A plug-in inductance winding process control method based on production parameter dynamic compensation

Abstract

The application discloses a plug-in inductance winding process control method based on production parameter dynamic compensation, aiming at the problems of unstable inductance and uneven structure in the traditional process, and proposes a closed-loop compensation control strategy with a physical field evolution model as the core. The method first constructs an ideal physical field benchmark model according to target inductance parameters and structure, describes the evolution law of characteristics such as magnetic field, electric field or impedance in the ideal state. In the actual winding process, the physical field signal is collected through the sensor, and the wire tension, spindle speed and other production parameters are obtained to generate a physical field distribution snapshot of the current winding state. The system compares the extracted actual characteristics with the benchmark model, identifies the dominant disturbance source causing the deviation in combination with real-time production parameters, and outputs a disturbance type label carrying causal information. According to the label and the production parameter change trend, the compensation instruction is dynamically generated and executed to fine-tune the tension, wire arrangement track or density, and ensure the consistency and accuracy of the inductance finished product.

Description

Technical Field

[0001] This invention relates to the field of inductor process control technology, and in particular to a method for controlling the winding process of plug-in inductors based on dynamic compensation of production parameters. Background Technology

[0002] In existing technologies, the winding process of plug-in inductors typically employs fixed parameter settings or a tension-based closed-loop control method. The inductance of the finished inductor is controlled by adjusting the wire tension, the number of turns, and the winding speed. Some high-end equipment incorporates tension monitoring and feedback mechanisms to correct local tension anomalies, thereby improving the repeatability of the inductance. However, these methods largely rely on empirical formulas and static settings, primarily using the number of turns and tension as input variables, making it difficult to respond promptly to dynamic disturbances in the inductance evolution process during actual winding.

[0003] The main problem with existing technologies is that their control logic typically lacks direct sensing means for changes in the electromagnetic field structure inside the inductor, making it impossible to assess in real time the changes in spatial electromagnetic distribution and structural consistency deviations during the winding process. Furthermore, most current systems do not model multi-source disturbances such as temperature and humidity fluctuations, wire diameter variations, and equipment vibrations during the production process as inputs. This leads to problems such as inductance drift, non-uniform distribution, and local interlayer overlap that easily occur under material batch variations or environmental condition fluctuations, thus affecting product consistency and yield.

[0004] Based on the above background, there is an urgent need to propose a plug-in inductor winding control method that can combine production parameters and has dynamic compensation capabilities, so as to improve control accuracy and process adaptability. Summary of the Invention

[0005] This application provides a method for controlling the winding process of plug-in inductors based on dynamic compensation of production parameters, so as to improve control accuracy.

[0006] This application provides a method for controlling the winding process of plug-in inductors based on dynamic compensation of production parameters, including:

[0007] Based on the target inductor parameters and plug-in inductor structure, a benchmark model describing the evolution characteristics of the ideal physical field is constructed. The benchmark model is used to characterize the physical field change law of the core region of the inductor with the number of turns or time under ideal winding conditions.

[0008] A non-contact sensor array deployed in the winding area is used to collect physical field signals, including magnetic field, electric field or impedance, and combined with the sampling data of current production parameters, including wire tension, spindle speed, wire pitch, ambient temperature and humidity, to generate a snapshot of the physical field distribution at the current level and extract evolutionary feature parameters that match the benchmark model.

[0009] The extracted evolutionary feature parameters are compared with the ideal feature parameters under the same winding progress in the benchmark model. Combined with the collected production parameters, a mapping relationship between physical field deviation and production disturbance is established. The dominant disturbance source causing the deviation is identified through the preset deviation attribution mechanism, and the disturbance type label is obtained.

[0010] Based on the disturbance type label and the corresponding production parameter change trend, a compensation instruction for fine-tuning the winding behavior is generated. The compensation instruction includes: adjusting the tension curve of the wire feeding mechanism to offset the influence of tension disturbance, correcting the wire laying rhythm to alleviate displacement deviation, optimizing the wire laying path to restore distribution uniformity, or adjusting the wire laying density to compensate for wire diameter changes; the compensation instruction is executed to compensate for the winding behavior.

[0011] The beneficial effects of the technical solution provided in this application include:

[0012] (1) By constructing a benchmark model of the evolution characteristics of ideal physical fields, it is possible to provide an accurate reference for electromagnetic field behavior in the inductor winding process, improve the ability to judge the structural consistency in the winding process, and provide a theoretical basis for subsequent deviation control. (2) Using a non-contact sensor array to collect multi-dimensional physical field signals in real time, and combining them with production parameters such as tension, rotation speed, temperature and humidity, it is possible to comprehensively reflect the actual winding state and significantly enhance the process's ability to perceive and respond to disturbance changes. (3) By establishing a causal mapping relationship between physical field deviation and production disturbance, and by using disturbance type labels to guide compensation control, the compensation measures are targeted and interpretable, overcoming the problem of inaccuracy in traditional empirical parameter adjustment methods. (4) The generated compensation instructions can achieve multi-dimensional fine-tuning of tension, winding rhythm, path and density, thereby achieving a dual improvement in sensitivity accuracy control and finished product consistency without changing the mechanical structure. Attached Figure Description

[0013] Figure 1 This is a flowchart of a plug-in inductor winding process control method based on dynamic compensation of production parameters provided in the first embodiment of this application. Detailed Implementation

[0014] Many specific details are set forth in the following description to provide a full understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application; therefore, this application is not limited to the specific embodiments disclosed below.

[0015] The first embodiment of this application provides a method for controlling the winding process of plug-in inductors based on dynamic compensation of production parameters. Please refer to... Figure 1 This figure is a schematic diagram of the first embodiment of this application. The following is in conjunction with... Figure 1The first embodiment of this application provides a detailed description of a method for controlling the winding process of plug-in inductors based on dynamic compensation of production parameters.

[0016] Step S101: In the initial stage of the winding task, a benchmark model describing the evolution characteristics of the ideal physical field is constructed based on the target inductor parameters and the plug-in inductor structure. The benchmark model is used to characterize the physical field change law of the core region of the inductor under ideal winding conditions with the number of turns or time, and serves as a reference for subsequent winding deviation judgment.

[0017] In the initial stage of the winding task, in order to achieve real-time monitoring and compensation of dynamic deviations during the winding of plug-in inductors, it is necessary to first construct a benchmark model of the evolution characteristics of an ideal physical field. This model aims to characterize the changing trends of key physical quantities such as magnetic field, electric field, or impedance in the core region of the plug-in inductor under ideal winding conditions, as the number of winding coils or time progresses, providing an accurate and quantifiable reference for subsequent feature extraction, deviation judgment, and compensation control.

[0018] The input information required to build this benchmark model includes performance parameters of the target inductor, such as its inductance value, tolerance, and operating frequency range, as well as structural information such as the inductor's frame type, core size, wire diameter, insulation thickness, and wiring method. These parameters can be obtained from the inductor product design drawings. In addition, electromagnetic property data of relevant materials are required, such as the initial permeability and saturation flux density of the core, the resistivity of the wire, and the dielectric constant of the dielectric. This data can be obtained from material supplier specifications or experimental measurements. To simulate the operating conditions in a real winding environment, initial boundary conditions should also be introduced, such as the current excitation method, winding speed setting, and assumptions about stable ambient temperature, to ensure that the modeling process is consistent with real operating conditions.

[0019] The model can be implemented in two ways. The first is numerical simulation, which involves constructing a three-dimensional geometric model of the inductor to be wound on an electromagnetic field simulation platform. Combining the aforementioned input parameters and physical constraints, a winding stepping strategy is set, allowing the simulation system to simulate the inductor winding process turn by turn. At each stepping node, the system calculates the physical field distribution inside and around the inductor, extracting representative characteristic quantities such as the magnetic field strength at the inductor center, field gradient, magnetic flux density per unit volume, or impedance phase angle change, and recording their continuous evolution trend over time or the number of turns. The second method is sample inversion modeling, which uses several sets of stable plug-in inductor samples as a basis. Non-contact testing equipment such as miniature Hall sensors, electric field probes, or impedance analyzers is used to collect physical field response data at different winding levels of the samples. The correspondence between the inductor's inductance and the physical field is established through timestamps or turn number markers. Then, ideal evolution curves are constructed using algorithms such as data fitting and principal component analysis.

[0020] Regardless of whether simulation or inversion methods are used, the model's output should be structured and reusable. The output primarily includes a sequence of expected physical field eigenvalues ​​at multiple winding progress nodes, as well as alignment criteria and difference quantification methods between actual measured data and baseline eigenvalues. For ease of subsequent use, the model can be saved as a database, function expression, or normalized table, and a unified data interface should be provided for retrieval and comparison. To improve the model's adaptability and robustness, it is recommended to perform noise filtering, standardization, and spatial interpolation optimization on the original data during the modeling process to ensure good versatility across different material batches and environmental conditions.

[0021] By constructing this benchmark model, the system can establish a complete set of desired physical field trajectories in the early stages of winding, and use it as the basis for deviation judgment and compensation control of the actual winding state in subsequent processes. This significantly enhances the precision control capability and environmental adaptability of the plug-in inductor winding process, providing key support for the dynamic compensation mechanism of this invention.

[0022] Furthermore, the benchmark model is established through electromagnetic field finite element simulation. The simulation process uses the magnetic flux density distribution, magnetic field strength distribution, and local electric field response in the capacitive coupling surface after each turn of wire in the plug-in inductor structure is completed as reference indicators. The simulation input includes target inductor parameters, plug-in inductor structure, wire arrangement, and introduces material permeability, saturation magnetic flux density, and dielectric constant of insulating material as modeling parameters.

[0023] During the simulation process, multiple feature points are extracted from the impedance frequency response curve, including the maximum phase difference point, the minimum impedance point, the impedance abrupt change inflection point, and the critical frequency band boundary. This set of frequency response features is used as an auxiliary reference parameter to describe the winding evolution process under ideal conditions. Together with the magnetic flux density distribution and electric field distribution, they are used to form the physical field evolution path of the benchmark model.

[0024] The output of the benchmark model is in the form of a structured vector. The structured vector includes a winding turn index, an evolution feature parameter index corresponding to each turn, and a tolerance range field corresponding to each feature parameter. The structured vector is aligned and compared turn by turn with the real-time extracted evolution feature parameters during the actual winding of the plug-in inductor.

[0025] The structured vector serves as the sole reference for comparing the extracted evolutionary feature parameters with the ideal feature parameters under the same winding progress in the benchmark model in subsequent steps, and is consistent with the deviation input data structure required to identify the dominant disturbance source causing the deviation through a preset deviation attribution mechanism.

