Recall hard constraint double-threshold screening and sorting method for permanent magnet synchronous motor parameter design

Abstract

The invention provides a recall hard constraint dual-threshold screening and sorting method for permanent magnet synchronous motor parameter design, and belongs to the technical field of motor design and optimization. The method comprises the steps that a sample set is obtained based on parameterized finite element modeling; generating a feasibility label according to the motor engineering constraint; in the training stage, a classifier is adopted, and boundary sensitive weight and probability calibration are adopted; on the premise of recalling the hard constraint on the verification set, selecting a first threshold value by utilizing a criterion; a second threshold value is generated according to the self-adaptive quantile on the left side, and a high-confidence straight-through area and a candidate pool are formed; and for samples in the straight-through region or the candidate pool, constructing local preference pairs based on normalized multiple targets by using a stage II pairwise sorting model, training the local preference pairs, outputting ranking scores, and giving a Top-k priority simulation list. The method is controllable in missed judgment risk, adjustable in candidate set scale and high in optimal hit rate, and is suitable for motor parameter design, batch parameter screening and rapid parameter optimization.

Description

TECHNICAL FIELD

[0001] The application relates to a recall hard constraint double threshold screening and sorting method for parameter design of a permanent magnet synchronous motor, and belongs to the technical field of motor design and optimization. BACKGROUND

[0002] Permanent magnet synchronous motors are widely used in electric vehicles, electric tools and household compressor scenes. The electromagnetic performance of such motors is affected by the rotor and stator size, slot and pole matching, magnet size and arrangement, air gap, lamination thickness, winding parameters and other multi-variable coupling. Performance indicators usually include average electromagnetic torque, torque ripple / tooth slot torque, total harmonic distortion (THD), efficiency / copper and iron loss, temperature rise and manufacturability constraints. In engineering, parameterized modeling combined with finite element (such as Maxwell) simulation is often used to evaluate the design point, but single-point simulation is time-consuming and costly for batch scanning, and the proportion of "feasible solutions" is often very low under multi-index constraints.

[0003] The typical process is: sampling in the design space according to the design of experiment → electromagnetic simulation → building a surrogate model based on the data → searching for the optimal structure in the surrogate or simulation loop using an optimizer. The motor optimization process of the prior art does not set a constraint judgment module, and cannot determine whether the sample is 'qualified / feasible'. The lack of this process exists the following common problems:

[0004] Sparse feasible region and extremely unbalanced categories: under multi-constraint parallel judgment, the proportion of samples meeting all indicators is often significantly less than 10%, or even lower. Direct training of a classifier can easily be biased towards the "infeasible" class, resulting in low recall rate and missing high-quality design points, limiting the subsequent optimization space;

[0005] Sampling and computing resources cannot be allocated according to value: the prior art generally "cuts off" at the same threshold to retain or discard samples, lacks the concept of "candidate pool" and resource grading strategy, and cannot guarantee high recall while preferentially consuming limited simulation / experiment quotas on more valuable design points through the ordering of suboptimal samples;

[0006] Limitations of "direct optimal" instead of "order optimal": multi-objective optimization often weights or constrains several indicators and then performs global search, but within the "feasible" samples, how to stably sort and select the top-k according to the comprehensive performance, the existing solutions mostly use point estimation of single-point regression, which is easily affected by scale difference and noise, lacks learning sorting mechanism of pairwise comparison, resulting in insufficient ranking robustness and hit rate. The existing improvement direction is extremely insufficient: some researches jointly use "feasibility classification + multi-objective optimization", but lack the hierarchical strategy of "high recall hard constraint + candidate pool + sorting learning"; another work introduces Bayesian optimization to control the simulation budget, but when the constraint feasibility is extremely sparse, it is easy to converge early or insufficient exploration. SUMMARY

[0007] The purpose of the present application is to solve the problems existing in the prior art, and further provide a recall hard constraint double threshold screening and sorting method for permanent magnet synchronous motor parameter design. The method can explicitly impose high recall hard constraint and provide a transferable threshold in the feasibility identification stage through a data-driven method constructed around the motor design decision target; form a "candidate pool" in the resource-limited scene and sort the feasible samples robustly to improve the Top-k hit rate; calibrate the probability after the event to ensure stable decision-making across batches; and can be connected with the existing simulation / test closed loop to continuously update the model and threshold. The above needs provide a real background and technical motivation for the present application.

