A method and system for position servo control of permanent magnet synchronous motors based on electromechanical closed residual drive strategy and permission projection.
By constructing electromechanical kinematic closed residuals and using the method of authority projection, the problem of managing multi-source errors in high dynamic position servo systems is solved, enabling direct observation and stable control of error combinations, and improving the control accuracy and robustness of the system.
Patent Information
- Application Number
- CN202611361579.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-03
- Publication Date
- 2026-11-10
AI Technical Summary
In high dynamic position servo systems, multi-source errors are difficult to identify and process separately, resulting in transmission errors, position observation errors, current reconstruction errors, and learning strategy extrapolation errors that are difficult to manage effectively. Existing technologies cannot effectively utilize limited hardware information to directly observe and control error combinations.
By constructing electromechanical kinematic closed residuals and combining them with positionless observers and single-bus current reconstruction, the combination of observable errors can be directly obtained. By using PWM sampling windows and disturbance boundaries to form permission projections, permission control of the learning strategy can be achieved, avoiding the estimation of multi-source errors one by one and the identification of redundant states.
It achieves unified processing of unobservable errors, reduces the complexity of error perception and learning strategies, forms a single causal closed loop, avoids online multivariate optimization, and improves the stability and accuracy of control.
Smart Images

Figure CN123074348A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of permanent magnet synchronous motor position servo control technology, and in particular to a permanent magnet synchronous motor position servo control method and system based on the permission projection of an electromechanical closed residual drive strategy. Background Technology
[0002] In high-dynamic position servo systems, to reduce costs and wiring complexity, a high-precision encoder can be placed at the load end, a position observer-less device can be used on the motor side, and the stator current can be reconstructed using a single DC bus current sensor. If deep reinforcement learning or other online adaptive strategies are introduced to improve tracking performance under conditions of reducer nonlinearity, load variation, and model uncertainty, transmission errors, position observation errors, current reconstruction errors, and extrapolation errors of the learning strategy will all exist simultaneously.
[0003] Existing technologies have disclosed reinforcement learning-based servo PID control, PMSM single-current sensor reconfiguration / predictive control, position-observation-free control, load-side position closed-loop control, and reinforcement learning safety layers. Therefore, simply inputting multi-source errors into TD3 and then superimposing safety penalties on the reward easily results in a parallel combination of several known modules, and cannot answer whether these errors can be identified separately under a given sensor topology.
[0004] The key problem that this invention aims to solve is not to add another error estimator or a safety filter, but rather: when the dimensionality of information that existing hardware can directly provide is limited, first identify the truly measurable error combinations that directly affect control, and then let these combinations determine how much control authority the learning strategy can still have along the physical energy / motion chain, so that error perception and policy constraint become the two ends of the same closed relationship. Summary of the Invention
[0005] In view of this, the purpose of the present invention is to provide a position servo control method and system for permanent magnet synchronous motors with electromechanical closed residual drive strategy permission projection, which avoids identifying multiple unobservable errors separately and does not use specific deep reinforcement learning algorithms, fixed filtering or reward weights as a basis for security.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: a position servo control method for a permanent magnet synchronous motor based on the permission projection of an electromechanical closed-loop residual drive strategy, applied to a position servo system in which a permanent magnet synchronous motor drives a load via a reducer, a position encoder is installed at the load end, the rotor position on the motor side is obtained without a position observer, and the stator current is at least partially reconstructed by sampling a single current from the DC bus, comprising:
[0007] Step 1: Obtain the encoder angle θ at the load end L and rotor electrical angle output without position observer Based on the number of pole pairs p of the motor, the reduction ratio N, and the installation zero position θ0, the electromechanical kinematic closed residual is constructed as follows:
[0008] (1)
[0009] Where, r θ The electromechanical kinematic closed residual is in electrical angle (rad); wrap(·) represents mapping the angle to the principal value interval (-π,π]; this electromechanical kinematic closed residual is obtained by directly closing the mechanical angle at the load end and the observed electrical angle at the motor side, without requiring separate identification of reducer error and position observation error;
[0010] The closed residual is not decomposed into two independent estimates of reducer transmission error and position observation error;
[0011] Step 2: Determine the effective sampling window t required for single bus current reconstruction based on the current PWM switching state. win The effective sampling window is relative to the minimum required sampling time t. req Missing quantity t miss The upper bound of the current reconstruction error is formed by the maximum slope boundary of the motor current:
[0012] (2)
[0013] Among them, t win t is the effective sampling window length available for single bus current reconfiguration under the current PWM switching state. req The minimum time required to complete sample setup, hold, and analog-to-digital conversion; t miss The missing value is relative to the minimum sampling time, and max(0,·) is guaranteed not to take a negative value;
[0014] (3)
[0015] Among them, S i,max This is a conservative upper bound for the rate of change of stator current; V dc R is the DC bus voltage; s I is the stator phase resistance; max E represents the upper limit of the phase current amplitude in the considered operating region. max L is the upper bound of the back electromotive force amplitude. min This is the lower bound of the stator equivalent inductance within the operating area;
[0016] (4)
[0017] in, ε is the upper bound of the amplitude of the single bus current reconstruction error; ADC This is the upper bound of the equivalent current error after ADC quantization, bias, and sampling chain adjustments; S i,max·t miss This indicates the additional uncertainty caused by the maximum current slope during periods of insufficient sampling window;
[0018] Step 3: Based on the closed residual, the preset reducer angle error boundary, and the current reconstruction error upper boundary from Step 2, the rotor electrical angle error boundary, the q-axis current error boundary, and the position state disturbance boundary radius ρ are sequentially formed. k ;where ρ k It is obtained from the same closed residual along the physical transmission path of electrical angle, current, torque and load position;
[0019] (5)
[0020] in, This is the upper limit of the main value error of the rotor electrical angle, which shall not exceed π. This is a known engineering upper bound for the equivalent mechanical angular error of the reducer; this bound is based on the principal value distance around the circle and |Δθ|. g |≤ get;
[0021] (6)
[0022] Where Î represents the magnitude of the reconstructed stator current vector; î d î q To reconstruct the d-axis and q-axis currents; Δi q The underlined term represents the upper bound of the q-axis current comprehensive error; the second term represents the upper bound of the current vector chord length caused by angular mismatch, and... The conservative q-axis error bound is obtained by superposition;
[0023] (7)
[0024] in, This is a conservative upper bound vector for the state disturbance components, where the first component is the load position disturbance bound and the second component is the load angular velocity disturbance bound; T p For position control period; Δr θ ,k represents the principal difference of the closed residuals of adjacent periods; G τ Map the q-axis current error boundary to the load angular acceleration error;
[0025] (8)
[0026] Where, ρ k Let be the radius of the P-weighted state perturbation boundary for the k-th period; The symmetric positive definite weighted matrix is determined offline; since x k Containing two components with different dimensions, position and velocity, P simultaneously serves as a scale normalization / weighting factor; therefore, ρ kshall be interpreted as the certificate radius under the P-norm, rather than rad, rad / s or A alone.
