A permanent magnet synchronous motor trusted parameter online identification and prediction control method

Abstract

The application relates to a permanent magnet synchronous motor trusted parameter online identification and prediction control method, and belongs to the technical field of permanent magnet synchronous motor digital control and model prediction control. The method triggers multi-rate recursive least squares online identification by Active model mismatch, forms Candidate candidate parameters; Candidate obtains trusted qualifications through four trusted judgments of physical rationality, incentive sufficiency, continuous convergence and model residual, meets the submission gate, and freezes the parameter package to be atomically locked as Shadow on the PL side; in the VERIFY stage, Active keeps real control, Shadow only bypasses prediction, verifies Shadow by using the prediction difference between new and old models and the prediction error of the two models on the real next beat current; after verification, direct, fast ramp, slow ramp takeover or takeover after extended verification are selected according to the prediction difference level, the parameter change rate is limited during takeover, and a double model robust cost evaluation is adopted; four-stage request / ack handshake is adopted between PS and PL to transmit the parameter package, and the high-speed control cycle determinacy is maintained.

Description

Technical Field

[0001] This invention belongs to the field of digital control, online parameter identification and model predictive control technology of permanent magnet synchronous motors, and relates to a method for online identification and predictive control of reliable parameters of permanent magnet synchronous motors. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs) offer advantages such as high power density, high efficiency, and fast torque response, making them widely used in high-performance drive applications such as servo drives, new energy vehicles, and industrial robots. Among various current control strategies, Finite Control Set Model Predictive Control (FCS-MPC) directly enumerates a finite number of candidate switching states of the inverter, uses a discretized predictive model of the motor to calculate future current, and selects the optimal output based on a cost function. This eliminates the need for pulse width modulation, resulting in fast dynamic response and ease of handling multiple constraints, and has garnered continuous attention in recent years.

[0003] However, FCS-MPC is a typical model-based control method, and its control effect is highly dependent on the accuracy of the predictive model parameters. In actual operation, the stator resistance changes significantly with temperature rise, and the stator inductor magnetic saturation and operating point influence deviate from the nominal value. Model parameter mismatch will directly translate into current prediction deviation, distorting the cost ranking of candidate switching states. This can lead to deterioration of current and torque quality, or even induce oscillations or overcurrent.

[0004] To address the parameter mismatch problem, existing technologies generally employ online parameter identification to correct the control model in real time. For example, Chinese patent CN116915103A uses recursive least squares online identification of self-tuning model parameters and directly applies them to FCS-MPC; Chinese patent CN121546955B uses a model reference adaptive observer to jointly estimate stator resistance and speed and inputs them into a deadbeat predictive controller; Chinese patent CN114679095A uses a Luneburger observer for load disturbance feedforward and mismatch parameter correction; some schemes also add filtering or convergence judgment before the identification results are put into use. However, the common feature of the above approaches is that the identification output is directly or periodically written into the predictive model after simple processing, which essentially assumes an equivalence between "algorithm output convergence" and "parameters can be put into control," a assumption that is not actually valid.

[0005] Specifically, firstly, the numerical convergence of the identification algorithm is not equivalent to the physical reliability of the parameters: insufficient regression information under low-excitation conditions may cause the identification results to deviate from the physically reasonable range; measurement noise and transient anomalies may cause output spikes or drift; using the same sample for both fitting and evaluation may also mask the true residual level. If such parameters are directly granted control without discrimination, it may introduce model errors that are more severe than the slow drift of the parameters. Secondly, even if the new parameters themselves are reliable, replacing the model parameters all at once between high-speed control cycles will still cause a step change in the predicted current trajectory and the order of switching states within a single cycle, triggering current surges. "Parameter reliability" and "parameters should take full effect immediately" are two different issues. Existing technologies lack a quantitative assessment of the differences in the predictive behavior of the old and new models and a mechanism for graded control of the parameter activation process based on the degree of difference. Third, FCS-MPC is increasingly deployed on PS / PL heterogeneous SoC platforms such as Zynq. The difference in time scale between high-speed control on the PL side and slow identification on the PS side leads to problems such as cross-cycle misshooting of observation data, half-update state of parameter writing process, and slow task blocking high-speed control, making it difficult to reliably implement the identification parameters on heterogeneous platforms.

[0006] In summary, existing technologies have not yet solved the problems of reliability evaluation before identification parameters are given control authority, verification of differences in predictive behavior between new and old models, hierarchical control of the timing and rate of change of reliable parameters, and consistency of observations across processing domains and atomic updates of parameters on heterogeneous platforms.

[0007] Therefore, it is necessary to provide a permanent magnet synchronous motor control method that can uniformly manage the entire process of online identification parameters from generation, confirmation, transmission, verification to effectiveness, and is compatible with PS / PL heterogeneous real-time control architecture. Summary of the Invention

[0008] In view of this, to address the problems that may arise when existing online identification results are directly or periodically entered into the FCS-MPC prediction model, such as unreliable parameter erroneous updates, control shocks caused by full parameter switching, and cross-processing domain observation errors, parameter half-updates, and slow tasks blocking high-speed control under heterogeneous PS / PL platforms such as Zynq, a reliable parameter online identification and predictive control method for permanent magnet synchronous motors is provided. This method provides unified management of the entire process of online identification parameters from generation, confirmation, transmission, verification to effectiveness.

