Anti-interference speed control method for permanent magnet synchronous motor

Abstract

The invention relates to an anti-interference speed control method for a permanent magnet synchronous motor, in particular to the technical field of electrical engineering, and effectively solves the problem that the control performance of the permanent magnet synchronous motor is degraded due to model parameter drift and external load sudden change under complex working conditions. According to the method, the high robustness of sliding mode control and the rolling optimization characteristic of model prediction control are creatively fused, and an adaptive observer is introduced to estimate the system state and lumped disturbance online, so that the real-time feedforward compensation and active suppression of parameter uncertainty and disturbance are realized; according to the method, the dynamic response speed and the steady-state precision of the system are remarkably improved, the rotating speed fluctuation and the torque ripple are effectively inhibited, meanwhile, controller parameters are adaptively adjusted through an intelligent learning mechanism, the stability and the adaptability in long-term operation are ensured, and the method is suitable for large-scale popularization and application. The limitation that a traditional control strategy depends on an accurate model and the disturbance boundary is unknown is overcome fundamentally, and high-performance and high-robustness speed control is achieved.

Description

Technical Field

[0001] This invention relates to the field of electrical engineering technology, and more specifically, to an anti-interference speed control method for a permanent magnet synchronous motor. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs), with their high power density, high efficiency, and superior dynamic performance, have become core actuators in electric vehicle drive systems and high-precision industrial servo systems. However, these applications typically operate in complex, variable, and difficult-to-model harsh environments. In electric vehicles, the motor not only needs to cope with the slow drift of key parameters such as permanent magnet flux decay, stator resistance, and inductance caused by internal temperature rise during long-term operation, but also needs to withstand the severe impact of external load torque from sudden changes in road gradient, emergency acceleration, or braking. In industrial automation, servo motors may face sudden changes in mechanical load during processing and disturbances such as grid voltage fluctuations. The combined effect of these time-varying internal parameters and external disturbances causes the actual operating conditions of the motor to deviate significantly from its pre-set ideal mathematical model, leading to system model mismatch problems, which pose a serious challenge to control strategies that rely on accurate models.

[0003] To address the aforementioned control challenges, existing technologies widely employ advanced strategies such as model predictive control (MMC) and sliding mode control (SMCC). MMC, through rolling optimization and feedback correction, exhibits excellent dynamic performance, but its control quality is highly dependent on the accuracy of the motor's mathematical model. When motor parameters drift, the mismatch in the predictive model directly leads to a significant decrease in control performance, manifested as increased torque ripple and exacerbated speed fluctuations. While SMC is renowned for its strong robustness to parameter disturbances and external interference, its full realization depends on accurate knowledge of the disturbance boundaries. In practical systems, disturbance boundaries are difficult to obtain precisely, often resulting in overly conservative control actions or the introduction of significant chattering to suppress uncertainties, thus deteriorating system performance. Although schemes such as model reference adaptive control attempt to identify parameters online, their algorithms are typically complex, difficult to implement in real-time, and challenging to effectively coordinate with predictive control frameworks. Therefore, the current technological bottleneck lies in the lack of a novel strategy capable of dynamically adjusting the predictive model online and effectively integrating a robust control mechanism to ensure high-performance and stable system operation even under conditions of model uncertainty. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing an anti-interference speed control method for permanent magnet synchronous motors, thereby solving the problems mentioned in the background art.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: a method for anti-interference speed control of a permanent magnet synchronous motor, comprising the following steps:

[0006] Step S1: In the control system of permanent magnet synchronous motor, the raw data of three-phase current, DC bus voltage, rotor position and radiator temperature of the motor are synchronously collected by sensor array at a fixed sampling frequency. The raw data is aligned in real time and filtered by an adaptive Kalman filter algorithm that dynamically adjusts the observation noise covariance matrix according to the real-time speed of the motor to obtain a clean synchronous multiphysics data packet.

[0007] Step S2: Construct a collaborative identification unit consisting of a recursive least squares parameter estimator and an adaptive sliding mode observer; using the clean synchronous multiphysics data packet obtained in step S1, the recursive least squares parameter estimator identifies the stator resistance and quadrature-direct axis inductance of the motor online, forming a real-time parameter matrix; simultaneously, based on the current and voltage data and the real-time parameter matrix obtained in step S1, the adaptive sliding mode observer generates an equivalent control quantity through a sliding mode variable structure control law, synchronously estimates the system state and the total disturbance of the unified lumped array, and outputs the observed value of the system state and the real-time estimated value of the total disturbance;

[0008] Step S3: Establish a finite control set model predictive controller with integrated sliding surface constraints. First, using the real-time estimate of the total disturbance obtained in step S2 as the feedforward compensation, a sliding surface with respect to the current tracking error is constructed. Then, during the rolling optimization process of the model predictive controller, the degree of convergence of the system state trajectory to the sliding surface is incorporated as a soft constraint into the performance index function. At the same time, the predictive model is constructed using the real-time parameter matrix obtained in step S2. By solving for the minimum value of the performance index function, the optimal voltage vector is obtained and applied.

[0009] Step S4: Continuously monitor the control performance indicators of the motor and store them in association with the operating conditions, real-time parameter matrix and total disturbance estimate; use an ensemble learning model to analyze historical data, and when performance changes are identified, adaptively adjust the parameters of the prediction model and the gain of the adaptive sliding mode observer.

[0010] In a preferred embodiment, the specific operation flow of real-time alignment of the original data and filtering processing using an adaptive Kalman filter algorithm in step S1 is as follows:

[0011] First, the three-phase current, DC bus voltage, rotor position and radiator temperature data synchronously collected by the sensor array are strictly aligned based on their corresponding timestamps.

[0012] Subsequently, based on coordinate transformation theory, the aligned three-phase current data are transformed from the natural coordinate system to the orthogonal coordinate system that rotates synchronously with the rotor magnetic field to obtain the direct-axis current component and the quadrature-axis current component.

[0013] Subsequently, the rotor position signal is numerically differentiated to calculate the real-time mechanical angular velocity of the motor. Based on this real-time mechanical angular velocity, the value of the observation noise covariance matrix used by the adaptive Kalman filter in the current sampling period is dynamically adjusted. The adjustment strategy is that the value of this matrix decreases as the real-time mechanical angular velocity increases and increases as the real-time mechanical angular velocity decreases.

[0014] Finally, using the aforementioned dynamically adjusted observation noise covariance matrix, Kalman filtering estimation was performed on the direct-axis current component, quadrature-axis current component, DC bus voltage, and radiator temperature signal to obtain the filtered estimates of each physical quantity.

[0015] In a preferred embodiment, the specific operation of dynamically adjusting the observation noise covariance matrix is ​​as follows:

[0016] A nonlinear scaling factor is constructed with real-time mechanical angular velocity and radiator temperature as independent variables. This nonlinear scaling factor is used to scale a preset benchmark observation noise covariance matrix. The nonlinear scaling factor consists of a constant term with a value of one, a velocity-related term, and a temperature-related term added together.

[0017] The velocity-related term is configured as a negative exponential function of the absolute value of the real-time mechanical angular velocity, and the temperature-related term is configured as a quadratic function of the difference between the real-time temperature and a preset reference temperature. The reference observation noise covariance matrix is ​​multiplied by a nonlinear scaling factor to generate the observation noise covariance matrix that is finally applied to the adaptive Kalman filter algorithm at the current sampling time.

[0018] In a preferred embodiment, in step S2, the specific operation of the recursive least squares parameter estimator is as follows: First, a discrete-time system regression model for motor parameter identification is constructed. The input variables of this model are the filtered direct-axis current component, the filtered quadrature-axis current component, the filtered DC bus voltage, and the real-time electrical angular velocity calculated by the differential of the filtered rotor position signal in the clean data packet obtained in step S1. The output variables are the direct-axis voltage component and the quadrature-axis voltage component. The parameters to be identified are the stator resistance, direct-axis inductance, quadrature-axis inductance, and permanent magnet flux linkage of the permanent magnet synchronous motor.

[0019] Subsequently, a recursive least squares algorithm with a forgetting factor is used for online parameter updates. The recursive least squares algorithm updates the parameter estimation vector at the current time by calculating the gain matrix at the current time, the parameter estimates at the previous time, and the error between the observed data and the model predictions at the current time, and also updates the covariance matrix of the algorithm.