[0027] Step 4: The reference position controller that meets the stability requirement in advance generates the reference q-axis current command i q,b * ; the learning controller or adaptive controller only generates the candidate additional action a π , and the candidate additional action is connected to the reference closed loop with a scalar authority coefficient γ∈[0,1]:
[0028] (9)
[0029] where A b is the discrete state matrix after the reference position controller is closed-loop; b is the discrete input mapping vector from the q-axis additional current to the state increment; a π k is the candidate additional q-axis current given by the learning / adaptive strategy; γ k ∈[0,1] is the strategy authority coefficient; w k is the actual unmodeled state disturbance, whose P-norm is required to be bounded by ρ k ;
[0030] Step 5: Using the same disturbance bound radius ρ obtained in Step 3 k to establish a P-weighted contraction constraint, converting this constraint into a quadratic inequality with respect to the scalar authority coefficient γ, and intersecting it with the q-axis current limit to obtain the feasible set Γ k ; take the maximum authority in this feasible set as γ k *, and set γ when no feasible authority exists k *=0:
[0031] (10)
[0032] where q is a preset contraction factor, 0<q<1; the first term on the left is the nominal next-step P-norm after the reference closed loop is superimposed with the candidate action, ρ k is the tube radius that must be reserved for unmodeled disturbances; when this inequality holds, the current candidate authority will not damage the preset P-norm contraction margin;
[0033] (11)
[0034] where Γ Q,k is the feasible authority set that satisfies the robust contraction quadratic constraint; Γ I,k is the feasible authority set that simultaneously satisfies 0≤γ≤1 and the q-axis current hard limit; Γ k is the intersection of the two; i q,b,k * is the q-axis current command given by the reference controller, Imax The upper limit of the allowable q-axis current amplitude;
[0035] Where Γ Q,k Γ is the quadratic constrained feasible set within the range 0 ≤ γ ≤ 1; I,k To simultaneously satisfy 0≤γ≤1 and |i q,b,k * +γa k π The feasible set of current limits ≤ Imax;
[0036] (12)
[0037] Where, γ k * R represents the maximum feasible strategy permission coefficient to be adopted; when R k >0 and Γ k If the set is not empty, take the maximum value in the feasible set; otherwise, set it to 0, causing the candidate learning / adaptive action to exit while retaining the baseline control.
[0038] Step 6: Output the final q-axis current command to the inner current loop:
[0039] (13)
[0040] Among them, i q,k * This is the q-axis current given to the final current loop; i q,b,k * For the reference controller output; γ k * a π k The candidate additional q-axis current after authorization projection.
[0041] In a preferred embodiment, the motor mechanical angle θ m , load end angle θ L Reducer transmission error Δθ g And the installation of zero position θ0 satisfies:
[0042] (14)
[0043] Where, θ m θ is the mechanical angle of the motor; L θ0 is the mechanical angle at the load end; N is the reduction ratio (defined as the ratio of the increment of the mechanical angle on the motor side to the increment of the mechanical angle on the load side, dimensionless); θ0 is the installation zero-position offset referred to the mechanical side of the motor; Δθ g This refers to the equivalent mechanical angular error caused by the reducer transmission.
[0044] Define the position observation error δ e back:
[0045] (15)
[0046] Where, θ e This represents the actual rotor electrical angle; This is the rotor electrical angle estimate output without a position observer; The position observation electrical angle error; p is the number of motor pole pairs;
[0047] The electromechanical closure residuals from step 1 satisfy the following identity:
[0048] (16)
[0049] Where, r θ Strictly speaking The modulo 2π equivalence class. Therefore, the subsequent angular error bound should be interpreted according to the principal angular distance. This equation shows that what is directly observable in the current sensor topology is the combination of the two types of errors, rather than two independent error states.
[0050] The identity relation allows two errors that cannot be directly measured separately to enter into subsequent control as a control-related, directly calculable combination.
[0051] In a preferred embodiment, the reducer transmission error has a known upper engineering bound. When the upper bound of the rotor electrical angle error and the upper bound of the q-axis current error are calculated according to Equation (5) and Equation (6), respectively, the current vector error caused by the angle mismatch adopts the upper bound of the chord length shown in Equation (6), and the position observation error is not treated as another independent learning state.
[0052] In a preferred embodiment, let K t For the torque constant of the permanent magnet synchronous motor, η g For reducer efficiency, J L,eq To calculate the equivalent moment of inertia referred to the load end, the load end current-angular acceleration gain is obtained from equation (17); during the position period T p Internally, it is constructed using the position equivalent term of the closed residual, the velocity equivalent term of the closed residual difference, and the q-axis current error boundary:
[0053] (17)
[0054] Among them, K t For surface-mount PMSM in i d The q-axis torque constant when ψ ≈ 0; f For permanent magnet fundamental flux linkage; η g For reducer efficiency; J L,eq The equivalent moment of inertia referred to the load end; G τThis is the gain from the q-axis current error to the load angular acceleration error;
[0055] (18)
[0056] in, This is a conservative upper bound vector for the state disturbance components, where the first component is the load position disturbance bound and the second component is the load angular velocity disturbance bound; T p For position control period; Δr θ,k G represents the principal difference between the closed residuals of adjacent periods; τ Map the q-axis current error boundary to the load angular acceleration error;
[0057] (19)
[0058] Where, ρ k Let P be the radius of the P-weighted state perturbation boundary for the k-th period; P = P T >0 is a symmetric positive definite weighted matrix determined offline. Since x k Containing two components with different dimensions, position and velocity, P simultaneously serves as a scale normalization / weighting factor; therefore, ρ k It should be interpreted as the certificate radius under the P-norm, rather than rad, rad / s, or A alone.
