Motor control method based on fuzzy RBF neural network sliding mode observer
By introducing a fuzzy RBF neural network sliding mode observer and using a hyperbolic tangent function and adaptive adjustment of the sliding mode gain, the chattering problem of permanent magnet synchronous motors under complex conditions is solved, thereby improving control accuracy and robustness.
Patent Information
- Application Number
- CN202411010989.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-26
- Publication Date
- 2024-11-12
AI Technical Summary
Existing permanent magnet synchronous motor control systems are difficult to control precisely under complex external conditions, and traditional sliding mode observers suffer from chattering problems, affecting robustness and accuracy.
A sliding mode observer based on a fuzzy RBF neural network is adopted, the switching function is replaced by a hyperbolic tangent function, and the optimal parameters of the membership function are calculated through fuzzy control. The sliding mode gain is adaptively adjusted by the RBF neural network to improve the robustness and chattering problem of the system.
It improves the observation accuracy and robustness of permanent magnet synchronous motors, effectively suppresses chattering, and achieves more accurate rotor position and speed estimation.
Smart Images

Figure CN118944508A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of motors, and in particular relates to a motor control method based on a fuzzy RBF neural network sliding mode observer. Background Art
[0002] With the development of science and technology, permanent magnet synchronous motors are increasingly used in aerospace, CNC machine tools, and industrial manufacturing industries. Permanent magnet synchronous motors have the characteristics of simple structure, power saving, high efficiency, small size, and high power factor, so they are widely used in various industries. Permanent magnet synchronous motors generally use sensors to measure the rotor speed and corresponding position information of the motor, but if the sensor device is added, the space of the motor will be reduced, the production cost of the motor will increase, and the motor system will become more complicated. Therefore, for a permanent magnet synchronous motor with good control performance, the research on position sensorless is increasingly valued. At present, the speed control of permanent magnet synchronous motors still uses traditional controller algorithms, such as PI controllers. The PI controller algorithm is simple, has few parameters, is easy to adjust, and can meet the control requirements under certain conditions. However, during the operation of permanent magnet synchronous motors, the external conditions are often very complex, accompanied by the influence of external factors such as noise, and are easily affected by external disturbances, making it difficult to accurately control. In order to make the control accuracy more accurate, it is necessary to introduce an observer in the control system. In general, the observer reconstructs another system based on the known signal to estimate the desired signal.
[0003] The sliding mode observer has strong adaptability, is insensitive to changes in motor parameters and load disturbances, and has the advantages of being simple and easy to implement, with low computational complexity and good dynamic performance. It is the first choice in practical engineering applications. Although the sliding mode variable structure control is highly robust, it has the problem of jitter. The traditional sliding mode includes approaching mode and sliding mode. High-frequency switching in the approaching mode will cause jitter, while the sliding mode will cause vibration and overshoot due to the discontinuity of the sign function, which directly affects the robustness of the entire system. The existence of jitter affects the performance of the sliding mode observer, so it must be weakened and suppressed. Although scholars have done a lot of work and obtained many solutions, these methods can suppress jitter to a certain extent, but they also have their own shortcomings and limitations. In actual engineering, specific analysis should be conducted for specific systems. Summary of the invention
[0004] In order to improve the robustness and rapidity of the sliding mode observer and further reduce the chattering problem, the present invention provides a motor control method based on a fuzzy RBF neural network sliding mode observer. Compared with the traditional sliding mode observer, the switching function is replaced by a hyperbolic tangent function under the premise of ensuring the robustness of the system, and on the basis of fuzzy control, the RBF neural network algorithm is used to calculate the optimal parameters of the membership function in the fuzzy controller. The new sliding mode observer has autonomous learning ability and good self-adaptation ability, thereby better improving the robustness and observation accuracy.