[0020] Among them, the forgetting factor is dynamically adjusted according to the system operating state. The adjustment strategy is that the value of the forgetting factor is negatively correlated with the norm of the system state estimation error.

[0021] Finally, the stator resistance estimate, direct-axis inductance estimate, quadrature-axis inductance estimate, and permanent magnet flux estimate obtained at each sampling time are combined into a diagonal real-time parameter matrix, providing accurate motor model parameters for the subsequent adaptive sliding mode observer.

[0022] In a preferred embodiment, the specific design and operation of the adaptive sliding mode observer in step S2 is as follows:

[0023] First, based on the real-time parameter matrix output by the recursive least squares parameter estimator, an extended state-space model containing the electrical state of the motor and lumped disturbances is constructed. This extended model treats the parameter uncertainties, unmodeled dynamics, and external disturbances in the system as lumped disturbance terms and as extended state variables.

[0024] Subsequently, a sliding mode observer structure is designed, which includes a model-based feedforward term, a linear feedback term to ensure the dynamic stability of the estimation error, and a discontinuous switching feedback term based on the sliding surface to suppress the influence of disturbances. The switching gain of the sliding mode observer is adjusted online using an adaptive law, which makes the magnitude of the sliding gain proportional to the norm of the sliding surface defined between the actual system output and the observer output, and also proportional to the integral term of the sliding surface norm.

[0025] The adaptive sliding mode observer eventually estimates the direct-axis current state, quadrature-axis current state, and unified lumped direct-axis and quadrature-axis disturbances of the motor, and outputs these state observations and disturbance estimates.

[0026] In a preferred embodiment, the collaborative working mechanism between the recursive least squares parameter estimator and the adaptive sliding mode observer in step S2 is specifically as follows:

[0027] A bidirectional data exchange and correction loop is established between the parameter estimator and the state observer. In the forward channel, the real-time parameter matrix updated and output by the recursive least squares parameter estimator in each sampling period is transmitted in real time to the adaptive sliding mode observer to update the coefficient matrix of its internal state-space expression based on the motor mathematical model, ensuring that the observer's dynamic model is consistent with the actual motor parameters. In the feedback channel, the norm of the state estimation error vector between the actual system output calculated by the adaptive sliding mode observer and the observer output is sent as a feedback signal to the recursive least squares parameter estimator to dynamically correct the magnitude of its forgetting factor. When the norm of the state estimation error vector increases, the forgetting factor is automatically reduced; when the norm of the state estimation error vector decreases, the forgetting factor is increased. Through this bidirectional collaborative mechanism of forward parameter transmission and feedback error correction, closed-loop optimization of parameter identification and state disturbance estimation is achieved.

[0028] In a preferred embodiment, the specific operation flow for establishing the finite control set model predictive controller with fused sliding surface constraints in step S3 is as follows:

[0029] First, the real-time estimate of the total disturbance obtained in step S2 is used as the feedforward compensation amount. Combined with the tracking error between the current reference value and the actual current measurement value at the current moment, a sliding surface for the current tracking error is constructed. In the process of constructing this sliding surface, the current tracking error is weighted by introducing a sliding surface error weight matrix, and the total disturbance estimate is weighted by introducing a disturbance compensation weight matrix, so that the constructed sliding surface can both reflect the tracking error of the system and actively compensate for the system disturbance.

[0030] Subsequently, during the rolling optimization process of the model predictive controller, the degree of convergence of the system state trajectory to the sliding surface is incorporated as a soft constraint into the performance index function. This soft constraint is weighted by the sliding surface weight matrix. At the same time, the real-time parameter matrix obtained in step S2 is used to construct a predictive model, which is a dynamic model based on the discrete state equation of a permanent magnet synchronous motor.

[0031] Finally, by solving for the minimum value of the performance index function that incorporates the soft constraint conditions of the sliding mode surface, the optimal voltage vector is obtained and applied under the constraint condition of a finite set of inverter voltage vectors.

[0032] In a preferred embodiment, the specific design and optimization solution of the performance index function in step S3 is as follows:

[0033] The performance index function consists of a weighted sum of three parts. The first part is the weighted sum of squared errors between the system output and the reference trajectory in the prediction time domain. The weighting is performed by the tracking error weight matrix to ensure the tracking performance of the system.

[0034] The second part is the weighted sum of squares of the control increment changes in the control time domain, which is weighted by the control increment weight matrix to suppress drastic changes in the control signal;

[0035] The third part is the weighted sum of squares of the predicted sliding surface function values ​​in the time domain. The weights are applied by the sliding surface weight matrix and used as soft constraints to drive the system state to approach the sliding surface.

[0036] When solving for the minimum value of the performance index function, a finite control set optimization strategy is adopted. By enumerating all possible combinations of voltage vectors of the inverter, the predicted value of the future state of the system and the corresponding performance index function value under each combination are calculated. Finally, the voltage vector that minimizes the performance index function value is selected as the optimal control quantity and applied to the next control cycle.

[0037] In a preferred embodiment, step S4, specifically the operation of continuously monitoring the motor's control performance indicators and associating and storing them with the operating conditions, real-time parameter matrix, and total disturbance estimate, is as follows:

[0038] Real-time monitoring of multiple control performance indicators of the motor, including the root mean square value of current tracking error, output torque ripple coefficient, equivalent switching frequency and Lyapunov exponent estimate, to form a comprehensive performance feature vector;

[0039] Simultaneously, the operating condition data of the motor is collected, including speed, load torque, radiator temperature, and running time, forming an operating condition feature vector. The comprehensive performance feature vector, operating condition feature vector, real-time parameter matrix from step S2, and total disturbance estimate at each sampling moment are correlated to construct a unified time series data record. Based on the importance and characteristics of the data, a hierarchical storage strategy is adopted, including saving short-term memory data of the most recent sampling periods, typical long-term memory data of operation when a new stable operating condition is identified, and abnormal state data of performance indicators exceeding preset thresholds, to establish a complete historical operating database.

[0040] In a preferred embodiment, step S4, which involves analyzing historical data using an ensemble learning model and adaptively adjusting parameters when performance changes are detected, specifically includes the following steps:

[0041] A hybrid ensemble learning model is constructed, which includes a random forest operating condition identifier, an online sequence extreme learning machine performance predictor, and a performance change detector. The model is trained and learned online using data from a historical operating database, enabling it to accurately identify the current operating condition category, predict future performance index values, and detect abnormal changes in system performance.

[0042] When the performance change detector determines that the change in the system performance index exceeds the preset threshold, it triggers the parameter adaptive adjustment mechanism. This mechanism adjusts the parameters of the prediction model in step S3 and the gain of the adaptive sliding mode observer in step S2 based on the analysis results of the ensemble learning model. The adjustment of the prediction model parameters is based on the gradient information of the performance index on the model parameters, and the adjustment of the adaptive sliding mode observer gain is based on the product of the current gain and the hyperbolic tangent function of the performance change magnitude.

[0043] After the adjustment is completed, assess the degree of improvement in system performance. If the improvement does not meet the expected requirements, roll back the parameter adjustment and enable the backup control strategy.

[0044] The beneficial effects of this invention are: it effectively solves the problem of control performance degradation of permanent magnet synchronous motors under complex working conditions caused by model parameter drift and sudden changes in external load. By innovatively integrating the strong robustness of sliding mode control with the rolling optimization characteristics of model predictive control, and introducing an adaptive observer to estimate the system state and lumped disturbances online, it achieves real-time feedforward compensation and active suppression of parameter uncertainties and disturbances. This method significantly improves the dynamic response speed and steady-state accuracy of the system, effectively suppresses speed fluctuations and torque pulsations, and at the same time, through an intelligent learning mechanism to adaptively adjust the controller parameters, it ensures stability and adaptability in long-term operation. It fundamentally overcomes the limitations of traditional control strategies that rely on accurate models and have unknown disturbance boundaries, and achieves high-performance and highly robust speed control. Attached Figure Description

[0045] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0047] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0048] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0049] This embodiment provides, for example Figure 1The method for anti-interference speed control of a permanent magnet synchronous motor, as shown, includes the following steps:

[0050] Step S1: In the control system of the permanent magnet synchronous motor, the raw data of the motor's three-phase current, DC bus voltage, rotor position, and radiator temperature are synchronously collected at a fixed sampling frequency through a sensor array. The raw data is aligned in real time and filtered using an adaptive Kalman filter algorithm that dynamically adjusts the observation noise covariance matrix according to the motor's real-time speed to obtain a clean synchronous multiphysics data packet. This step effectively suppresses measurement noise and synchronization errors between different sensors through dynamic adaptive filtering technology, providing a high-precision and high-reliability synchronous data foundation for subsequent control links, and improving the control quality of the system from the source.