[0059] In a preferred embodiment, the closed-loop discrete matrix of the reference controller is A. b , Given a pre-calculated weighted matrix, the position tracking state and the P-norm are defined as follows:
[0060] (20)
[0061] Where, x k The tracking state during the k-th position control cycle; e θ =θ L * -θ L For load position error; e ω = This represents the load angular velocity error; the superscript * indicates the command value, and k is the discrete position control cycle number.
[0062] (twenty one)
[0063] Among them, V k V is a nonnegative quadratic form of the certificate quantity based on the P matrix; z is an arbitrary vector with the same dimension as the state; ||z||P is the weighted norm of P; in engineering implementation, P includes position / velocity scale normalization, therefore V k The symbols ||·||P are only used to standardize the stability certificate and do not represent the actual energy storage of the motor.
[0064] When the disturbance satisfies ||w k ||P≤ρ k And when the permission coefficient satisfies equation (10), we have:
[0065] (twenty two)
[0066] Among them, the contractile chain is only in R k >0、Γ k Non-empty and the actual perturbation satisfies ||w k ||P≤ρ k The current sampling time holds true; it is a local / stepwise certificate for candidate additional actions, and does not replace the stability design of the reference controller itself, nor is it extrapolated to an unconditional global stability conclusion.
[0067] Equation (22) holds only for sampling moments where the feasible set is non-empty and the current perturbation bound is covered by the certificate; when the certificate is not feasible, the candidate learning action is withdrawn without claiming additional global stability beyond the baseline controller.
[0068] In a preferred embodiment, R is defined. k for:
[0069] (twenty three)
[0070] Among them, R k This is the usable shrinkage margin after deducting the perturbation boundary radius, and is related to ||x||P and ρ. k They are on the same certification scale; only R k Only when the value is greater than 0 can the equivalent square of equation (10) be a scalar quadratic constraint;
[0071] In R k When the value is greater than 0, square equation (19) and rearrange it into equation (24). The coefficients are then calculated directly from equation (25).
[0072] (twenty four)
[0073] Where γ is the dimensionless scalar authority to be determined; A k B k C k The quadratic constraint coefficients for the current period are determined by the state, candidate actions, and the P matrix, all of which correspond to the same certificate scale of the square of the P norm.
[0074] (25)
[0075] Among them, A k Characterize the secondary contribution of candidate actions to the P-weighted next state; B kCharacterize the cross-contribution between candidate actions and the baseline closed-loop state; C k The weighted value of P in the next state of the baseline closed loop is reduced by the available shrinkage margin R. k 2 ;
[0076] Since the decision variable is only a single scalar authority coefficient, the controller does not need to solve a multivariate quadratic programming problem online.
[0077] In a preferred embodiment, a current phase residual is further constructed when the current reference vector is non-zero:
[0078] (26)
[0079] Where, r i For current phase residual; i d * i q * These are the d-axis and q-axis current references, respectively; d î q These are the reconstructed currents along the d and q axes obtained in the observation coordinate system, respectively; the two-parameter arctangent function atan2(y,x) gives the angle between the directed principal values; this formula is only used when both the reference current vector and the reconstructed current vector are non-zero.
[0080] In the operating region where the current inner loop tracking error is relatively small, the current phase residual is added to the electromechanical closure residual to obtain:
[0081] (27)
[0082] Where, r Σ The auxiliary closure residual is the result of combining the electromechanical closure residual and the current phase residual; δ c The current phase disturbance is defined as the current reconstruction phase error δ. rec Phase error δ of dynamic tracking of the inner current loop dyn The sum; the approximate relationship is applicable to working areas with small phase errors and consistent coordinate sign conventions;
[0083] This causes the electrical angle error of the position observer to cancel out in the first-order approximation, and in terms of r Σ Auxiliary monitoring is performed on the combined drift of reducer error and current reconstruction / dynamic error.
[0084] In a preferred embodiment, the candidate additional action generator is any one of the offline trained TD3, SAC, or DDPG policy networks, or an online adaptive control law; its output does not directly cover the baseline controller, but only serves as aπ in the permission projection.
[0085] In a preferred embodiment, twin The current space vector PWM switching state, bus current sampling setup time, analog-to-digital conversion hold time, and dead time are all determined together; when t win ≥t req time t miss =0, when t win <t req The time error bound increases continuously with the missing time, thus allowing the same PWM sampling dead zone to pass through ρ k Continuously reduce the availability of candidate learning actions.
[0086] This invention also provides a permanent magnet synchronous motor position servo control system with electromechanical closed-loop residual drive strategy permission projection, including a permanent magnet synchronous motor, a reducer, a load, a load-end position encoder, a three-phase inverter, a DC bus single current sensor, a position observerless device, a processor, and a memory. When the processor executes the program in the memory, it implements the aforementioned permanent magnet synchronous motor position servo control method with electromechanical closed-loop residual drive strategy permission projection.
[0087] Compared with the prior art, the present invention has the following beneficial effects:
[0088] (1) Instead of estimating and splicing the multi-source errors one by one, the sensor topology is used to determine the observable combination quantity first, thus theoretically eliminating the unidentifiable redundant state.
[0089] (2) The PWM sampling blind zone, position observation uncertainty and reducer error are unified into a state disturbance tube with physical dimensions, instead of empirical risk scoring;
[0090] (3) The same perturbation tube directly determines the maximum available authority of the candidate learning strategy, so that error perception, learning enhancement and security constraints form a single causal closed loop;
[0091] (4) The permission calculation is just a scalar quadratic inequality intersection, avoiding online multivariate optimization; the learning algorithm is replaceable while the core of the patent remains unchanged;
[0092] (5) Provide a clear P-norm contraction relationship for feasible certificate coverage time, and return to baseline control when it is not feasible, so as to avoid claiming security based solely on the reward function. Attached Figure Description
[0093] Figure 1 This is a flowchart of a preferred embodiment of the present invention. Detailed Implementation
[0094] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0095] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.
