , represents the stator current, , represents the stator voltage, , represents the stator inductance, is the stator resistance, represents the electrical angular velocity, , represents the back electromotive force; The back EMF expanded equation is, ； Where: represents the permanent magnet flux, , represents the components of the stator current on the d-axis and q-axis of the dq rotating coordinate system, Indicates the rotor electrical angle.

3. The motor position sensorless control method based on neural network sliding mode observer according to claim 2 is characterized in that: In step S1, the current state equation is: ； in, , represents the stator current, , represents the stator voltage, , represents the stator inductance, is the stator resistance, represents the electrical angular velocity, , represents the back electromotive force; Based on the mathematical model of permanent magnet synchronous motor, the equation of the sliding mode observer is, ； in, , represents the estimated current value, represents the sliding mode control law; Subtract Formula 3 from Formula 4 to obtain the current error equation: ； in, , represents the current error value, , .

4. The motor position sensorless control method based on neural network sliding mode observer according to claim 1 is characterized in that: Based on the sliding mode control law, the state differential equation is obtained: ； in , represents the current error value, represents the state matrix, represents the design parameters, represents the sliding mode switching gain, represents the stator inductance, , represents the back electromotive force; Based on the equivalent control principle, 。 5. The motor position sensorless control method based on neural network sliding mode observer according to claim 1 is characterized in that: In step S3, it also includes using a low-pass filter to filter the high-frequency component of the control signal, and the expression is: ； in, is the cutoff frequency of the low-pass filter, represents a complex frequency variable, , represents the estimated value of the back electromotive force.

6. The motor position sensorless control method based on neural network sliding mode observer according to claim 1 is characterized in that: Step S3 also includes verifying the stability using the Lyapunov function stability criterion. Construct a positive definite Lyapunov function, ； in , represents the current error value; By taking the derivative of formula 12, we can obtain ； in, represents the sliding mode switching gain, is the stator resistance, represents the stator inductance, , represents the back electromotive force; According to Lyapunov's stability criterion, for the equilibrium state , if there exists a continuous function V, if it satisfies: ； Then the system is uniformly asymptotically stable over a large range at S=0; according to formulas 12 and 13, the stability condition is: ； in represents the sign function, this condition indicates that the sliding mode switching gain is greater than the back EMF , The sliding mode observer is stable when .

7. The motor position sensorless control method based on neural network sliding mode observer according to claim 1 is characterized in that: In step S2, the convolution layer of the convolutional neural network uses a one-dimensional convolution kernel to extract data features, and the formula is: ； in, represents the input signal, represents the hybrid filter matrix, For the The weights of the convolution kernels; For the The bias of the convolution kernel.

8. The motor position sensorless control method based on neural network sliding mode observer according to claim 7 is characterized in that: In step S2, the convolutional layer of the convolutional neural network uses a linear rectification function as an activation function, and the calculation formula of the linear rectification function mapping feature is: ； in, represents the mapping feature, Represents the activation function.

9. The motor position sensorless control method based on neural network sliding mode observer according to claim 6 is characterized in that: In step S2, the convolution layer of the convolutional neural network includes a first convolutional layer for inputting a feature vector for feature extraction and outputting a feature map and a second convolutional layer for encoding the feature. The convolutional neural network also includes an input layer connected to the first convolutional layer for inputting a feature vector; a pooling layer connected to the second convolutional layer for selecting features and filtering information; a fully connected layer connected to the pooling layer for integrating feature information; an output layer connected to the fully connected layer for outputting a prediction result; the training process of the convolutional neural network includes forward propagation and error back propagation, the forward propagation is a predicted value obtained after inputting the convolutional neural network; the error back propagation is to compare the predicted value obtained by the forward propagation with the true value, and use the cross entropy function as the loss function to calculate the gradient of the loss error for each weight parameter, update the weight parameter through the Adam optimizer, repeat the forward propagation and the error back propagation to continuously update the weight parameter, and make the loss function reach the optimal value as the gradient decreases.

Images