[0026] To ensure the baseline model possesses sufficient physical realism and control precision during the winding process of plug-in inductors, it is constructed using electromagnetic field finite element simulation. In this modeling process, considering the specific structural characteristics of the plug-in inductor, the physical field evolution state after the wire winding is completed needs to be simulated turn by turn. The magnetic flux density distribution, magnetic field strength distribution, and local electric field response within the capacitive coupling surface are used as the main simulation output indicators to characterize the spatial accumulation path of the inductance and the consistency of the structural electrical performance.

[0027] Specifically, during the model initialization phase, the key performance parameters of the target inductor, such as the target inductance value, operating frequency, and expected tolerance, should be input first. This should be combined with a detailed geometric model of the plug-in inductor skeleton structure, including the core shape, center post size, winding window size, and mounting method. Regarding conductor arrangement, information such as wiring density, layer crossing method, conductor routing sequence, and interlayer insulation thickness should be defined. Subsequently, at the material level, the initial permeability and saturation flux density of the magnetic conductor, the conductivity and surface roughness of the conductor, and physical parameters such as the dielectric constant and loss factor of the insulating material should be introduced. All input parameters must have sufficient resolution to meet the requirements for detailed reproduction of electromagnetic behavior during turn-by-turn modeling.

[0028] During the modeling process, after each simulation of a winding of the conductor, the system automatically records the magnetic flux density distribution map and magnetic field vector map under the corresponding structure, and generates a capacitive coupling distribution map by combining the potential gradient field of the dielectric region between the conductors. Simultaneously, to capture the detailed characteristics of the inductor's frequency domain response, AC sweep boundary conditions are applied to the model at each stage, and the corresponding impedance frequency response curve is extracted through simulation. Several typical feature points are then extracted from these curves, including the maximum phase difference point (corresponding to behavioral changes near the resonance peak), the minimum impedance point (corresponding to conduction characteristics in the low-frequency band), the impedance abrupt change inflection point (reflecting response discontinuities caused by structural or material abrupt changes), and the critical frequency band boundary (used to describe the inductor-capacitor switching threshold of the winding structure). These frequency feature points not only reflect the full-band performance of the inductor but also reveal its structural consistency and boundary behavior, serving as an indispensable auxiliary reference in the ideal evolution path.

[0029] The data obtained from the above simulation will be uniformly encoded and output in the form of a structured vector. In this structured vector, the number of winding turns is the main index, representing the physical progress of the position of each turn of the conductor in the evolution path; each turn contains several evolution characteristic parameter indices, used to identify core indicators such as magnetic flux intensity, electric field coupling rate, and impedance change amplitude; each characteristic parameter is accompanied by a preset tolerance range field, representing the allowable physical offset limit during the actual winding process, ensuring that the subsequent comparison process has engineering applicability and judgment tolerance.

[0030] The structured vector, serving as the sole ideal reference in the entire control system, is precisely aligned spatially and temporally with the evolutionary feature parameters extracted turn by turn during the real-time winding of the plug-in inductor. The deviation is then output through methods such as difference calculation and threshold judgment. Simultaneously, the field structure and indexing rules of this structured vector remain consistent with the deviation input data structure used in the subsequent deviation attribution module, ensuring that the comparison results can be directly used as input for attribution judgment without additional format conversion, thus achieving seamless integration of the data processing chain. This approach not only improves the overall response speed of the system but also enhances the physical interpretability and semantic correspondence between the model results and real-time data, making the entire compensation control mechanism rigorous, ensuring data compatibility between modules, and practical engineering feasibility.

[0031] Furthermore, during the simulation process, multiple feature points are extracted from the impedance frequency response curve, including:

[0032] In the electromagnetic field finite element simulation process, the impedance frequency response curve of each turn of wire after winding is first simulated in the whole frequency band, and the set of characteristic points including the maximum phase difference point, the minimum impedance point, the impedance change inflection point and the critical frequency band boundary are identified in the simulation results.

[0033] In the identified set of feature points, the frequency response slope change, impedance phase angle fluctuation range and amplitude gradient curvature of each feature point are further calculated, and a frequency response perturbation sensitivity matrix is ​​constructed accordingly to characterize the response amplification factor of each feature point to structural micro-perturbations.

[0034] The frequency response perturbation sensitivity matrix is ​​coupled with the local magnetic flux density distribution map of the plug-in inductor structure for analysis. Abnormal frequency abrupt points that appear in the weakly coupled section of the structure are screened out to ensure that only effective response feature points with structural physical correlation are retained.

[0035] Frequency band clustering is performed on the retained set of valid feature points. The points are grouped and merged based on bandwidth coverage and frequency proximity. The center frequency, main response mode and representative response feature value are calculated for each cluster segment to form a frequency band representation subset.

[0036] The frequency band representation subset is used as an auxiliary reference parameter to form the baseline model. After being aligned with the magnetic flux density distribution and electric field distribution in the number of turns dimension, it is encoded into the frequency domain evolution path field in the structured vector in a unified data format, so as to compare the evolution feature parameters extracted in real time during the actual winding process of the plug-in inductor turn by turn.

[0037] In the electromagnetic field finite element simulation described in this invention, the goal is to generate an impedance-frequency response curve for the structural state formed by each turn of the conductor. This curve typically represents the amplitude and phase information of the complex impedance at multiple frequency points, and formally includes the relationship between the real and imaginary parts and frequency. The simulation employs a three-dimensional finite element method based on magnetic vector potential, obtaining the frequency response results by applying a swept current source to the model. The frequency sweep range should cover at least two orders of magnitude of bandwidth, for example, 10kHz to 10MHz, to ensure that the main operating frequency band and boundary behavior of the plug-in inductor are encompassed.

[0038] After obtaining the frequency response curves for each turn of the structure, the system begins to identify feature points. First, it detects the derivative changes of the phase curve to find the point of maximum phase difference, which is usually located near resonance and is most sensitive to the winding structure. Second, it searches for the point of minimum impedance, indicating that certain paths in the structure exhibit approximate electrical shorting behavior. Further, it analyzes the regions of abrupt gradient changes at the intersection of the real and imaginary parts, extracting impedance abrupt change inflection points to reflect changes in the electrical coupling response at the structural boundaries. Finally, it searches for flattening transition regions in the curve's terminal segment or high-frequency portion, defining them as critical frequency band boundaries. The set of all these points constitutes the first layer of frequency response feature point identification results, stored as an original feature point set containing frequency values, complex impedance values, and turn number indices.

[0039] To improve the model's sensitivity to disturbances, the system enters a second stage: calculating the local response parameters for the aforementioned feature points. During this process, a local window is constructed around each feature point at a certain frequency interval (e.g., ±5% bandwidth). The first derivative of the complex impedance curve within this window (frequency response slope), the maximum range of change in the complex phase angle (phase fluctuation amplitude), and the second derivative of the amplitude with respect to frequency (curvature) are calculated to assess the point's amplification capability for small disturbances. All calculation results form a disturbance response sensitivity matrix, with the feature point as the primary index and column vectors including four dimensions: frequency, phase fluctuation, response slope, and curvature.

[0040] Next, the system performs a spatial correspondence analysis between the frequency response perturbation sensitivity matrix and the magnetic flux density distribution map under the same structure. Here, a three-dimensional structural mesh from the electromagnetic simulation model is introduced and its position is bound to the structural segment represented by the frequency response point. This determines whether the frequency response feature point is located in a region of concentrated magnetic flux density or a region of significant magnetic field gradient change. If the magnetic flux distribution at the structural location corresponding to a feature point is flat and there is no obvious response concentration, it is considered to have no direct physical coupling with the structure, and the system will identify it as a pseudo-response point or boundary error point and discard it.

[0041] After filtering effective feature points, the system clusters the remaining feature points according to their frequency values. The clustering algorithm can employ K-means or bandwidth-aware clustering mechanisms, combining bandwidth coverage (whether the frequency spacing between points is sufficiently dense to form a continuous band) and frequency proximity (the mean and variance of the distance between points) for grouping. Within each cluster segment, the system further calculates the center frequency (geometric mean), the dominant response pattern (primarily based on the point with the largest response slope), and the representative feature value (the maximum product of response strength and stability), ultimately outputting a representation subset for that frequency band. This representation subset not only compresses the original data dimensionality but also preserves the dominant frequency region most sensitive to structural disturbances.

[0042] Finally, the frequency band representation subset is encoded in a structured format as part of the baseline model, specifically as a "frequency domain evolution path field" in a structured vector. Its data structure should include a turn index, the center frequency of the frequency band, a response index triplet (slope, phase amplitude, curvature), and a tolerance range (defining the allowable comparison error window). This frequency domain evolution path field will be aligned with the flux density distribution field and the electric field distribution field in the turn dimension to unify the timing and structural index for comparison of evolutionary characteristic parameters extracted during the actual winding process of the plug-in inductor. The comparison logic is not limited to amplitude comparison but should also include dimensions such as frequency shift deviation, phase shift, and response rate of change to comprehensively support deviation diagnosis and disturbance attribution.

[0043] Step S102: During the actual winding process of the plug-in inductor, a non-contact sensor array is deployed in the winding area to collect physical field signals such as magnetic field, electric field or impedance. Combined with real-time sampling data of current production parameters, including wire tension, spindle speed, wire pitch, and ambient temperature and humidity, a snapshot of the physical field distribution at the current level is generated. Evolutionary characteristic parameters matching the benchmark model are extracted to reflect the local gradient, coupling pattern or rate of change of the wound area.

[0044] In the actual winding process of plug-in inductors, real-time physical field information acquisition and evolutionary characteristic parameter extraction are necessary to achieve dynamic sensing of the winding state and provide fundamental support for subsequent deviation judgment. Therefore, deploying a non-contact sensor array in the winding area is a key implementation method for this step. These sensors are typically installed in multiple spatial locations near the winding station, covering the main electromagnetically active areas of the inductor winding region, and possess multi-channel, high-time-resolution sensing capabilities. Commonly used sensor types include miniature Hall effect magnetic field sensors, high-frequency electric field probe arrays, and broadband impedance sampling probes; the specific selection depends on the structural dimensions of the target inductor, its operating frequency range, and the process cycle. The aforementioned sensor array must have non-contact measurement capabilities to avoid mechanical interference to the coil formation process and must be connected to the data acquisition and control system via a high-speed sampling interface.