[0096] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations according to this application; as used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise; furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.
[0097] A method and system for position servo control of a permanent magnet synchronous motor based on the permission projection of an electromechanical closed-loop residual drive strategy, referenced. Figure 1 This includes the following steps:
[0098] Step 1: Convert the electromechanical closed-loop residual from indivisible error into measurable spatial coordinates.
[0099] The mechanical angle of the motor and the angle at the load end satisfy equation (1). The electrical angle is given without a position observer. After defining the observation error, we obtain equation (2). By directly closing the angles available on both sides, we construct equation (3). Substituting equations (1) and (2) into equation (3), we obtain the electromechanical closure identity relation containing the wrap principal value mapping; this relation holds strictly in the sense of circular angles.
[0100] (1)
[0101] In the formula, θ m θ is the mechanical angle of the motor (rad). L θ0 is the mechanical angle at the load side (rad); N is the reduction ratio (defined as the ratio of the increment of the mechanical angle on the motor side to the increment of the mechanical angle on the load side, dimensionless); θ0 is the installation zero-position offset referred to the mechanical side of the motor (rad); Δθ g The equivalent mechanical angular error (rad) caused by the reducer transmission.
[0102] (2)
[0103] In the formula, θ e The actual rotor electrical angle (rad); The rotor electrical angle estimate (rad) is the output of the position observer-free system. denoted as electrical angle error of position observation (rad); p is the number of pole pairs of the motor (dimensionless).
[0104] (3)
[0105] In the formula, r θ The residual is the electromechanical kinematic closed loop, with the dimension of electrical angle (rad). wrap(·) represents mapping the angle to the principal value interval (-π, π). This residual is obtained by directly closing the loop between the mechanical angle at the load end and the observed electrical angle at the motor side, without requiring separate identification of reducer error and position observation error.
[0106] (4)
[0107] In the formula, r θ Strictly speaking The modulo 2π equivalence class; therefore, the subsequent angular error bound should be interpreted according to the principal angular distance. This equation shows that what is directly observable in the current sensor topology is the combination of the two types of errors, rather than two independent error states.
[0108] Equation (4) explains that for this sensor topology, the load encoder and motor side do not directly provide Δθ without a position observer. g The two independent errors are δe and δe. The modulo 2π equivalence class, by r θ Characterization. Forcibly training the network to output multiple errors separately would introduce unidentifiable degrees of freedom; therefore, this invention directly forms control uncertainty within the observable combination space.
[0109] Step 2: Optional second closed residual and observable rank
[0110] When the current reference vector is non-zero, the directed angular residual r can be formed from the reference current vector and the reconstructed current vector in the observation coordinate system. i .
[0111] (5)
[0112] In the formula, r i i represents the current phase residual (rad); d *、i q *Reference currents (A) for the d and q axes, respectively; d î q The d-axis and q-axis reconstructed currents (A) obtained in the observation coordinate system are given by the two-parameter arctangent function atan2(y,x), which provides the directed principal angle. This formula is only used when both the reference current vector and the reconstructed current vector are non-zero.
[0113] (6)
[0114] In the formula, r Σ The auxiliary closure residual (rad) is the result of combining the electromechanical closure residual and the current phase residual; δ cThe combined current phase disturbance (rad) is defined as the current reconstruction phase error δ. rec Phase error δ of dynamic tracking of the inner current loop dyn The sum; the approximate relationship is applicable to working areas with small phase errors and consistent coordinate sign conventions.
[0115] (7)
[0116] In the formula, the three columns of the matrix correspond to the equivalent angular error pΔθ of the reducer, respectively. g Position observation error δ e and combined current phase disturbance δ c ; rank=2 indicates that the two-way closed residual provides at most two independent observation directions, therefore this invention does not advocate unique separation of the three types of potential errors.
[0117] Equation (7) yields an important structural conclusion: the three potential errors—equivalent angular error of the reducer, position observation error, and combined current phase disturbance—have only two independent closed residual directions, and the rank of the approximate observation matrix is 2 instead of 3. Therefore, this invention does not advocate for the unique identification of each of the three types of errors, but rather directly maps their observable equivalence classes to control uncertainties. Σ As a dependent implementation, it is used to assist in monitoring combined drift, but cannot be used to uniquely separate the three types of potential errors; the independent claims do not depend on this approximation.
[0118] Step 3: The PWM effective sampling window transforms the current reconstruction quality into an upper bound of physical dimensions.
[0119] The quality of single-bus current reconstruction varies with the duration of the current effective PWM vector. Define t. win To meet the effective duration requirements for sample setup and analog-to-digital conversion hold, t req To minimize the required time, the missing time is calculated according to equation (8). From the phase voltage equation, the upper bound of the conservative current slope can be given by equation (9), and equation (10) is obtained.
[0120] (8)
[0121] In the formula, t win t is the effective sampling window length (s) available for single bus current reconfiguration under the current PWM switching state. req The minimum time (s) required to complete sample setup, hold, and analog-to-digital conversion; t miss The missing amount (s) is relative to the minimum sampling time, and max(0,·) guarantees that it will not take a negative value.
[0122] (9)
[0123] In the formula, S i,maxV is a conservative upper bound on the rate of change of stator current (A / s); dc DC bus voltage (V); R s Stator phase resistance (Ω); I max E represents the upper limit of the phase current amplitude (A) for the considered operating region. max Lmin is the upper limit of the back electromotive force amplitude (V); Lmin is the lower limit of the stator equivalent inductance in the operating region (H).
[0124] (10)
[0125] In the formula, ε is the upper bound (A) of the single bus current reconstruction error. ADC The upper bound (A) of the equivalent current error after ADC quantization, bias, and sampling chain conversion; S i,max ·t miss This represents the additional uncertainty (A) caused by the maximum current slope during the period of insufficient sampling window.