[0005] To achieve the above object, the present invention adopts the following technical solutions:
[0006] The motor control method based on fuzzy RBF neural network sliding mode observer is characterized by comprising:
[0007] A mathematical model of a permanent magnet synchronous motor is established, and the coordinates are transformed to the two-term stationary coordinate system α-β axis through Clark transformation; the control voltage on the α-β axis is input into the sliding mode observer to obtain the current observation value of the α-β axis, and the current error is obtained by subtracting the obtained current observation value from the actual current value. The current error is input into the fuzzy RBF neural network to obtain the sliding mode gain proportional factor under the synergistic effect of the RBF neural network and fuzzy control, the sliding mode gain is adjusted by the proportional factor, and the adjusted sliding mode gain is sent to the switching function to obtain the back electromotive force, and finally the position information and the estimated value of the speed of the motor are obtained through phase-locked loop demodulation.
[0008] To optimize the above technical solutions, the specific measures taken also include:
[0009] Furthermore, the control voltage on the α-β axis is:
[0010]
[0011] In the formula, u α and u β are the control voltages on the α and β axes, R s is the stator phase resistance, i α and i β are the currents on the α and β axes, Ψ α and β are the magnetic flux on the α and β axes respectively, and t represents time.
[0012] Furthermore, the equations of the magnetic flux on the α and β axes are:
[0013]
[0014] Where, L q and L dare the inductances of the q-axis and d-axis respectively, θ is the angle the motor has rotated, and Ψ f is the excitation flux.
[0015] Furthermore, the state equations of the currents on the α and β axes are:
[0016]
[0017] In the formula, R s and L s are the stator phase resistance and the motor stator inductance, Ψ α and β are the magnetic flux on the α-β axis respectively.
[0018] Furthermore, the fuzzy RBF neural network includes four layers, as follows:
[0019] The input-output relationship function f1(x) of the first layer nodes is:
[0020]
[0021] In the formula, x is the input state variable of the network, Q is the current error, is the derivative of the current error;
[0022] The second layer is the fuzzy layer, and its membership function f2(i, j) uses a Gaussian function:
[0023]
[0024] In the formula, β ij and c ij are the width and center of the fuzzy set Gaussian function respectively;
[0025] The third layer is the fuzzy reasoning layer, and the fitness of each node f3(j) is:
[0026]
[0027] Where N i is the number of fuzzy segmentations of the i-th input;
[0028] The fourth layer is the output layer, and the output function of the output layer is:
[0029]
[0030] Where W j is the connection weight matrix between the fourth layer and each node of the fuzzy reasoning layer, and u is the parameter value of the membership function.
[0031] Furthermore, the fuzzy RBF neural network learning weights W j , the width of the membership function bij and center c ij ; Using supervised learning method, the objective function J is defined as:
[0032]
[0033] In the formula, i(t) is the current observation value of the sliding mode observer and the actual current value respectively.
[0034] Furthermore, the gradient descent method is used to solve the minimum value of J, and the gradient descent method is used again to correct the weight W j 、b ij and c ij , through iteration and calculation, the optimal parameter value of the membership function is obtained, the proportional factor c of the sliding mode gain is obtained through the membership function, and the optimal value of the sliding mode gain k is obtained.
[0035] Furthermore, the switching function is a hyperbolic tangent function, and the sliding mode observer using the hyperbolic tangent function is as follows:
[0036]
[0037] In the formula, and is the current observation value on the α and β axes, R s and L s are the stator phase resistance and the motor stator inductance, u α and u β are the control voltages on the α and β axes respectively, t represents time, m is the expansion coefficient of the hyperbolic tangent function, k is the sliding mode gain, f represents the hyperbolic tangent function, e α and e β are the back EMF on the α and β axes respectively;
[0038] The current error equation is:
[0039]
[0040] In the formula, Q α and Q β They are the deviations between the current observation value of the sliding mode observer and the actual current value, representing the current error Q, i α and i β They are the current on the α and β axes respectively, indicating the actual value of the current.
[0041] Furthermore, the back electromotive force on the α and β axes is:
[0042]
[0043] In the formula, Ψ f is the excitation flux, ω e is the angular velocity of the motor, and θ is the angle through which the motor rotates.