[0051] Step S2: Construct a collaborative identification unit consisting of a recursive least squares parameter estimator and an adaptive sliding mode observer. Using the clean synchronous multiphysics data package obtained in Step S1, the recursive least squares parameter estimator identifies the stator resistance and quadrature-axis inductance of the motor online, forming a real-time parameter matrix. Simultaneously, the adaptive sliding mode observer, based on the current and voltage data and the real-time parameter matrix obtained in Step S1, generates an equivalent control quantity through a sliding mode variable structure control law, synchronously estimating the system state and the total disturbance of the unified lumped disturbance, and outputting the observed value of the system state and the real-time estimated value of the total disturbance. This step, through the deep collaboration between the parameter estimator and the state observer, achieves accurate online identification of key motor parameters and real-time observation of lumped disturbances, effectively overcoming the model mismatch problem caused by time-varying motor parameters and external disturbances, and providing accurate model and disturbance feedforward information for robust control.

[0052] Step S3: Establish a finite control set model predictive controller with integrated sliding surface constraints. First, using the real-time estimate of the total disturbance obtained in step S2 as the feedforward compensation, a sliding surface with respect to the current tracking error is constructed. Subsequently, during the rolling optimization process of the model predictive controller, the degree of convergence of the system state trajectory to the sliding surface is incorporated as a soft constraint into the performance index function. At the same time, the predictive model is constructed using the real-time parameter matrix obtained in step S2. By solving for the minimum value of the performance index function, the optimal voltage vector is obtained and applied. This step deeply integrates the strong robustness of sliding mode control with the rolling optimization characteristics of predictive control. Through feedforward compensation and sliding surface soft constraints, the system possesses excellent dynamic performance of predictive control while also having inherent robustness to cope with model uncertainties and disturbances, achieving a unity of high performance and high robustness.

[0053] Step S4: Continuously monitor the control performance indicators of the motor and store them in association with the operating conditions, real-time parameter matrix, and total disturbance estimate; use an integrated learning model to analyze historical data, and when performance changes are identified, adaptively adjust the parameters of the prediction model and the gain of the adaptive sliding mode observer; this step constitutes the intelligent optimization layer of the system. Through continuous monitoring and learning, the controller parameters have online self-tuning capabilities, which can adapt to the characteristic changes and environmental changes of the motor during long-term operation, thereby significantly improving the adaptive capability and long-term operational stability of the control system, and realizing the leap from "passive suppression" to "active adaptation".

[0054] In this embodiment, it is particularly important to note that in step S1, as the first step of data perception and preprocessing, the core task is to provide high-quality, time-synchronized, and physically meaningful input data, laying a solid foundation for accurate parameter identification, state and disturbance observation, and optimized control in subsequent steps. The innovation of this step lies in the adoption of a dynamic filtering mechanism that is deeply coupled with the motor's operating state, which surpasses conventional static filtering.

[0055] This step receives the raw data stream from the sensor array, and its processing logic proceeds sequentially, ultimately producing a standardized data packet that can be directly used by step S2.

[0056] In step S1, the specific operation process of real-time alignment of the original data and filtering using the adaptive Kalman filter algorithm is as follows:

[0057] First, the three-phase current is synchronously acquired by a sensor array at a fixed sampling frequency (e.g., 10kHz). DC bus voltage Rotor position and radiator temperature The data, based on its corresponding timestamp, is aligned to a preset synchronization time node (e.g., every 0, 1ms) using an interpolation algorithm (such as linear interpolation or spline interpolation) to eliminate the transmission delay, which is usually less than [e.g. 50μs], caused by differences in sensor hardware response time and sample-and-hold circuits, thus forming a time-aligned original data sequence.

[0058] Subsequently, based on coordinate transformation theory, the aligned three-phase current data are transformed from the natural coordinate system to an orthogonal coordinate system that rotates synchronously with the rotor magnetic field. This transformation specifically includes the Clark transformation and the Park transformation. First, the three-phase current... Transformation to two-phase stationary Coordinate system, to obtain current components and Then, based on the rotor position Calculate electrical angle (in (where the number of pole pairs is the motor pole pair) and Transforming to the dq rotating coordinate system, the direct-axis current component is finally obtained. With cross-axis current component This step converts the time-varying AC quantity into DC quantity, which facilitates subsequent processing.

[0059] Subsequently, the rotor position signal Numerical differentiation is performed to calculate the real-time mechanical angular velocity of the motor. Based on this real-time mechanical angular velocity The value of the observation noise covariance matrix used by the adaptive Kalman filter (KF) within the current sampling period is dynamically adjusted. The value of is adjusted by the following strategy: when the real-time mechanical angular velocity... Speeds below a first speed threshold (e.g., 100 rpm, corresponding to 5% of the motor's rated speed) are considered low-speed zones, at which point the observation noise covariance matrix is ​​significantly increased. The value of (e.g., adjusted to 3-5 times the base value) is used to enhance the filtering and smoothing effect; when If the speed exceeds the second speed threshold (e.g., 80% of the rated speed), it is considered a high-speed zone, and the speed is reduced. The value of (e.g., adjusted to 0, 5-0, 8 times the base value) is used to improve the system response speed. This strategy aims to enhance filtering to suppress noise at low speeds and improve response speed at high speeds.

[0060] Finally, the dynamically adjusted observation noise covariance matrix described above is adopted. For the direct-axis current components respectively quadrature axis current components DC bus voltage and radiator temperature The signal undergoes the prediction and update steps of standard Kalman filtering (KF) estimation to obtain the optimal filtered estimate of each physical quantity at the current time, denoted as... These filtered estimates together constitute a clean synchronous multiphysics data packet.

[0061] The specific operation for dynamically adjusting the observation noise covariance matrix is ​​as follows:

[0062] Construct a system based on real-time mechanical angular velocity and radiator temperature Nonlinear scaling factor for the independent variable This non-linear scaling factor Used to evaluate a pre-defined baseline observation noise covariance matrix Scaling, non-linear scaling factor It consists of three parts: a constant term with a value of one, a velocity-related term, and a temperature-related term. Its specific mathematical expression is as follows:

[0063]

[0064] in, The adjustment coefficient is greater than zero and is determined experimentally. Its specific value is determined based on the rated parameters of the motor and the expected dynamic response speed. In practice, by performing parameter sweep on a permanent magnet synchronous motor with a rated power of [e.g., 5kW] under typical operating conditions (e.g., speed range of 0-3000rpm), the value range of α is determined to be [e.g., 0, 1-1, 0], the value range of β is [e.g., 0, 01-0, 1s / rad], and the value range of γ is [e.g., 1e-4-1e-31 / ℃²]. Those skilled in the art can adjust it within this range according to the actual motor specifications. The preset reference temperature is usually set to the standard ambient temperature under the rated operating conditions of the motor (such as 25℃) or the average radiator temperature when the motor is working normally (such as 60℃), as the benchmark point for temperature compensation.

[0065] In the expression: is a nonlinear scaling factor, a dimensionless number, used to dynamically scale the reference observation noise covariance matrix; This is the real-time mechanical angular velocity, usually measured in rad / s, and is used as a core adjustment variable to reflect the motor's operating speed. This is the real-time temperature of the radiator, in °C, used to sense changes in system temperature. This is a dimensionless adjustment coefficient for the speed term, used to control the influence of speed-related terms during low-speed operation. This is the velocity attenuation coefficient, measured in seconds per rad, used to control how quickly velocity affects the rate of attenuation. This is the temperature adjustment factor, with the unit 1 / ℃², used to control the weight of the impact when the temperature deviates from the reference value; This is the preset reference temperature, in °C, used as a reference point for temperature adjustment. It is a negative exponential function of the absolute value of velocity, used to realize the velocity-related terms as... It increases but decays exponentially; It is a quadratic function of the temperature difference, used to symmetrically handle the effects of positive and negative temperature shifts.

[0066] Among them, the speed-related terms are configured to be related to the real-time mechanical angular velocity. The negative exponential function of absolute value, i.e. This makes the contribution significant in the low-speed operating region, driving the observation noise covariance matrix. Increased to enhance the filtering and smoothing effect, this contribution decreases exponentially with increasing speed. The temperature-related term is configured to be relevant to the real-time temperature. With a preset reference temperature The quadratic function of the difference, i.e. This contribution increases as the real-time temperature deviates from the reference temperature, thus moderately increasing the observation noise covariance matrix. The value of is chosen to reduce the confidence level of sensor data that may be affected by temperature drift.