[0045] During the winding process, the system uses the number of winding turns or time as a synchronization index to collect data on the transient physical field distribution generated during the formation of each layer or each turn of the coil. The acquired data is then filtered, denoised, and normalized. To ensure timing consistency, this physical field acquisition process must be precisely synchronized with the spindle speed of the winding equipment, the wire feed speed, and the winding rhythm. The system typically uses encoder signals, tension closed-loop feedback signals, and physical field data acquisition commands to achieve hardware-software time binding, ensuring that the measured electromagnetic signals correspond one-to-one with the specific winding stage in the physical process. After data acquisition, the system combines the structural state of the current layer to generate a snapshot of the spatial field distribution reflecting the physical characteristics of that layer. The so-called "physical field distribution snapshot" refers to reconstructing the two-dimensional or three-dimensional electromagnetic field spectrum of the inductor winding region by using spatial interpolation and structural mapping to reconstruct the physical quantities measured at each sensing point at a specific moment, such as a magnetic induction intensity distribution map, an electric field coupling intensity map, or an equivalent impedance phase map. This snapshot not only presents the spatial physical characteristic distribution of the wound inductor region but can also be used to further extract core evolution parameters that match the benchmark model.

[0046] The evolutionary characteristic parameters corresponding to the benchmark model refer to quantitative indicators extracted from snapshots of the actual physical field distribution that accurately reflect the differences between the winding state and the evolution trend of the ideal inductor. These parameters are consistent with the predefined physical quantity paths in the benchmark model in terms of temporal sequence, spatial structure, and physical dimensions. Specifically, they include, but are not limited to: the evolution curve of the axial magnetic induction intensity at the inductor center, the local magnetic field gradient distribution, the decay rate of magnetic flux density at the winding edge, the energy density of electric field coupling between lines, the displacement trajectory of the maximum electric field response point, the rate of change of the capacitive coupling path, and the trend of impedance phase angle with the number of turns. These characteristic parameters can be extracted from the snapshot spectrum using algorithms such as spatial weighted averaging, principal component analysis, template fitting, and Fourier transform, and uniformly output as structured vectors or tensors for high-precision comparison with the benchmark model in the next step.

[0047] To improve the stability and environmental adaptability of feature extraction results, real-time acquired production parameters need to be introduced as synchronous input variables during the extraction process. These parameters include conductor tension, spindle speed, cable pitch, ambient temperature, and humidity. These production parameters directly affect the microstructure quality and electromagnetic response of the inductor. For example, tension fluctuations may cause abnormal magnetic field gradients, humidity changes may affect the dielectric properties of the insulating medium, and uneven pitch may lead to capacitive coupling imbalance. Therefore, during feature parameter extraction, the system jointly models the production parameters and field data, and sets dynamic weighting coefficients or sensitivity factors in the algorithm. This ensures that the final extracted evolutionary feature parameters possess both physical interpretability and robustness to perturbed environments.

[0048] Ultimately, the system will output a set of evolutionary characteristic parameters with clear structure, explicit physical meaning, and one-to-one correspondence with the benchmark model within each winding cycle. These parameters describe the key physical response trends of the currently wound inductor region. These parameters are not only the basis for deviation calculation and disturbance attribution, but also the direct basis for the generation of subsequent compensation control strategies, constituting the core perception layer of the plug-in inductor intelligent winding control system.

[0049] Furthermore, the sampled data combining current production parameters, including wire tension, spindle speed, wire pitch, and ambient temperature and humidity, generates a snapshot of the physical field distribution at the current level, and extracts evolutionary feature parameters that match the benchmark model, including:

[0050] The raw physical field signals collected by the non-contact sensor array are processed synchronously with the sampled data. Based on the reference physical response value corresponding to the number of winding turns, the outputs of different sensing channels are normalized to obtain standardized multi-channel equivalent physical response curves for subsequent analysis.

[0051] Based on the standardized equivalent physical response curve, combined with the arrangement coordinates of each coil of wire and the trend of magnetic flux density change, the local physical field boundary in the current inductor winding is identified, and based on this boundary information, the entire winding area is divided into multiple local areas.

[0052] Within each local region, based on the spatial distribution of the equivalent physical response curve, the magnetic field response path, capacitive coupling trajectory, and impedance phase change curve are extracted. Through fitting processing, magnetic field gradient orbits, capacitive energy orbits, and complex impedance orbits are generated, and a set of orbit vectors characterizing the spatial physical evolution trend is output.

[0053] Based on this set of orbit vectors, and combined with the aforementioned standard deviation of conductor tension, first derivative of environmental humidity change, fluctuation rate of wire pitch and rate of change of spindle speed, the dynamic threshold parameters used to adjust the comparison tolerance of each type of evolution orbit are calculated, and the allowable offset range of magnetic field orbit, response anomaly tolerance of capacitance orbit and phase error threshold of impedance orbit are reconstructed.

[0054] The set of orbital vectors adjusted by dynamic thresholds is used as the final evolutionary feature parameters.

[0055] During the actual winding process of the plug-in inductor, a non-contact sensor array continuously collects raw physical field signals covering the winding space, including raw physical quantities such as magnetic induction intensity, electric field response, capacitive coupling, and current path impedance. Simultaneously, the control system samples key production parameters such as current conductor tension, spindle speed, wire pitch, and ambient temperature and humidity at high frequency in parallel. These production parameters are real-time and subject to disturbances, and must be processed synchronously with the physical field data to ensure physical consistency and time registration accuracy in subsequent inference processes.

[0056] During synchronous processing, the system uses the current cycle number as the time index to establish a binding relationship between the sensor channel number and the cycle sequence for each sampling point. It then combines the various physical field response values ​​collected in that cycle with the tension curve, line laying rhythm changes, humidity change slope, etc., within the corresponding time period, and performs unified processing of the multi-channel response using a normalization function. The normalization process not only standardizes the dimensions but also corrects the dynamic sensitivity offset of each channel, ensuring that the final output is a standardized multi-channel response curve, which can be used for subsequent structure reconstruction and physical trajectory fitting.

[0057] After generating the normalized response, the system enters the winding local structure identification stage. This stage requires reading the spatial arrangement of the conductor in the skeleton coordinate system and coupling it with the magnetic flux density variation trend at that location to identify the boundary position of the current winding in the physical response map. Through spatial differences and gradient variation trends between multiple sets of boundary points, the system can delineate local response regions, each exhibiting essentially consistent electromagnetic response behavior. These local regions serve as the spatial basis for the physical field distribution snapshot, providing boundary constraints for subsequent orbit vector extraction.

[0058] Within each local region, the system calculates the spatial distribution of the magnetic induction vector field, electric potential gradient field, and equivalent impedance phase change function within that region. It then fits the phase change trajectories reflecting the main gradient path of the magnetic field transmission path, the energy path of the capacitive coupling accumulation path, and the impedance trajectory, respectively. Each trajectory is fitted by a set of continuous coordinate points, whose physical meaning corresponds to the contribution trajectory of the coil structure to the inductance, parasitic capacitance, and equivalent impedance performance, respectively. These trajectory vectors not only express the trend of physical state changes with position but also possess parameters such as spatial directionality, amplitude change rate, and response distribution width, used to quantify the orderliness and offset of the physical structure within space.

[0059] After track extraction is complete, the system enters the dynamic threshold calculation stage. In this stage, the system introduces higher-order statistical characteristics of the aforementioned production parameters as adjustment factors, including the standard deviation of conductor tension (reflecting mechanical stability), the first derivative of environmental humidity changes (reflecting medium environmental fluctuations), the winding pitch fluctuation rate (reflecting winding control accuracy), and the spindle speed change rate (reflecting winding rate stability). Based on these parameters, the system constructs a set of physical offset sensitivity functions and applies them to the extracted track vectors, setting different upper limits for comparison tolerances for magnetic field tracks, capacitive tracks, and impedance tracks. The calculation of these tolerance thresholds follows the logic of "the greater the disturbance, the tighter the tolerance" or "the steeper the offset trend, the more dynamically the threshold is strengthened," thereby achieving dynamic adaptation between the deviation comparison rules and the actual disturbance scenario.

[0060] Finally, the system outputs the set of all dynamically tolerant adjusted orbital vectors as evolutionary characteristic parameters for the current revolution number or level. These evolutionary characteristic parameters not only fully record the structural trend and response behavior of the current winding region in electromagnetic space, but also possess numerical stability after perturbation adaptation. They are the only high-fidelity input required for subsequent comparison with ideal characteristic parameters under the same winding progress in the benchmark model. Through this processing path, the system can establish a clear and operable mapping relationship between the physical field space and changes in production perturbations, ensuring that the entire deviation diagnosis and compensation path still possesses high precision control capability and evolutionary consistency under complex operating conditions.

[0061] Furthermore, the calculation method for the dynamic threshold parameter used to adjust the alignment tolerance of each type of evolutionary orbit includes a weighted fusion model based on the perturbation attribution factor and orbital response sensitivity, and the dynamic threshold ΔT(i,j) is calculated as follows:

[0062]

[0063] Where ΔT(i,j) represents the alignment tolerance threshold of the i-th orbit in the j-th region, and the unit is the normalized response quantity; S(i,j) represents the perturbation sensitivity coefficient of the orbit vector in the orbit number i and the spatial region j, which is derived from the perturbation response gradient in the previous orbit feature modeling process; This represents the variance of tension fluctuation in the tension sensor during the current winding cycle.

[0064] The first derivative of the magnetic field as a function of the number of revolutions reflects the trend of abrupt changes in local magnetic flux.

[0065] Var(P j ) represents the historical variance of the wiring pitch in the j-th spatial region, reflecting the fluctuation range of structural density; α represents the spatial gradient of the electric field in the j-th spatial region, reflecting the steepness of the boundary of the dielectric coupling response in that region; α is the basic scaling factor used to adjust the overall tolerance level, with a value range of [0.5, 2.0]; β1, β2, and β3 are disturbance type attribution weighting factors, which are dynamically selected according to the disturbance type indicated in the disturbance label, satisfying the normalization constraint of β1 + β2 + β3 = 1.

[0066] The calculated result ΔT(i,j) will be used as the basis for determining the offset of each trajectory node in the orbit vector, and will be used to set the comparison thresholds for the main magnetic field path, the capacitor energy path and the complex impedance path, so as to improve the ability to suppress false alarms of low amplitude disturbances while ensuring detection sensitivity.

[0067] In the plug-in inductor winding process control method proposed in this invention, in order to achieve high-precision comparison of different types of physical evolution tracks and effectively deal with the local error amplification effect caused by factors such as tension fluctuations, environmental disturbances or structural inhomogeneities, the system introduces a dynamic threshold parameter calculation method based on the weighted fusion of disturbance attribution factor and track response sensitivity, which is used to determine the allowable response offset tolerance range for each type of track during the comparison process.

[0068] This dynamic threshold uses both the number of cycles and the spatial region as dual indices to construct a two-dimensional tolerance matrix, and its core calculation formula is as follows:

[0069]

[0070] Wherein, ΔT(i,j) represents the alignment tolerance threshold of the i-th orbit in the j-th spatial region. The unit is the normalized response amplitude, which is a positive real value. Its physical meaning is the range within which the offset of the orbit vector at this position can be regarded as an invalid disturbance or normal fluctuation, and is used to determine whether to trigger the compensation mechanism.

