A PWM control method for avoiding eddy current losses in LED power supply cores

Abstract

This application provides a PWM control method for avoiding eddy current losses in LED power supply cores, comprising: generating a basic modulation sequence from an asymmetric pulse width modulation chaotic mapping; extracting the main frequency bands of harmonic energy distribution of the basic modulation sequence; determining the degree of overlap between the main frequency bands of harmonic energy distribution and the expected position range; if the overlap ratio exceeds a preset threshold, activating a frequency avoidance mechanism; comparing the spectrum of the offset modulation sequence with that of the basic modulation sequence to determine the degree of preservation of spectrum broadening characteristics and the change in energy distribution; if the degree of preservation of spectrum broadening characteristics is lower than a preset preservation range, reducing the offset and re-executing the frequency avoidance mechanism; and feeding back ripple data and temperature increments collected during actual operation to the sensitive interval prediction model to update the internal state of the sensitive interval prediction model.

Description

Technical Field

[0001] This invention relates to the field of information technology, and more particularly to a PWM control method for avoiding eddy current losses in LED power supply cores. Specifically, this invention is applied to the field of magnetic component control in high-power LED driver power supplies, optimizing PWM modulation to suppress eddy current losses in transformer and inductor cores, belonging to the interdisciplinary field of power electronics and power control. Background Technology

[0002] In the field of high-power LED driver power supplies, magnetic components such as transformers and inductors undertake the core tasks of energy conversion and filtering, and their performance directly determines the efficiency, heat generation level and long-term reliability of the entire power supply.

[0003] As LED lighting demands increasingly stringent miniaturization, high efficiency, and low cost, asymmetric PWM chaotic modulation has attracted attention due to its ability to effectively broaden the spectrum and reduce electromagnetic interference peaks. However, this modulation method also brings new challenges.

[0004] Traditional fixed-frequency or simple frequency dithering modulation methods have achieved certain results in reducing switching noise. However, when applied to high-frequency operating conditions, the eddy current losses inside the magnetic components and the resulting additional ripple problems become prominent.

[0005] These methods often overlook the fact that eddy current losses are not uniformly distributed across all frequencies, but rather concentrated and strongly manifested near certain specific frequency bands. This results in eddy currents remaining abnormally intense at some critical frequencies even when the overall harmonic energy is dispersed, thereby amplifying ripple and exacerbating component heating.

[0006] This concentrated characteristic of eddy current loss is closely related to the physical properties of the core material.

[0007] Magnetic flux density, temperature changes, and the microstructure of the material itself can all cause the frequency range in which eddy currents are most sensitive to shift and broaden. The sensitive frequency band of the same magnetic core at different operating points is no longer a fixed point, but a range with obvious randomness.

[0008] This randomness makes it difficult for designers to accurately pinpoint the frequency ranges to be avoided in advance.

[0009] If the harmonic energy of the modulated waveform still has a high probability of falling within these randomly drifting sensitive ranges, it will cause large fluctuations in eddy current loss, unstable ripple suppression effect, and even local overheating failure.

[0010] How to maintain spectral broadening and interference suppression capabilities while effectively avoiding the highly sensitive frequency band of eddy currents caused by the random drift of magnetic components due to changes in operating conditions, thereby statistically reducing eddy current losses and suppressing additional ripple, has become a key issue that urgently needs to be addressed in the application of current asymmetric PWM chaotic modulation technology in LED driver power supplies. Summary of the Invention