[0126] This invention does not consider the single bus current reconstruction algorithm itself as an original point; its function is to transform the information gap of any existing reconstruction algorithm in the current switching state into a verifiable error bound in the dimension of amperes, and to allow this bound to enter the subsequent same disturbance bound.
[0127] Step 4: Map the closed residuals to disturbance boundaries along the angle, current, torque, and load position.
[0128] If the transmission error of the reducer can be given an upper limit by manufacturing precision, backlash calibration, or offline testing. Then, under the meaning of the principal angle distance on the circumference, the upper bound of the rotor electrical angle error of equation (11) can be obtained from equation (4); the maximum distance of the principal angle is explicitly processed by min(π,·). The current vector rotates due to the mismatch of coordinate angles, and its maximum change is obtained by using the current vector chord length boundary, which is then superimposed with the single bus current reconstruction error boundary to obtain the corresponding upper bound of the q-axis comprehensive current error.
[0129] (11)
[0130] In the formula, This is the upper bound (rad) of the main value error of the rotor electrical angle, and the maximum value shall not exceed π. The known engineering upper bound (rad) for the equivalent mechanical angular error of the reducer is given. This bound is based on the principal value distance of the circle and |Δθ| in equation (4). g |≤ get.
[0131] (12)
[0132] In the formula, Î represents the magnitude of the reconstructed stator current vector (A); î dî q To reconstruct the d-axis and q-axis currents (A); Δi q The overlined quantity is the upper bound of the q-axis current comprehensive error (A). The second term is the upper bound of the current vector chord length caused by angular mismatch, and... The conservative q-axis error bound is obtained after superposition.
[0133] (13)
[0134] In the formula, K t The q-axis torque constant (N·m / A) of the surface-mounted PMSM when id≈0; ψ f For permanent magnet fundamental flux linkage (Wb); η g The efficiency of the reducer (dimensionless); J L,eq The equivalent moment of inertia (kg·m²) referred to the load end; G τ This is the gain from the q-axis current error to the load angular acceleration error, expressed in (rad / s²) / A.
[0135] The q-axis current error is mapped to the angular acceleration within a position period through the torque constant, reducer and load inertia, and the closed residual and its difference are used as the direct equivalent inconsistency of position / velocity. The conservative state disturbance vector of equation (15) can be constructed. Then the disturbance boundary radius is calculated using the same P matrix as the stability certificate.
[0136] (14)
[0137] In the formula, x k The tracking state during the k-th position control cycle; e θ =θ L * -θ L For load position error (rad); e ω = The load angular velocity error is represented in rad / s; the superscript * indicates the command value, and k is the discrete position control cycle number.
[0138] (15)
[0139] In the formula, This is a conservative upper bound vector for the state disturbance components, where the first component is the load position disturbance bound (rad) and the second component is the load angular velocity disturbance bound (rad / s); T p The position control period (s); Δr θ,k G represents the principal difference between the closed residuals of adjacent periods; τ The q-axis current error boundary is mapped to the load angular acceleration error.
[0140] (16)
[0141] In the formula, ρ k Let be the radius of the P-weighted state perturbation boundary for the k-th period; This is a symmetric positive definite weighting matrix determined offline. Since x k Containing two components with different dimensions, position and velocity, P simultaneously serves as a scale normalization / weighting factor; therefore, ρ k It should be interpreted as the certificate radius under the P-norm, rather than rad, rad / s, or A alone.
[0142] The key uniformity here is that ρ is not another risk score learned by a neural network, but rather a physical chain of electromechanical closure residuals and PWM effective sampling window edge angles, currents, torques, and positions, ultimately forming a disturbance bound radius that can be used for stability certification on a P-weighted state scale.
[0143] Step 5: Candidate learning actions are not directly controlled; the maximum available permissions are determined by the same perturbation boundary.
[0144] First, design a stable reference position controller to obtain the reference closed-loop matrix A. b and reference q-axis current i q,b * TD3, SAC, or other learning / adaptive algorithms only propose candidate additional q-axis currents a. π The actual closed loop is written as equation (17). After defining the P norm, the robust contraction requirement of equation (19) is applied to the candidate actions.
[0145] (17)
[0146] In the formula, A b b is the discrete state matrix after the reference position controller closes the loop; a is the discrete input mapping vector from the q-axis additional current to the state increment; a π k Candidate additional q-axis current (A) given for the learning / adaptive policy; γ k ∈[0,1] represents the policy permission coefficient; w k For actual unmodeled state perturbations, their P-norm is required to be derived from ρ. k cover.
[0147] (18)
[0148] In the formula, V k Let V be a nonnegative quadratic form of the certificate quantity based on the P matrix; z is an arbitrary vector with the same dimension as the state; ||z||P is the weighted norm of P. In engineering implementation, P includes position / velocity scale normalization, therefore V k The values ||·||P are used only to standardize the stability certificate and do not represent the actual energy storage of the motor.
[0149] (19)
[0150] In the formula, q is a preset contraction factor, 0<q<1; the first term on the left side is the nominal next-step P-norm after the reference closed-loop superimposes the candidate action, ρ k is the tube radius that must be reserved for unmodeled disturbances; when this inequality holds, the current candidate authority will not damage the preset P-norm contraction margin.
[0151] (20)
[0152] In the formula, R k is the available contraction margin after deducting the disturbance bound radius, which is on the same certification scale as ||x||P, ρ k ; only when R k >0 can formula (19) be equivalently squared into a scalar quadratic constraint.
[0153] Since the only unknown quantity is the authority coefficient γ, when R k >0, formula (19) can be refined into a scalar quadratic inequality, instead of solving a multi-dimensional optimization problem online.
[0154] (21)
[0155] In the formula, γ is the dimensionless scalar authority to be solved; A k , B k , C k are the quadratic constraint coefficients determined by the state, the candidate action and the P matrix in the current cycle, and all three correspond to the same certification scale of the P-norm square.