[0067] Benchmark observation noise covariance matrix With nonlinear scaling factor Multiplication, that is This generates the observation noise covariance matrix that is ultimately applied to the adaptive Kalman filter algorithm at the current sampling time. This formula, by integrating the effects of velocity and temperature, achieves the observation noise covariance matrix. It adaptively adjusts across all operating conditions to optimize the filtering effect.

[0068] Step S1 achieves adaptive filtering by introducing a dynamic observation noise covariance adjustment mechanism that is nonlinearly related to rotational speed and temperature.

[0069] In this embodiment, it is particularly important to note that the core innovation of step S2 lies in constructing a deep collaborative mechanism between the parameter estimator and the state observer. This enables the synchronous and accurate identification of motor parameters and disturbances, providing an accurate model foundation for subsequent robust predictive control. In the aforementioned implementation, step S1 provides a clean synchronous multiphysics data package. .

[0070] In step S2, the specific operation of the recursive least squares parameter estimator (RLSE) is as follows:

[0071] First, a discrete-time system regression model for motor parameter identification is constructed. The input variable of this model is the filtered direct-axis current component in the clean data packet obtained in step S1. Filtered quadrature-axis current component DC bus voltage after filtering and the filtered rotor position signal Real-time electrical angular velocity obtained by differential calculation The output variable is the direct-axis voltage component. and cross-axis voltage components The parameter to be identified is the stator resistance of the permanent magnet synchronous motor. Direct-axis inductor quadrature axis inductance and permanent magnet magnetic flux The mathematical expression for constructing the discrete-time system regression model for motor parameter identification is as follows:

[0072]

[0073] in, Sampling time, The sampling period is and To observe noise, the model discretizes the continuous-time motor equations, providing a mathematical basis for parameter estimation.

[0074] Subsequently, a forgetting factor was used. The Recursive Least Squares (RLS) algorithm is used for online parameter updates. The RLS algorithm calculates the gain matrix at the current time step. Based on the parameter estimates from the previous time step and the observation data at the current moment Compared with model predictions The error is used to update the parameter estimation vector at the current time step. And update the covariance matrix of the algorithm. Its recursive formula is:

[0075]

[0076]

[0077]

[0078] in, The Kalman gain matrix at time k is used to weigh the confidence levels between new observations and historical estimates. This represents the parameter estimation vector at time k, containing... The estimated value; The covariance matrix of the estimation error at time k represents the uncertainty of the parameter estimation. The regression matrix at time k is composed of observed data such as current and angular velocity. This represents the output vector at time k, i.e. ; This represents the forgetting factor, used to control the degree to which the algorithm remembers historical data; The identity matrix is ​​represented; the algorithm updates the motor parameters in real time through recursion.

[0079] Among them, the forgetting factor The system is dynamically adjusted based on its operating status. The adjustment strategy involves a negative correlation between the value of the forgetting factor and the norm of the system state estimation error. Specifically, the forgetting factor... The dynamic adjustment formula is:

[0080]

[0081] In the above formula: This is the nominal value of the forgetting factor. To adjust the coefficient, This provides the state estimation error vector for the Adaptive Sliding Mode Observer (ASMO). When the state estimation error norm... An increase indicates a worsening model mismatch, prompting an automatic decrease. To accelerate parameter tracking; when the error norm decreases, it increases. To ensure the smoothness of the estimation.

[0082] Finally, the stator resistance estimate obtained at each sampling time is updated. Direct-axis inductance estimation value Quadrature axis inductance estimate and permanent magnet flux linkage estimation value Combining into a diagonal form of the real-time parameter matrix This provides accurate motor model parameters for the subsequent Adaptive Sliding Mode Observer (ASMO).

[0083] In step S2, the specific design and operation of the Adaptive Sliding Mode Observer (ASMO) are as follows:

[0084] First, the real-time parameter matrix is ​​based on the output of the recursive least squares parameter estimator (RLSE). An extended state-space model is constructed that includes the electrical state of the motor and lumped disturbances. This extended model treats parameter uncertainties, unmodeled dynamics, and external disturbances in the system as a unified lumped disturbance term. And as an extended state variable, the extended state vector is Its state-space equation is:

[0085]

[0086]

[0087] in, For dependent on real-time parameter matrix The system matrix, For the input matrix, For the input voltage vector, This is the output matrix. The model makes the perturbations explicit by extending the states, which facilitates observer design.

[0088] Subsequently, a sliding mode observer (ASMO) structure is designed, which includes a model-based feedforward term, a linear feedback term to ensure the dynamic stability of the estimation error, and a discontinuous switching feedback term based on the sliding surface to suppress the effects of disturbances. The dynamic equations of the sliding mode observer (ASMO) are as follows:

[0089]

[0090] in, This is the state estimation vector. To output the estimated value, For linear feedback gain matrix, To switch the gain matrix, For sliding surface functions, It is a symbolic function.

[0091] Switching gain of sliding mode observer (ASMO) An adaptive law is used for online adjustment, which makes the magnitude of the sliding mode gain relative to the sliding surface defined between the actual system output and the observer output. The norm is proportional to the integral term of the sliding surface norm, and the mathematical expression of the adaptive law is:

[0092]

[0093] In the above expression: This represents the adaptive sliding mode gain matrix at time k; Represents the basic sliding mode gain matrix; This represents the adaptive adjustment coefficient, indicating the sensitivity of the control gain to the sliding surface and its integral. Let k represent the sliding mode surface vector at time k. This formula achieves a performance balance by adaptively adjusting the gain to enhance robustness when the disturbance is large and suppress chattering when the disturbance is small.

[0094] The Adaptive Sliding Mode Observer (ASMO) ultimately estimates the direct-axis current state of the motor synchronously. Quadrature axis current state and unified lumped direct-axis disturbance and cross-axis disturbance It outputs these state observations and disturbance estimates.

[0095] To further improve the observation speed and accuracy of broadband disturbances (especially rapid load impacts under conditions such as emergency braking) and avoid untimely feedforward compensation due to the dynamic lag of the observer, this invention optimizes the design of the adaptive sliding mode observer (ASMO). Its core lies in the parallel processing architecture that combines a high-precision identification module for slow-varying disturbances with a high-speed response module for fast-varying disturbances. Through prediction mechanisms and hardware acceleration technology, a leapfrog improvement in disturbance observation performance is achieved.

[0096] Modular Design and Collaborative Mechanism: A high-precision slow-disturbance identification module focuses on lumped disturbances caused by slow drift of motor parameters (such as resistance changes and flux decay due to temperature rise) and gradually varying loads. This module works in deep collaboration with the recursive least squares parameter estimator (RLSE), utilizing its output real-time parameter matrix for precise model feedforward compensation. This module employs long-term time-series data (typically with a time window of 100ms to 1s) for fitting and smoothing estimation, achieving a parameter identification accuracy within ≤1.5%, thus providing a high-precision steady-state and quasi-steady-state model foundation for the system.

[0097] The high-speed response module for rapid disturbances is specifically designed to handle rapid disturbances caused by sudden changes in external load (such as emergency braking or mechanical jamming). This module employs a high-gain adaptive law, and its sliding mode gain can be dynamically adjusted according to the derivative of the estimation error, thus achieving extremely high response speed. To meet real-time requirements, the algorithm of this module (including state prediction, sliding mode surface calculation, and gain update) is implemented in a hardware parallel pipeline on a field-programmable gate array (FPGA), ensuring that the entire process from the occurrence of the disturbance to the convergence of the estimation value is completed within 0.3ms.

[0098] To fundamentally compensate for the inherent computation and convergence delay of the observer itself, a Long Short-Term Memory (LSTM) network is introduced to construct a disturbance trend predictor. This LSTM network uses historical operating data of the motor (including current, voltage, speed, and disturbance values ​​estimated by the fast-change disturbance module) as training samples and is trained offline to learn the dynamic characteristics of the system.