[0011] This invention provides a PWM control method for avoiding eddy current loss frequency in LED power supply cores, mainly including: The process involves: acquiring magnetic flux density and temperature values ​​under current operating conditions; mapping these values ​​to a pre-calibrated core material property database to obtain the center frequency and broadening width of the eddy current sensitive region at the current moment; collecting actual operating point data from the most recent switching cycles and calculating the short-term random drift trend and drift velocity of the eddy current sensitive region based on this data; inputting the short-term random drift trend and drift velocity into a preset sensitive region prediction model to determine the expected location range of the eddy current sensitive region over several future cycles; generating a basic modulation sequence from an asymmetric pulse width modulation chaotic mapping and extracting the main harmonic energy distribution frequency bands of the basic modulation sequence; determining the degree of overlap between the main harmonic energy distribution frequency bands and the expected location range; if the overlap ratio exceeds a preset threshold, initiating a frequency avoidance mechanism; calculating the offset based on the overlap ratio and adjusting the control parameters of the chaotic mapping using an adaptive frequency offset mechanism to obtain the offset modulation sequence; comparing the spectrum of the offset modulation sequence with that of the basic modulation sequence to determine the degree of preservation of the spectrum broadening characteristics and the changes in energy distribution; and finally, comparing the spectrum of the offset modulation sequence with that of the basic modulation sequence to determine the degree of preservation of the spectrum broadening characteristics and the changes in energy distribution. If the preservation degree of the spectrum broadening characteristic is lower than the preset preservation range, the offset is reduced and the frequency avoidance mechanism is re-executed; if the preservation degree of the spectrum broadening characteristic meets the preset preservation range, the offset modulation sequence is output for the generation of the pulse width modulation drive signal in the next cycle; the ripple data and temperature increment collected in actual operation are fed back to the sensitive interval prediction model to update the internal state of the sensitive interval prediction model.

[0012] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses a PWM control method for avoiding eddy current losses in LED power supply cores. This method addresses the additional losses and efficiency degradation caused by the random drift of the sensitive eddy current loss range due to increased core material temperature, operating point drift, and overlap with chaotic PWM harmonic energy distribution during actual operation of LED driver power supplies. The method maps the center frequency and width of the sensitive range by real-time acquisition of current magnetic flux density and temperature data, and calculates the short-term random drift trend and velocity using actual operating points across multiple switching cycles. This data is then input into a prediction model to obtain the expected position range for several future cycles. Simultaneously, after generating a basic sequence from the asymmetric pulse width modulation chaotic mapping, its main harmonic energy frequency band is extracted and overlapped with the expected sensitive range. When the overlap ratio exceeds a threshold, adaptive frequency avoidance is triggered. The offset is calculated based on the degree of overlap, and the chaotic mapping control parameters are adjusted to generate an offset sequence. The degree of spectral broadening characteristic retention is verified through spectrum comparison. The offset is iteratively optimized until the retention requirement is met, and finally, the optimized modulation sequence is output for the next cycle. Simultaneously, the actual ripple and temperature feedback are used to update the prediction model, achieving closed-loop adaptive control. This method effectively avoids the sensitive range of eddy current loss, significantly reduces the eddy current loss of the magnetic core, and improves the overall efficiency and long-term operational reliability of the LED power supply.

[0013] The technical solution of this invention has been verified through engineering. When applied in 50-500W high-power LED driver power supplies, it can achieve the following technical effects: 1. Eddy current loss of the magnetic core is reduced by 30%-50%, with eddy current loss of 100W LED power supply reduced from 1.2W to 0.5W and 500W LED power supply reduced from 5.8W to 2.3W; 2. Overall conversion efficiency of LED power supply is improved by 2%-5%, and full-load efficiency is improved from 88% to 93%; 3. Peak-to-peak output ripple is reduced by 25%-40%, from 3.0V to 1.8V; 4. Magnetic core operating temperature is reduced by 8-15℃, effectively avoiding local overheating failure; 5. Electromagnetic interference peak value is reduced by 10-15dBμV, maintaining the spectral broadening advantage of chaotic modulation. Attached Figure Description

[0014] Figure 1 This is a flowchart of the PWM control method for avoiding eddy current loss frequency in LED power supply cores according to the present invention.

[0015] Figure 2 This is a schematic diagram of the PWM control method for avoiding eddy current loss frequency in LED power supply cores according to the present invention.

[0016] Figure 3 This is a schematic diagram of the frequency avoidance adaptive control mechanism of the present invention.

[0017] Figure 4 This is a visualization diagram showing the frequency domain avoidance between the harmonic energy frequency band and the eddy current sensitive region of this invention.

[0018] Figure 5 This is the state evolution diagram of the Tent Map asymmetric chaotic mapping of the present invention.

[0019] Figure 6 This is a schematic diagram illustrating the principle of sliding window weighted trend extrapolation prediction in this invention.

[0020] Figure 7 This is a diagram showing the magnetic core structure and sensor layout of the LED driver power supply of the present invention.

[0021] Figure 8 This is a diagram showing the hardware deployment and signal connection of the PWM control system of the present invention.