[0071] S(i,j) is the track disturbance sensitivity coefficient, which is the quantized result of the local disturbance response gradient extracted after fitting the historical disturbance response during track modeling. It is typically based on the rate of change of the magnetic field gradient track, capacitance track, and complex impedance track to the disturbance input, with a unit of 1, and is a dimensionless ratio. S(i,j) reflects the susceptibility of the response at that position in that loop to the disturbance input. A suggested value range is [0.8, 1.5]. When the winding area is at the edge of the structure or across layers, this value should be increased to improve monitoring sensitivity. The track disturbance sensitivity coefficient S(i,j) is determined based on the magnitude of the sensor response change at that position during disturbances in historical samples; the more severe the response, the higher the coefficient. For structural edges, across layers, or areas prone to tension fluctuations, the system automatically sets a higher value (e.g., 1.3); for the central area with a stable response, a lower value (e.g., 0.8) is set. The system establishes a position-correspondence coefficient table through pre-calibration, and directly looks up the table by loop number and region during the actual winding process.

[0072] This represents the variance of conductor tension fluctuation during the current winding cycle, expressed in N·m. 2 The variance is calculated using continuous real-time sampling from the tension sensor and a sliding window method, with a recommended time window of ±2 turns for the current number of turns. A larger value indicates significant tension fluctuations during winding, which can easily lead to changes in winding density; in such cases, the local comparison threshold should be relaxed.

[0073] It is the square of the first derivative of the magnetic flux density as a function of the number of turns, used to measure the abrupt change in the magnetic field along the winding time direction. This quantity is calculated by analyzing the difference of the B-value sequence along the time axis of the non-contact Hall sensor, and its unit is (T / s). 2 If a local magnetic flux jump occurs, this value will increase significantly.

[0074] Var(P j ) represents the historical variance of the cable pitch in the j-th spatial region, in mm. 2 This is an indicator of the consistency of the wiring. This value is obtained by calculating the statistical variance by reading the wiring step distance of the corresponding area across multiple winding cycles. A larger variance indicates significant fluctuations in the wiring density in that area, poor structural stability, and a tendency to accumulate capacitive coupling errors.

[0075] This represents the squared magnitude of the electric field gradient within this region, expressed in V / m. 2 This value characterizes the spatial drasticness of changes in the electric field of the dielectric layer. It is obtained by acquiring data using a high-frequency electric field probe array, reconstructing it into a continuous field distribution through interpolation, and then calculating the local gradient norm. A large value may indicate the presence of enhanced interlayer coupling, requiring increased tolerance to avoid false triggering.

[0076] α is the basic scaling factor, a global scaling factor set by the user to adjust the overall tolerance. The recommended value is 1.0, and it can be adjusted between 0.5 and 2.0 during the commissioning phase to adapt to different equipment and product specifications.

[0077] β1, β2, and β3 are normalized weighted coefficients for three types of physical deviation factors, determining the proportion of each factor's influence in the dynamic tolerance. The sum of the three must be 1, and they should be dynamically set according to the currently identified disturbance type label. For example, when the label indicates that the main cause of the disturbance is tension fluctuation, it is recommended to set β1 to 0.5 or higher; if the disturbance is dominated by inter-layer structural offset, then it is recommended to increase β2 to 0.4–0.6; if the main cause is abnormal electric field coupling, then increasing the value of β3 to 0.5 is more effective.

[0078] The calculated ΔT(i,j) is then applied to the comparison process of each type of orbit vector to determine whether the current orbit node offset exceeds a reasonable range. If the offset values ​​of multiple consecutive nodes exceed this threshold, the system determines that an actual disturbance exists and triggers the corresponding compensation action. Compared with the fixed threshold mechanism, this method has stronger environmental adaptability and disturbance robustness, significantly reduces the false alarm rate, and improves the consistency of actual winding.

[0079] Furthermore, within each local region, based on the spatial distribution of the equivalent physical response curve, the magnetic field response path, capacitive coupling trajectory, and impedance phase change curve are extracted. These are then used to generate magnetic field gradient orbits, capacitive energy orbits, and complex impedance orbits through fitting processing. The resulting set of orbit vectors characterizing the spatial physical evolution trend is output, including:

[0080] Within each local region, based on the multi-channel equivalent physical response curves acquired and normalized by the non-contact sensor array, directional projection is performed in the structural coordinate system to project the magnetic field response, electric field coupling, and complex impedance phase change onto the axial, radial, and planar transverse vector channels of the inductor winding, thereby constructing a set of response matrices with directional labels.

[0081] Using the aforementioned set of response matrices with directional labels, the main response path identification is performed along the three-dimensional spatial grid nodes in each local region. The main path identification algorithm adopts the principle of local maximum response growth rate, selects the continuous chain of points with the largest absolute value of the first derivative of the response change as the candidate segment of the main response trajectory, and combines the conditions of magnetic flux continuity, electric field enclosure and impedance admittance consistency to eliminate abnormal segments with inconsistent physical logic, and finally forms three types of spatial response trajectories: magnetic field main gradient trajectory, capacitance energy aggregation trajectory and complex impedance offset path.

[0082] Based on the three types of spatial response trajectories obtained, trajectory fitting models are established respectively based on the number of conduction coils and the spatial hierarchy. The fitting models use B-spline curves or polynomial fitting to smooth the trajectory curves, and each type of trajectory path is discretized into a set of trajectory nodes with a coil index, path direction vector, response amplitude and rate of change.

[0083] For abnormal change points in the track node set, combined with the aforementioned production parameter sampling data, including the standard deviation of conductor tension, the fluctuation rate of cable pitch and the first derivative of humidity change, the weight of disturbance influence factor is set on the fitted track, and dynamic compensation fitting is performed on the local section of the track, so that each track not only reflects the physical response itself, but also has structural sensitivity strongly correlated with external disturbance sources.

[0084] The three types of track node sets are merged and encoded into a track vector set. The track vector set uses the number of laps as the main index, the spatial path segment as the sub-index, and the track type as the classification label. It also includes a disturbance sensitivity field, a location confidence field, and a spatial scale standardization factor field, which serve as the physical structure reference basis for subsequent deviation judgment, dynamic threshold calculation, and disturbance attribution comparison.

[0085] During the actual winding process of plug-in inductors, a sensor array continuously collects spatially distributed magnetic field, electric field, and complex impedance response signals in the winding area in a non-contact manner, and performs synchronous sampling in conjunction with the number of turns and the current winding rhythm. All sensor channel output data undergoes unified normalization processing to form multi-channel equivalent physical response curves. Each channel corresponds to a spatial location, and this response curve exhibits significant directional differences at different winding stages; therefore, spatial decomposition in the structural coordinate system is necessary.

[0086] Specifically, the system performs three-directional projection processing on the acquired response curves within each local region (e.g., within a range of ± a few millimeters around the current winding center). The magnetic field response is projected onto the inductor winding axis, i.e., along the winding centerline; the capacitive coupling characteristics are projected radially, reflecting the insulation density and dielectric response between the conductors; while the complex impedance phase change is more reflected in the winding cross-sectional area and is projected in a planar direction. After three-dimensional projection, a response matrix set with directional labels is formed. Each column of the matrix corresponds to a response type, and each row corresponds to a spatial sampling point, while retaining the turn index and timestamp.

[0087] Based on the aforementioned response matrix set, the system performs a main response path identification operation. For each response dimension, the spatial grid nodes in the corresponding direction are sorted according to the absolute value of the first derivative of the response value between adjacent points. The continuous chain of points with the largest response growth rate is identified as a candidate segment of the main response path. To eliminate erroneous paths caused by sensing noise or isolated disturbances, three physical consistency constraints are introduced: the magnetic flux path should be continuous and satisfy the closed loop condition; the electric field trajectory should conform to the consistency of the potential gradient direction change in space; and the complex impedance path should follow the admittance continuity principle in phase change. Only response paths that meet the above physical conditions are retained, ultimately outputting three types of spatial trajectories: the main gradient trajectory of the magnetic field, the convergence trajectory of the capacitance energy, and the complex impedance offset path.

[0088] For these three types of trajectories, further trajectory fitting models are established. Considering the discreteness and inherent error of actual sensor data, B-spline or polynomial fitting methods are used to ensure that the path is physically continuous and smooth in structural space. The fitted trajectory is discretized into a set of trajectory nodes along the revolution dimension. Each node records the revolution index of the current position, the direction vector of the trajectory at that point (representing the direction of trajectory propagation), the instantaneous response amplitude (such as magnetic field strength, capacitance density, complex impedance modulus), and the rate of change between it and adjacent nodes, forming a four-tuple data structure. The number of trajectory nodes is proportional to the spatial resolution of the sensor array to ensure fitting accuracy.

[0089] Because disturbances during the production process can cause nonlinear deformations in certain trajectory segments, the system further performs dynamic disturbance compensation processing on the track fitting process to improve the track's expressive ability in abnormal sections. This processing calls upon previously collected production parameters, including the standard deviation of conductor tension, the fluctuation rate of cable pitch, and the first derivative of changes in ambient humidity, to apply a disturbance factor weighted correction at points of abrupt changes in the track's rate of change. For example, if a sudden change in the magnetic field direction occurs in a certain segment of the track, and a local sudden drop in tension is detected simultaneously, the track fitting weight at that node will be increased, and the change amplitude of the fitted curve at that point will be appropriately enhanced, thus maintaining the track's sensitivity to disturbances.

[0090] All processed orbit node sets will be fused and encoded into an orbit vector set. This set is indexed by the orbit number, with each orbit containing several path segments as sub-indexes. Each path segment is categorized according to orbit type: magnetic, capacitive, and impedance. Each orbit segment contains the following fields: orbit type label, path direction vector, response amplitude, rate of change, perturbation sensitivity score, spatial location confidence of the node, and scale normalization factor per unit space. This orbit vector set will be used in subsequent steps to compare with a baseline model to determine if evolutionary deviations exceed the tolerance range, and will serve as a physical structure reference in tasks such as dynamic threshold calculation and perturbation attribution comparison.

[0091] Through the aforementioned continuous, nested, and layer-dependent processing paths, the system not only achieves spatial behavior modeling of multi-source physical signals during the winding process of plug-in inductors, but also constructs a highly expressive physical track vector representation system.

[0092] Step S103: Compare the extracted evolutionary feature parameters with the ideal feature parameters under the same winding progress in the benchmark model. Combine the collected production parameters to establish a mapping relationship between physical field deviation and production disturbance. Identify the dominant disturbance source causing the deviation through a preset deviation attribution mechanism and obtain a disturbance type label. The disturbance type label carries causal information related to changes in production parameters and is used to guide subsequent compensation control.