[0044] The beneficial effects of the present invention are as follows: the present invention can effectively improve the chattering problem caused by the asymmetric effect on both sides of the sliding surface by introducing the hyperbolic tangent function as the switching function of the sliding mode observer. On the basis of adding the hyperbolic tangent function, the self-learning function of the fuzzy RBF neural network algorithm can be used to calculate the optimal value of the fuzzy membership function parameter. The estimation accuracy of the motor rotor position and speed by the sliding mode observer of the fuzzy RBF neural network is more accurate than that of the traditional sliding mode observer, and the chattering of the system is effectively suppressed. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 This is a system block diagram of the motor control based on the fuzzy RBF neural network sliding mode observer of the present invention.
[0046] Figure 2 It is the switching function used by the sliding mode observer of the present invention.
[0047] Figure 3 It is a structural diagram of the fuzzy RBF neural network sliding mode observer of the present invention.
[0048] Figure 4 It is a schematic diagram of the neural network structure of the fuzzy RBF neural network sliding mode observer of the present invention.
[0049] Figure 5 Schematic diagram of the phase-locked loop structure of the present invention.
[0050] Figure 6 is the fuzzy control rule of the fuzzy controller of the present invention. DETAILED DESCRIPTION
[0051] The present invention will now be described in further detail with reference to the accompanying drawings.
[0052] The present invention proposes a motor control method based on fuzzy RBF neural network sliding mode observer, and its control system block diagram is as follows: Figure 1 As shown, the control method includes: establishing a mathematical model of the permanent magnet synchronous motor and performing corresponding coordinate transformation. Due to the need to decouple the complex current relationship, it is necessary to transform the three-term stationary coordinate system into a two-term stationary coordinate system. The two-term stationary coordinate system α-β axis is obtained by coordinate transformation. The voltage u on the α-β axis obtained by Clark transformation is α 、u βThe current error is obtained by subtracting the actual value from the observed value, and the current error is sent to the designed fuzzy neural network controller. The sliding mode gain proportional factor c is obtained through the synergistic effect of the RBF neural network and fuzzy control. The sliding mode gain k is adjusted by the proportional factor c, and the obtained k' is sent to the switching function to further obtain the back electromotive force. The position information and the estimated value of the speed of the motor are obtained through phase-locked loop demodulation.
[0053] The voltage of the α-β axis obtained by coordinate transformation is:
[0054]
[0055] Among them, u α and u β are the control voltages on the α and β axes, R s is the stator phase resistance, i α and i β are the currents on the α and β axes, Ψ α and β are the magnetic flux on the α and β axes respectively, and t represents time.
[0056] The magnetic flux equation is:
[0057]
[0058] Among them, α and β are the magnetic flux on the α and β axes, L q and L d are the inductances of the q-axis and d-axis respectively, θ is the angle of rotation of the rotor, i α and i β are the currents on the α and β axes, Ψ f is the excitation flux.
[0059] The state equation for current is:
[0060]
[0061] Among them, R s and L s They are stator phase resistance, motor stator inductance, Ψ α and β are the magnetic flux on the α-β axis respectively.
[0062] The voltage of the α-β axis is used to obtain the back electromotive force through the fuzzy RBF neural network sliding mode observer:
[0063]
[0064] Among them, e α and e β are the back EMF on the α and β axes, Ψ f is the excitation flux, ω e is the angular velocity of the motor, and θ is the angle through which the motor rotates.
[0065] Set the switching function to the hyperbolic tangent function, such as Figure 2 As shown, its specific mathematical expression is:
[0066]
[0067] In order to achieve the goal of reducing jitter and obtaining better performance, the switching function meets the following requirements: (1) the function is continuous; (2) with reference to the saturation function, the upper limit of the function is 1 and the lower limit is -1; (3) the slope within the boundary layer is nonlinear; and (4) the function has no delay characteristics.
[0068] In the sliding mode observer, the expression of the hyperbolic tangent function is as follows:
[0069]
[0070] Where m is the constant that adjusts the hyperbolic tangent function. are the current observation values of α-β axis respectively.
[0071] When the hyperbolic tangent function is used as the switching function, the sliding mode observer can have switching characteristics, and can effectively reduce discontinuity and the high-frequency components generated by the symbol, thereby achieving the purpose of reducing chattering. The sliding mode observer designed using the hyperbolic tangent function is as follows:
[0072]
[0073] in, and is the current observation value on the α and β axes; R s and L s are the stator phase resistance and the motor stator inductance respectively; k is the sliding mode gain, which is generally taken as f represents the hyperbolic tangent function.