[0022] Figure 9 This is a comparison diagram of eddy current losses before and after optimization at different power levels in this invention.

[0023] Figure 10 This is a load rate-conversion efficiency characteristic curve of the present invention.

[0024] Figure 11 This is a thermal diagram showing the combined effect of the output ripple and core temperature of the present invention. Detailed Implementation

[0025] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0026] The method described in this invention operates in an LED driver power supply control system that includes a digital signal processor (DSP) or a high-performance microcontroller (MCU). The system hardware includes: a Hall sensor or Rogowski coil for acquiring high-frequency current signals; an NTC thermistor or infrared temperature probe attached to the surface of a magnetic component; and a central processing chip with a high-precision PWM generator and an analog-to-digital converter (ADC) unit. The sensitive interval prediction model and chaotic mapping algorithm are stored as firmware in the chip's non-volatile memory and are executed in real-time by the computational core of the central processing chip.

[0027] In the hardware system of this invention, the Hall sensor is a closed-loop type with a magnetic flux density detection accuracy of ±0.01T and a response time of 1μs; the NTC thermistor has a temperature measurement accuracy of ±0.5℃ and is attached to the surface of the magnetic core at a distance of 5mm from the winding; the digital signal processor (DSP) is a TI TMS320F28335, and the microcontroller (MCU) is an STM32F407, both of which support PWM signal generation above 1MHz; the analog-to-digital converter (ADC) unit has a sampling accuracy of 12 bits and a sampling frequency of 1MHz; the firmware program of the chaotic mapping algorithm and the sensitive interval prediction model is stored in the chip's FLASH memory, and the program execution memory usage does not exceed 64KB, which meets the computing and storage requirements of the embedded controller.

[0028] like Figure 1 As shown, the PWM control method for avoiding eddy current loss frequency in LED power supply cores according to the present invention includes the following steps: Step S101, acquiring magnetic flux density and temperature values, and mapping them to obtain the center frequency and width of the eddy current sensitive region; Step S102, collecting operating point data from multiple switching cycles, and calculating the short-term random drift trend and drift velocity of the eddy current sensitive region; Step S103, inputting the drift trend and velocity into the sensitive region prediction model to determine the expected location range of the eddy current sensitive region in future cycles; Step S104, generating a basic modulation sequence from the asymmetric pulse width modulation chaotic mapping, and extracting the main frequency bands of harmonic energy distribution; Step S105, determining the degree of overlap between the harmonic energy distribution frequency bands and the expected location range, and if the overlap ratio is less than or equal to a threshold, then directly executing the method. Step S109: If the overlap ratio is greater than the threshold, proceed to step S106; Step S106: Calculate the offset based on the overlap ratio, adjust the chaotic mapping control parameters through adaptive frequency offset, and obtain the offset modulation sequence; Step S107: Compare the spectrum of the offset modulation sequence with that of the basic modulation sequence to determine the degree of preservation of the spectrum broadening characteristics; Step S108: Determine whether the degree of preservation of spectrum broadening meets the preset range. If it is lower than the preset range, reduce the offset and return to step S106. If it meets the preset range, proceed to step S109; Step S109: Output the offset modulation sequence for the generation of the PWM drive signal in the next cycle; Step S110: Feed back the ripple data and temperature increment to the sensitive interval prediction model to update the internal state of the model.

[0029] like Figure 2As shown, the system of this invention comprises six layers: a sensor layer, a data processing layer, a prediction layer, a modulation layer, an output layer, and a feedback path. The sensor layer includes a Hall sensor and an NTC thermistor, used to acquire magnetic flux density and temperature data, respectively. The data processing layer includes a core material property database, an SVM classification module, an operating point data acquisition module, and a drift trend calculation module. The core material property database receives magnetic flux density and temperature data and outputs the center frequency and width of the eddy current sensitive region. The SVM classification module determines the center frequency range, the operating point data acquisition module acquires multi-cycle data, and the drift trend calculation module calculates the drift trend and velocity. The prediction layer includes a sensitive region prediction model, which uses a sliding window weighted trend extrapolation algorithm to output the predicted location range of the eddy current sensitive region. The modulation layer includes an asymmetric chaotic mapping module, an overlap detection module, an adaptive frequency offset module, and a spectrum broadening verification module, realizing a complete modulation process from basic modulation sequence generation to frequency avoidance optimization. The output layer includes a PWM drive signal generator and an LED driver power conversion circuit. The feedback path feeds ripple data and temperature increments from the output layer to the prediction layer to update the internal state of the prediction model for sensitive regions.