[0156] (22)
[0157] In the formula, A k characterizes the quadratic contribution of the candidate action to the P-weighted next-step state; B k characterizes the cross contribution of the candidate action and the reference closed-loop state; C k is the P-weighted quantity of the reference closed-loop next-step state minus the available contraction margin R k 2 .
[0158] (23)
[0159] In the formula, Γ Q,k is the feasible set of authorities that satisfies the robust contraction quadratic constraint; Γ I,k is the feasible set of authorities that simultaneously satisfies 0≤γ≤1 and the q-axis current hard limit; Γ k is the intersection of the two; i q,b,k* The q-axis current command (A) given by the reference controller, I max The upper limit of the allowable q-axis current amplitude (A).
[0160] Where Γ Q,k By taking the intersection of equation (21) and 0≤γ≤1, Γ I,k From the q-axis current hard limit |i q,b,k * +γa π k |≤I max Intersects with 0≤γ≤1.
[0161] (twenty four)
[0162] In the formula, γ k * R is the maximum feasible strategy authority coefficient (dimensionless); when R k >0 and Γ k If the set is not empty, take the maximum value in the feasible set; otherwise, set it to 0, causing the candidate learning / adaptive action to exit while retaining the baseline control.
[0163] (25)
[0164] In the formula, i q,k * The final q-axis current (A) is given to the inner current loop; i q,b,k * For the reference controller output (A); γ k * a π k The candidate additional q-axis current (A) after authorization projection.
[0165] This makes "error perception" and "learning control security" no longer two spliced modules: the same ρ is both the quantitative result of the sensor and transmission closure degree and directly becomes the shrinkage amount of the learning strategy authority.
[0166] VI. Verifiable local contraction properties
[0167] If the actual unmodeled disturbance satisfies ||w k ||P≤ρ k And Γ k If it is not empty, then we can immediately obtain it from the triangle inequality and equation (19):
[0168] (26)
[0169] In the formula, the contractile chain only applies to R. k >0, Γk is non-empty and the actual perturbation satisfies ||wk ||P≤ρ k The current sampling time holds true; it is a local / stepwise certificate for candidate additional actions, and does not replace the stability design of the reference controller itself, nor is it extrapolated to an unconditional global stability conclusion.
[0170] Therefore, at the sampling time covered by the certificate, the learning action will not violate the preset P-norm contraction rate q. When R k When γ ≤ 0 or the feasible set is empty, * =0, only the baseline controller is executed; in this case, no additional claim is made regarding global stability beyond the design conditions of the baseline controller provided by this invention. This boundary declaration is used to avoid substituting mathematical guarantees with training rewards.
[0171] The embodiment employs a servo object with a 48V DC bus, a 4-pole surface-mount PMSM, a reduction ratio of 10:1, and a position outer loop period of 1 ms. An absolute encoder is installed at the load end; there is no mechanical position sensor on the motor side, with θ̂e provided by an existing position observer; the inverter has only one DC bus current sensor, reconstructing the stator current based on the PWM switching state. The reference position controller is pre-tuned to a stable closed loop, and the learning strategy only outputs candidate additional q-axis currents.
[0172] R is used in numerical verification s =0.35 Ω, L min =0.35 mH, I max =7 A、E max =32 V, ADC equivalent current error 25 mA, t req =3μs. From equation (9), the maximum current slope is approximately 2.356×10⁻⁶. 5 A / s. For 20,000 sets of random effective windows and random reconstruction error samples, the maximum upper bound of equation (10) is about 0.661 A, the maximum generation error is about 0.645 A, and no out-of-bounds samples appear; the maximum numerical error of the electromechanical closed relation (4) is about 7.73 × 10⁻⁶. -14 rad.
[0173] For 5000 sets of random closed-loop states, candidate actions, and disturbance tube radii, γ is calculated using equations (21)-(24). * Then, exhaustively check the grid using γ∈[0,1] at 20001 points; the maximum difference in γ among all comparable samples is approximately 5.0×10⁻⁶. -5 The average difference is approximately 1.24 × 10 -5 The corresponding grid resolution indicates that the scalar closed feasible region calculation is consistent with direct search.
[0174] Further, a mechanism-level discrete servo simulation was established, intentionally superimposing aggressive behaviors and sudden uncertainties into candidate learning actions. After 150 ms, the maximum position error of the candidate control without permission projection was approximately 11.68°, while after permission projection according to this invention, it was approximately 2.53°; the nominal interval average permission was approximately 0.723, the uncertainty sudden interval was approximately 0.628, and the minimum permission was reduced to 0; all samples whose candidate actions were allowed did not violate the set shrinkage certificate. This simulation only verifies the implementation of the 'residual tube-permission' mechanism and formula, and is not equivalent to the performance data of the trained TD3 real machine.
[0175] During engineering implementation, θ0 and It can be obtained from the factory calibration; E max With L min A conservative value is taken for the working region; P and q are obtained offline from the benchmark closed-loop discrete model. Only a few scalar operations, a 2×2 quadratic form, and the intersection of a scalar quadratic feasible region are required within the control cycle, making it suitable for DSP / MCU implementation.
