PMSM position sensorless control method based on PINN
By introducing a physical information neural network (PINN) to construct labelless learning constraints, the parameters of the PMSM motor are identified in real time, which solves the position and speed estimation errors caused by parameter mismatch, realizes high-precision sensorless control, and improves the robustness and real-time performance of the control system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HANGZHOU DIANZI UNIV
- Filing Date
- 2026-01-30
- Publication Date
- 2026-04-21
AI Technical Summary
In existing sensorless PMSM control systems, motor parameter mismatch leads to flux linkage trajectory distortion, low position and speed estimation accuracy, and degraded control performance. Furthermore, online parameter identification methods are computationally complex or unstable.
A physical information neural network (PINN) is used to construct label-free learning constraints. A loss function is constructed based on the physical voltage equation of the motor to identify stator resistance and inductance in real time. By combining Sigmoid interval constraints and smooth updates, a nonlinear flux observer is injected to achieve high-precision estimation of position and velocity.
It improves the accuracy of position and velocity estimation under all operating conditions, enhances control robustness, reduces noise sensitivity and computational complexity, and realizes low-cost sensorless closed-loop control.
Smart Images
Figure CN121907084A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of motor control technology, specifically relating to a sensorless control method for PMSM based on PINN. Background Technology
[0002] Permanent magnet synchronous motors (PMSMs) are widely used in electric vehicles, industrial robots, and home appliances due to their high efficiency, high power density, and wide speed range. PMSM vector control (FOC) relies on precise rotor position information. While mechanical position sensors can directly measure position, they increase system cost, size, and assembly complexity, and their reliability decreases under harsh conditions such as vibration, dust, and humidity.
[0003] In sensorless control technology, flux linkage observers are widely used due to their relatively simple structure and fast response. However, traditional flux linkage observers are usually based on fixed motor parameters (stator resistance). ,inductance The design involves... In actual operation, motor temperature rise will lead to... Increasing the load current will cause magnetic circuit saturation. The nonlinear changes in parameters lead to parameter mismatch. This mismatch causes distortion of the flux linkage trajectory, resulting in deviations in position and velocity estimations, and in severe cases, significant deterioration or even instability in control performance.
[0004] To address the parameter mismatch problem, existing technologies often employ recursive least squares (RLS), model reference adaptive systems (MRAS), or extended Kalman filters (EKF) for online parameter identification. However, these methods may suffer from issues such as sensitivity to noise, high computational complexity, or difficulty in directly introducing physical constraints on the parameters. When implemented on embedded controllers, they are prone to numerical divergence or outputting non-physical parameters under transient conditions. Summary of the Invention
[0005] In view of the above-mentioned problems, the technical problem to be solved by the present invention is to provide a sensorless control method for PMSM based on PINN. By introducing a Physical Information Neural Network (PINN), label-free learning constraints are constructed using the physical voltage equation of the motor to achieve online identification. The identification results are then injected into the nonlinear flux observer after being constrained by physical intervals and smoothed, thereby improving the accuracy of position and velocity estimation and control robustness under all operating conditions.
[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A sensorless control method for PMSM based on PINN includes the following steps: S10, Data Construction: Real-time acquisition of three-phase current in the vector control loop of the permanent magnet synchronous motor (PMSM). The two-phase stationary coordinate system current is obtained through the stationary coordinate transformation unit. And based on the estimated rotor electrical angle Synchronous rotating coordinate system current feedback is obtained through the rotating coordinate transformation unit. The current loop controller outputs a voltage command in a synchronous rotating coordinate system. The voltage command in the two-phase stationary coordinate system is obtained through the inverse rotation coordinate transformation unit. Construct state data for parameter identification and observation; the state data includes ; S20, PINN Parameter Identification: Constructing a Physical Information Neural Network (PINN), and building physical residuals and loss functions based on the voltage balance equation of the PMSM in a synchronously rotating coordinate system; wherein the calculation of the physical residuals includes a current differential term. The current differential term is calculated from the sampled current sequence using a finite difference or sliding window differential filter; the state data is constructed as input feature vector data and input into PINN; the loss function is minimized through an online backpropagation algorithm to obtain the stator resistance estimate in real time. With inductance estimate ; S30, Physical Constraints and Smooth Updates: The PINN output parameters are mapped to a preset physical range using Sigmoid interval constraints, and the constrained parameters are then low-pass filtered or progressively updated to obtain the parameters used for injection. ; S40, Parameter Injection and Flux Observation: The flux linkage discrete update equation of the injected nonlinear flux linkage observer, which is based on... and Update stator flux linkage estimate ; and will Used for decoupling flux linkage components to obtain permanent magnet flux linkage components. ; S50, Closed-loop observation and control: The rotor electrical angle is calculated from the permanent magnet flux linkage component. The estimated rotational speed is obtained via a phase-locked loop (PLL) or differential rotation. ,Will and Feedback is sent to the vector control loop to achieve sensorless closed-loop control.