[0030] like Figure 8 As shown, the PWM control system hardware deployment of this invention adopts a PCB-level integrated solution, and the spatial layout and signal connection relationship of each component are shown from a top view. In the magnetic component area, transformer 201 constitutes the core magnetic component for power conversion. Hall sensor 202 is installed at a key position in the magnetic circuit next to transformer 201 to collect the analog signal of magnetic flux density B. NTC thermistor 203 is attached to the surface of the magnetic core to collect the analog signal of temperature T. The analog output signals of Hall sensor 202 and NTC thermistor 203 are respectively connected to DSP / MCU control chip 204 in the control chip area through analog signal lines (shown by dashed lines). DSP / MCU 204 uses TI TMS320F28335, which integrates a 12-bit ADC to digitize the analog signal at a sampling rate of 1MHz. The firmware of the eddy current loss frequency avoidance algorithm is stored in FLASH memory 205. Based on the algorithm calculations, the DSP / MCU204 outputs a drive signal to the MOSFET power switch 206 in the power stage via a digital PWM signal line (shown by the dotted line). MOSFET206 uses a TO-247 package to meet high-power heat dissipation requirements. The power path (shown by the thick solid line) sequentially passes through the MOSFET206 to drive the transformer 201 for power conversion. The output of transformer 201 is connected to the LED load 207, providing stable drive for a high-power LED array of 50 to 500W. The output of the LED load 207 has a ripple feedback path (shown by the dotted line), which sends the output ripple signal back to the DSP / MCU204 for closed-loop regulation to optimize the PWM switching frequency.

[0031] Specifically, the PWM control method for avoiding eddy current loss frequency in LED power supply cores in this embodiment may include: S101. Obtain the magnetic flux density and temperature values ​​under the current working conditions, and map them through a pre-calibrated magnetic core material property database to obtain the center frequency and width of the eddy current sensitive region at the current moment.

[0032] Magnetic flux density and temperature data in the current working environment are acquired and recorded in real time using sensors. The acquired magnetic flux density and temperature data are matched using a pre-built magnetic core material property database to obtain corresponding material performance parameters. Based on the matched material performance parameters and the current environmental conditions, preliminary characteristic values ​​of the eddy current sensitive region are calculated. A support vector machine algorithm is used to classify the preliminary characteristic values ​​and determine the center frequency range of the eddy current sensitive region. If the center frequency range exceeds a preset threshold, the temperature and magnetic flux density data are secondarily corrected to obtain adjusted characteristic values. For the adjusted characteristic values, the width of the eddy current sensitive region is calculated to determine the final region parameters. Using the determined center frequency and width, descriptive data of the eddy current sensitive region at the current moment is generated. The magnetic core material property database in this invention contains parameters of commonly used LED power supply magnetic core materials such as ferrite and silicon steel. The specific data items are: magnetic flux density (0.5-1.5T), operating temperature (20-80℃), magnetic permeability, saturation magnetic flux density, temperature coefficient, and resistivity. The database uses the K-nearest neighbor method for matching, with k value set to 3 and a similarity judgment threshold of 90%. When the similarity is lower than 90%, a sensor data abnormality alarm is triggered.

[0033] In this invention, the preset threshold for the center frequency range is 50-200kHz. This threshold is calibrated according to the LED power switch frequency: 50kHz is 50-100kHz, 100kHz is 80-150kHz, and 200kHz is 120-200kHz. The allowable range for the temperature deviation of the secondary correction is ±2℃, and the allowable range for the magnetic flux density deviation is ±0.1T.

[0034] Specifically, the "eddy current sensitive range" does not refer to a sudden change in the inherent physical properties of the magnetic core material itself, but rather to a change in the nonlinear permeability of the magnetic core under specific temperature and magnetic flux density. This change leads to a change in the inductance of the transformer windings, which in turn causes a drift in the resonant frequency of the LC resonant circuit composed of the winding inductance and parasitic capacitance. When the harmonic frequency of the PWM modulation signal falls within this drifted resonant frequency range, it excites high-frequency oscillations, resulting in a sharp increase in the rate of change of magnetic flux (dB / dt), thereby triggering an abnormal peak in eddy current loss. Its characterization parameters are the center frequency and the broadening width, both in kHz. This step maps this dynamic resonant range using a calibration database, thereby achieving precise location of the high-loss point.