[0074] The current error equation is obtained:
[0075]
[0076] in, are the deviations between the estimated value and the actual value of the sliding mode observer current.
[0077] Since the sliding mode gain needs to be large enough to meet the stability condition, the sliding mode gain is usually a fixed value designed by the designer. However, the current value observed in the sliding mode observation changes dynamically over time, so the fixed sliding mode gain may cause jitter to the system, and it is necessary to dynamically adjust the sliding mode gain. The present invention uses fuzzy control to adaptively adjust the sliding mode gain, thereby eliminating the jitter of the system. When the system parameters are perturbed, the advantages of RBF neural network and fuzzy control are combined to achieve more accurate and faster adjustment of the sliding mode gain.
[0078] The principle of fuzzy RBF neural network sliding mode observer is as follows Figure 3 As shown in the figure, the two inputs of the fuzzy RBF neural network sliding mode observer are the current state error Q and the derivative of the current state error dQ, and its output is the proportional factor c of the sliding mode gain. According to the designed fuzzy inference rules, different proportional factors will be obtained by inputting different Q and dQ, and the defuzzified proportional factor c will be multiplied by the sliding mode gain k to obtain the adjusted sliding mode gain k'.
[0079] The fuzzy language of input and output is [NB (negative large), NM (negative medium), NS (negative small), ZO (zero), PS (positive small), PM (positive medium), PM (positive large)]. The membership functions of input and output are overlapping triangular membership functions. The fuzzy control rules are as follows: Figure 6 shown.
[0080] Since the membership parameters of traditional fuzzy control are determined by human experience, they cannot adapt to the dynamic changes of the system, which in turn affects the control accuracy. Therefore, the RBF neural network has the advantages of better local approximation and faster convergence, and a fuzzy neural network is established. The structure is as follows Figure 4 shown.
[0081] The function of the input-output relationship of the first layer of nodes in the RBF neural network is:
[0082]
[0083] Where Q is the current error of the system, is the derivative of the current error of the system.
[0084] The second layer is the fuzzy layer, and its membership function uses the Gaussian function:
[0085]
[0086] Among them, b ij and c ij is the width and center of the fuzzy set Gaussian function.
[0087] The third layer is the fuzzy reasoning layer, and the applicability of each node is:
[0088]
[0089] Among them, N i is the number of fuzzy segmentations of the ith input.
[0090] The fourth layer is the output layer, and the output function of the output layer is:
[0091]
[0092] Among them, W j is the connection weight matrix between the fourth layer and each node of the fuzzy reasoning layer, f3(j) is the fitness of each node in the third layer, and u is the optimal parameter value of the membership function.
[0093] What the fuzzy RBF neural network needs to learn is the weight W j , the width of the membership function b ij and the center value c ij . Using supervised learning method, the objective function is defined as:
[0094]
[0095] in, i(t) is the observed value of current and the actual value of current respectively.
[0096] The gradient descent method is used to solve the minimum value of J, so that the estimated value of the current is closest to the actual current value, and then the gradient descent method is used again to correct the weight W and width b ij and the center value c ij Through continuous iteration and calculation, the optimal parameters of the optimal membership function are obtained, and the proportional factor c of the sliding mode gain is obtained through the fuzzy membership function, and the optimal value of the sliding mode gain k is further obtained.
[0097] The sliding mode observer will produce phase lag after filtering. Usually, the rotor position information is obtained by inverse tangent operation. Figure 5 The PLL (phase-locked loop) shown in the figure is used to estimate the position and speed of the rotor. The PLL is a feedback system that keeps the frequency tracking the input and can realize automatic tracking of the output signal frequency.
[0098] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.