[0099] In practical implementation, the LSTM network operates online, predicting the trend of rapidly changing disturbances 5ms in advance. This predicted value serves as the initial amount for feedforward compensation. Subsequently, the real-time high-precision estimate generated by the high-speed response module for rapidly changing disturbances is used to quickly correct the residual between the LSTM prediction value and the actual disturbance. Finally, the disturbance prediction value corrected in real time by the rapidly changing module is fused with the high-precision parameter compensation amount output by the slowly changing module to form the final total disturbance feedforward compensation amount. This fusion mechanism effectively compresses the execution lag of feedforward compensation to within 0.2ms, thereby solving the key problem of control performance degradation or even instability caused by insufficient observer response speed under dynamic extreme conditions (such as emergency braking).

[0100] Both the fast-change perturbation high-speed response module and the forward inference process of the LSTM network are implemented on the FPGA chip. By designing dedicated parallel computing units and pipeline structures, tasks such as matrix operations and nonlinear function calculations are decomposed into multiple steps that can be executed simultaneously, making full use of hardware parallelism and meeting the requirements of the above algorithms for extremely high computing speed.

[0101] In step S2, the collaborative working mechanism between the recursive least squares parameter estimator (RLSE) and the adaptive sliding mode observer (ASMO) is as follows:

[0102] A bidirectional data exchange and correction loop is established between the parameter estimator (RLSE) and the state observer (ASMO); in the forward channel, the recursive least squares parameter estimator updates and outputs the real-time parameter matrix in each sampling period. The coefficient matrix is ​​transmitted in real time to the adaptive sliding mode observer to update its internal state-space expression based on the motor mathematical model. This ensures that the observer's dynamic model remains consistent with the actual parameters of the motor.

[0103] In the feedback channel, the state estimation error vector between the actual system output calculated by the adaptive sliding mode observer and the observer output is represented. The norm of the function is sent as a feedback signal to the recursive least squares parameter estimator to dynamically correct its forgetting factor. The magnitude of the error vector is determined by the correction strategy: when the norm of the state estimation error vector is... When the value increases, the forgetting factor automatically decreases. To accelerate the tracking speed of parameter changes by the parameter estimator (RLSE), the forgetting factor is increased when the norm of the state estimation error vector decreases. To enhance the smoothness and noise resistance of parameter estimation;

[0104] Through this two-way collaborative mechanism of forward parameter transfer and feedback error correction, the recursive least squares parameter estimator (RLSE) and the adaptive sliding mode observer (ASMO) form a closed-loop optimization system. The accuracy of parameter estimation improves the accuracy of state observation, while the state estimation error in turn optimizes the adaptability of parameter estimation. Thus, the overall collaborative optimization of parameter identification and state disturbance estimation is achieved, providing a robust model foundation for subsequent control steps.

[0105] Step S2 achieves synchronous and accurate identification of motor parameters and lumped disturbances through deep collaboration between recursive least squares method and adaptive sliding mode observer. Its innovation lies in the closed-loop collaborative mechanism of parameter estimation and state observation; adaptive sliding mode gain design; and real-time correction strategy for bidirectional data exchange.

[0106] The real-time parameter matrix output by this step The disturbance estimates provide accurate basic models and look-ahead compensation information for the robust predictive control in step S3.

[0107] In this embodiment, it should be specifically explained that the specific operation process of establishing the finite control set model predictive controller with fused sliding surface constraints in step S3 is as follows:

[0108] First, using the real-time estimate of the total disturbance obtained in step S2 as the feedforward compensation, and combining it with the tracking error between the current reference value and the actual current measurement value at the current moment, a sliding mode surface about the current tracking error is constructed. Specifically, the real-time estimate of the total disturbance is obtained from step S2. Simultaneously, it obtains the current reference value at the current moment from the upper-level controller of the system. and actual current measurement value Sliding surface The following formula is used for its construction:

[0109]

[0110] in, The sliding surface vector at time k is used to comprehensively reflect the current tracking error and the disturbance compensation amount, serving as a constraint condition for subsequent optimization. is the sliding surface error weight matrix (error weight matrix), which is a positive definite diagonal matrix used to adjust the weight of current tracking error in the sliding surface; Let be the current tracking error vector at time k; is the disturbance compensation weight matrix (compensation weight matrix), which is a positive definite diagonal matrix used to adjust the intensity of disturbance feedforward compensation in the sliding surface; The total disturbance estimate at time k is derived from step S2; among the parameters mentioned above, the error weight matrix... and compensation weight matrix The value of the error weight matrix is ​​determined based on the system's dynamic performance requirements and is tuned through simulation experiments. The diagonal elements take values ​​in the range [0, 5, 5, 0], and the compensation weight matrix... The diagonal elements take values ​​in the range [0, 1, 2, 0], and their specific values ​​are determined as follows: First, an open-loop test is performed under typical operating conditions to collect system response data. Then, a particle swarm optimization algorithm is used to search for the parameter combination that minimizes the absolute error of the integral within a specified range; current reference value. and actual value The sampling frequency is consistent with the main frequency of the control system, set to 10kHz, with a sampling precision of 12 bits and a sliding surface. Calculated once per control cycle (0, 1ms). By introducing disturbance feedforward compensation, the sliding surface not only reflects the tracking error but also actively cancels the effects of disturbances, improving the robustness of the system. The introduction of the weight matrix allows for priority adjustment of different control objectives.

[0111] In the process of constructing the sliding surface, the current tracking error is weighted by introducing a sliding surface error weight matrix, and the total disturbance estimate is weighted by introducing a disturbance compensation weight matrix, so that the constructed sliding surface can both reflect the tracking error of the system and actively compensate for the system disturbance.

[0112] Subsequently, during the rolling optimization process of the model predictive controller, the degree to which the system state trajectory approaches the sliding surface is incorporated as a soft constraint into the performance index function. This soft constraint is weighted using the sliding surface weight matrix. Simultaneously, a predictive model is constructed using the real-time parameter matrix obtained in step S2. This predictive model is a dynamic model based on the discrete state equations of a permanent magnet synchronous motor. (Include Construct a discrete state-space prediction model for a permanent magnet synchronous motor:

[0113]

[0114] in, Represents the system state vector (direct-axis and quadrature-axis currents) at time k; Represents the control input vector (direct-axis and quadrature-axis voltages) at time k. The state transition matrix (transition matrix) depends on the real-time parameter matrix. Its elements are calculated from the motor parameters, for example ; The input matrix (control input matrix) depends on the real-time parameter matrix. Its elements are calculated from the motor parameters, for example ; The sampling period (control period) is 0.0001 seconds. Among the parameters above, the transition matrix... and control input matrix The elements are directly calculated from the parameters identified in real time, ensuring the consistency between the prediction model and the actual motor model, and the sampling period. The value is determined based on the hardware performance of the control system and the electrical time constant of the motor. The value must be less than one-tenth of the electrical time constant of the motor. This process is used to make the online updated prediction model adaptable to changes in motor parameters, improve prediction accuracy, and the discretization process facilitates implementation by the digital controller.

[0115] Performance metrics function for rolling optimization Designed as follows:

[0116]

[0117] in, This is represented as the prediction time domain (prediction step size), with a value of 10, used to define the number of future steps considered in the optimization. This is represented as the control time domain (control step size), with a value of 5, used to limit the number of steps for control input optimization; Represented as Reference current value at any given time; Represented as The predicted state value at time; This represents the control input increment; The tracking error weight matrix (error weight) is a positive definite diagonal matrix used to penalize tracking errors. This represents the control increment weight matrix (increment weight), a positive definite diagonal matrix used to suppress sudden changes in the control quantity; This represents the sliding surface weight matrix (sliding surface weight), a positive definite diagonal matrix used as a soft constraint to drive the system state to approach the sliding surface; Denotes the weighted norm, i.e. Among the above parameters, the prediction time domain and control time domain The value is selected based on a trade-off between system response speed and computational burden, and is determined to be 2-3 times the electrical time constant of the covered motor, according to the error weight matrix. Incremental weight matrix Sliding mode weight matrix The diagonal elements are tuned through simulation, with values ​​ranging from [0, 5, 5, 0], [0, 1, 1, 0], to [0, 2, 3, 0]; performance index function Calculated once in each control cycle, prediction time domain =10 corresponds to a time length of 1 millisecond. This process incorporates the sliding surface convergence as a soft constraint into the performance index, enabling the optimization process to simultaneously consider tracking performance, control effort, and robustness, thus achieving multi-objective optimization. The weight matrix provides adjustment means.