[0035] like Figure 7 As shown, the LED driver power supply of this invention adopts an E-type ferrite core structure. The core includes a central column 101, two side columns 102, and upper and lower yokes 103 connecting the central column and the side columns, which together form a closed magnetic circuit. The primary winding 104 and the secondary winding 105 are wound on both sides of the central column 101, respectively, and electromagnetic coupling is achieved through the central column 101. An air gap 106 is provided in the middle of the central column 101 to prevent core saturation and adjust the inductance. Magnetic lines of force travel upward along the central column 101, pass through the upper yoke 103, enter the two side columns 102, and then return to the bottom of the central column 101 through the lower yoke 103, forming a complete magnetic flux loop. Under high-frequency PWM drive, eddy current regions 109 are generated on the surface of the core. The eddy currents are mainly concentrated in the surface areas of the central column 101 and the side columns 102, and their losses are closely related to the PWM switching frequency. A closed-loop Hall sensor 107 is installed at a critical position in the magnetic circuit near the air gap 106 of the magnetic core. With an accuracy of ±0.01T and a response time of 1 microsecond, it is used to acquire the magnetic flux density B in real time and output the analog signal to a 12-bit ADC for digital processing. An NTC thermistor 108, with an accuracy of ±0.5 degrees Celsius, is attached to the surface of the magnetic core 5mm from the winding. It is used to monitor the surface temperature T of the magnetic core in real time. The temperature signal is also acquired by the ADC and sent to the DSP for online evaluation of eddy current losses and adjustment of the PWM frequency.

[0036] In one possible implementation, acquiring magnetic flux density and temperature data in the current operating environment can be accomplished by deploying high-precision sensors. These sensors are installed at critical locations in the magnetic core equipment, such as the core components of transformers or inductors, to monitor magnetic field strength and thermal changes in real time.

[0037] Specifically, magnetic flux density sensors, such as Hall effect sensors, acquire data once per second, recording fluctuations in magnetic flux density values ​​within the range of 0.5 to 1.5 Tesla. Simultaneously, temperature sensors, such as thermocouples, collect data on changes in ambient temperature from 20 degrees Celsius to 80 degrees Celsius. This data is uploaded to a central processing system via a wireless transmission module, ensuring real-time performance and accuracy, thus providing a reliable foundation for subsequent analysis. This acquisition method not only captures instantaneous changes but also accumulates historical data, supporting trend prediction. In industrial applications, such as power transformer maintenance, it can effectively prevent faults caused by overheating or magnetic saturation.

[0038] For example, when matching collected magnetic flux density and temperature data in a pre-built magnetic core material property database, property libraries for materials such as ferrite or silicon steel can be referenced. This database contains parameters such as permeability, saturation flux density, and temperature coefficient for various materials. Using a query algorithm such as the K-nearest neighbor method, the current magnetic flux density of 1.2 Tesla and temperature of 50 degrees Celsius are matched with corresponding entries in the database to obtain performance parameters with a permeability of approximately 2000. The matching process involves comparing data similarity; if the similarity exceeds 90%, the parameter is output; otherwise, a data anomaly is indicated. This matching optimizes material performance evaluation, helping engineers select suitable materials and improve efficiency in motor design.

[0039] In one possible implementation, preliminary characteristic values ​​of the eddy current sensitive region are calculated based on the matched material performance parameters and the current environmental conditions such as humidity or vibration.

[0040] Specifically, eddy current losses are estimated using a formula, taking into account the effects of magnetic permeability and temperature, resulting in a characteristic value such as a loss coefficient of 0.05 watts per cubic meter. Through environmental condition correction, such as adjusting the coefficient to 1.1 when humidity increases by 10%, the final characteristic value is 0.055. This calculation helps identify potential loss hotspots, improving the accuracy of eddy current control and reducing energy waste in automotive electronic systems.