[0008] The purpose of the present application is realized by the following technical solutions:

[0009] A recall hard constraint double threshold screening and sorting method for permanent magnet synchronous motor parameter design, comprising the following steps:

[0010] Step one, data acquisition and standardization:

[0011] Generating data samples based on the parameterized motor model through experimental design or adaptive sampling Obtaining permanent magnet synchronous motor performance indicators by finite element simulation , performing numerical standardization on , and dimensionless or interval scaling on ;

[0012] Step two, engineering constraint and label generation:

[0013] According to the design red line, the threshold of the permanent magnet synchronous motor performance indicators in step one is set to generate a feasibility label y;

[0014] Step three, probability estimation and calibration:

[0015] Selecting a probability classifier , combining class imbalance and boundary sensitivity weight training through the threshold obtained in step two, the probability classifier first outputs a score , and obtains a feasible probability , is the feature vector of the motor scheme;

[0016] Step four, threshold strategy determination:

[0017] Determine the high confidence straight-through threshold on an independent validation set and the candidate pool threshold in step three ;

[0018] Step five, sample diversion:

[0019] In left search low threshold , according to the fourth step has been determined and threshold calibration probability The sample is divided into three domains: form three domains, S5A straight-through area: , direct determination of feasible; S5B candidate pool: , reserved for review and sorting; S5C discard area: , not into the subsequent expensive simulation;

[0020] Step six, high confidence learning sorting:

[0021] In step five, the construction of the straight-through area and candidate pool sub-domain sorting supervision signal:

[0022] Step seven, closed-loop feedback and data back:

[0023] The new simulation or test results are fed back to the training data set, and the threshold is periodically calibrated and retrained according to steps three to six; when the motor working condition platform or red line is changed, the threshold is reselected and ;

[0024] Step eight, output decision results.

[0025] Compared with the prior art, the beneficial effects of the present application are as follows:

[0026] The present application is based on parameterized finite element modeling to obtain the sample set; according to the motor engineering constraints (output torque lower limit, tooth slot torque upper limit, torque fluctuation upper limit, total harmonic distortion upper limit) to generate feasibility label; training stage - classifier, and using boundary sensitive weight and probability calibration; on the premise of recall hard constraint on the validation set, the first threshold is selected by using criterion The second threshold is generated by the adaptive quantile on the left side , forming a "high confidence pass zone" and a "candidate pool"; for samples in the pass zone or the candidate pool, a stage two pair-wise ranking model (RankSVM or equivalent model) is used to construct local preference pairs based on normalized multi-objectives (torque, THD, cogging, ripple) and is trained, and a ranking score is outputted to give a Top-k priority simulation list. The method of the present application significantly reduces simulation overhead while maintaining high recall, and realizes stable priority ranking in high confidence samples. Unlike existing single-stage classification or pure threshold / regression agents, the present application integrates recall hard constraints + double threshold shunting + local pair-wise ranking into an integrated process, and the technical effects are: controllable misjudgment risk, adjustable candidate set size, improved preferred hit rate, suitable for motor parameter design, batch screening parameters and rapid optimization parameters.

[0027] The present application can explicitly impose a high recall hard constraint in the feasibility identification stage and provide a transferable threshold; form a "candidate pool" in a resource-limited scenario and robustly rank the feasible samples to improve the Top-k hit rate; perform posterior calibration on the probability to ensure stable decision-making across batches; and can be connected with existing simulation / test closed loops to continuously update the model and threshold.

[0028] On a typical data set, the "total set" recall rate of the present application is ≥0.98, the pass zone recall is ≥0.97 and the precision is ≥0.55; the Top-10 of the pass zone is ≈0.40, and the Top-20 is ≈0.50. Compared with single threshold or "regression agent + empirical threshold", the Top-k hit is improved by about 1.5-2.0 times under the same budget, and the number of simulation points is reduced by about 40%-60%, reducing the research and development cost and shortening the iteration cycle without sacrificing feasible coverage.