[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for online identification and predictive control of reliable parameters of a permanent magnet synchronous motor, applied to a Zynq heterogeneous SoC platform including a processing system PS and a programmable logic PL, wherein the platform drives a permanent magnet synchronous motor connected via a three-phase two-level voltage source inverter, comprising the following steps: S1. On the PL side, obtain the motor current and electric angular velocity with a fixed high-speed control cycle; use the Active parameters as model parameters for the finite control set model predictive control FCS-MPC, predict the current of candidate switch states and select the actual switch state for output; calculate the model mismatch index based on the prediction of the next step of the actual executed switch state by the Active model and the actual current; when the model mismatch meets the entry condition, enter the IDENTIFY state from the HOLD state; when the exit condition is met, return to the HOLD state. S2 and PL sides form interval-aligned observation snapshots using adjacent control intervals as a unified time reference and upload them to PS; PS uses low-frequency supervision beats in HOLD state and high-frequency identification beats in IDENTIFY state. It constructs regression samples based on the current difference between adjacent high-speed control cycles and the actual applied voltage. It identifies stator resistance and stator inductance online through recursive least squares (RLS) with forgetting factor and forms candidate parameters through low-pass filtering. S3 and PS assess the credibility of the Candidate based on four criteria: physical rationality, sufficiency of incentives, continuous convergence, and model residuals. The Candidate is considered credible when all four criteria are met. S4. When the PS is in the IDENTIFY state, the Candidate is trusted, the Candidate's relative Active parameter change reaches the submission threshold, and the PL currently has no unprocessed Shadows, the PS freezes the parameter packet and sends it to the PL; the PL atomically latches the parameter packet as a Shadow at the security control cycle boundary. S5. After entering the VERIFY phase, Active continues to determine the actual switch state, while Shadow only performs bypass predictions for candidate states and actual executed switches that are the same as Active, and does not participate in the optimal switch selection; the differences in predictions between the old and new models are statistically analyzed. And the prediction error of Active and Shadow on the actual next clock cycle current. , ; S6. After Shadow passes VERIFY, select the takeover method based on the difference level predicted by the old and new models: small difference, direct takeover; medium difference, fast ramp or slow ramp with limited rate of change takeover; large difference, enter extended validation, and slow ramp takeover is allowed only after the extended validation is satisfied; during the rate of change limit takeover, the robust cost of the Active and Shadow dual models is evaluated and the rate of change is limited in each period. After completion, make Active exactly equal to Shadow. S7, PS and PL use a four-stage request / ack handshake to transmit parameter packets. PL does not wait for PS to calculate, maintaining the determinism of the high-speed control cycle.

[0010] Furthermore, the permanent magnet synchronous motor is a surface-mounted permanent magnet synchronous motor, which satisfies... = = For the j-th candidate switch state, based on the parameter ( , The single-step current prediction is as follows: (k+1)= (k)+( / )[ (k) · (k)+ (k)· · (k)], (k+1)= (k)+( / )[ (k) · (k) (k)· · (k) (k)· ]; in, For high-speed control cycle, , Let be the dq-axis candidate voltage corresponding to the j-th candidate switching state. , For the dq axis stator current, For stator resistance, For stator inductance, Electric angular velocity, It is a permanent magnet flux linkage.

[0011] Furthermore, the interval-aligned observation snapshot includes the current of the k-th cycle. (k) (k), the current in the (k-1)th period (k 1) (k 1) The actual applied voltage in the previous control interval (k 1) (k 1) Electric angular velocity (k 1) and PL sample counter (k), or the regression quantity calculated from the above quantities. , and counter; the HOLD low-frequency supervisory beat and IDENTIFY high-frequency identification beat are respectively = · and = · ,and < Each recursive least squares (RLS) sample is always composed of data from two adjacent high-speed PL control cycles, and the current differential is used. Forgetting factor of PS-side recursive least squares Low-pass filter coefficients and the forgetting coefficient of the recursive information matrix The physical time constant is uniformly set as follows: = , = , = , where x∈{mon,ident}.

[0012] Furthermore, the four credibility judgments described in S3 include: physical rationality judgment, when... ≤ ≤ and ≤ ≤ The timing and physical conditions are reasonable and effective; the sufficiency of the excitation is judged, and a recursive information matrix is ​​constructed after normalizing the regression matrix by columns. ,when =det( ) / ([tr( ) + )≥ And tr( )≥ Time-based incentives are valid; continuous convergence is checked by incrementing the count when the update amounts of adjacent Candidates do not exceed a stable threshold, otherwise the count is reset to zero. The count is incremented according to physical time. The threshold for conversion is set to be stable and effective; for model residual judgment, the residual is calculated by performing a one-step pre-update on the new sample of the kth group using the Candidate of the previous time step, and the residual is set to be effective when the average residual window does not exceed the threshold.

[0013] Furthermore, the submission threshold mentioned in S4 is: | |≥ or | |≥ When the IDENTIFY state is satisfied, =1. When at least one parameter change reaches the submission threshold and there are no unprocessed Shadows, set =1, freeze , , , Form a complete parameter package; do not commit Shadow in HOLD state.

[0014] Furthermore, the model mismatch described in S1 specifically satisfies the entry and exit conditions as follows: if the normalized model mismatch index... (k) continuous Each high-speed control cycle is no less than the entry threshold. , place =1 and enter IDENTIFY from HOLD; if (k) continuous Each control cycle should not exceed the exit threshold. If there are no locked or currently being processed Shadows, exit IDENTIFY and restore HOLD; < .

[0015] Furthermore, the condition for Shadow to pass the initial VERIFY in S5 is: =1; This is the new version after Active; Age≤ ; ≤ + ;and ≤ ;in For the allowable noise fluctuation amount, This represents the upper limit of the Shadow prediction error.

[0016] Furthermore, in S6 according to The specific takeover method is as follows: Set 0< < < ;when < Direct takeover is used at times; when ≤ < Fast-Ramp quick ramp connectors are used; when ≤ < When using a slow-ramp connection; ≥ When the Extended-Verify step is entered, the takeover method is fixed once COMMIT is entered; the Extended-Verify step includes: when the prediction quality of the initial VERIFY meets the requirements but... ≥ At that time, continue to maintain active real control, in The maximum prediction difference was recalculated over each period. and the mean of prediction error , Only when < , =1、Age≤ , ≤ + and ≤ Only when a slow-ramp is allowed to enter COMMIT, otherwise the Shadow is discarded and IDENTIFY is returned.

[0017] Furthermore, the change rates of the Fast-Ramp and Slow-Ramp limits are updated as follows: (k+1)= (k)+sat( (k), , + ), (k+1)= (k)+sat( (k), , + ), where m∈{fast, slow}, and > >0, > >0; the robust cost of the two models during COMMIT is =max( , )+ The dual-model overcurrent hard constraint requires that the predicted current of both Active and Shadow models not exceed [a certain value]. If all candidate states violate the hard constraints, the current protection branch is entered; the takeover completion condition is | |≤ And| |≤ After satisfying the condition, set Active to exactly equal Shadow, and update. = juxtaposition =COMPLETE.