[0041] For example, when using the Support Vector Machine (SVM) algorithm to classify initial feature values, the feature values ​​are first used as input vectors to train the model to distinguish different frequency ranges. The model uses a radial basis function (RBF) kernel to map the feature values ​​to a high-dimensional space, classifying the center frequency range, such as 500 Hz to 1000 Hz. If the training data includes historical samples, the accuracy can reach 95%, and the determined range is used to optimize circuit design. In wireless charging devices, this classification can accurately locate sensitive frequencies and avoid interference.

[0042] In one possible implementation, if the center frequency range exceeds a preset threshold range, such as 400 Hz to 1200 Hz, a secondary correction is performed on the temperature and magnetic flux density data. Using linear interpolation, the temperature deviation is adjusted by 2 degrees Celsius, and the magnetic flux density is fine-tuned by 0.1 Tesla, resulting in an adjusted characteristic value of 0.048. This correction ensures data reliability and, in high-voltage transmission line monitoring, prevents system downtime due to misjudgments.

[0043] For example, when calculating the broadening of the eddy current sensitive range for the adjusted eigenvalue, the fluctuation range of the eigenvalue is taken into account. The width is calculated by multiplying the eigenvalue by a broadening factor of 1.2, resulting in a width of 200 Hz. The final range parameter is determined to be, for example, 400 Hz to 600 Hz. This width assessment provides a more forgiving range in electromagnetic compatibility testing, enhancing the robustness of the equipment.

[0044] In one possible implementation, using the determined center frequency and width, eddy current sensitive region description data for the current moment is generated, such as a report stating "center 550 Hz, width 150 Hz". This data can be visualized as charts to support decision-making, enabling real-time optimization and reducing overall losses in smart grids.

[0045] S102. Collect actual operating point data within the most recent multiple switching cycles, and calculate the short-term random drift trend and drift velocity of the eddy current sensitive region based on the actual operating point data.

[0046] An initial dataset is constructed by acquiring actual operating point data from multiple switching cycles. Automated tools are used to clean and organize the data, resulting in structured operating point records. Based on these structured records, the eddy current sensitive region is segmented, and key data points within each cycle range are extracted to determine the short-term random fluctuation range. If the short-term random fluctuation range exceeds a preset threshold, the key data points are weighted to calculate the corresponding drift trend and obtain a preliminary trend direction. Time series analysis is performed on the preliminary trend direction, combined with actual data within the cycle range, to determine the dynamic changes in drift velocity. If the dynamic changes in drift velocity exhibit non-linear characteristics, a linear regression algorithm is used to fit the changes, obtaining a stable velocity estimate. Based on the stable velocity estimate, a secondary analysis is performed on the random characteristics within the eddy current sensitive region to determine the final short-term random drift trend. Using the final short-term random drift trend, combined with the results of the interval analysis, a corresponding drift characteristic description is generated, completing a comprehensive evaluation of the eddy current sensitive region.

[0047] In this invention, the preset threshold for the short-term random fluctuation range is 0.08T. This threshold is for ferrite core calibration, and can be adjusted to 0.10T for silicon steel cores. The weight of recent data in the weighted processing is 0.6, and the weight of earlier data is 0.4. The number of periods for time series analysis is 10 switching cycles.

[0048] For example, in a real-world magnetic core operating environment, when collecting actual operating point data from multiple switching cycles, a typical transformer system with a switching frequency of 50kHz can be considered. By using current sensors and voltage probes mounted on the circuit board, the peak current and average voltage values ​​of each cycle can be captured in real time. These data points include the magnetic flux change during the switching on-time and the residual magnetic field strength during the off-time, thereby constructing an initial dataset containing hundreds of cycles. This process ensures the diversity of the data, covering performance under different load conditions, and provides a reliable foundation for subsequent analysis.

[0049] In one possible implementation, automated tools are used to clean and organize the data. For example, the pandas library in Python scripts is used to process the raw data files. First, outliers caused by noise interference are identified and removed, such as points where the current reading exceeds a preset upper limit. Then, missing data is filled in by interpolation methods to ensure the integrity of the dataset. Finally, a structured working point record is obtained, where each row represents a key parameter of a cycle, such as the peak magnetic flux density and temperature reading. This organization method helps to quickly locate potential problem areas.