[0093] After completing the real-time acquisition and evolution characteristic parameter extraction of the physical field of the plug-in inductor, the system needs to compare the extracted actual feature values ​​with the previously constructed ideal benchmark model one by one to identify the sources of deviation and disturbance mechanisms in the winding process. The core of this step is to establish a stable and generalizable deviation judgment and attribution mechanism, enabling the system to automatically identify the main causes of the current winding deviation without human intervention, and generate information labels to guide subsequent compensation decisions.

[0094] To ensure physical comparability, the system must align evolutionary characteristic parameters with ideal values ​​in the baseline model in terms of revolution number, time, or hierarchical index. This means that the magnetic field distribution, electric field characteristics, or impedance response of each revolution or winding segment must be compared with corresponding indicators in the baseline model at the same location or progress. The comparison process can be performed using various methods, such as setting deviation tolerance intervals, Euclidean distance metrics, cosine similarity analysis, and principal component offset trajectories. The system can select appropriate comparison strategies based on different modeling methods and quantify the comparison results.

[0095] After feature matching, the system obtains a set of data describing the degree of deviation in each physical quantity direction. These results are merely reflections of static differences and cannot directly explain the physical causes behind the deviations. Therefore, it is necessary to introduce a disturbance attribution mechanism based on experience and simulation. This mechanism identifies the dominant physical disturbance source behind the deviation by matching deviation features with a set of preset disturbance response patterns. The disturbance attribution mechanism can be implemented based on expert systems, rule matching algorithms, or trained models (such as discriminative models based on decision trees or fuzzy logic). Its knowledge base comes from a "deviation morphology-disturbance type" mapping library established from historical process data, experimental research, or simulation analysis.

[0096] In practical applications, when the system identifies phenomena such as an abnormal increase in the magnetic field gradient, asymmetry in the electric field coupling flux, or impedance spectrum drift during a certain winding cycle, it combines these features and matches them with typical disturbance patterns in the mapping library. For example, if the system detects multi-order distortion of the magnetic field along the axial direction, it can determine that the source is a short-term fluctuation in the winding tension; if it detects that the electric field energy is concentrated off-center and accompanied by a decrease in the local conductor density, it may be attributed to instability in the winding rhythm or cross-layer displacement errors caused by mechanical vibration. Each identification result corresponds to a disturbance type label. The label not only indicates the physical source of the disturbance, such as "tension fluctuation," "pitch deviation," "surge in ambient humidity," or "uneven wire diameter," but also carries causal chain information between the disturbance and a certain type of production parameter change.

[0097] To ensure the operability of the labels, the disturbance type labels need to be organized in a structured manner, including the disturbance type name, potentially controllable process parameters, affected evolutionary characteristic categories, recommended compensation direction, and logical association with production parameters. For example, a label might indicate: "The current deviation pattern corresponds to a tension disturbance; it is recommended to increase the initial tension setpoint of the wire feeding mechanism in the following cycles; the deviation amount is positively correlated with the tension change trend over the past five cycles." Such labels facilitate automatic interpretation by the control system and provide direct decision support for subsequent fine-tuning strategies.

[0098] Through this comparison and attribution mechanism, the system no longer merely focuses on judging whether the inductor deviates from the target state, but is able to further explore the causes of the deviation and form a structured representation.

[0099] Furthermore, the extracted evolutionary feature parameters are compared with the ideal feature parameters under the same winding progress in the benchmark model. Combined with the collected production parameters, a mapping relationship between physical field deviations and production disturbances is established. A preset deviation attribution mechanism is used to identify the dominant disturbance source causing the deviation, and disturbance type labels are obtained, including:

[0100] After comparing the extracted evolutionary feature parameters with the ideal feature parameters under the same winding progress in the benchmark model, a set of multi-dimensional deviation index vectors is output. The deviation index vectors are mainly composed of magnetic field gradient difference, capacitive coupling energy difference, and impedance phase angle offset, and are accompanied by confidence weights and spatial position index information of each feature.

[0101] The deviation index vector is jointly analyzed with the synchronous production parameter set composed of the currently collected conductor tension, spindle speed, wire pitch, and ambient temperature and humidity. The joint data is used as input to call a pre-established deviation disturbance mapping database for matching inference. The database is obtained by fusing experimental data, finite element simulation data and historical process data. The mapping database is constructed according to the disturbance feature space hierarchical index and presets the causal structure between disturbance type and deviation mode.

[0102] During the mapping inference process, a combined attribution path calculation based on fuzzy rule inference and gradient decision tree algorithm is performed. For each set of deviation input data, the corresponding possible disturbance type, the causal trigger chain of the corresponding production parameter, the direction of physical field evolution interference, and the response time window prediction result are output.

[0103] After obtaining the attribution inference results, a structured perturbation type label is generated. The perturbation type label includes at least a perturbation type field, a source parameter field, a speculative causal chain field, a suggested intervention direction field, and a perturbation judgment confidence score field. Each field is structurally independent and has an interface call identifier.

[0104] The disturbance type label is encoded into a standardized JSON data structure and embedded in the process control instruction frame of the current winding control cycle. This serves as the input for generating compensation instructions for fine-tuning the winding behavior based on the disturbance type label and the corresponding production parameter change trend.

[0105] During the winding process, the evolutionary characteristic parameters extracted by the system are generally expressed in the form of multi-dimensional vectors, including but not limited to the magnetic field spatial gradient curvature distribution, the changing trend of capacitive coupling energy density with the evolution of layers, and the phase shift of complex impedance at a specified test frequency. These parameters should have corresponding turn indexes, spatial location coordinate indexes, and sampling timestamps to ensure a one-to-one correspondence with the ideal characteristic parameters in the benchmark model.

[0106] During comparison, the system calculates the difference between the actual evolutionary characteristic parameters of each turn or time segment and the ideal characteristic parameters of the same winding progress in the benchmark model, forming a set of deviation indicators. For example, if the gradient of the main magnetic field direction corresponding to a certain turn of winding shows a discontinuous abrupt change in actual measurement, the system identifies it as the magnetic field gradient difference exceeding the preset offset tolerance; similarly, if the extracted capacitive coupling flux in the wiring structure shows an abnormal increase in concentration intensity in a nearby region, accompanied by an increase in conductor spacing, the system will use this as an important quantitative indicator of capacitive coupling offset; similarly, if the phase angle of the complex impedance corresponding to a certain turn shows a positive or negative inflection point compared to the ideal curve under AC testing, the system can determine that the impedance behavior is abnormal. All these deviation indicators will be organized into a unified data structure, forming a multi-dimensional deviation indicator vector. Each dimension not only contains the difference of the physical quantity, but also needs to include the confidence score of the corresponding feature (which can be calculated from the sampling signal-to-noise ratio or fitting residual) and its spatial location index to support the subsequent location and analysis of spatial attribution paths.

[0107] After obtaining the aforementioned deviation index vector, the system synchronizes and integrates it with the currently collected production parameters. The required production parameters are not limited to single-moment values ​​but should include their short-term trends, such as the instantaneous fluctuation amplitude of conductor tension, the short-term acceleration of the spindle speed, the variance of the cable pitch, and dynamic information such as the first derivative of ambient temperature and humidity. These parameters need to be normalized and input along with the deviation vector into the attribution analysis logic module.

[0108] The core of this attribution analysis module lies in the use of a pre-established deviation disturbance mapping database. This database contains data from at least three sources: first, a large amount of historical process data collected from actual winding equipment, recording disturbance events that occurred during production and their corresponding deviation patterns; second, a dataset obtained by simulating plug-in inductor structures under different disturbance conditions through a finite element electromagnetic simulation platform, specifically used to cover deviation response patterns under extreme operating conditions; and finally, disturbance-response matching samples obtained through laboratory controlled variable experiments, used to verify model stability. The entire database is indexed and classified according to a multi-dimensional feature space, including disturbance type, disturbance occurrence area, disturbance intensity level, and deviation dimension index, ensuring query efficiency.

[0109] During mapping inference, the system calls the aforementioned database and, based on the input deviation index vector and synchronous production parameters, makes a joint judgment using a fuzzy rule-based inference system and a gradient decision tree model. The fuzzy inference system primarily addresses problems where physical relationships have continuity and fuzzy boundaries, such as whether a slight tension fluctuation is sufficient to trigger outlier behavior in coupled energy. The gradient decision tree algorithm handles the classification of disturbance types, identifies dominant disturbance parameters, and provides branch decision paths. Joint inference outputs disturbance attribution results, including the identified disturbance type (e.g., tension disturbance, pitch fluctuation, structural vibration, or temperature and humidity shock), possible triggering causal chains (e.g., tension disturbance → density decrease within winding layers → magnetic field gradient shift → inductance decrease), the main direction of the physical field disturbance (e.g., axial degradation or radial skew), and the estimated duration of the disturbance's impact.

[0110] After obtaining the above attribution inference results, the system will construct a set of structured perturbation type labels. Each label must include a perturbation type field (e.g., "local capacitive coupling peak drift caused by uneven cross-layer cabling"), a source parameter field (e.g., "instantaneous fluctuation rate of cabling pitch exceeds the mean ±8%)", a hypothetical causal chain field (e.g., "pitch fluctuation → increased inter-line spacing → enhanced capacitive coupling → local impedance anomaly"), a suggested intervention direction field (e.g., "tighten the current layer cabling pitch, compress the lateral redundancy of the cabling window"), and a perturbation judgment confidence score field (usually using a probability output in the 0–1 range, or a consistency score from multi-model fusion). The data type of each field needs to be predefined; for example, the type field uses string enumeration, and the confidence score uses floating-point numbers.

[0111] After the aforementioned disturbance type tags are structurally completed, they need to be uniformly encapsulated. It is recommended to use a JSON format structure to ensure that the fields have key-value correspondences and can be called and distributed by the control platform's standard protocol. In the actual system, this JSON-formatted disturbance tag will be encapsulated and embedded in the process control instruction frame generated in the current winding cycle, aligned with the control platform's digital communication protocol. This ensures that the lower-level driver layer can directly read the tag information and execute subsequent compensation instructions for fine-tuning the winding behavior without increasing the control communication burden.

[0112] Through the above complete process, the system can accurately identify the type of disturbance and the triggering mechanism before the physical field deviation just appears or before it exceeds the process tolerance, providing the control system with advance feedback and a basis for targeted intervention, significantly improving the process adaptability, finished product consistency and process closed-loop regulation capability of inductor products.

[0113] Furthermore, the process of performing mapping inference, including the combined attribution path calculation based on fuzzy rule inference and gradient decision tree algorithm, includes:

[0114] The deviation index vector is subjected to feature selection and importance ranking. By introducing a set of dynamic feature correlation evaluation mechanisms, the feature weight vector of the current sample is generated based on the contribution of disturbance response fluctuations of each dimension of features (including magnetic field gradient difference, capacitive coupling energy difference, and impedance phase angle offset) in historical disturbance samples. This vector is used to guide the priority of the main cause fields in the subsequent inference path.