[0018] Furthermore, the four-phase request / ack handshake includes: idle state =0 and =0; PS: First write the complete parameter package, then set... =1 and maintain data stability; PL only =1、 =0, No Shadows to be processed and parameter package version is newer than At that time, the entire system is latched as Shadow at the boundary of the safety control cycle; after processing, PL outputs... juxtaposition =1; PS clears after detecting ack. ;PL clears after detecting request zeroing. ; At least distinguish between ACCEPT, , and .

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: To address the issue of erroneous updates caused by unverified parameters being directly incorporated into the prediction model, this invention employs four credibility criteria: physical rationality, sufficiency of normalized excitation, continuous convergence, and one-step pre-update model residuals. These criteria constrain four types of risks: abnormal values, unidentifiable low-excitation conditions, non-convergent transients, and stable biases. The excitation information matrix is ​​constructed based on dimensionless regression quantities and simultaneously constrains directional richness and overall information intensity, ensuring that the excitation criteria do not change with unit selection. The model residuals are evaluated using new samples that the candidate has not yet participated in fitting, avoiding optimistic bias caused by fitting and validating the same sample. This clearly distinguishes between "numerical convergence" and "physical credibility," allowing only credible parameters to enter the subsequent effective process.

[0020] To address the control shock caused by full parameter switching (i.e., parameter reliability), this invention establishes a three-level parameter isolation mechanism (Candidate, Shadow, Active) and a two-stage activation mechanism (VERIFY, COMMIT): During the VERIFY stage, Active maintains true control throughout, while Shadow only performs bypass prediction, using the difference in prediction between the old and new models. Error compared to the actual current prediction based on the same actual operating switch , Evaluate model behavior differences and Shadow realism separately; COMMIT phase according to The system selects direct, fast ramp, slow ramp, or extended validation. During the rate-of-change-limited validation period, a dual-model robust cost evaluation is used, and a clear protection branch is entered when all candidate states are overcurrent. This ensures that the speed and timing of parameter activation match the degree of difference between the old and new models, avoiding step abrupt changes in the prediction model within a single control cycle.

[0021] To address the conflict between the time scale of identification supervision and high-speed control, this invention employs a model mismatch-triggered HOLD and IDENTIFY dual-rate scheduling, where each RLS sample is always constructed from adjacent high-speed control cycles, and the current difference is calculated using... The update frequency is decoupled from the regression time base; the forgetting factor, low-pass filter coefficient and information matrix forgetting coefficient are set uniformly according to the physical time constant, so that the real time memory length of the algorithm remains consistent before and after multi-rate switching.

[0022] To address the consistency and real-time issues across processing domains in heterogeneous PS / PL platforms, this invention employs interval-aligned observation snapshots to ensure that uplink observation data is taken from the same time base. It also uses frozen parameter packets and secure control cycle boundary atomic latches to avoid the coexistence of new stator resistance and old stator inductance in a half-update state. Furthermore, it employs a four-stage request / ack handshake combined with version number, sample count, and Shadow freshness Age management to ensure that the PL does not wait or block during PS calculations, maintaining the determinism of high-speed control cycles.

[0023] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0024] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 A schematic diagram illustrating the functional division of PS / PL in the Zynq chip, as well as the uplink of observation data and the downlink of reliable parameters. Figure 2 Flowchart for Candidate parameter credibility assessment and submission gating; Figure 3 This diagram illustrates the lifecycle of the Candidate, Shadow, and Active parameters. Figure 4 This diagram illustrates model mismatch, reliability assessment, and state transition. Figure 5 This is a schematic diagram of VERIFY bypass verification and parameter takeover. Figure 6 A diagram illustrating PS / PL parameter package handshake, version synchronization, and Shadow freshness. Figure 7 This is a schematic diagram of the VERIFY event-level diagnostic results. Detailed Implementation

[0025] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0026] The specific embodiments of the present invention will be described below with reference to the accompanying drawings. The following embodiments take a surface-mounted permanent magnet synchronous motor (SPMSM) driven by a three-phase two-level voltage source inverter as an example, and implement the FCS-MPC current inner loop and reliable parameter lifecycle mechanism on the Zynq PS / PL heterogeneous SoC platform.

[0027] Example 1 This example demonstrates the overall system architecture and the lifecycle of the three-level parameters. For example... Figure 1 As shown, the system in this embodiment includes a permanent magnet synchronous motor, a three-phase two-level voltage source inverter, a current / voltage sampling unit, a rotor position or electrical angle acquisition unit, a Zynq processing platform, and an inverter drive interface.

[0028] Zynq is internally divided into a processing system (PS) side and a programmable logic (PL) side. The PS side primarily includes observation snapshot reception and buffering, multi-rate RLS online parameter identification, candidate buffering, parameter reliability judgment, submission determination, parameter version and sampling count management, and AXI parameter transmission / handshake management. The PL side primarily includes ADC sampling, Clark / Park coordinate transformation, actual applied dq-axis voltage acquisition or reconstruction, candidate dq-axis voltage generation for eight candidate switching states, interval-aligned RLS observation snapshots, Active Bank, Shadow Bank, FCS-MPC parallel prediction, model mismatch monitoring, VERIFY bypass verification, four-state management, Age and version checking, COMMIT dual-model robustness evaluation, parameter takeover, and optimal switching state output.

[0029] This embodiment separates the five stages of parameter generation, trust confirmation, transmission latching, bypass verification, and control activation, and sets three levels of parameters: Candidate is the candidate parameter identified and filtered by the PS-side RLS, which only has the right to make suggestions; Shadow is the target parameter to be taken over that is fully latched by the PL after the Candidate passes the trust and submission judgment; Active is the FCS-MPC model parameter that is currently officially effective.

[0030] The system is configured with four main states: HOLD, IDENTIFY, VERIFY, and COMMIT. Extended-Verify serves as an extended subprocess of VERIFY. In the HOLD, IDENTIFY, and VERIFY stages, the Active state determines the actual inverter switching state. In the VERIFY stage, the Shadow only bypasses the prediction and does not change the actual switching decision. Only after entering COMMIT does the Shadow participate in the dual-model robustness cost evaluation or act as a parameter takeover target. At no stage can the Candidate directly override the Active state. If a Candidate subsequently becomes invalid again, it will not automatically cancel a Shadow that has been fully latched and entered the verification process; the Shadow is handled independently. , and manage.