Claims
1. A position servo control method for a permanent magnet synchronous motor based on the permission projection of an electromechanical closed-loop residual drive strategy, applied to a position servo system in which the permanent magnet synchronous motor drives a load via a reducer, a position encoder is installed at the load end, the rotor position on the motor side is obtained without a position observer, and the stator current is at least partially reconstructed from a single DC bus current sample, characterized in that... include: Step 1: Obtain the encoder angle θ at the load end L and rotor electrical angle output without position observer Based on the number of pole pairs p of the motor, the reduction ratio N, and the installation zero position θ0, the electromechanical kinematic closed residual is constructed as follows: (1) Where, r θ The electromechanical kinematic closure residual is expressed in electrical angle (rad); wrap(·) represents mapping the angle to the principal value interval (-π,π]; this electromechanical kinematic closure residual is obtained by directly closing the mechanical angle at the load end and the observed electrical angle at the motor side. Step 2: Determine the effective sampling window t required for single bus current reconstruction based on the current PWM switching state. win The effective sampling window is relative to the minimum required sampling time t. req Missing quantity t miss The upper bound of the current reconstruction error is formed by the maximum slope boundary of the motor current: (2) Among them, t win t is the effective sampling window length available for single bus current reconfiguration under the current PWM switching state. req The minimum time required to complete sample setup, hold, and analog-to-digital conversion; t miss The missing value is relative to the minimum sampling time, and max(0,·) is guaranteed not to take a negative value; (3) Among them, S i,max This is a conservative upper bound for the rate of change of stator current; V dc R is the DC bus voltage; s I is the stator phase resistance; max E represents the upper limit of the phase current amplitude in the considered operating region. max L is the upper bound of the back electromotive force amplitude. min This is the lower bound of the stator equivalent inductance within the operating area; (4) in, ε is the upper bound of the amplitude of the single bus current reconstruction error; ADC This is the upper bound of the equivalent current error after ADC quantization, bias, and sampling chain adjustments; S i,max ·t miss This indicates the additional uncertainty caused by the maximum current slope during periods of insufficient sampling window; Step 3: Based on the closed residual, the preset reducer angle error boundary, and the current reconstruction error upper boundary from Step 2, the rotor electrical angle error boundary, the q-axis current error boundary, and the position state disturbance boundary radius ρ are sequentially formed. k ;where ρ k It is obtained from the same closed residual along the physical transmission path of electrical angle, current, torque and load position; (5) in, This is the upper limit of the main value error of the rotor electrical angle, which shall not exceed π. This is a known engineering upper bound for the equivalent mechanical angular error of the reducer; this bound is based on the principal value distance around the circle and |Δθ|. g |≤ get; (6) in, To reconstruct the stator current vector magnitude; d î q To reconstruct the d-axis and q-axis currents; Δi q The underlined term represents the upper bound of the q-axis current comprehensive error; the second term represents the upper bound of the current vector chord length caused by angular mismatch, and... The conservative q-axis error bound is obtained by superposition; (7) in, This is a conservative upper bound vector for the state disturbance components, where the first component is the load position disturbance bound and the second component is the load angular velocity disturbance bound; T p For position control period; Δr θ ,k represents the principal difference of the closed residuals of adjacent periods; G τ Map the q-axis current error boundary to the load angular acceleration error; (8) Where, ρ k Let be the radius of the P-weighted state perturbation boundary for the k-th period; It is a symmetric positive definite weighted matrix determined offline; Step 4: The reference q-axis current i is generated by the reference position controller that meets the stability requirements in advance. q,b * The learning controller or adaptive controller generates only candidate additional actions a. π And the candidate additional actions are connected to the baseline closed loop with scalar permission coefficient γ∈[0,1]: (9) Among them, A b b is the discrete state matrix after the reference position controller closes the loop; a is the discrete input mapping vector from the q-axis additional current to the state increment; a π k Candidate additional q-axis currents given for the learning / adaptive policy; γ k ∈[0,1] represents the policy permission coefficient; w k For actual unmodeled state perturbations, their P-norm is required to be derived from ρ. k cover; Step 5: Use the same perturbation boundary radius ρ obtained in Step 3 k Establish a P-weighted contraction constraint, transform this constraint into a quadratic inequality with respect to the scalar limit coefficient γ, and intersect it with the q-axis current limit to obtain the feasible set Γ. k Take the maximum permission in the feasible set as γ. k * No feasible permission command γ exists. k *=0: (10) where q is a preset contraction factor, 0<q<1; the first term on the left is the nominal next-step P-norm after the reference closed-loop superimposes the candidate action, ρ k is the tube radius that must be reserved for unmodeled disturbances; when this inequality holds, the current candidate action will not damage the preset P-norm contraction margin; (11) Among them, Γ Q,k To satisfy the feasible set of permissions for robust contraction quadratic constraints; Γ I,k For the set of feasible permissions that simultaneously satisfy 0≤γ≤1 and the hard limit of q-axis current; Γ k The intersection of the two; i q,b,k * For the q-axis current command given by the reference controller, I max The upper limit of the allowable q-axis current amplitude; Where Γ Q,k Γ is the quadratic constrained feasible set within the range 0 ≤ γ ≤ 1; I,k To simultaneously satisfy 0≤γ≤1 and |i q,b,k * +γa k π The feasible set of current limits ≤ Imax; (12) Where, γ k * R represents the maximum feasible strategy permission coefficient to be adopted; when R k >0 and Γ k If the set is not empty, take the maximum value in the feasible set; otherwise, set it to 0, causing the candidate learning / adaptive action to exit while retaining the baseline control. Step 6: Output the final q-axis current command to the inner current loop: (13) Among them, i q,k * This is the q-axis current given to the final current loop; i q,b,k * For the reference controller output; γ k * a π k The candidate additional q-axis current after authorization projection.
2. The permanent magnet synchronous motor position servo control method based on the electromechanical closed residual drive strategy and permission projection according to claim 1, characterized in that, Motor mechanical angle θ m , load end angle θ L Reducer transmission error Δθ g And the installation of zero position θ0 satisfies: (14) Where, θ m θ is the mechanical angle of the motor; L θ is the mechanical angle at the load end; N is the reduction ratio; θ0 is the installation zero offset referred to the mechanical side of the motor; Δθ g This refers to the equivalent mechanical angular error caused by the reducer transmission. Define the position observation error δ e back: (15) Where, θ e This represents the actual rotor electrical angle; This is the rotor electrical angle estimate output without a position observer; The position observation electrical angle error; p is the number of motor pole pairs; The electromechanical closure residuals from step 1 satisfy the following identity: (16) Where, r θ Strictly speaking The modulo 2π equivalence class.
3. A permanent magnet synchronous motor position servo control method based on electromechanical closed-loop residual drive strategy permission projection according to claim 1 or 2, characterized in that, The transmission error of the reducer has a known upper engineering limit. When the upper bound of the rotor electrical angle error and the upper bound of the q-axis current error are calculated according to Equation (5) and Equation (6), respectively, the current vector error caused by the angle mismatch adopts the upper bound of the chord length shown in Equation (6), and the position observation error is not treated as another independent learning state.