[0115] Based on the feature weight vector, a multi-condition fuzzy logic rule set constructed by fusing expert experience and training data is loaded into the fuzzy rule reasoning system. Each rule is based on the joint membership relationship between one or more physical deviation dimensions and their corresponding production parameters, and outputs an intermediate causal trigger level label to initially narrow down the candidate set of disturbance types.

[0116] The intermediate perturbation labels and their confidence scores output by the fuzzy inference results are used as prior conditions to input into the gradient decision tree model. The gradient decision tree model uses the perturbation causal graph as the training structure and uses the perturbation label confidence, the rate of change of the first derivative of the production parameter time series curve, and the multi-scale features of the perturbation occurrence time window as input nodes to determine the most likely perturbation type and its spatial-temporal causal path step by step.

[0117] During model inference, the structure mapping table in the perturbation feature space index database is called in real time to verify whether there is a spatiotemporal overlap between the current inference path interruption point and the physical field evolution trajectory. If a match is found, the confidence of the inference path is enhanced; otherwise, the weight is reduced and a second-best branch is triggered for retry.

[0118] The output of the combined inference, including the disturbance type, the corresponding production parameter name and fluctuation direction, the predicted physical field interference propagation direction, and the predicted duration window of the disturbance impact, is output in a unified data structure. At the same time, the feature dimension, model branch sequence, and maximum information gain point number called by the current inference path are recorded for subsequent disturbance identification verification based on path consistency.

[0119] During the actual inductor winding process, the system continuously extracts physical field evolution characteristic parameters obtained by comparing with the benchmark model. Combined with real-time acquired production parameters such as conductor tension, cable pitch, spindle speed, and ambient temperature and humidity, it generates a multi-dimensional deviation index vector for deviation identification. This deviation index vector serves as input for disturbance attribution analysis and first enters the feature selection and importance ranking stage.

[0120] To effectively avoid information dilution caused by indiscriminate participation in calculations across a multidimensional parameter space, this method introduces a dynamic feature correlation evaluation mechanism. This mechanism, based on a pre-built perturbation label sample library, statistically analyzes the contribution of deviation features of each dimension to the perturbation type determination result in historical perturbation events. For example, if the magnetic field gradient difference shows a significant response in all tension perturbation samples, while the phase angle offset changes little in such perturbations, the magnetic field gradient will receive a higher weight. This correlation evaluation is typically quantified using information gain or mutual information values, and the mean weight within each category is extracted based on the perturbation sample clustering results, outputting a set of feature weight vectors. This vector not only reflects the deviation dimensions that should be prioritized for the current sample but also determines the priority matching order of subsequent fuzzy rules and decision paths.

[0121] After completing the feature weight sorting, the system inputs the vector into the fuzzy rule inference system for the first stage of attribution calculation. The fuzzy rule base consists of two parts: one is a set of rules constructed by experts based on long-term experience, such as "when the magnetic field gradient is high, the tension fluctuation is moderate, and the environmental humidity change is high, there may be winding loosening disturbance caused by a sudden drop in tension"; the other is the membership rules automatically summarized from historical labeled data through supervised machine learning. Each rule maps one or more physical deviation dimensions and their corresponding production parameters to fuzzy membership degrees (such as "low", "medium", "high"), and then outputs an intermediate label for the disturbance type (such as "suspected tension disturbance"), along with a confidence score. This score is determined based on the strength of the membership degree overlap interval of the current sample matching rule and the historical hit rate statistics.

[0122] The intermediate perturbation labels and their confidence scores from the fuzzy inference output serve as prior inputs to the gradient decision tree model for the second stage of refined path determination. This gradient tree model is trained based on a perturbation causal graph structure, where nodes represent the combined features of physical deviation dimensions and production parameters, and edges represent logical causal relationships. During the prediction phase, the model uses the perturbation label confidence score, the rate of change of the first derivative of each production parameter's time-series curve (e.g., the instantaneous decrease slope of tension), and multi-scale windows (e.g., mean shifts within 0.5s, 1.0s, and 2.0s time periods) as input variables to progressively evaluate information gain and automatically select the most discriminative decision path. If multiple candidate branches with similar outputs exist, the system prioritizes the optimal main path based on historical path hit rate, reserving the remaining paths for subsequent retries.

[0123] To enhance the structural consistency of attribution inference, this method invokes a built-in perturbation feature spatial index database in real time during model inference execution. This database uses spatial-temporal labels to map each perturbation mode to its corresponding physical field evolution trajectory. When the time node of the current inference path coincides with the physical field trajectory—for example, if an anomaly in the magnetic field gradient in the trajectory matches the perturbation outbreak time point predicted by the model—the system increases the overall confidence of the inference path. If there is no obvious match, the model automatically reduces the weight of that path and triggers a retry mechanism for the next alternative branch to prevent erroneous paths from becoming the final result.

[0124] Ultimately, the output of the combined inference will include: the type of disturbance (e.g., tension disturbance, pitch drift disturbance, etc.), the name of the corresponding main production parameter (e.g., conductor tension) and its fluctuation direction (e.g., continuous decrease), the predicted direction of physical field interference propagation (e.g., winding radial expansion), and the predicted duration window of the disturbance's impact (e.g., 6–10 turns or 0.8–1.2 seconds). Simultaneously, the system records the deviation index dimension number called by the current inference path, the branch sequence number of the gradient tree model traversed, and the feature point number where the maximum information gain occurs in the path, and saves these to the disturbance identification cache for subsequent verification or path consistency comparison in cases of continuous disturbance superposition.

[0125] Step S104: Based on the disturbance type label and the corresponding production parameter change trend, generate compensation instructions for fine-tuning the winding behavior. The compensation instructions include: adjusting the tension curve of the wire feeding mechanism to offset the influence of tension disturbance, correcting the wire laying rhythm to alleviate displacement deviation, optimizing the wire laying path to restore distribution uniformity, or adjusting the wire laying density to compensate for wire diameter changes; execute the compensation instructions to compensate for the winding behavior.

[0126] After generating disturbance type labels, the system enters the compensation control phase. Its core task is to generate and execute a set of targeted winding compensation instructions in real time, based on the disturbance types identified in the previous phase and their corresponding production parameter change trends. These compensation instructions must not only reflect the response to the disturbance but also demonstrate the logical compatibility between the specific execution actions and the disturbance characteristics. This ensures that the winding behavior is finely corrected without halting operation, thereby guiding the physical field evolution state to gradually return to the target trajectory defined by the ideal model.

[0127] The generation of compensation commands takes a disturbance type label as input. This label typically includes information such as the disturbance category (e.g., tension fluctuation, pitch anomaly, sudden change in ambient humidity, wire diameter deviation, etc.), the affected area (corresponding rotation range or structural location), the direction of influence (e.g., magnetic field distribution pattern, electric field coupling path, or impedance phase trajectory), and suggested intervention method (active compensation or delayed adjustment). The system first parses the label content and combines it with real-time updated production parameter curves (e.g., tension history sequence, spindle speed trend, wire feed rate, etc.) to assess whether the disturbance effect is continuing, expanding, or stabilizing, thereby determining the dynamic strength and execution range of the compensation command.

[0128] For example, when the system detects a brief drop in tension during a certain section of winding, and this tension disturbance has been identified as the main source of abnormal magnetic field gradient, the control system will immediately generate a compensation command to adjust the tension curve of the wire feeding mechanism. This command is not executed simply by "increasing the tension setpoint," but rather by fine-tuning the instantaneous acceleration of the wire feeding mechanism in the form of a curve, or by locally modifying the closed-loop PID parameters of the constant tension motor, to ensure that the tension recovery action matches the local density of the winding, thereby compensating for the linear density gap turn by turn without causing a reverse impact.

[0129] If the disturbance type label indicates instability in the wiring rhythm, such as misalignment of conductor layers due to mechanical backlash fluctuations, the system will fine-tune by correcting the wiring stepping rhythm. These instructions will act on the drive pulse logic of the wiring stepper motor, adjusting the lateral displacement per revolution, or optimizing the conductor laying path through interpolation algorithms to achieve a smoother, more consistent wiring rhythm per unit time, thereby suppressing discontinuous deformation of the winding structure.

[0130] In certain types of disturbances, such as abnormal electric field coupling caused by a sudden increase in ambient humidity or changes in conductor batches, the system may need to implement compensation strategies such as "optimizing the wiring path" or "adjusting the wiring density." These strategies typically do not involve direct acceleration of moving parts, but rather achieve corrections to the micro-morphology of the local structure by controlling the starting position of the wiring, carry interval, and transposition method between layers. For example, the system can make slight offsets to the wiring coordinates to create a more uniform spacing distribution at the coil interlayer interface, or actively insert compensation pitch and reduce edge unwinding when a slight deviation in wire diameter is detected, balancing the impact of volume errors on inductance parameters.

[0131] All compensation commands are transmitted to the actuator drive control units in the form of structured control data and are executed in real time without causing winding interruption. The winding behavior after execution is then re-acquired by the system and fed back to the physical field distribution snapshot extraction module, thus forming a closed-loop process of perception, judgment, execution, and feedback. The system will continuously evaluate whether the compensated physical field converges towards the baseline model, and will converge the compensation behavior after reaching the tolerance threshold, or trigger the disturbance identification and strategy update process again when a compensation failure trend appears.

[0132] Through this closed-loop dynamic compensation mechanism—triggered by disturbance tags, modulated by production parameter trends, and fine-tuned by control curves—this invention enables precise control over various uncertainties during the winding process, such as inductance deviation, structural non-uniformity, and electrical performance drift. It does not rely on static rules or introduce mechanical interruptions, significantly improving the adaptability and product consistency of the plug-in inductor manufacturing process. This compensation method can be implemented through software upgrades of existing winding equipment and is suitable for the precision manufacturing of plug-in I-shaped inductors in automated mass production lines.

[0133] Furthermore, the step of generating compensation instructions for fine-tuning winding behavior based on the disturbance type label and the corresponding production parameter change trend includes:

[0134] After identifying the disturbance type label, a specific fine-tuning control method is selected based on the disturbance type field and source parameter field in the label, including reconstructing the tension change trajectory for tension disturbances, correcting the lateral step trajectory for rigging rhythm disturbances, and adjusting the path interpolation for rigging path uniformity deviations.

[0135] For tension disturbances, the starting number of disturbances and the direction of disturbances are obtained. Based on the suggested intervention direction field in the disturbance type label, the target range for tension adjustment is set, and a time interval of three to five cycles is constructed with the starting circle of the disturbance as the base point. Within this interval, the rate of change of tension output by the wire feeding mechanism is adjusted so that the tension increases or decreases at a fixed slope in the dimension of the number of cycles, so that it falls into the target tension range at the end of the compensation cycle. The tension adjustment process does not have a step change and the first derivative remains continuous.