[0031] like Figure 3 As shown, stator resistance and stator inductor After a change occurs, the Candidate first tracks the parameter change, forms a Shadow, and then the Active takes over. This reflects the separation and update process of parameters at the three levels: Candidate, Shadow, and Active. This decouples the parameter's credibility qualification from its formal effectiveness.

[0032] Example 2 This embodiment is used to demonstrate the online identification of permanent magnet synchronous motor model, FCS-MPC basic prediction model and interval-aligned multi-rate RLS candidate parameters.

[0033] This embodiment uses a surface-mounted permanent magnet synchronous motor (SPMSM) as the object, which satisfies... = = The voltage equation for the synchronously rotating dq coordinate system is: , ;in, , This refers to the dq-axis stator voltage. , For the dq axis stator current, For stator resistance, For stator inductance, Electric angular velocity, This is a permanent magnet flux linkage. This embodiment uses online identification. and , Use nominal values ​​or pre-calibrated values.

[0034] For the j-th candidate switching state, the candidate voltage is obtained based on the current DC bus voltage, switching combination, and electrical angle. (k) (k). Forward Euler discretization is employed, based on any model parameter ( , The single-step prediction is: , ; Substituting the Active or Shadow parameters into the above prediction formulas yields two sets of prediction results. It should be noted that the actual applied voltage and the FCS-MPC candidate voltage are conceptually strictly distinguished: , This represents the actual dq-axis voltage applied to the motor within the previous high-speed control interval, used for RLS regression and model mismatch evaluation; , This represents the candidate voltage corresponding to the j-th candidate switch state, used only for FCS-MPC candidate prediction. The actual voltage can be obtained from voltage sampling or reconstructed from the DC bus voltage, the actual switch state, and coordinate transformation relationships.

[0035] This embodiment uses SPMSM to clearly illustrate the two-parameter trusted update mechanism; for ≠ The built-in permanent magnet synchronous motor can expand the parameter vector to [ , , The regression matrix and Active and ShadowBank are expanded accordingly, while the Candidate credibility judgment, bypass verification, hierarchical takeover and atomic update mechanisms remain unchanged.

[0036] Based on the above prediction model, such as Figure 1 As shown, PL uses a fixed high-speed control cycle. Continuous operation; the PS-side RLS uses a low-frequency monitoring beat in the HOLD state and a high-frequency identification beat in the IDENTIFY state: = · , = · ,and < .

[0037] Multi-rate only changes the frequency at which RLS samples are sent to the PS and updated, without altering the current differential time base used to construct individual motor regression samples; each RLS sample is always composed of data from two adjacent PL high-speed control cycles, and its current differential uses... This avoids the problem that the low-frequency RLS update interval and the instantaneous voltage are not on the same time scale.

[0038] To ensure that the voltage, current, and rate of change of current in the RLS regression equation correspond to the same physical interval, sampling is performed in the k-th control cycle. (k) After (k), use the actual applied control within the previous control interval [k-1,k]. (k 1) (k 1) and (k 1) (k 1) (k 1) Constructing the regression sample. From the SPMSM voltage equation, we can obtain: , , , ; The two sets of equations for the d and q axes are used together in the same RLS update, defined as follows: =[ (k), (k) , =[ (k), (k) , =[ (k), (k) Then there is = · .

[0039] The PL can directly upload the above-mentioned original interval data, or an equivalent one can be pre-generated on the PL side. and Uploaded later. HOLD only per Each control cycle selects a set of interval-aligned samples, IDENTIFY each Each control cycle selects one set; regardless of the selection interval, the current difference in the regression always uses the adjacent high-speed control cycles. instead of using .

[0040] For the joint regression form, matrix RLS with a forgetting factor is used: , , ; Where x∈{mon,ident} represents the current RLS tick. To avoid changes in the memory length of the forgetting factor, low-pass filter, and information matrix in real time after changes in the update frequencies of HOLD and IDENTIFY, this embodiment uniformly sets multiple rate coefficients according to the physical time constant: = , = , = .

[0041] The raw RLS output is filtered by a first-order low-pass filter: (k)= · (k 1)+(1 )· The low-pass filtered parameter vector is then saved as a candidate parameter for Candidate. (k)= (k)=[ (k), (k) The HOLD state allows RLS to update the Candidate infrequently for oversight, but it closes the Shadow commit gate; a Candidate can only be eligible to commit after entering IDENTIFY.

[0042] Example 3 This example demonstrates model mismatch detection, switching between HOLD and IDENTIFY states, Candidate credibility assessment, and submission gating. like Figure 4 As shown, PL utilizes the Active model in the (k-1)th control cycle based on the actual execution switch state. (k 1) Compare the obtained next-cycle prediction with the actual sampled current in the k-th control cycle to construct a normalized model mismatch index: ; in As a current reference, To prevent positive numbers with excessively small denominators, an entry threshold is set. and exit threshold ,satisfy < .like (k) continuous The high-speed control cycle is no less than Then place =1 and enter IDENTIFY from HOLD; if (k) continuous Each control cycle should not exceed If no Shadow is latched or being processed, exit IDENTIFY and resume HOLD. This hysteresis and continuous counting are used to avoid noise causing repeated state switching.

[0043] It should be noted that the model mismatch index only indicates that the Active Prediction Model is inconsistent with the actual system, and does not pre-equate the cause of the error to be... or Therefore, parameters are only eligible for submission after subsequent RLS and four credibility checks have passed.

[0044] RLS outputs parameter values ​​with each update, but the existence of these values ​​does not equate to their suitability for inclusion in the control model. For example... Figure 2 As shown, this embodiment establishes a credibility judgment from four dimensions: physical feasibility, regression data identifiability, parameter continuous convergence, and candidate model data consistency. Only when all four conditions are met simultaneously can the Candidate obtain credibility.

[0045] (1) Physical rationality judgment. Boundaries are set based on the motor's rated parameters, allowable temperature rise, magnetic saturation range, offline calibration results, and engineering margins: ≤ ≤ and ≤ ≤ When both conditions are met =1, otherwise =0. This judgment only excludes obviously impossible estimates and does not prove that the parameters have converged or are consistent with the actual motor.