[0029] The present application divides the overall samples into three regions with calibrated probability + double threshold, first determines on the validation set with constraint optimization to suppress misjudgment from the mechanism, and the actual pass zone recall is ≥0.95, which can still guarantee "high recall" in extremely unbalanced scenarios; then selects in the to control the candidate pool proportion to 20%-40%, forming a pass / candidate / discard three-domain, and only the candidate pool is reviewed and ranked, therefore, the simulation or test overhead is focused on value, saving about 30% of the second stage calculation and manual review amount, at the same time, the samples in the pass zone can directly enter the subsequent design process, significantly shortening the iteration cycle; the learning ranking model is introduced in the candidate and pass zone (weighted with the boundary, learning is "relative advantages and disadvantages" rather than direct regression of absolute indicators, avoiding the mismatch of "direct optimal" in cross-working condition and cross-machine type, and the Top-k hit rate can reach 0.4-0.5 when k=10~20, with less review quota to preferentially hit high-quality schemes; the supporting closed-loop feedback re-trains and re-selects the threshold with the newly added simulation / test results, so that the working point Auditable, reproducible and steady-state data improvement, high recall, value-based resource allocation, and stable ranking. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 A flowchart of a recall hard constraint double threshold screening and ranking method for permanent magnet synchronous motor parameter design.

[0031] Figure 2 A test set Precision-Recall curve chart of the present application.

[0032] Figure 3 A test set and a threshold line containing a feasible probability histogram of the present application.

[0033] Figure 4 A Pass∪Pool selection set confusion matrix chart of the present application.

[0034] Figure 5 A confusion matrix chart of the "high confidence straight-through area" of the present application.

[0035] Figure 6 A "candidate pool" confusion matrix chart of the present application.

[0036] Figure 7 A Top-k hit rate curve chart of the present application. DETAILED DESCRIPTION

[0037] The present application will be further described in detail below with reference to the accompanying drawings: The present embodiment is implemented on the premise of the technical solution of the present application, and a detailed implementation manner is given, but the protection scope of the present application is not limited to the following embodiments.

[0038] As shown in Figure 1 , a recall hard constraint double threshold screening and ranking method for permanent magnet synchronous motor parameter design involved in the present embodiment includes:

[0039] S1. Data acquisition and standardization: based on the parameterized motor model (stator tooth width, magnet thickness, pole arc ratio, air gap, rotor inner diameter, winding parameter, etc.), data samples are generated through experimental design or adaptive sampling: Performance indicators are obtained through finite element simulation , are expressed as feature matrices is a matrix composed of real numbers , wherein is the number of samples, i.e. the number of rows, is the number of features, i.e. the number of columns. , wherein is torque, is total harmonic distortion, for cogging torque, for torque ripple. The is numerically standardized, and the is dimensionless or interval scaled.

[0040] S2. Engineering constraints and label generation: Set thresholds according to design redlines:

[0041]

[0042] Either fixed thresholds or quantile adaptive thresholds can be used. Generate a feasibility label y, if

[0043]

[0044] y = 1, otherwise y = 0. Here, y = 1 means feasible, y = 0 means infeasible.

[0045] S3. Probability estimation and calibration: Choose a probabilistic classifier (gradient boosting, random forest, logistic regression, light-weight neural network, etc. In the example, GBDT is chosen), combined with class imbalance and boundary sensitive weighting training. The sample weight and feasibility label are both based on the redline in S2, and the weight can be constructed according to the minimum margin:

[0046]

[0047] where, is the boundary margin, controlling the weight amplitude (1-2 is taken in practice); is the value of the natural exponential function; is the minimum normalized margin; is the scale factor of the torque, used to normalize different dimensions; is the scale factor of the total harmonic distortion; is the scale factor of the cogging torque; is the scale factor of the torque ripple;

[0048] “Pay more attention to samples near the boundary / easy to make mistakes” when training the probabilistic classifier. Here, is the scale, . The probabilistic classifier first outputs the score , and then the output score is isotonic or Platt calibrated to get a reliable feasibility probability:

[0049]

[0050] where, is the feature vector of the motor scheme, i.e. the input of a sample. In this data, the feature vector of the motor scheme , a set of independent variables expressed as motor geometry, material, winding or control settings, etc. always refer to the same set of motor structure parameter vectors.