[0136] To address the disturbance of the wiring rhythm, the lateral wiring position index of the disturbance location is used to adjust the lateral step of the wiring driver turn by turn within the remaining number of turns before the winding is changed. The adjustment amount is determined by the product of the disturbance confidence score and the root mean square error of the current wiring pitch, so that the lateral spacing of the conductors in the layer gradually returns to the historical average pitch, avoiding the formation of uneven coil density in the cross-layer edge area.

[0137] To address the abnormal capacitance distribution caused by non-uniform wiring paths, the spatial location index of the deviation is read to locate the specific disturbance area. Two to three transition point coordinates are inserted near this area, and linear interpolation or equal-interval interpolation is used to adjust the subsequent wiring trajectory, causing a slight spatial shift in the conductor centerline, thereby expanding the capacitance accumulation area in space and smoothing the change in coupling strength.

[0138] During the compensation process, the evolution characteristic parameters corresponding to the disturbance type label are continuously monitored, and the changes in magnetic field gradient, capacitive coupling energy and impedance phase angle offset in the comparison results are used to determine whether the disturbance has naturally subsided. When the variation of physical deviation is detected to remain within the tolerance range of the reference model for two consecutive winding cycles, and the disturbance type label has not been updated for two consecutive cycles, the compensation is considered to be completed and the compensation process is terminated.

[0139] During the compensation period, the parameters of all adjustment paths will be written into the dynamic record stack and used in the next disturbance identification process to determine the correlation between historical compensation actions and the current disturbance, thereby realizing the historical closed-loop reinforcement between the deviation path and the compensation trajectory.

[0140] In this method, after the system generates a disturbance type label through the preceding steps, it enters the compensation instruction generation stage. The system first parses multiple fields in the disturbance type label, especially the "disturbance type field" and the "source parameter field". The disturbance type field identifies the type of physical or structural anomaly to which the current deviation belongs, such as tension disturbance, wiring rhythm disturbance, or wiring path uniformity disturbance; while the source parameter field indicates which type of production parameter fluctuation triggered the disturbance, such as a short-term drop in tension, an abnormal increase in wiring pitch, or a change in insulation performance caused by a sudden change in humidity.

[0141] Based on the combined information from these two fields, the system will select the corresponding control strategy and generate fine-tuning instructions that can be directly sent to the drive layer. Specifically, for tension disturbances, the system will first read the number of winding turns at the time of the disturbance and identify whether the disturbance presents "tension lower than expected" or "tension higher than expected". Then, the system will refer to the adjustment suggestions (such as incremental compensation or incremental deceleration) provided in the suggested intervention direction field in the label to set a clear tension target range, such as a target window within ±5%. The system then sets a compensation cycle starting from the current number of turns and extending three to five turns backward. Within this cycle, by adjusting the control signal of the wire feeding motor, the conductor tension gradually approaches the target tension window at an approximately linear rate as the number of turns increases, rather than abruptly changing. The control logic requires the rate of change of tension to remain continuous, that is, the first derivative of the tension compensation curve should not show abrupt changes to avoid inertial overshoot and secondary disturbances.

[0142] When a wiring rhythm disturbance is detected, the system first locates the lateral wiring position index where the disturbance occurs, i.e., the lateral coordinate interval of the conductor on the wiring platform. Based on this, the system assesses whether the remaining number of turns in the current inductor structure is sufficient for rhythm correction and selects a reasonable number of remaining turns as the compensation zone. Within this compensation zone, the system adjusts the lateral step size of the wiring motor turn by turn. The adjustment range is not fixed but is obtained by multiplying the statistical volatility of the current wiring pitch (usually calculated using the root mean square deviation) by the confidence score in the disturbance label. This strategy allows for a larger correction response in areas with high disturbance intensity, while minimizing intervention in areas with low disturbance intensity. The goal is to gradually return the lateral conductor spacing within the current layer to the historical average wiring pitch to restore structural uniformity and avoid cross-layer overloading or winding skew caused by rhythm misalignment.

[0143] For cases of abnormal capacitance distribution caused by local non-uniformity in the wiring path, the system will locate the disturbance to a specific segment of the conductor path based on the deviation spatial location index provided in the disturbance tag. Then, two to three transition points are inserted before and after this local area, and the conductor path is reconstructed using linear interpolation or equidistant interpolation. This means that without changing the winding layer or number of turns, the spatial orientation of the conductor is slightly offset, causing the wiring to exhibit slight expansion or compression in the disturbance area, thereby alleviating capacitance concentration and reducing abrupt changes in electric field coupling strength in the electromagnetic space.

[0144] During the execution of any of the above compensation strategies, the system does not default to immediately ending the compensation action, but instead activates a real-time monitoring mechanism. This monitoring mechanism continuously compares the evolutionary characteristic parameters corresponding to the compensation process with the ideal characteristic values ​​of that loop or region in the benchmark model. Characteristic parameters include the rate of change of the magnetic field gradient, the rate of change of the capacitive coupling energy density, and the trajectory of the impedance phase angle change, etc. Once the system detects that the deviations in all corresponding dimensions are within the tolerance range of the benchmark model for two consecutive winding cycles, and the disturbance type label has not been updated for two consecutive cycles, it considers that the disturbance has naturally subsided, the compensation has achieved its purpose, and then terminates the current compensation behavior to avoid ineffective or excessive intervention.

[0145] Simultaneously, the system writes all adjustment path parameters used in this compensation process, including tension adjustment path, step correction coefficient, interpolation coordinate offset, etc., into a dynamic record stack using time index and disturbance type as the primary key. This record stack will be retrieved in the next round of disturbance identification to analyze whether the currently identified new disturbance is similar to a previously compensated path. If the similarity meets a set threshold, the system can directly reference the historical compensation path, further reducing response time and adjustment costs, thereby forming a reinforced closed-loop matching mechanism between the disturbance path and the compensation trajectory.

[0146] Through the above steps, the system can not only accurately classify and grade specific disturbances, but also achieve continuous, nonlinear, and target-controllable compensation behavior throughout the entire process through multi-cycle range, dynamic curvature, feedback convergence, and historical reinforcement mechanisms.

[0147] Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims of this application.

Claims

1. A method for controlling the winding process of plug-in inductors based on dynamic compensation of production parameters, characterized in that, include: Based on the target inductor parameters and plug-in inductor structure, a benchmark model describing the evolution characteristics of the ideal physical field is constructed. The benchmark model is used to characterize the physical field change law of the core region of the inductor with the number of turns or time under ideal winding conditions. A non-contact sensor array deployed in the winding area is used to collect physical field signals, including magnetic field, electric field or impedance, and combined with the sampling data of current production parameters, including wire tension, spindle speed, wire pitch, ambient temperature and humidity, to generate a snapshot of the physical field distribution at the current level and extract evolutionary feature parameters that match the benchmark model. The extracted evolutionary feature parameters are compared with the ideal feature parameters under the same winding progress in the benchmark model. Combined with the collected production parameters, a mapping relationship between physical field deviation and production disturbance is established. The dominant disturbance source causing the deviation is identified through the preset deviation attribution mechanism, and the disturbance type label is obtained. Based on the disturbance type label and the corresponding production parameter change trend, a compensation instruction for fine-tuning the winding behavior is generated. The compensation instruction includes: adjusting the tension curve of the wire feeding mechanism to offset the influence of tension disturbance, correcting the wire laying rhythm to alleviate displacement deviation, optimizing the wire laying path to restore distribution uniformity, or adjusting the wire laying density to compensate for wire diameter changes; and executing the compensation instruction to compensate for the winding behavior. Specifically, the extracted evolutionary feature parameters are compared with the ideal feature parameters under the same winding progress in the benchmark model. Combined with the collected production parameters, a mapping relationship between physical field deviations and production disturbances is established. Furthermore, a preset deviation attribution mechanism is used to identify the dominant disturbance source causing the deviation, obtaining a disturbance type label, including: After comparing the extracted evolutionary feature parameters with the ideal feature parameters under the same winding progress in the benchmark model, a set of multi-dimensional deviation index vectors is output. The deviation index vectors are mainly composed of magnetic field gradient difference, capacitive coupling energy difference, and impedance phase angle offset, and are accompanied by confidence weights and spatial position index information of each feature. The deviation index vector is jointly analyzed with the synchronous production parameter set composed of the currently collected conductor tension, spindle speed, wire pitch, and ambient temperature and humidity. The set of joint data is used as input to call a pre-established deviation disturbance mapping database for matching inference. The database is obtained by fusing experimental data, finite element simulation data and historical process data. The mapping database is constructed according to the disturbance feature space hierarchical index and presets the causal structure between disturbance type and deviation mode. During the mapping inference process, a combined attribution path calculation based on fuzzy rule inference and gradient decision tree algorithm is performed. For each set of deviation input data, the corresponding possible disturbance type, the causal trigger chain of the corresponding production parameter, the direction of physical field evolution interference, and the response time window prediction result are output. After obtaining the attribution inference results, a structured perturbation type label is generated. The perturbation type label includes at least a perturbation type field, a source parameter field, a speculative causal chain field, a suggested intervention direction field, and a perturbation judgment confidence score field. Each field is structurally independent and has an interface call identifier. The disturbance type label is encoded into a standardized JSON data structure and embedded in the process control instruction frame of the current winding control cycle. This serves as the input for generating compensation instructions for fine-tuning the winding behavior based on the disturbance type label and the corresponding production parameter change trend.

2. The method for controlling the winding process of plug-in inductors based on dynamic compensation of production parameters according to claim 1, characterized in that, The baseline model is established through electromagnetic field finite element simulation. The simulation process uses the magnetic flux density distribution, magnetic field strength distribution, and local electric field response in the capacitive coupling surface after each turn of wire in the plug-in inductor structure is completed as reference indicators. The simulation input includes target inductor parameters, plug-in inductor structure, wire arrangement, and introduces material permeability, saturation magnetic flux density and dielectric constant of insulating material as modeling parameters. During the simulation process, multiple feature points are extracted from the impedance frequency response curve, including the maximum phase difference point, the minimum impedance point, the impedance abrupt change inflection point, and the critical frequency band boundary. This set of frequency response features is used as an auxiliary reference parameter to describe the winding evolution process under ideal conditions. Together with the magnetic flux density distribution and electric field distribution, they are used to form the physical field evolution path of the benchmark model. The output of the benchmark model is in the form of a structured vector. The structured vector includes a winding turn index, an evolution feature parameter index corresponding to each turn, and a tolerance range field corresponding to each feature parameter. The structured vector is aligned and compared turn by turn with the real-time extracted evolution feature parameters during the actual winding of the plug-in inductor. The structured vector serves as the sole reference for comparing the extracted evolutionary feature parameters with the ideal feature parameters under the same winding progress in the benchmark model in subsequent steps, and is consistent with the deviation input data structure required to identify the dominant disturbance source causing the deviation through a preset deviation attribution mechanism.