[0046] (2) Determination of excitation sufficiency. Since the two columns of the regression matrix correspond to the resistance and inductance terms respectively, and their physical dimensions and numerical scales are different, directly constructing an information matrix from the unnormalized regression vector would cause the criterion to depend on unit selection. In this embodiment, the regression matrix is ​​first normalized column by column: =diag( , ), = · ,in As a current reference, Using the current change rate as a benchmark, a recursive information matrix is ​​constructed using a normalized regression matrix: = · + · and calculate =det( ) / ([tr( ) + ). It mainly reflects whether the information in the direction of the two parameters degenerates, tr( This reflects the overall information strength after normalization. To simultaneously avoid both "one-dimensional direction" and "weak absolute incentive," only when... ≥ And tr( )≥ At the same time, time position was established =1, otherwise =0. Because... Based on the construction of dimensionless regression quantities and It does not change with the switching of units such as ampere / milliampere or henry / millihenry.

[0047] (3) Continuous convergence judgment. Define the amount of data updated between two consecutive Candidate updates: (k)= (k) (k 1), (k)= (k) (k 1). When | (k)|≤ And| (k)|≤ hour Add 1; if any condition is not met Reset immediately and Set to 0 immediately. To ensure that HOLD and IDENTIFY have the same actual settling time, the counting threshold is set according to physical time. Conversion: =ceil( / );when To reach the corresponding Time setting =1. This judgment prevents the parameter from being mistakenly judged as stable due to a small, accidental change during oscillation, spikes, or convergence.

[0048] (4) Model residual judgment. To avoid using the same set of samples that just participated in the RLS update to evaluate the updated Candidate and obtain an overly optimistic fitting residual, this embodiment adopts a one-step pre-update to verify the residual: after the kth group of new regression samples arrives, the Candidate that was frozen at the previous RLS time step is used first. (k 1) Predict the new sample, then complete the k-th RLS update. Its prediction is... = · (k 1), the residual is (k)=‖ / (‖ + ),in This serves as a physically meaningful voltage normalization reference to prevent the relative error from being infinitely amplified when the observations approach zero. The residual uses new samples that the candidate has not yet participated in the fitting, reflecting the consistency of parameter predictions for subsequent observations; after this residual recording is completed, sample k is allowed to enter the RLS update. The residual window is set to a fixed physical duration. set up: =ceil( / ), (k) represents the window. The mean; when (k)≤ hour =1, otherwise =0. If a Candidate fails to maintain stability again within the residual evaluation window, the window is discarded and restarted to avoid accumulating old residuals with significantly changed Candidates.

[0049] The four-item comprehensive judgment uses logical AND: only if... , , and When all four flags are 1, set =1; set immediately if any flag is 0. =0.

[0050] Candidate is considered trustworthy based on four criteria, but updates are not mandatory. To establish a parameter update dead zone, further comparisons are made between Candidate and Active: | |≥ or | |≥ Only when the system is in IDENTIFY state... =1. When at least one parameter change reaches the submission threshold, and there are currently no unprocessed Shadows in the PL, set =1, freeze at this time , , and This forms a complete parameter packet to be sent. In the HOLD state, regardless of the Candidate value, Shadow is not submitted directly.

[0051] Example 4 This embodiment demonstrates VERIFY bypass verification, Shadow authenticity evaluation, Extended-Verify extended verification, and COMMIT hierarchical takeover.

[0052] Upon entering the VERIFY phase, Shadow does not alter the actual switching decision. The actual switching decisions in the HOLD, IDENTIFY, and VERIFY phases all use only the Active model: for the j-th candidate state, the Active current tracking cost is defined. (k)=‖ (k) (k+1) Define Active overcurrent hard constraints :when The amplitude does not exceed hour =0, otherwise = Total cost = + Optimal switching state = The actual inverter switch combination corresponding to the PL output.

[0053] If all eight candidate states are deemed invalid by hard constraints, the controller enters a predefined current protection branch, such as outputting a safety zero vector and setting a protection flag, instead of forcibly executing argmin for the invalid candidates. The value is a sufficiently large positive number, which is much larger than the maximum possible value of the current tracking cost of any candidate state, so that the candidate state that is judged as invalid by the hard constraint will not be selected when finding the minimum total cost.

[0054] like Figure 5 As shown, during VERIFY, PL simultaneously calculates Active and Shadow predictions for each candidate state: (k+1)=f(i(k), (k), (k), ), (k+1)=f(i(k), (k), (k), Define the difference in predictions between the old and new models for a single period as D(k) = || (k+1) (k+1) After normalization (k)=D(k) / ( + ), and in continuous Within each high-speed control cycle, take =max (k). This only measures the impact of the old and new models on the predicted behavior of the candidate motor, and is not used to determine which set of parameters is closer to the actual motor. To truly verify Shadow, it is selected and actually executed by Active in the k-th control cycle. After (k), the actual current i(k+1) is obtained in the next control cycle, and the prediction errors of the two models for the same actual switching state are calculated respectively. (k+1)=‖i(k+1) (k+1) / ( + )and (k+1)=‖i(k+1) (k+1) / ( + ), and take the window mean. and .

[0055] Shadow must satisfy the following conditions simultaneously for the initial VERIFY: =1; This is the latest version after the current Active version; Age≤ ; ≤ + ;and ≤ .in Small statistical fluctuations are permissible within the noise range. Limit the absolute level of Shadow's own prediction error. If any condition is not met, reject Shadow and return IDENTIFY.

[0056] like Figure 7 As shown, in the two VERIFY decisions in this embodiment, Located respectively to interval sum to The intervals correspond to the Fast-Ramp and Slow-Ramp takeovers; simultaneously Below and satisfy The restriction indicates that Shadow's authenticity verification has passed.

[0057] Only after the authenticity verification is passed can it be based on Decide on the speed of takeover. For example... Figure 5 As shown, set 0 < < < : < Use Direct; ≤ < Fast-Ramp was used; ≤ < Use Slow-Ramp; ≥ Enter Extended-Verify. Once the takeover mode enters COMMIT, it remains fixed and will not switch repeatedly on each beat.