[0118] Finally, by solving for the minimum value of the performance index function that incorporates the soft constraint conditions of the sliding surface, and considering the constraint of a finite set of inverter voltage vectors, the optimal voltage vector is obtained and applied. Specifically, this involves enumerating the basic voltage vectors corresponding to all possible switching states of the inverter (7 vectors for a two-level inverter), and for each candidate voltage vector... Using predictive models to calculate the future The state prediction value of each step is calculated, and the corresponding performance index function value is also calculated. , choose to The voltage vector with the smallest value is taken as the optimal voltage vector. This method is applied to the inverter to generate corresponding switching signals to drive the permanent magnet synchronous motor. The inverter voltage vector finite set U contains 7 vectors (6 non-zero vectors and 1 zero vector). The optimization solution process is completed within each control cycle, with a maximum calculation time of no more than 0.05 milliseconds. This process finite control set method directly considers discrete switching states, avoids modulators, has a fast dynamic response, ensures global optimality through enumeration optimization, and guarantees real-time performance due to computation time constraints.

[0119] In step S3, the specific design and optimization solution of the performance index function are as follows:

[0120] The performance index function consists of a weighted sum of three parts. The first part is the weighted sum of squared errors between the system output and the reference trajectory in the prediction time domain. The weighting is performed by the tracking error weight matrix to ensure the tracking performance of the system.

[0121] The first part corresponds to the performance index function. The purpose is to minimize the deviation between the actual current value and the reference value in the future prediction time domain, ensuring good tracking performance. The error weight matrix... Set as a diagonal matrix, the size of its diagonal elements determines the importance of the tracking error of the corresponding current components (d-axis and q-axis). The d-axis error weight is usually set to a larger value (e.g., 3, 0) to ensure the magnetic field orientation accuracy, and the q-axis error weight is set to a value (e.g., 2, 0) to ensure torque tracking performance.

[0122] The second part is the weighted sum of squares of the control increment changes in the control time domain, which is weighted by the control increment weight matrix to suppress drastic changes in the control signal;

[0123] The second part corresponds to the performance index function. Its purpose is to punish drastic changes in control input, smooth control actions, reduce the impact on the system, and improve system stability. The incremental weight matrix... The values ​​of the diagonal elements are determined based on the maximum allowable rate of voltage change of the system, and are usually 0 or 5. The larger the value, the stronger the suppression of changes in the control quantity.

[0124] The third part is the weighted sum of squares of the predicted sliding surface function values ​​in the time domain. The weights are applied by the sliding surface weight matrix and used as soft constraints to drive the system state to approach the sliding surface.

[0125] To further optimize the overall performance of the system under different operating conditions and solve the problem that fixed weight coefficients are difficult to achieve the optimal trade-off across the entire operating range, this invention dynamically and adaptively adjusts the weight matrix in the performance index function.

[0126] Specifically: using the real-time operating condition feature vectors provided in steps S1 and S2 Using this as input, a fuzzy inference system is used to calculate the optimization weights for four performance dimensions in real time online. This fuzzy inference system adopts a Mamdani-type fuzzy inference architecture. The fuzzy subsets and membership functions of its input variables (speed, load, temperature) and output variables (weighting coefficients) are predefined based on motor characteristics. The fuzzy rule base consists of a series of IF-THEN fuzzy rules, representing the culmination of expert experience and experimental optimization. An example of the rule base is shown below:

[0127] Rule 1, For "low" AND load torque , The steady-state mass weight is "very large". (Aimed at enhancing dynamic response capabilities).

[0128] Rule 2, IF speed For "high" AND load torque For "light", THEN For "small", For "big", "Very large" It is rated as "Medium" (aimed at optimizing steady-state accuracy and energy efficiency).

[0129] Rule 3: IF Temperature To achieve a "high" rating, THEN should appropriately increase the stability weight. (Intended to prevent overheating and oscillation).

[0130] Subsequently, the weights output by the fuzzy inference system Embedded performance metric functions Used to dynamically adjust its internal weight matrix. The mapping relationship is defined as follows:

[0131]

[0132]

[0133]

[0134] in, This is the initial weight matrix determined based on the system's baseline performance. Through this mechanism, the system can prioritize dynamic response speed under low speed and high load conditions, and prioritize energy efficiency optimization under high speed and light load conditions, thereby achieving intelligent optimization and performance balance across the entire operating range.

[0135] The third part corresponds to the performance index function. The aim is to integrate the concept of sliding mode control into predictive control. By penalizing future sliding surface function values, it drives the system state trajectory to approach the sliding surface, thereby enhancing the system's robustness to parameter changes and disturbances. The sliding mode weight matrix... The value of is determined based on the system's robustness requirements, and is usually 1 or 0. A larger value indicates a higher requirement for robustness.

[0136] When solving for the minimum value of the performance index function, a finite control set optimization strategy is adopted. By enumerating all possible combinations of voltage vectors of the inverter, the predicted value of the future state of the system and the corresponding performance index function value under each combination are calculated. Finally, the voltage vector that minimizes the performance index function value is selected as the optimal control quantity and applied to the next control cycle. The optimization solution process is performed in a rolling manner. At each sampling time k, the following steps are executed:

[0137] 1. Obtain the current state measurement value and reference trajectory ;

[0138] 2. For each candidate voltage vector in the finite control set U Using predictive models Predicting future state sequences Then calculate the corresponding sliding surface sequence. Finally, based on the predicted state sequence and sliding surface sequence, the performance index function value is calculated. ;

[0139] 3. Compare the corresponding values ​​of all candidate voltage vectors. Value, selection The minimum voltage vector is used as the optimal control quantity. ;

[0140] 4. Used in inverters to drive motors;

[0141] 5. Proceed to the next sampling time k+1 and repeat the above process.

[0142] In the above, the size of the finite control set U is fixed at 7 vectors, the optimization calculation is completed within 0.05 milliseconds, and the prediction time domain is... Control Time Domain .

[0143] Although the computational complexity of this enumeration method increases with the number of voltage vectors, it guarantees finding the global optimum for systems with finite control sets, such as inverters, and is easy to implement without requiring gradient calculations. The rolling optimization strategy can continuously modify the control strategy to adapt to system changes.

[0144] In this embodiment, it should be specifically explained that the specific operation of continuously monitoring the control performance indicators of the motor and associating and storing them with the operating conditions, real-time parameter matrix and total disturbance estimate in step S4 is as follows:

[0145] Real-time monitoring of multiple control performance indicators of the motor, including the root mean square value of the current tracking error. Output torque ripple coefficient Equivalent switching frequency and Lyapunov index estimates To form a comprehensive performance feature vector .

[0146] Specifically, the root mean square value of the current tracking error In each control cycle ( Calculated once every 0, 1ms, with the calculation window being the most recent value. =100 sampling points; output torque ripple coefficient The equivalent switching frequency is obtained by calculating the ratio of the standard deviation to the mean of the torque signal, with a sampling frequency of 10kHz. Calculated by counting the number of switch state changes per unit time; Lyapunov index estimate. The estimation was performed using a small amount of data, with the embedding dimension set to [value missing]. =6, with a time delay of τ=10 sampling points.

[0147] Simultaneously, it collects operating condition data of the motor, including speed. Load torque Radiator temperature and running time Forming a working condition feature vector .

[0148] By synchronously collecting multi-dimensional performance indicators and operating condition data, the system's operating status can be comprehensively reflected, providing a complete data foundation for subsequent intelligent analysis and comprehensive performance feature vectors. The system performance is comprehensively evaluated from four dimensions: dynamic performance, steady-state quality, energy consumption characteristics, and stability.

[0149] The comprehensive performance feature vector at each sampling time Operating condition feature vector Real-time parameter matrix from step S2 Total disturbance estimate Perform correlation to construct a unified time series data record. .

[0150] Specifically, the data association uses timestamp alignment to ensure that data from different sources are strictly synchronized in time. Each data record contains five parts: timestamp, performance indicators, operating data, system parameters, and disturbance information, forming a complete data packet.

[0151] Based on the importance and characteristics of the data, a hierarchical storage strategy is adopted, including saving short-term memory data from the most recent sampling periods, typical long-term memory data of operation when a new stable operating condition is identified, and abnormal state data when performance indicators exceed preset thresholds, thus establishing a complete historical operation database. .

[0152] Specifically, short-term memory data is stored most recently. The data from each sampling period (corresponding to 0, 1s) is based on covering the main dynamic processes of the system; long-term memory data is continuously stored in the system. Save is triggered when the cycle is in a steady state (performance index fluctuation less than 5%); the threshold for abnormal state data is set to ±3 of the historical average of the performance index. (Three standard deviations), this threshold is determined based on the statistical characteristics of historical normal operation data.