3. The method for controlling the winding process of plug-in inductors based on dynamic compensation of production parameters according to claim 1, characterized in that, The sampled data, combined with current production parameters including conductor tension, spindle speed, wire pitch, and ambient temperature and humidity, generates a snapshot of the physical field distribution at the current level. Evolutionary feature parameters matching the baseline model are then extracted, including: The raw physical field signals collected by the non-contact sensor array are processed synchronously with the sampled data. Based on the reference physical response value corresponding to the number of winding turns, the outputs of different sensing channels are normalized to obtain standardized multi-channel equivalent physical response curves for subsequent analysis. Based on the standardized multi-channel equivalent physical response curve, combined with the arrangement coordinates of each coil of wire and the trend of magnetic flux density change, the local physical field boundary in the current inductor winding is identified, and based on this boundary information, the entire winding area is divided into multiple local areas. Within each local region, based on the spatial distribution of the equivalent physical response curve, the magnetic field response path, capacitive coupling trajectory, and impedance phase change curve are extracted. Through fitting processing, magnetic field gradient orbits, capacitive energy orbits, and complex impedance orbits are generated, and a set of orbit vectors characterizing the spatial physical evolution trend is output. Based on this set of orbit vectors, and combined with the aforementioned standard deviation of conductor tension, first derivative of environmental humidity change, fluctuation rate of wire pitch and rate of change of spindle speed, the dynamic threshold parameters used to adjust the comparison tolerance of each type of evolution orbit are calculated, and the allowable offset range of magnetic field orbit, response anomaly tolerance of capacitance orbit and phase error threshold of impedance orbit are reconstructed. The set of orbital vectors adjusted by dynamic thresholds is used as the final evolutionary feature parameters.

4. The method for controlling the winding process of plug-in inductors based on dynamic compensation of production parameters according to claim 1, characterized in that, The step of generating compensation instructions for fine-tuning winding behavior based on the disturbance type label and the corresponding production parameter change trend includes: After identifying the disturbance type label, a specific fine-tuning control method is selected based on the disturbance type field and source parameter field in the label, including reconstructing the tension change trajectory for tension disturbances, correcting the lateral step trajectory for rigging rhythm disturbances, and adjusting the path interpolation for rigging path uniformity deviations. For tension disturbances, the starting number of disturbances and the direction of disturbances are obtained. Based on the suggested intervention direction field in the disturbance type label, the target range for tension adjustment is set, and a time interval of three to five cycles is constructed with the starting circle of the disturbance as the base point. Within this interval, the rate of change of tension output by the wire feeding mechanism is adjusted so that the tension increases or decreases at a fixed slope in the dimension of the number of cycles, so that it falls into the target tension range at the end of the compensation cycle. The tension adjustment process does not have a step change and the first derivative remains continuous. To address the disturbance of the wiring rhythm, the lateral wiring position index of the disturbance location is used to adjust the lateral step of the wiring driver turn by turn within the remaining number of turns before the winding is changed. The adjustment amount is determined by the product of the disturbance confidence score and the root mean square error of the current wiring pitch, so that the lateral spacing of the conductors in the layer gradually returns to the historical average pitch, avoiding the formation of uneven coil density in the cross-layer edge area. To address the abnormal capacitance distribution caused by non-uniform wiring paths, the spatial location index of the deviation is read to locate the specific disturbance area. Two to three transition point coordinates are inserted near this area, and linear interpolation or equal-interval interpolation is used to adjust the subsequent wiring trajectory, causing a slight spatial shift in the conductor centerline, thereby expanding the capacitance accumulation area in space and smoothing out changes in coupling strength.

5. The method for controlling the winding process of plug-in inductors based on dynamic compensation of production parameters according to claim 4, characterized in that, The step of generating compensation instructions for fine-tuning winding behavior based on disturbance type labels and corresponding production parameter change trends also includes: During the compensation process, the evolution characteristic parameters corresponding to the disturbance type label are continuously monitored, and the changes in magnetic field gradient, capacitive coupling energy and impedance phase angle offset in the comparison results are used to determine whether the disturbance has naturally subsided. When the variation of physical deviation is detected to remain within the tolerance range of the reference model for two consecutive winding cycles, and the disturbance type label has not been updated for two consecutive cycles, the compensation is considered to be completed and the compensation process is terminated. During the compensation period, the parameters of all adjustment paths will be written into the dynamic record stack and used in the next disturbance identification process to determine the correlation between historical compensation actions and the current disturbance, thereby realizing the historical closed-loop reinforcement between the deviation path and the compensation trajectory.

6. The method for controlling the winding process of plug-in inductors based on dynamic compensation of production parameters according to claim 2, characterized in that, The simulation process extracts multiple feature points from the impedance frequency response curve, including: In the electromagnetic field finite element simulation process, the impedance frequency response curve of each turn of wire after winding is first simulated in the whole frequency band, and the set of characteristic points including the maximum phase difference point, the minimum impedance point, the impedance change inflection point and the critical frequency band boundary are identified in the simulation results. In the identified set of feature points, the frequency response slope change, impedance phase angle fluctuation range and amplitude gradient curvature of each feature point are further calculated, and a frequency response perturbation sensitivity matrix is ​​constructed accordingly to characterize the response amplification factor of each feature point to structural micro-perturbations. The frequency response perturbation sensitivity matrix is ​​coupled with the local magnetic flux density distribution map of the plug-in inductor structure for analysis. Abnormal frequency abrupt points that appear in the weakly coupled section of the structure are screened out to ensure that only effective response feature points with structural physical correlation are retained. Frequency band clustering is performed on the retained set of valid feature points. The points are grouped and merged based on bandwidth coverage and frequency proximity. The center frequency, main response mode and representative response feature value are calculated for each cluster segment to form a frequency band representation subset. The frequency band representation subset is used as an auxiliary reference parameter to form the baseline model. After being aligned with the magnetic flux density distribution and electric field distribution in the number of turns dimension, it is encoded into the frequency domain evolution path field in the structured vector in a unified data format, so as to compare the evolution feature parameters extracted in real time during the actual winding process of the plug-in inductor turn by turn.

7. The method for controlling the winding process of plug-in inductors based on dynamic compensation of production parameters according to claim 3, characterized in that, Within each local region, based on the spatial distribution of the equivalent physical response curve, the magnetic field response path, capacitive coupling trajectory, and impedance phase change curve are extracted. These are then fitted to generate magnetic field gradient orbits, capacitive energy orbits, and complex impedance orbits, outputting a set of orbit vectors characterizing the spatial physical evolution trend, including: Within each local region, based on the multi-channel equivalent physical response curves acquired and normalized by the non-contact sensor array, directional projection is performed in the structural coordinate system to project the magnetic field response, electric field coupling, and complex impedance phase change onto the axial, radial, and planar transverse vector channels of the inductor winding, thereby constructing a set of response matrices with directional labels. Using the aforementioned set of response matrices with directional labels, the main response path identification is performed along the three-dimensional spatial grid nodes in each local region. The main path identification algorithm adopts the principle of local maximum response growth rate, selects the continuous chain of points with the largest absolute value of the first derivative of the response change as the candidate segment of the main response trajectory, and combines the conditions of magnetic flux continuity, electric field enclosure and impedance admittance consistency to eliminate abnormal segments with inconsistent physical logic, and finally forms three types of spatial response trajectories: magnetic field main gradient trajectory, capacitance energy aggregation trajectory and complex impedance offset path. Based on the three types of spatial response trajectories obtained, trajectory fitting models are established respectively based on the number of conduction coils and the spatial hierarchy. The fitting models use B-spline curves or polynomial fitting to smooth the trajectory curves, and each type of trajectory path is discretized into a set of trajectory nodes with a coil index, path direction vector, response amplitude and rate of change. For abnormal change points in the track node set, combined with the aforementioned production parameter sampling data, including the standard deviation of conductor tension, the fluctuation rate of cable pitch and the first derivative of humidity change, the weight of disturbance influence factor is set on the fitted track, and dynamic compensation fitting is performed on the local section of the track, so that each track not only reflects the physical response itself, but also has structural sensitivity strongly correlated with external disturbance sources. The three types of track node sets are merged and encoded into a track vector set. The track vector set uses the number of laps as the main index, the spatial path segment as the sub-index, and the track type as the classification label. It also includes a disturbance sensitivity field, a location confidence field, and a spatial scale standardization factor field, which serve as the physical structure reference basis for subsequent deviation judgment, dynamic threshold calculation, and disturbance attribution comparison.

8. The method for controlling the winding process of plug-in inductors based on dynamic compensation of production parameters according to claim 1, characterized in that, The process of performing mapping inference includes calculating a combined attribution path based on fuzzy rule inference and gradient decision tree algorithm, including: The deviation index vector is subjected to feature selection and importance ranking. By introducing a set of dynamic feature correlation evaluation mechanisms, the feature weight vector of the current sample is generated based on the contribution of the perturbation response fluctuation of each dimension of features in the historical perturbation samples. This vector is used to guide the priority of the main cause field in the subsequent inference path. Based on the feature weight vector, a multi-condition fuzzy logic rule set constructed by fusing expert experience and training data is loaded into the fuzzy rule reasoning system. Each rule is based on the joint membership relationship between one or more physical deviation dimensions and their corresponding production parameters, and outputs an intermediate causal trigger level label to initially narrow down the candidate set of disturbance types. The intermediate perturbation labels and their confidence scores output by the fuzzy inference results are used as prior conditions to input into the gradient decision tree model. The gradient decision tree model uses the perturbation causal graph as the training structure and uses the perturbation label confidence, the rate of change of the first derivative of the production parameter time series curve, and the multi-scale features of the perturbation occurrence time window as input nodes to determine the most likely perturbation type and its spatial-temporal causal path step by step. During model inference, the structure mapping table in the perturbation feature space index database is called in real time to verify whether there is a spatiotemporal overlap between the current inference path interruption point and the physical field evolution trajectory. If a match is found, the confidence of the inference path is enhanced; otherwise, the weight is reduced and a second-best branch is triggered for retry. The output of the combined inference, including the disturbance type, the corresponding production parameter name and fluctuation direction, the predicted physical field interference propagation direction, and the predicted duration window of the disturbance impact, is output in a unified data structure. At the same time, the feature dimension, model branch sequence, and maximum information gain point number called by the current inference path are recorded for subsequent disturbance identification verification based on path consistency.

Images