[0051] S4. Threshold strategy determination: to reduce false positives and control the size of the candidate pool while ensuring high recall, determine a high-confidence pass-through threshold on the independent validation set with candidate pool threshold .

[0052] Set the lower limit of recall and precision on the independent validation set , Pick one from all thresholds on the validation set as the high threshold , which must meet the following conditions:

[0053]

[0054] That is and among these qualified thresholds, the largest.

[0055] The comprehensive index uses score as follows:

[0056]

[0057] Note: Determine the value through the comprehensive index, in the example more emphasis on precision, more emphasis on recall; preferred .

[0058] Determine the lower limit of the candidate pool on the validation set in the interval, which is used to "determine the lower boundary of the candidate pool". The idea is to make the candidate pool size controllable and as small as possible without sacrificing quality.

[0059] where, is the positive judgment threshold for the calibrated probability , both calculated on the independent validation set; represents , is the precision, of all the feasible solutions judged by the model, how many are really feasible, used to control the false positive rate and review cost; represents , is the recall rate, of all the real feasible solutions, how many are found by the model, used to control the false negative; is the control recall weight.

[0060] S5. Sample triage (pass / fail / candidate): in left search low threshold (determined by right-side quantile of calibration probability or by "selected proportion" constraint), forming three regions: S5A pass region: , directly determined as feasible; S5B candidate pool: , reserved for review and ranking; S5C discard region: , not entering subsequent expensive simulation. The determination satisfies: (1) the recall of the total selected set (pass and candidate) is not lower than the target; (2) the size of the candidate pool is within the target interval (e.g. 20%-40% of the total); (3) the accuracy of the total selected set is not lower than the preset lower limit, which can be adjusted within the interval (0.55, 0.65) according to the project; if the false positive cost is high, the preset lower limit can be increased to the interval (0.65, 0.75).

[0061] S6. High-confidence learning ranking: construct ranking supervision signals within the pass and candidate sub-regions: (1) target fusion: normalize the output , define a weighted comprehensive advantage (or use Pareto dominance relationship). (2) Pairwise samples: if or is superior in most targets, then is positive; to enhance stability, only sample within the neighborhood and assign boundary weights:

[0062]

[0063] (3) Train ranker: use RankSVM, Pairwise-LR, LambdaMART or RankNet to train the score function . (4) Inference and Top-k: calculate for the pass or pass ∪ candidate set, and output the top-k as the priority simulation list.

[0064] S7. Closed-loop feedback and data back: feedback the new simulation or test results to the training data set, the validation set is kept independent and only used for threshold selection and evaluation; periodically calibrate and retrain according to S3-S6; when the motor working condition platform or red line changes, reselect the threshold to ensure portability and stability. The motor working condition platform refers to significant changes in slot-pole ratio, material system, cooling structure, load / speed working condition, etc.

[0065] S8. Output decision results: output includes: operating point PR curve and operating point coordinates , probability histogram , three-region confusion matrix, pass region ranking scatter plot, Top-k hit curve, priority list. Among them, for That is, at a high threshold Precision at the location; for That is, at a high threshold Recall rate at each location.

[0066] In this invention, a recall hard constraint dual threshold screening and sorting system for permanent magnet synchronous motor parameter design includes: a data generation module, a label and constraint module, a stage-screening module, a dual threshold generation module, a stage-two sorting module, and a result output module.