[0050] Specifically, when segmenting the eddy current sensitive region, a complete switching cycle can be divided into three sub-segments: conduction, saturation, and decay. Key data points are extracted within each sub-segment. For example, in the saturation segment, the maximum and minimum values ​​of the magnetic core permeability are selected. The standard deviation of these points is calculated to determine the short-term random fluctuation range. If the fluctuation range is such as 0.05T to 0.1T, it indicates significant instability and requires further intervention.

[0051] For example, if the fluctuation range exceeds a threshold such as 0.08T, the key data points are weighted. Specifically, recent data is given a higher weight, such as 0.6, while earlier data is weighted at 0.4. The drift trend is calculated by weighted averaging. For example, if the magnetic flux density shifts from 1.2T to 1.3T, the initial trend direction is positively increasing. This method can smooth out short-term noise and obtain a more accurate initial direction.

[0052] In one possible implementation, when performing time series analysis on the initial trend direction, the autocorrelation function is used to check the dependence of the sequence by combining actual data such as magnetic flux readings for 10 consecutive cycles, and to determine the dynamic changes in drift velocity. For example, if the velocity accelerates from 0.01T / s to 0.03T / s, it exhibits acceleration characteristics, which helps to understand the dynamic response of the system.

[0053] Specifically, if the change exhibits nonlinear characteristics, a linear regression algorithm is used for fitting. For example, time is used as the independent variable and drift velocity as the dependent variable. The slope of the regression line is calculated using the least squares method to obtain a stable velocity estimate, such as an average of 0.02T / s. This fitting can simplify complex changes and provide a reliable estimation basis.

[0054] For example, when performing secondary analysis based on the estimated values, Monte Carlo simulations are used to generate various scenarios for random characteristics within the eddy current sensitive range, such as resistivity fluctuations in the core material, to determine the final short-term random drift trend. For example, if the trend shows a shift towards higher frequencies, this can reveal the potential thermal effects.

[0055] In one possible implementation, a drift characteristic description is generated by combining the final trend with interval analysis. For example, it can be described as "the magnetic flux drift shows a positive trend in the short term, with a speed of 0.02T / s, and it is recommended to adjust the cooling system." This completes a comprehensive evaluation and provides guidance for optimizing the core performance, achieving higher efficiency and stability.

[0056] S103. Input the short-term random drift trend and the drift speed into the preset sensitive interval prediction model to determine the expected location range of the eddy current sensitive interval in the future several periods.

[0057] The short-term random drift trend and drift velocity are obtained. A pre-defined sensitive interval prediction model is used to process the short-term random drift trend and drift velocity to obtain the predicted location range of the eddy current sensitive interval within several future periods. For the predicted location range, an interval boundary extraction method is used to obtain the upper and lower bounds of the eddy current sensitive interval. Based on the upper and lower bounds, it is determined whether the current position is within the eddy current sensitive interval. If it is, the current period is marked as a sensitive period. From the marked sensitive period sequence, the number of consecutive sensitive period segments and the length of each segment are counted to obtain the sensitive segment distribution information. Based on the sensitive segment distribution information and drift velocity, the movement direction and predicted drift distance of the sensitive interval are determined. A linear extrapolation method is used to process the movement direction and predicted drift distance of the sensitive interval to obtain the corrected location range of the eddy current sensitive interval for the next period.

[0058] Specifically, the preset sensitive interval prediction model is constructed using a "sliding window weighted trend extrapolation algorithm," rather than a general neural network black box. The internal processing logic of this algorithm is as follows: First, a time sliding window of length N is constructed to store the center frequency values ​​of the sensitive interval for the most recent N periods. Second, the difference between the center frequencies of adjacent periods within the window is calculated to obtain a frequency drift velocity sequence. Third, time decay weights are assigned to the drift velocity sequence; the closer the velocity value is to the current moment, the greater the weight, and the farther away, the smaller the weight. The current equivalent drift acceleration is calculated through weighted summation. Fourth, based on the current center frequency, the weighted drift velocity, and the drift acceleration, the center frequency position for the next moment is extrapolated using kinematic principles. Fifth, using the extrapolated center frequency as a benchmark, combined with the current expansion width, the expected position range for the next few periods is defined. This deterministic algorithm logic avoids the convergence risk of complex neural network training and is adaptable to the computing capabilities of embedded controllers. In this invention, the sliding window length N is 10-20 switching cycles, which can be adaptively adjusted according to the LED power supply switching frequency; the time decay weight adopts the rule of larger weight for near periods and smaller weight for far periods, with a weight of 0.8-1.0 for the first 1-5 periods of the current period, a weight of 0.4-0.7 for the 6-10 periods, and a weight of 0.1-0.3 for 11 periods and above; the kinematic extrapolation is based on the current center frequency, weighted drift velocity, and equivalent drift acceleration, and is applicable to LED power supplies with switching frequencies of 50-200kHz.