[0153] In step S4, the specific operation of using the ensemble learning model to analyze historical data and adaptively adjust parameters when performance changes are detected is as follows:

[0154] Building Hybrid Ensemble Learning Models The model includes a random forest condition identifier. Online Sequence Extreme Learning Machine Performance Predictor and performance change detector ;

[0155] Specifically, the random forest condition identifier Include Decision trees, with a maximum depth of Using the Gini coefficient as the splitting criterion; Online Sequence Extreme Learning Machine Performance Predictor The number of hidden layer nodes is The activation function used is the Sigmoid function; performance change detector Based on the sliding window mechanism, the window size is... One sampling period.

[0156] Utilizing historical operation database The data in the database is used for training and online learning, enabling the ensemble learning model to accurately identify the current operating condition category. Predicting future performance metrics And abnormal changes in the performance of the detection system;

[0157] Specifically, the model training uses an incremental learning approach, each... Updated once per sampling period, training data comes from historical operational databases. China recently The number of records, determined based on model convergence testing, ensures model accuracy while controlling computational complexity.

[0158] When the performance change detector determines that the change in the system performance index exceeds the preset threshold, the parameter adaptive adjustment mechanism is triggered.

[0159] Judgment Logic: The performance change detector calculates the rate of change of the overall performance index.

[0160]

[0161] in, These are the weighting coefficients. and For historical mean and standard deviation; when Adjustments are triggered at specific times.

[0162] This mechanism adjusts the parameters of the prediction model in step S3 based on the analysis results of the ensemble learning model. and the gain of the adaptive sliding mode observer in step S2 The adjustment of the prediction model parameters is based on the gradient information of the model parameters by the performance index, and the adjustment of the adaptive sliding mode observer gain is based on the product of the current gain and the hyperbolic tangent function of the performance change magnitude.

[0163] Specifically, adjusting the parameters of the prediction model:

[0164]

[0165] in, The learning rate (or step size) is an important control parameter. The learning rate determines the magnitude of each parameter adjustment. A value of 0 or 01 is an empirical setting designed to balance the speed and stability of parameter updates (avoiding overshoot or oscillation). For performance index functions For the parameter matrix The gradient is used to indicate the optimal approach to improve system performance (i.e., reduce performance metrics) under the current parameters. In which direction (increase or decrease) should the parameter be adjusted? They represent the first The real-time parameter matrix before and after the adjustment contains key motor parameters identified online from step S2, such as stator resistance, direct-axis inductance, quadrature-axis inductance, and permanent magnet flux linkage. This parameter is the core output of step S2.

[0166] Sliding mode observer gain adjustment:

[0167]

[0168] in, and They represent the first The sliding mode observer gain matrix before and after time adjustment. This gain directly affects the observer's ability to estimate disturbances and uncertainties and its convergence speed. This represents the sensitivity coefficient for gain adjustment. It controls the degree to which performance changes affect the magnitude of gain adjustment; the larger the value, the more drastic the gain response to performance changes. For adjustment coefficients; The hyperbolic tangent function is a smooth sigmoid function that maps input values ​​to the range of (-1, 1). Here, it is used to ensure that the gain adjustment is smooth and finite, avoiding abrupt changes and enhancing system stability. Indicates the first The change in the comprehensive performance index at any given time is used to quantify the degree of deviation of the system performance from its normal level; A reference baseline value representing performance changes, used to... Normalization is performed to make the adjustment algorithm more adaptable to performance changes of different orders of magnitude. For reference.

[0169] After the adjustment is completed, assess the degree of improvement in system performance. If the improvement does not meet the expected requirements, roll back the parameter adjustment and enable the backup control strategy.

[0170] Evaluation criteria: The degree of performance improvement is determined by comparing the performance before and after adjustments. The average performance metrics calculated over a period of time must show an improvement exceeding [a certain percentage]. Only when the condition is met is the system considered valid. The backup control strategy uses conservative PID control parameters to ensure the system operates stably.

[0171] Through intelligent analysis using an integrated learning model, online self-tuning of controller parameters is achieved, enabling the system to adapt to parameter changes and environmental disturbances during long-term operation and maintain optimal control performance. The hierarchical storage strategy ensures data utilization efficiency, while the backup control mechanism provides the final guarantee for system reliability.

[0172] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0173] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0174] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0175] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0176] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for anti-interference speed control of a permanent magnet synchronous motor, characterized in that, The method comprises the following steps: Step S1, in the control system of the permanent magnet synchronous motor, the original data of the three-phase current, the DC bus voltage, the rotor position and the radiator temperature of the motor are synchronously collected by a sensor array at a fixed sampling frequency, the original data is real-time aligned, and an adaptive Kalman filtering algorithm based on a motor real-time speed dynamic adjustment observation noise covariance matrix is used for filtering processing to obtain clean synchronous multi-physical field data packets; Step S2, a cooperative identification unit composed of a recursive least squares parameter estimator and an adaptive sliding mode observer is constructed; The clean synchronous multi-physical field data packets obtained in step S1 are used to identify the stator resistance and cross-axis inductance of the motor online by the recursive least squares parameter estimator to form a real-time parameter matrix; at the same time, the adaptive sliding mode observer generates equivalent control quantities through a sliding mode variable structure control law based on the current and voltage data obtained in step S1 and the real-time parameter matrix to synchronously estimate the system state and the total disturbance, and outputs the observation value of the system state and the real-time estimated value of the total disturbance; Step S3, a finite control set model predictive controller with a fusion sliding mode surface constraint is established; first, the real-time estimated value of the total disturbance obtained in step S2 is used as a feedforward compensation to construct a sliding mode surface about the current tracking error; Subsequently, in the rolling optimization process of the model predictive controller, the approaching degree of the system state trajectory to the sliding mode surface is taken as a soft constraint condition into the performance index function, and the real-time parameter matrix obtained in step S2 is used to construct a prediction model; by solving the minimum value of the performance index function, the optimal voltage vector is obtained and applied; Step S4, the control performance indicators of the motor are continuously monitored and stored in association with the operating conditions, the real-time parameter matrix and the total disturbance estimated value; the integrated learning model is used to analyze the historical data, and when the performance change is identified, the parameters of the prediction model and the gain of the adaptive sliding mode observer are adaptively adjusted.

2. The anti-interference speed control method of a permanent magnet synchronous motor according to claim 1, characterized in that: In step S1, the specific operation process of real-time alignment of the original data and filtering processing by the adaptive Kalman filtering algorithm is as follows: First, the three-phase current, DC bus voltage, rotor position and radiator temperature data synchronously collected by the sensor array are strictly aligned based on their corresponding time stamps; Subsequently, according to the coordinate transformation theory, the aligned three-phase current data is transformed from the natural coordinate system to the orthogonal coordinate system synchronous with the rotor magnetic field to obtain the direct-axis current component and the cross-axis current component; Then, the rotor position signal is numerically differentiated to calculate the real-time mechanical angular velocity of the motor, and the value of the observation noise covariance matrix used by the adaptive Kalman filter in the current sampling period is dynamically adjusted according to the numerical value of the real-time mechanical angular velocity, and the adjustment strategy is that the value of the matrix decreases with the increase of the real-time mechanical angular velocity and increases with the decrease of the real-time mechanical angular velocity; Finally, the observation noise covariance matrix after the aforementioned dynamic adjustment is used to perform Kalman filtering estimation on the direct-axis current component, the cross-axis current component, the DC bus voltage and the radiator temperature signal to obtain the filtered estimated value of each physical quantity.

3. The anti-interference speed control method of a permanent magnet synchronous motor according to claim 2, characterized in that: The specific operation of dynamically adjusting the observation noise covariance matrix is as follows: A nonlinear scaling factor is constructed with the real-time mechanical angular velocity and the radiator temperature as the independent variables, and the nonlinear scaling factor is used to scale a preset reference observation noise covariance matrix; The nonlinear scaling factor is composed of a constant term with a value of one, a speed-related term, and a temperature-related term; The speed-related term is configured as a negative exponential function of the absolute value of the real-time mechanical angular velocity, and the temperature-related term is configured as a quadratic function of the difference between the real-time temperature and a preset reference temperature; The reference observation noise covariance matrix is multiplied by the nonlinear scaling factor to generate the observation noise covariance matrix finally applied to the adaptive Kalman filtering algorithm at the current sampling time.