[0067] The system operates in the following order: "Data → Label → Probability → Threshold → Triage → Sorting → Feedback → Output": First, the data generation module generates sample parameter vectors. The labeling and constraint module generates a feasibility label y based on the engineering boundary line; the phase one screening module trains a probability classifier and obtains the result using Platt / Isotonic calibration. The dual threshold generation module searches for thresholds on independent validation sets. : take first make Maximum and satisfy And then Internal selection The sample splitting module divides the samples into three regions based on the target range (e.g., 20%–40%) of the candidate pool: For direct access area, Enter the candidate pool, Discard; the Phase 2 sorting module ranks candidates within the pass zone and candidate pool according to normalized multi-indicator merits. Constructing nearest neighbor directed pairs The model is then trained by assigning boundary weights to obtain ranking scores. The module outputs Top-k hits; the closed-loop feedback and data feedback module will add simulation / experimental results to the training set, periodically repeating probability training and threshold updates; the results output module will uniformly generate PR curves and operating points. Probability histograms, three-domain confusion matrices, sorting scatter plots, Top-k curves, and priority lists are used to form an iterative and adaptive closed loop.

[0068] like Figure 2 The image shows the Precision-Recall (PR) curve (test set). Operating points are marked on the curve, corresponding to those satisfying... and of The selection criteria are given in the figure caption. With threshold.

[0069] like Figure 3The figure shows the feasible probability histogram (test set). The probability axis is labeled simultaneously With , showing a three-domain division: among them, : high-confidence straight-through area; : candidate pool; : discard area. It is explained that it can be determined according to quantile or candidate size constraints, and it is guaranteed .

[0070] As shown in Figure 4 , it is the confusion matrix diagram of the "selection set (Pass∪Pool)". The four quadrants are labeled TN, FP, FN, and TP; this matrix reflects the end-to-end effect of "recall hard constraint + candidate strategy" (overall recall, overall precision).

[0071] As shown in Figure 5 , it is the confusion matrix diagram of the "high-confidence straight-through area". It is used to explain the effect of the straight-through decision under high recall (reducing FN) and controlled precision (tolerating certain FP), and to reflect the core technical benefits of the recall hard constraint.

[0072] As shown in Figure 6 , it is the "candidate pool (Pool-only)" confusion matrix diagram. It is used to show the objective phenomenon that there are fewer true feasible samples in the candidate pool when the threshold is close to ; as an example or comparative example diagram, it is not placed in the core effect comparison.

[0073] As shown in Figure 7 , it is the Top-k hit rate curve diagram. On the test set, the hit rate of the ranking module is about 0.40 when k=5, 10, 15, and increases to about 0.50 when k=20, indicating that the present application can output a priority simulation list with high credibility, thereby reducing the number of tests and simulations.

[0074] The above is only a preferred specific embodiment of the present application, these specific embodiments are different implementations based on the overall concept of the present application, and the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A recall hard constraint double threshold screening and sorting method for permanent magnet synchronous motor parameter design, characterized in that, Comprising the following steps: Step one, data acquisition and standardization: Generating data samples based on a parameterized motor model by design of experiments or adaptive sampling Obtaining performance indicators of permanent magnet synchronous motor by finite element simulation Numerical standardization is performed on Dimensionless or interval scaling is performed on ​ Step two, engineering constraints and label generation: According to the design red line, set the threshold value of the permanent magnet synchronous motor performance index in step one, generate a feasibility label y; Step three, probability estimation and calibration: Selecting a probabilistic classifier The probabilistic classifier outputs scores through the threshold obtained in step two, combined with class imbalance and boundary sensitive weights training Obtain feasible probabilities through isotonic or Platt calibration , is the feature vector for the motor scheme Step four, threshold strategy determination: Determining a high-confidence pass threshold on an independent validation set As in step three Candidate pool threshold ; Step five, sample diversion: In left search low threshold , according to the fourth step has been determined and threshold calibration probability The sample is divided into three domains: the formation of three domains, S5A straight through area: , direct determination of feasible; S5B candidate pool: , reserved for review and sorting; S5C discard area: , not to follow the subsequent expensive simulation; Step six, high confidence learning ranking: Construct a ranking supervision signal in the straight-through area and candidate pool sub-domain formed in step five: Step seven, closed-loop feedback and data back: The newly added simulation or test result is fed back to the training data set, and periodic calibration and retraining are performed according to steps three to six; when the motor working condition platform or the red line is changed, the threshold is reselected and ; Step eight, output decision result.