[0058] When the initial VERIFY's Shadow prediction quality meets the requirements, but ≥ At that time, instead of immediately taking over, the Active entity continues to maintain real control, and in the new... Re-statistical analysis of maximum forecast difference within a periodic window =max and the mean of prediction error , When extended validation ends, only when =1、Age≤ , ≤ + , ≤ and < Only when the condition is met is COMMIT allowed and Slow-Ramp consistently used; otherwise, the Shadow is deemed unsuitable for takeover under the current conditions, discarded, and the process returns to IDENTIFY. Therefore, Extended-Verify is not a simple delay, but rather a re-evaluation of whether the original cause of the significant difference has disappeared.

[0059] During the COMMIT phase, Direct mode is only applicable to those who have passed VERIFY and < The Shadow, at the boundary of the security control cycle, directly... = And immediately complete the COMMIT. Fast-Ramp and Slow-Ramp use limit rate of change updates: , ; Where m∈{fast,slow}, and satisfies > >0、 > >0. Fast-Ramp and Slow-Ramp make predictions simultaneously during COMMIT using the currently gradually changing Active and the final Shadow, defined as follows: =‖ , =‖ The dual-model overcurrent hard constraint requires that neither model predicts overcurrent, i.e., only when... and The amplitudes do not exceed hour =0, otherwise = Dual-model robustness cost =max( , )+ , = This worst-case cost is used to avoid making a switching choice that is extremely unfavorable to another model based on only one model during parameter transitions; if all candidate states violate hard constraints, the current protection branch is entered.

[0060] The conditions for Ramp takeover to be completed are | |≤ And| |≤ After satisfying the condition, set Active to exactly equal Shadow, and update. = juxtaposition =COMPLETE. The Age check has been completed before entering COMMIT, therefore the natural growth of Age is not used as a condition for mid-rollback; however, if a hardware overcurrent, parameter packet integrity failure, or continuous Shadow prediction error occurs... More than one cycle If a hard safety condition is met, immediately stop further ramping, retain the current Active parameter, and set... = and = Clear the current Shadow and return IDENTIFY. This is only used as a sequence number for parameter packet processing. This is used to indicate whether the version has fully reached the Shadow target, and to avoid... This avoids the misconception that the parameter value must be equal to the corresponding Shadow, and also prevents a sudden rollback from causing a second parameter jump.

[0061] Example 5 This example demonstrates PS / PL bidirectional multi-rate communication, version management, and a four-phase atomic handshake. For example... Figure 6 As shown, PS / PL communication is divided into two directions.

[0062] (1) Uplink of PL→PS observation data. After the PL completes current sampling in the high-speed period k, it forms an RLS observation snapshot with the adjacent control interval [k-1,k] as a unified time base. The snapshot can at least uniquely recover the data. (k) (k) (k 1) (k 1) (k 1) (k 1) (k 1) and (k), or directly including the quantities calculated from these quantities , and counter. HOLD button Extract, IDENTIFY Extraction; PL writes via AXI FIFO, AXI BRAM double buffer, DMA, or an interface with equivalent full snapshot semantics. PS only retrieves data when the entire packet is readable. PL does not wait for PS to complete RLS computation.

[0063] (2) PS→PL trusted parameter downlink. Use the latest PL observation snapshot count used to generate the current Candidate, instead of the local PS software time. When the Candidate meets the commit conditions in IDENTIFY... Generate a new version according to a monotonically increasing rule, and freeze the parameter package { , , , The PS must not modify the parameter packet before the handshake is complete. The version number is used to determine the order of parameter packets and to identify duplicate requests; whether Active has fully reached the Shadow target corresponding to this version is determined by... =COMPLETE or To be marked separately.

[0064] (3) A four-phase request / ack handshake is used to prevent the same request from being latched repeatedly. Idle state =0 and =0; PS: First write the complete parameter package, then set... =1 and maintain data stability; PL only =1、 =0, There are currently no Shadows to be processed and the parameter package version is newer than 0. At that time, the entire system is latched as Shadow at the security control cycle boundary, and then the VERIFY, COMMIT, or rejection procedure is executed; after processing, PL outputs... juxtaposition =1; PS clears after detecting ack. ;PL clears the request only after detecting that it has been cleared. This concludes the handshake session.

[0065] (4) PL calculates Shadow freshness: Age = Before entering VERIFY and at the end of Extended-Verify, Age must be ≤ Otherwise, Shadow is rejected; after entering COMMIT, Age is no longer used as a rollback condition. If the counter may wrap around during a single device operation, the old and new comparison rules of unsigned modulo arithmetic are used; this embodiment preferably ensures that the version count and sample count do not wrap around during a single continuous operation.

[0066] (5) Before VERIFY or COMMIT ends, PS must not overwrite the current Shadow. New Candidates generated subsequently are only cached on the PS side, and only the previous one is cached. The next version can only be sent after the four-phase handshake is completed. At least distinguish between ACCEPT, , and This is so that Photoshop can properly clean or re-identify the image.

[0067] like Figure 6 As shown, in this embodiment, both parameter updates completed the PS / PL parameter packet handshake, and the Candidate, Shadow, and Active versions were synchronized sequentially, with the Shadow freshness Age remaining unchanged. Within this timeframe, the parameter packets are processed sequentially.

[0068] Example 6 This embodiment demonstrates the algorithm simulation experiment and results. Using a three-phase two-level voltage source inverter-driven SPMSM as the object, this embodiment verifies the FCS-MPC current inner loop and the reliable parameter lifecycle mechanism. The algorithm simulation uses a given mechanical speed and dq-axis current reference.

[0069] Simulation settings: Initial stator resistance of permanent magnet synchronous motor =0.30Ω, initial stator inductance =1.50mH, mechanical speed 1000r / min, PL high-speed control cycle =50μs, the d-axis current reference value is 0A, and the q-axis current reference value is set to 50A after startup. At t=0.040s, the actual stator resistance is changed from 0.30Ω to 0.60Ω, and at t=0.090s, the actual stator inductance is changed from 1.50mH to 1.20mH to simulate the situation where the prediction model parameters are mismatched.