4. The anti-interference speed control method of a permanent magnet synchronous motor according to claim 3, characterized in that: In step S2, the specific operation of the recursive least squares parameter estimator is as follows: first, a discrete-time system regression model for motor parameter identification is constructed, the input variables of the model are the filtered direct-axis current component, the filtered quadrature-axis current component, the filtered DC bus voltage in the clean data packet obtained in step S1, and the real-time electrical angular velocity calculated by differentiating the filtered rotor position signal, the output variables are the direct-axis voltage component and the quadrature-axis voltage component, and the parameters to be identified are the stator resistance, the direct-axis inductance, the quadrature-axis inductance, and the permanent magnet flux of the permanent magnet synchronous motor; Then, the recursive least squares algorithm with a forgetting factor is used for online parameter updating, the recursive least squares algorithm updates the parameter estimation vector at the current time by calculating the gain matrix at the current time, based on the parameter estimation value at the previous time and the error between the current observation data and the model prediction value, and updates the covariance matrix of the algorithm; The forgetting factor is dynamically adjusted according to the system state, and the adjustment strategy is that the value of the forgetting factor is negatively related to the norm of the system state estimation error; Finally, the stator resistance estimation value, the direct-axis inductance estimation value, the quadrature-axis inductance estimation value, and the permanent magnet flux estimation value updated at each sampling time are combined into a real-time parameter matrix in diagonal form, which provides accurate motor model parameters for the subsequent adaptive sliding mode observer.

5. The anti-interference speed control method of a permanent magnet synchronous motor according to claim 4, characterized in that: In step S2, the specific design and operation of the adaptive sliding mode observer are as follows: First, based on the real-time parameter matrix output by the recursive least squares parameter estimator, an extended state space model containing motor electrical states and lumped disturbances is constructed, which unifies the parameter uncertainty, unmodeled dynamics, and external disturbances in the system as lumped disturbance terms and as extended state variables; Then, the structure of the sliding mode observer is designed, which includes a model-based feedforward term, a linear feedback term for ensuring the dynamic stability of the estimation error, and a discontinuous switching feedback term based on the sliding surface for suppressing the influence of disturbances; the switching gain of the sliding mode observer is adjusted online by using an adaptive law, which makes the size of the sliding mode gain proportional to the norm of the sliding surface defined between the actual output of the system and the output of the observer, and proportional to the integral term of the norm of the sliding surface; The adaptive sliding mode observer estimates the direct-axis current state, the quadrature-axis current state and the unified lumped direct-axis and quadrature-axis disturbance of the motor finally, and outputs the state observation values and the disturbance estimation values.

6. The anti-interference speed control method of a permanent magnet synchronous motor according to claim 5, characterized in that: In step S2, the cooperative working mechanism between the recursive least squares parameter estimator and the adaptive sliding mode observer is as follows: A bidirectional data exchange and correction loop is established between the parameter estimator and the state observer; in the forward channel, the real-time parameter matrix updated and output by the recursive least squares parameter estimator at each sampling period is transmitted to the adaptive sliding mode observer in real time, so as to update the coefficient matrix of the state space expression based on the mathematical model of the motor in the observer, and ensure that the dynamic model of the observer is consistent with the actual parameters of the motor; In the feedback channel, the norm of the state estimation error vector between the actual output of the system and the output of the observer calculated by the adaptive sliding mode observer is taken as the feedback signal to the recursive least squares parameter estimator, so as to dynamically correct the size of the forgetting factor; when the norm of the state estimation error vector increases, the forgetting factor is automatically reduced; when the norm of the state estimation error vector decreases, the forgetting factor is increased; through the bidirectional cooperative mechanism of the forward parameter transmission and the feedback error correction, the closed-loop optimization of the parameter identification and the state disturbance estimation is realized.

7. The anti-interference speed control method of a permanent magnet synchronous motor according to claim 6, characterized in that: In step S3, the specific operation process of establishing the model predictive controller with the constraint of the fused sliding mode surface is as follows: Firstly, the real-time estimation value of the total disturbance obtained in step S2 is used as a feedforward compensation, and a sliding mode surface about the tracking error between the current reference value and the actual current measurement value at the current time is constructed; in the construction process of the sliding mode surface, the current tracking error is weighted by introducing a sliding mode surface error weight matrix, and the total disturbance estimation value is weighted by introducing a disturbance compensation weight matrix, so that the constructed sliding mode surface can reflect the tracking error of the system and actively compensate the disturbance of the system; Subsequently, in the rolling optimization process of the model predictive controller, the approaching degree of the system state trajectory to the sliding mode surface is taken as a soft constraint condition into the performance index function, the soft constraint condition is weighted by a sliding mode surface weight matrix, and a prediction model is constructed by using the real-time parameter matrix obtained in step S2, the prediction model is a dynamic model based on the discrete state equation of the permanent magnet synchronous motor; Finally, by solving the minimum value of the performance index function fused with the soft constraint condition of the sliding mode surface, the optimal voltage vector is obtained and applied under the constraint condition of the finite set of inverter voltage vectors.

8. The anti-interference speed control method of a permanent magnet synchronous motor according to claim 7, characterized in that: In step S3, the specific design and optimization solving of the performance index function are as follows: The performance index function is composed of three parts weighted and summed, the first part is the weighted square sum of the tracking error of the system output and the reference trajectory in the prediction time domain, which is weighted by a tracking error weight matrix, and is used to ensure the tracking performance of the system; the second part is the weighted square sum of the control increment change in the control time domain, which is weighted by a control increment weight matrix, and is used to suppress the drastic change of the control signal; The third part is a weighted square sum of the function value of the sliding mode surface in the prediction time domain, which is weighted by a sliding mode surface weight matrix and used as a soft constraint condition to drive the system state to approach the sliding mode surface; In solving the minimum value of the performance index function, a finite control set optimization strategy is adopted, the future state of the system under each possible combination of inverter voltage vectors is calculated, and the performance index function value corresponding to each combination is calculated, and finally the voltage vector that makes the performance index function value minimum is selected as the optimal control quantity applied in the next control period.

9. The anti-interference speed control method of a permanent magnet synchronous motor according to claim 8, characterized in that: In step S4, the specific operation of continuously monitoring the control performance index of the motor and storing it in association with the operating condition, real-time parameter matrix and total disturbance estimate value is as follows: Real-time monitoring of multiple control performance indexes of the motor, including the root mean square value of the current tracking error, the output torque ripple coefficient, the equivalent switching frequency and the Lyapunov index estimate value, to form a comprehensive performance characteristic vector; At the same time, the operating condition data of the motor are collected, including the speed, load torque, radiator temperature and running time, to form a working condition characteristic vector; the comprehensive performance characteristic vector, the working condition characteristic vector, the real-time parameter matrix and the total disturbance estimate value at each sampling time are associated to construct a unified time series data record; according to the importance and characteristics of the data, a hierarchical storage strategy is adopted, including saving short-term memory data for a certain number of sampling periods, saving typical running long-term memory data when a new stable working condition is identified, and saving abnormal state data when the performance index exceeds the preset threshold, to establish a complete historical running database.

10. The anti-interference speed control method of a permanent magnet synchronous motor according to claim 9, characterized in that: In step S4, the specific operation of using the integrated learning model to analyze the historical data and adaptively adjusting the parameters when a performance change is identified is as follows: A hybrid integrated learning model is constructed, which includes a random forest working condition identifier, an online sequence extreme learning machine performance predictor and a performance change detector; the data in the historical running database are used for training and online learning, so that the integrated learning model can accurately identify the current working condition category, predict the performance index value at the future time, and detect abnormal changes in system performance; When the performance change detector determines that the change amplitude of the system performance index exceeds the preset threshold, the parameter adaptive adjustment mechanism is triggered; this mechanism adjusts the parameters of the prediction model and the gain of the adaptive sliding mode observer in step S2 according to the analysis results of the integrated learning model, where the adjustment of the prediction model parameters is based on the gradient information of the performance index to the model parameters, and the adjustment of the adaptive sliding mode observer gain is based on the hyperbolic tangent function product of the current gain and the performance change amplitude; After the adjustment is completed, the improvement degree of the adjusted system performance is evaluated, and if the improvement effect does not meet the expected requirements, the parameter adjustment is rolled back and a backup control strategy is enabled.

Images