[0007] In one possible implementation, the input feature vector of the PINN includes: .
[0008] In one possible implementation, the current differential term in S20 is calculated using a finite difference or sliding window differential filter:
[0009] in This is a combined operator of sliding window difference and smoothing filter.
[0010] In one possible implementation, the physical residual in S20 is defined as: And construct the loss function: in To preset the flux linkage constant of the permanent magnet, These are the weighting coefficients. Electric angular velocity, This represents the length of the sliding window.
[0011] In one possible implementation, the electric angular velocity Estimated rotational speed The conversion yields:
[0012] in This represents the number of pole pairs of the motor.
[0013] In one possible implementation, the Sigmoid interval constraint in S30 applies to any parameter to be identified. The mapping is: ,
[0014] in This is the original output from the network. This is the preset physical range.
[0015] The present invention has the following beneficial effects: (1) The physical residual is constructed using the motor voltage balance equation, which can be adapted online without offline annotation.
[0016] (2) To Temperature rise and Online correction of saturation changes reduces flux estimation error and improves position / velocity estimation accuracy.
[0017] (3) Sigmoid interval constraints avoid outputting non-physical parameters, and smooth updates reduce the risk of transient disturbances.
[0018] (4) The fast and slow loop dual time scale and FIFO triggering mechanism take into account both real-time performance and computational overhead.
[0019] (5) PINN parameter identification and flux observation are performed using voltage commands output by the controller. and its coordinate transformation quantity As an equivalent voltage input, it avoids high-frequency noise and phase delay introduced by the additional voltage sampling link; with the core objective of achieving sensorless closed-loop control of permanent magnet synchronous motor, it balances low cost and high robustness. Attached Figure Description
[0020] Figure 1 This is a flowchart illustrating the steps of a sensorless PMSM control method based on PINN, according to an embodiment of the present invention. Detailed Implementation
[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0022] See Figure 1 The image shows a sensorless control method for a PMSM based on PINN according to an embodiment of the present invention, comprising the following steps: S10, Data Construction: Real-time acquisition of three-phase current in the vector control loop of the permanent magnet synchronous motor (PMSM). The two-phase stationary coordinate system current is obtained through the stationary coordinate transformation unit. And based on the estimated rotor electrical angle Synchronous rotating coordinate system current feedback is obtained through the rotating coordinate transformation unit. The current loop controller outputs a voltage command in a synchronous rotating coordinate system. The voltage command in the two-phase stationary coordinate system is obtained through the inverse rotation coordinate transformation unit. Construct state data for parameter identification and observation; the state data includes ; S20, PINN Parameter Identification: Constructing a Physical Information Neural Network (PINN), and building physical residuals and loss functions based on the voltage balance equation of the PMSM in a synchronously rotating coordinate system; wherein the calculation of the physical residuals includes a current differential term. The current differential term is calculated from the sampled current sequence using a finite difference or sliding window differential filter; the state data is constructed as input feature vector data and input into PINN; the loss function is minimized through an online backpropagation algorithm to obtain the stator resistance estimate in real time. With inductance estimate ; S30, Physical Constraints and Smooth Updates: The PINN output parameters are mapped to a preset physical range using Sigmoid interval constraints, and the constrained parameters are then smoothly updated to obtain the parameters used for injection. ; S40, Parameter Injection and Flux Observation: The flux linkage discrete update equation of the injected nonlinear flux linkage observer, which is based on... and Update stator flux linkage estimate ; and will Used for decoupling flux linkage components to obtain permanent magnet flux linkage components. ; S50, Closed-loop observation and control: The rotor electrical angle is calculated from the permanent magnet flux linkage component. The estimated rotational speed is obtained via a phase-locked loop (PLL) or differential rotation. ,Will and Feedback is sent to the vector control loop to achieve sensorless closed-loop control.
[0023] In a specific application example, the input feature vector of PINN includes: .
[0024] In a specific application example, the current differential term in S20 is calculated using a finite difference or sliding window differential filter:
[0025] in This is a combined operator of sliding window difference and smoothing filter.
[0026] In a specific application example, the physical residual in S20 is defined as: And construct the loss function:
[0027] in To preset the flux linkage constant of the permanent magnet, These are the weighting coefficients. Electric angular velocity, This represents the length of the sliding window.
[0028] Furthermore, electric angular velocity Estimated rotational speed The conversion yields:
[0029] in This represents the number of pole pairs of the motor.