[0070] The simulation process is as follows: Establish SPMSM, three-phase two-level inverter and FCS-MPC models, formed from actual switching states. , It is formed independently by eight candidate switch states. , ; In PL, interval-aligned observation snapshots are implemented, and HOLD is used for each... One pack is drawn periodically, IDENTIFY each One packet is drawn periodically; PS constructs a joint d- and q-axis regression matrix using adjacent high-speed cycle data from the same snapshot, with the current difference fixed and divided by... Simultaneous update using matrix RLS and and according to Calculated using physical time constant , and ; After each new RLS sample arrives, the previous Candidate is used to calculate a one-step pre-update residual for the new sample. Then, matrix RLS is performed on the sample and a new Candidate is formed through a low-pass filter. Subsequently, four confidence judgments are completed. PL uses the previous clock prediction and the current actual current calculation. When the hysteresis condition is met, it switches between HOLD and IDENTIFY. When the commit conditions are met in IDENTIFY, the parameter package is frozen and a four-phase handshake is initiated. After the version and Age are pre-checked, PL latches Shadow as a whole. During VERIFY, the actual inverter switching is still selected by the Active single-model cost function; Shadow only performs parallel bypass prediction, continuously... Statistics , and ; After authenticity verification, according to and , , The relationship determines whether it is Direct, Fast-Ramp, Slow-Ramp, or Extended-Verify; In COMMIT, Direct updates Active directly at the security lifecycle boundary, while Fast-Ramp and Slow-Ramp updates according to their respective... , The parameters are limited to a frame-by-frame variation, and the max( of the two models is used) is utilized. , )+ Select the actual switch status; After completing or rejecting a Shadow, a four-phase handshake is performed to conclude the process. During the Verify and Commit phases, new Candidates are only cached in the PS and do not overwrite the current Shadow.

[0071] Record model mismatch indices during simulation Candidate, Shadow, Active parameter trajectories, four credibility judgment flags, state machine transition process, and the VERIFY phase. , and The takeover mode, PS / PL parameter package handshake, version number, and Shadow freshness Age are among the variables.

[0072] like Figure 3 As shown, after the parameters change, the Candidate first tracks the parameter changes, forms a Shadow, and then the Active class takes over; for example... Figure 4 As shown, the mismatch index of the Active model increases after parameter changes, and the state sequentially enters IDENTIFY, VERIFY, and COMMIT, returning to HOLD after takeover is complete; as... Figure 5 As shown, the two verifications used Fast-Ramp and Slow-Ramp respectively, and Active then took over from Shadow; as Figure 6As shown, both parameter updates completed the handshake, the versions were synchronized sequentially, and the Age remained unchanged. Within; such as Figure 7 As shown, the two VERIFY decisions They fall into the Fast-Ramp and Slow-Ramp intervals respectively, and Below And it satisfies the absolute upper limit constraint. The above results verify the effectiveness of the trusted parameter lifetime and the secure takeover mechanism.

[0073] It should be noted that the above thresholds and parameters can be adjusted according to the following principles: entry threshold Model mismatch index under normal operating conditions The statistical upper limit is set with a margin, and the exit threshold is set. Pick 50% to 80% to form a hysteresis; continuous counting threshold , Based on the maximum allowable mismatch duration and high-speed control cycle Conversion; Calculated based on the physical duration corresponding to the monitoring cycle. Based on the convergence speed requirements of the parameters and the computational power of PS, and < The time constants of the forgetting factor, low-pass filter coefficients, and information matrix forgetting coefficients. , , Set the desired algorithm memory duration; Values ​​between 0 and 0.25 are assigned based on the tolerance for directional degradation of the parameter. Calculated based on the typical amplitude of the normalized regressor and the required cumulative information; stability threshold. , Set the statistical upper limit based on the Candidate's steady-state fluctuations, physical time. Set based on the shortest dwell time under operating conditions; residual threshold The residual window physical duration is set based on the normal operating residual level plus a margin. Choose a duration that can cover a sufficient number of samples.

[0074] Submission threshold , The parameter update dead zone is formed by setting a preset percentage of the parameter nominal value or the sensitivity of the control performance to the parameter error. Calculated based on the duration for which the expected operating conditions remain unchanged after the Shadow is generated; The prediction error is set according to the statistical fluctuation caused by current measurement noise. Set according to the upper limit of acceptable absolute model prediction error; tiered thresholds. Set based on the maximum forecast difference where the takeover shock is negligible. Expand the lower limit of the validation difference setting as needed. exist and The range is set at a compromise between response speed and nozzle impact; the rate of change is limited. , Set according to the maximum parameter change that will not cause a perceptible current surge in each control cycle. , Take a value less than the corresponding fast ramp; count the number of periods. , Select a number of periods that can cover a sufficient statistical sample; complete the takeover criterion. , Set according to the allowable residual parameter error; , The settings are based on the continuous over-limit tolerance of hard safety conditions. The above setting principles do not depend on the specific motor model; those skilled in the art can determine the specific values ​​based on the motor parameters, control cycle, and performance requirements of the actual application.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for online identification and predictive control of reliable parameters of a permanent magnet synchronous motor, applied to a Zynq heterogeneous SoC platform including a processing system (PS) and programmable logic (PL), wherein the platform drives a permanent magnet synchronous motor connected via a three-phase two-level voltage source inverter, characterized in that, Includes the following steps: S1. On the PL side, obtain the motor current and electric angular velocity with a fixed high-speed control cycle; use the Active parameters as model parameters for the finite control set model predictive control FCS-MPC, predict the current of candidate switch states and select the actual switch state for output; calculate the model mismatch index based on the prediction of the next step of the actual executed switch state by the Active model and the actual current; when the model mismatch meets the entry condition, enter the IDENTIFY state from the HOLD state; when the exit condition is met, return to the HOLD state. S2 and PL sides form interval-aligned observation snapshots using adjacent control intervals as a unified time reference and upload them to PS; PS uses low-frequency supervision beats in HOLD state and high-frequency identification beats in IDENTIFY state. It constructs regression samples based on the current difference between adjacent high-speed control cycles and the actual applied voltage. It identifies stator resistance and stator inductance online through recursive least squares (RLS) with forgetting factor and forms candidate parameters through low-pass filtering. S3 and PS assess the credibility of the Candidate based on four criteria: physical rationality, sufficiency of incentives, continuous convergence, and model residuals. The Candidate is considered credible when all four criteria are met. S4. When the PS is in the IDENTIFY state, the Candidate is trusted, the Candidate's relative Active parameter change reaches the submission threshold, and the PL currently has no unprocessed Shadows, the PS freezes the parameter packet and sends it to the PL; the PL atomically latches the parameter packet as a Shadow at the security control cycle boundary. S5. After entering the VERIFY phase, Active continues to determine the actual switch state, while Shadow only performs bypass predictions for candidate states and actual executed switches that are the same as Active, and does not participate in the optimal switch selection; the differences in predictions between the old and new models are statistically analyzed. And the prediction error of Active and Shadow on the actual next clock cycle current. , ; S6. After Shadow passes VERIFY, select the takeover method based on the difference level predicted by the old and new models: for small differences, use direct takeover; for moderate differences, use fast ramp or slow ramp with limited rate of change takeover; for large differences, enter extended validation, and only after the extended validation is satisfied can slow ramp takeover be allowed; during the rate of change limit takeover, use Active and Shadow dual-model robust cost evaluation and limit the rate of change in each period, and after completion, make Active exactly equal to Shadow. S7, PS and PL use a four-stage request / ack handshake to transmit parameter packets. PL does not wait for PS to calculate, maintaining the determinism of the high-speed control cycle.