4. The permanent magnet synchronous motor position servo control method based on electromechanical closed residual drive strategy permission projection according to claim 1, characterized in that, Let K t For the torque constant of the permanent magnet synchronous motor, η g For reducer efficiency, J L,eq To calculate the equivalent moment of inertia referred to the load end, the load end current-angular acceleration gain is obtained from equation (17); during the position period T p Internally, it is constructed using the position equivalent term of the closed residual, the velocity equivalent term of the closed residual difference, and the q-axis current error boundary: (17) Among them, K t For surface-mount PMSM in i d The q-axis torque constant when ψ ≈ 0; f For permanent magnet fundamental flux linkage; η g For reducer efficiency; J L,eq The equivalent moment of inertia referred to the load end; G τ This is the gain from the q-axis current error to the load angular acceleration error; (18) in, This is a conservative upper bound vector for the state disturbance components, where the first component is the load position disturbance bound and the second component is the load angular velocity disturbance bound; T p For position control period; Δr θ,k G represents the principal difference between the closed residuals of adjacent periods; τ Map the q-axis current error boundary to the load angular acceleration error; (19) Where, ρ k Let be the radius of the P-weighted state perturbation boundary for the k-th period; It is a symmetric positive definite weighted matrix determined offline.
5. The permanent magnet synchronous motor position servo control method based on the electromechanical closed residual drive strategy and permission projection according to claim 1, characterized in that, The closed-loop discrete matrix of the reference controller is A b P=P ᵀ >0 represents the pre-calculated weighting matrix, and the position tracking state and P-norm are defined as follows: (20) Where, x k The tracking state during the k-th position control cycle; e θ =θ L * -θ L For load position error; e ω = This represents the load angular velocity error; the superscript * indicates the command value, and k is the discrete position control cycle number. (21) Among them, V k V is a nonnegative quadratic form of the certificate quantity based on the P matrix; z is an arbitrary vector with the same dimension as the state; ||z||P is the weighted norm of P; in engineering implementation, P includes position / velocity scale normalization, therefore V k The symbols ||·||P are only used to standardize the stability certificate and do not represent the actual energy storage of the motor. When the disturbance satisfies ||w k ||P≤ρ k And when the permission coefficient satisfies equation (10), we have: (22) Among them, the contractile chain is only in R k >0、Γ k Non-empty and the actual perturbation satisfies ||w k ||P≤ρ k The current sampling time is valid; Equation (22) holds only for sampling moments where the feasible set is non-empty and the current perturbation bound is covered by the certificate; when the certificate is not feasible, the candidate learning action is withdrawn without claiming additional global stability beyond the baseline controller.
6. A permanent magnet synchronous motor position servo control method based on electromechanical closed residual drive strategy permission projection according to claim 1 or 5, characterized in that, Define R k for: (23) Among them, R k This is the usable shrinkage margin after deducting the perturbation boundary radius, and is related to ||x||P and ρ. k They are on the same certification scale; only R k Only when the value is greater than 0 can the equivalent square of equation (10) be a scalar quadratic constraint; In R k When the value is greater than 0, square equation (19) and rearrange it into equation (24). The coefficients are then calculated directly from equation (25). (24) Where γ is the dimensionless scalar authority to be determined; A k B k C k The quadratic constraint coefficients for the current period are determined by the state, candidate actions, and the P matrix, all of which correspond to the same certificate scale of the square of the P norm. (25) Among them, A k Characterize the secondary contribution of candidate actions to the P-weighted next state; B k Characterize the cross-contribution between candidate actions and the baseline closed-loop state; C k The weighted value of P in the next state of the baseline closed loop is reduced by the available shrinkage margin R. k 2 ; Since the decision variable is only a single scalar authority coefficient, the controller does not need to solve a multivariate quadratic programming problem online.
7. The permanent magnet synchronous motor position servo control method based on the electromechanical closed residual drive strategy and permission projection according to claim 1, characterized in that, When the current reference vector is non-zero, the current phase residual is further constructed: (26) Where, r i For current phase residual; i d * i q * These are the d-axis and q-axis current references, respectively; d î q These are the reconstructed currents along the d and q axes obtained in the observation coordinate system, respectively; the two-parameter arctangent function atan2(y,x) gives the directed principal value angle; In the operating region where the current inner loop tracking error is relatively small, the current phase residual is added to the electromechanical closure residual to obtain: (27) Where, r Σ The auxiliary closure residual is the result of combining the electromechanical closure residual and the current phase residual; δ c The current phase disturbance is defined as the current reconstruction phase error δ. rec Phase error δ of dynamic tracking of the inner current loop dyn The sum; the approximate relationship is applicable to working areas with small phase errors and consistent coordinate sign conventions; This causes the electrical angle error of the position observer to cancel out in the first-order approximation, and in terms of r Σ Auxiliary monitoring is performed on the combined drift of reducer error and current reconstruction / dynamic error.
8. The permanent magnet synchronous motor position servo control method based on the electromechanical closed residual drive strategy and permission projection according to claim 1, characterized in that, The candidate additional action generator is any one of the offline trained TD3, SAC, or DDPG policy networks, or an online adaptive control law; its output does not directly cover the baseline controller, but only serves as aπ in the permission projection.
9. The permanent magnet synchronous motor position servo control method based on the electromechanical closed residual drive strategy and permission projection according to claim 1, characterized in that, t win The current space vector PWM switching state, bus current sampling setup time, analog-to-digital conversion hold time, and dead time are all determined together; when t win ≥t req time t miss =0, when t win <t req The time error bound increases continuously with the missing time, thus allowing the same PWM sampling dead zone to pass through ρ k Continuously reduce the availability of candidate learning actions.
10. A permanent magnet synchronous motor position servo control system with electromechanical closed-loop residual drive strategy permission projection, comprising a permanent magnet synchronous motor, a reducer, a load, a load-end position encoder, a three-phase inverter, a DC bus single current sensor, a position observerless device, a processor, and a memory, characterized in that, When the processor executes the program in the memory, it implements the permanent magnet synchronous motor position servo control method according to any one of claims 1 to 9, which is based on the electromechanical closed residual drive strategy permission projection.