[0030] A specific application example: the Sigmoid interval constraint in S30 for any parameter to be identified. The mapping is: ,
[0031] in This is the original output from the network. This is the preset physical range.
[0032] In a specific application example, smooth updates employ low-pass filtering or an asymptotic update law. The asymptotic update law is as follows:
[0033] in The parameters of PINN's current output after interval constraints. This is for updating coefficients.
[0034] In a specific application example, the nonlinear flux observer employs forward Euler discretization and includes at least the following flux update formula:
[0035] in The sampling period is For observer gain, It is a nonlinear function.
[0036] The error signal is constructed using flux linkage amplitude deviation:
[0037] And the nonlinear function Take the smooth form of the saturated function or the sign function: ,
[0038] in This is the smoothing coefficient.
[0039] The flux linkage component of the permanent magnet is obtained through inductive decoupling:
[0040] For surface-mount PMSMs For salient-pole PMSMs, preferably take
[0041] The rotor electrical angle is calculated as follows:
[0042] And through a phase-locked loop (PLL) Tracked Or obtained through angle difference and filtering Furthermore, during the startup or low-speed phase, PINN updates are paused and the parameters are kept at their nominal values or the previous valid estimates, pending further action. Online identification is enabled once convergence or the physical residual falls below a threshold.
[0043] Specific Example 1 The system adopts The vector control strategy. The speed loop controller operates based on a given speed. With feedback speed Output The current loop controller is based on... and Output voltage command The voltage command is obtained through the inverse rotation coordinate transformation unit. The inverter is driven by the SVPWM pulse generation unit to control the motor.
[0044] To achieve sensorless closed-loop operation, the system is equipped with an adaptive sensorless observation subsystem, which includes: a FIFO sliding window buffer unit, a PINN parameter identification unit, a physical interval constraint and smooth update unit, a nonlinear flux linkage observer unit, and a phase-locked loop (PLL) unit. Its overall workflow is as follows: 1) Sampling three-phase current The result obtained after static coordinate transformation unit .
[0045] 2) The nonlinear flux linkage observer unit is based on and renew and solve The phase-locked loop (PLL) unit or the angle differential speed measurement unit is composed of... get .
[0046] 3) Based on After being transformed by the rotating coordinate unit Converted to It is used for current loop feedback and PINN input construction.
[0047] 4) Write to the FIFO sliding window buffer unit; construct a state data sequence containing the above variables; when the buffer length reaches... One or more PINN training iterations are triggered at each sampling point.
[0048] 5) PINN constructs an input feature vector based on cached state data, and performs forward and backward propagation in conjunction with the calculated current differential term, outputting the feature vector. After applying sigmoid interval constraints and low-pass filtering or asymptotic updates, we obtain And inject the nonlinear flux observer module (update the resistance term and use it for inductor decoupling).
[0049] 6) Observer output The feedback is fed back into the FOC closed loop to form a sensorless closed-loop control.
[0050] Furthermore, the aforementioned It can be the equivalent voltage quantity of the controller output voltage command in the stationary coordinate system, or it can be obtained by equivalent reconstruction of the SVPWM duty cycle and the DC bus voltage, and is used as the input of the flux linkage observer.
[0051] Specific Example 2 PINN input can be used as follows:
[0052] The network can consist of several fully connected layers (e.g., 2 layers, each with 32 or 64 neurons), and the activation function can be Tanh or ReLU. To ensure the output satisfies physical plausibility, a Sigmoid scaling layer is set at the output to map the network's original output to a preset physical region.
[0053] Electric angular velocity is converted from estimated rotational speed:
[0054] use Constructing physical residuals from voltage balance equations: in To preset the flux linkage constant of the permanent magnet, It can be obtained from nameplate parameters, factory calibration, or offline identification, and remains unchanged during online identification. The current differential term is obtained through sliding window differential and smoothing filtering to reduce noise amplification. The loss function is: ,
[0055] An optimizer (such as Adam) is used to perform online backpropagation to update network parameters for the loss, so that the output... convergence.
[0056] Specific Example 3 To operate in a digital controller, the flux linkage state equation is discretized using the forward Euler method, with a sampling period of [missing information]. The flux linkage update equation for the nonlinear flux linkage observer in the stationary coordinate system is:
[0057] in For observer gain, It is a nonlinear function.
[0058] Preferably, the error signal is constructed using flux linkage amplitude deviation:
[0059] The nonlinear function It can take the smooth form of a saturated function or a sign function, for example: ,
[0060] in It is a saturation function. This is the smoothing coefficient.
[0061] In the above discrete update formula The magnetic flux integral term is obtained and injected by real-time output from PINN through interval constraints and smooth updates, thereby enabling the magnetic flux integral term to adapt to the resistance changes caused by temperature rise.