2. The method according to claim 1, characterized in that, The permanent magnet synchronous motor is a surface-mounted permanent magnet synchronous motor that meets the requirements. = = ; For the j-th candidate switch state, based on the parameter ( , The single-step current prediction is as follows: ， ； in, For high-speed control cycle, , Let be the dq-axis candidate voltage corresponding to the j-th candidate switching state. , For the dq axis stator current, For stator resistance, For stator inductance, Electric angular velocity, It is a permanent magnet flux linkage.

3. The method according to claim 1, characterized in that, The interval-aligned observation snapshot includes the current of the k-th cycle. (k) (k), the current in the (k-1)th period (k 1) (k 1) The actual applied voltage in the previous control interval (k 1) (k 1) Electric angular velocity (k 1) and PL sample counter (k), or the regression quantity calculated from the above quantities. , and counter; the HOLD low-frequency supervisory beat and IDENTIFY high-frequency identification beat are respectively = · and = · ,and < ; Each RLS sample is always composed of data from two adjacent PL high-speed control cycles, and the current differential is used. ; PS-side recursive least squares forgetting factor Low-pass filter coefficients and the forgetting coefficient of the recursive information matrix The physical time constant is uniformly set as follows: = , = , = , where x∈{mon,ident}.

4. The method according to claim 1, characterized in that, The four credibility judgments mentioned in S3 include: Physical rationality judgment: When ≤ ≤ and ≤ ≤ The timing and physical location are reasonable and effective; Motivation sufficiency assessment: Construct a recursive information matrix after normalizing the regression matrix by columns. ,when =det( ) / ([tr( ) + )≥ And tr( )≥ Time-based incentives are effective; Continuous convergence test: The count is incremented when the update amounts of adjacent Candidates do not exceed a stable threshold; otherwise, it is reset to zero. The count is determined by physical time. The conversion threshold is stable and effective at the time of setting; Model residual judgment: Use the Candidate from the previous time step to perform a one-step pre-update residual calculation on the new sample of the kth group. The residual is set to valid when the average residual window does not exceed the threshold.

5. The method according to claim 1, characterized in that, The submission threshold mentioned in S4 is: | |≥ or | |≥ When the IDENTIFY state is satisfied, =1. When at least one parameter change reaches the submission threshold and there are no unprocessed Shadows, set =1, freeze , , , Form a complete parameter package; do not commit Shadow in HOLD state.

6. The method according to claim 1, characterized in that, The model mismatch described in S1 specifically satisfies the entry and exit conditions as follows: if the normalized model mismatch index... (k) continuous Each high-speed control cycle is no less than the entry threshold. , place =1 and enter IDENTIFY from HOLD; if (k) continuous Each control cycle should not exceed the exit threshold. If there are no locked or currently being processed Shadows, exit IDENTIFY and restore HOLD; < .

7. The method according to claim 1, characterized in that, The condition for Shadow to pass the initial VERIFY in S5 is: =1; This is the new version after Active; Age≤ ; ≤ + ;and ≤ ; in, For the allowable noise fluctuation amount, This represents the upper limit of the Shadow prediction error.

8. The method according to claim 1, characterized in that, According to S6 The specific method of takeover is as follows: Set 0< < < ;when < Direct takeover is used at this time; when ≤ < Fast-Ramp rapid ramp connectors were used at that time; when ≤ < Slow-Ramp connectors are used at this time; when ≥ When prompted, proceed to Extended-Verify. The takeover method is fixed once it enters the COMMIT process; The Extended-Verify includes: when the prediction quality of the initial VERIFY meets the requirements but ≥ At that time, continue to maintain active real control, in The maximum prediction difference was recalculated over each period. and the mean of prediction error , ; Only when < , =1、Age≤ , ≤ + and ≤ Only allow entry into COMMIT with Slow-Ramp when the condition is met; otherwise, discard Shadow and return IDENTIFY.

9. The method according to claim 8, characterized in that, The Fast-Ramp and Slow-Ramp limit rates of change are updated as follows: ， ； Where m∈{fast,slow}, and > >0, > >0; the robust cost of the dual-model approach during COMMIT is: ; Among them, the dual-model overcurrent hard constraint requires that the predicted current of both Active and Shadow models does not exceed [a certain value]. ; If all candidate states violate the hard constraints, the current protection branch is entered; the takeover completion condition is | |≤ And| |≤ After satisfying the condition, set Active to exactly equal Shadow, and update. = juxtaposition =COMPLETE.

10. The method according to claim 1, characterized in that, The four-phase request / ack handshake includes: idle state =0 and =0; PS: First write the complete parameter package, then set... =1 and keep the data stable; PL only =1、 =0, No Shadows to be processed and parameter package version is newer than At that time, the entire system is latched as Shadow at the boundary of the security control cycle; After processing, PL outputs juxtaposition =1; PS detects ack and clears ; PL clears after detecting request zeroing. ; At least distinguish between ACCEPT, , and .

Images