2. The recall hard constraint double threshold screening and sorting method for parameter design of permanent magnet synchronous motor according to claim 1, characterized in that, the data samples in step one the permanent magnet synchronous motor performance index wherein, is torque, is total harmonic distortion, is cogging torque, is torque ripple.

3. The recall hard constraint double threshold screening and sorting method for parameter design of permanent magnet synchronous motor according to claim 1, characterized in that, The threshold value of the permanent magnet synchronous motor performance index in step two is set by using a fixed threshold value or a quantile adaptive threshold value method, and the threshold value of the permanent magnet synchronous motor performance index is as follows: Generate a feasibility label y, if y=1, otherwise y=0; y=1 indicates feasibility, and y=0 indicates infeasibility.

4. The recall hard constraint double threshold screening and sorting method for parameter design of permanent magnet synchronous motor according to claim 1, characterized in that, the combining class imbalance with boundary sensitive weight training in step three by minimum margin construction, weights the formula is as follows: wherein, is a boundary placement amount, control weight amplitude; is a value of a natural exponential function; is a minimum normalization margin; is a scale factor for torque, for normalizing different dimensions; is a scale factor for total harmonic distortion; is a scale factor for cogging torque; is a scale factor for torque ripple; In training the probability classifier, more attention is paid to samples near the boundary / prone to make mistakes; wherein, For scale, .

5. The recall hard constraint double threshold screening and sorting method for parameter design of permanent magnet synchronous motor according to claim 1, characterized in that, Determining a high confidence pass-through threshold in step four The steps are as follows: Set a lower bound on recall and precision on an independent validation set , Pick one of the thresholds from the validation set as the high threshold , requiring that it simultaneously satisfy the following conditions: The composite index is calculated using scores, The scores are as follows: Using validation sets The lower boundary of the candidate pool is determined within the specified interval. where t is the decision threshold for the calibration probability of the calibration probability, and the decision threshold t, are both computed on an independent validation set; denotes , the precision, denotes , the recall, and the control recall weight.

6. The recall hard constraint double threshold screening and sorting method for parameter design of permanent magnet synchronous motor according to claim 1, characterized in that, The determination in step five is performed by using the right side quantile according to the calibration probability or by using the selected proportion constraint method. Left side search low threshold The determination is performed by using the right side quantile according to the calibration probability or by using the selected proportion constraint method. The determination satisfies: (1) the recall of the pass-through area and the candidate pool is not lower than the target; (2) the size of the candidate pool is located in the target interval; (3) the precision of the pass-through area and the candidate pool is not lower than the preset lower limit.

7. The recall hard constraint double threshold screening and sorting method for parameter design of permanent magnet synchronous motor according to claim 1, characterized in that, The step of constructing a ranking supervision signal in step six is as follows: Step six, target fusion: the output is normalized, defining a weighted aggregate merit ; Step six two, pair-wise samples: if the overall merit obtained in step six one is or superior on most targets, then is positive; for increased stability, sample only within the immediate neighborhood and assign boundary weights as follows: wherein, are feature vectors of the i-th sample, respectively; are feature vectors of the i-th sample, respectively; are feasible probabilities of the i-th sample, respectively; is a pair-wise weight function for weighting the ranking loss; is a stability term to prevent the denominator from being zero; Step six three, training ranker: train score function using RankSVM, Pairwise-LR, LambdaMART or RankNet ; Step six four, reasoning and top-k: compute the score function in step six three on the straight or straight ∪ candidate set , output the top-k as the prioritized simulation list.

8. The recall hard constraint double threshold screening and sorting method for parameter design of permanent magnet synchronous motor according to claim 1, characterized in that, The output decision result in step eight includes: a working point PR curve and working point coordinates thereof annotation, probability histogram a three-domain confusion matrix, a straight-through domain ranking scatter plot, a Top-k hit curve, and a priority list; wherein, is the precision at high threshold ; is the recall at high threshold .

Images