[0062] After obtaining the stator flux linkage, the permanent magnet flux linkage components are decoupled by combining the PINN identification inductor:
[0063] Surface-mounted PMSM is acceptable. ; salient-pole PMSM can be derived from Equivalent processing is performed to obtain Preferably, the following can be adopted:
[0064] Specific Example 4 The rotor electrical angle is calculated from the permanent magnet flux linkage component:
[0065] Subsequently, PLL was used. The system tracks the electrical angular velocity to obtain an estimated electrical angular velocity, which is then converted to an estimated rotational speed. Alternatively, angular differential scanning spectroscopy combined with low-pass filtering can be used to obtain... .
[0066] Specific Example 5 To ensure real-time performance in embedded systems, the system adopts a dual-time-scale architecture: The fast loop operates at a first frequency (e.g., 10 kHz), performing sampling, coordinate transformation, flux linkage observer updates, and FOC control; the slow loop operates at a second frequency (e.g., 1 kHz) or uses a FIFO-triggered mode: when the FIFO buffer accumulates... After sampling points, the PINN training iteration is triggered. After training, the output is updated by interval constraints and smoothing. Fast loops are injected. This strategy reduces the computational burden of online learning while ensuring control dynamics.
[0067] Furthermore, during the startup or low-speed phase, PINN updates can be paused and the parameters kept at their nominal values or the previous valid estimates, pending further action. Online identification will only be enabled after convergence or the physical residual falls below the threshold.
[0068] It should be understood that the exemplary embodiments described herein are illustrative and not restrictive. Although one or more embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art will understand that various changes in form and detail may be made without departing from the spirit and scope of the invention as defined by the appended claims.
Claims
1. A sensorless control method for PMSM based on PINN, characterized in that, Includes the following steps: S10, Data Construction: Real-time acquisition of three-phase current in the vector control loop of the permanent magnet synchronous motor (PMSM). The two-phase stationary coordinate system current is obtained through the stationary coordinate transformation unit. And based on the estimated rotor electrical angle Synchronous rotating coordinate system current feedback is obtained through the rotating coordinate transformation unit. The current loop controller outputs a voltage command in a synchronous rotating coordinate system. The voltage command in the two-phase stationary coordinate system is obtained through the inverse rotation coordinate transformation unit. Construct state data for parameter identification and observation; the state data includes ; S20, PINN Parameter Identification: Constructing a Physical Information Neural Network (PINN), and building physical residuals and loss functions based on the voltage balance equation of the PMSM in a synchronously rotating coordinate system; wherein the calculation of the physical residuals includes a current differential term. The current differential term is calculated from the sampled current sequence using a finite difference or sliding window differential filter; the state data is constructed as input feature vector data and input into PINN; the loss function is minimized through an online backpropagation algorithm to obtain the stator resistance estimate in real time. With inductance estimate ; S30, Physical Constraints and Smooth Updates: The PINN output parameters are mapped to a preset physical range using Sigmoid interval constraints, and the constrained parameters are then smoothly updated to obtain the parameters used for injection. ; S40, Parameter Injection and Flux Observation: [The following text appears to be incomplete and requires further context: The flux linkage discrete update equation of the injected nonlinear flux linkage observer, which is based on... and Update stator flux linkage estimate ; and will Used for decoupling flux linkage components to obtain permanent magnet flux linkage components. ; S50, Closed-loop observation and control: The rotor electrical angle is calculated from the permanent magnet flux linkage component. The estimated rotational speed is obtained via a phase-locked loop (PLL) or differential rotation. ,Will and Feedback is sent to the vector control loop to achieve sensorless closed-loop control.
2. The PMSM sensorless control method based on PINN as described in claim 1, characterized in that, The input feature vector of the PINN includes: 。 3. The PMSM sensorless control method based on PINN as described in claim 1, characterized in that, The current differential term in S20 is calculated using a finite difference or sliding window differential filter: ， in This is a combined operator of sliding window difference and smoothing filter.
4. The PMSM sensorless control method based on PINN as described in claim 1, characterized in that, The physical residual in S20 is defined as follows: And construct the loss function: in To preset the flux linkage constant of the permanent magnet, These are the weighting coefficients. Electric angular velocity, This represents the length of the sliding window.
5. The PMSM sensorless control method based on PINN as described in claim 4, characterized in that, The electric angular velocity Estimated rotational speed The conversion yields: ， in This represents the number of pole pairs of the motor.
6. The PMSM sensorless control method based on PINN as described in claim 1, characterized in that, The Sigmoid interval constraint described in S30 applies to any parameter to be identified. The mapping is: ， in This is the original output from the network. This is the preset physical range.