Claims
1. A motor control method based on a fuzzy RBF neural network sliding mode observer, characterized in that: include: The mathematical model of the permanent magnet synchronous motor is established, and the coordinates are transformed to the two-term stationary coordinate system α-β axis through Clark transformation; The control voltage on the α-β axis is input into the sliding mode observer to obtain the current observation value of the α-β axis. The current observation value is subtracted from the actual current value to obtain the current error. The current error is input into the fuzzy RBF neural network to obtain the sliding mode gain proportional factor under the synergistic effect of the RBF neural network and fuzzy control. The sliding mode gain is adjusted by the proportional factor, and the adjusted sliding mode gain is sent to the switching function to obtain the back electromotive force. Finally, the position information of the motor and the estimated value of the speed are obtained through phase-locked loop demodulation.
2. The motor control method based on fuzzy RBF neural network sliding mode observer according to claim 1, characterized in that: The control voltage on the α-β axis is: In the formula, u α and u β are the control voltages on the α and β axes, R s is the stator phase resistance, i α and i β are the currents on the α and β axes, Ψ α and β are the magnetic flux on the α and β axes respectively, and t represents time.
3. The motor control method based on fuzzy RBF neural network sliding mode observer as claimed in claim 2, characterized in that: The equations for the flux linkage on the α and β axes are: Where, L q and L d are the inductances of the q-axis and d-axis respectively, θ is the angle the motor has rotated, and Ψ f is the excitation flux.
4. The motor control method based on fuzzy RBF neural network sliding mode observer as claimed in claim 2, characterized in that: The state equations of the currents on the α and β axes are: In the formula, R s and L s are the stator phase resistance and the motor stator inductance, Ψ α and β are the magnetic flux on the α-β axis respectively.
5. The motor control method based on fuzzy RBF neural network sliding mode observer as claimed in claim 1, characterized in that: The fuzzy RBF neural network includes four layers, as follows: The input-output relationship function f1(x) of the first layer nodes is: In the formula, x is the input state variable of the network, Q is the current error, is the derivative of the current error; The second layer is the fuzzy layer, and its membership function f2(i, j) uses a Gaussian function: Where b ij and c ij are the width and center of the fuzzy set Gaussian function respectively; The third layer is the fuzzy reasoning layer, and the fitness of each node f3(j) is: Where Ni is the fuzzy segmentation number of the i-th input; The fourth layer is the output layer, and the output function of the output layer is: Where W j is the connection weight matrix between the fourth layer and each node of the fuzzy reasoning layer, and u is the parameter value of the membership function.
6. The motor control method based on fuzzy RBF neural network sliding mode observer as claimed in claim 5, characterized in that: The fuzzy RBF neural network learning weights W j , the width of the membership function b ij and center c ij ; Using supervised learning method, the objective function J is defined as: In the formula, i(t) is the current observation value of the sliding mode observer and the actual current value respectively.
7. The motor control method based on fuzzy RBF neural network sliding mode observer as claimed in claim 6, characterized in that:Use the gradient descent method to find the minimum value of J, and use the gradient descent method again to correct the weight W j , b ij and c ij , through iteration and calculation, the optimal parameter value of the membership function is obtained, the proportional factor c of the sliding mode gain is obtained through the membership function, and the optimal value of the sliding mode gain k is obtained.
8. The motor control method based on fuzzy RBF neural network sliding mode observer according to claim 1, characterized in that: The switching function is a hyperbolic tangent function, and the sliding mode observer using the hyperbolic tangent function is as follows: In the formula, and is the current observation value on the α and β axes, R s and L s are the stator phase resistance and the motor stator inductance, u α and u β are the control voltages on the α and β axes respectively, t represents time, m is the expansion coefficient of the hyperbolic tangent function, k is the sliding mode gain, f represents the hyperbolic tangent function, e α and e β are the back EMF on the α and β axes respectively; The current error equation is: In the formula, Q α and Q β They are the deviations between the current observation value of the sliding mode observer and the actual current value, representing the current error Q, i α and i β They are the current on the α and β axes respectively, indicating the actual value of the current.
9. The motor control method based on fuzzy RBF neural network sliding mode observer as claimed in claim 8, characterized in that: The back EMF on the α and β axes is: In the formula, Ψ f is the excitation flux, ω e is the angular velocity of the motor, and θ is the angle through which the motor rotates.
Citation Information
Cited By
Permanent magnet synchronous motor fault detection method based on self-learning boundary reconstruction
CN119805213A