[0059] For example, in obtaining short-term random drift trends and drift velocities, a base sequence can be constructed by collecting actual eddy current data over the most recent 10 switching cycles.

[0060] Specifically, a sliding window averaging method is first applied to this data to calculate the drift offset for each cycle. For example, with a window size of three cycles, the offset is obtained by subtracting the window average from the current value, thus capturing random short-term fluctuation trends. If the drift velocity needs to be quantified, the rate of change of the offset can be further calculated, such as by dividing the difference in offset between adjacent cycles by the cycle length, to obtain the drift velocity value per second. In this way, a preliminary dynamic assessment of the eddy current sensitive area is formed, providing data support for subsequent predictions.

[0061] like Figure 6As shown, the horizontal axis represents the switching cycle number, and the vertical axis represents the center frequency of the eddy current sensitive region. Solid dots in the figure represent historically collected center frequency data, while dashed boxes mark the sliding window range covering the most recent N=10 cycles. The size and grayscale of each data point within the window change with time decay weights; recent data points are larger and darker, while older data points are smaller and lighter, reflecting an exponential decay time-weighted strategy. The inset in the lower right corner shows the weight distribution of each cycle within the window in bar chart form. The dashed line is the trend line fitted using the weighted least squares method, and the dotted line is the extrapolation prediction curve for the next 5 cycles. The prediction area is represented by a gray shaded band indicating the upper and lower bounds of the prediction width, which gradually expands with increasing prediction step size, reflecting the increase in prediction uncertainty. The figure labels the estimated drift velocity and drift acceleration parameters, as well as the final predicted location range, providing a basis for subsequent frequency avoidance decisions.

[0062] In one possible implementation, when processing these trends and velocities using a pre-defined sensitive interval prediction model, this model can be based on a neural network trained on historical data, such as a simple multilayer perceptron model. The input includes a drift trend vector and velocity values, and the output is the predicted location range for the next five periods. Specifically, the trend data is normalized and input into the model. Internally, the model calculates weights through hidden layers to predict the shift in the center position of the interval. For example, if the current trend is drifting upwards, the model will calculate that the interval may move upwards by 0.5 units based on the velocity, thus obtaining a predicted range expanded from the original position. For instance, if the original interval is [2.0, 3.0], the predicted range becomes [2.3, 3.4]. This processing ensures the continuity and accuracy of the prediction, helping to identify potential risk areas in advance during actual eddy current monitoring operations.

[0063] For example, when using the interval boundary extraction method for the expected location range, the maximum and minimum value filtering method can be used to obtain the upper and lower bound positions.

[0064] Specifically, the highest value is extracted from the data points within the prediction range as the upper bound. For example, if the data range is 2.3, 2.5, 3.0, and 3.4, then the upper bound is 3.4 and the lower bound is 2.3. This method is simple and efficient, connects to the output of the prediction model, and directly provides a boundary benchmark for subsequent judgments.

[0065] In one possible implementation, the current position is determined based on upper and lower bounds to see if it falls within the eddy current sensitive range. For example, if the current position value is 2.8, and it falls within [2.3, 3.4], then the current period is marked as a sensitive period. This determination is achieved through simple numerical comparison. If it falls within the range, a sensitive label is added to the period sequence, such as marking the 4th period in the sequence as "sensitive," thus forming a labeled sequence for statistical analysis.

[0066] For example, when counting the number of consecutive sensitive periods and the length of each segment from a labeled sensitive periodic sequence, one can iterate through the sequence and count consecutive "sensitive" labels. For instance, if the sequence is non-sensitive, sensitive, sensitive, non-sensitive, sensitive, then the number of segments is 2, the length of the first segment is 2, and the length of the second segment is 1, yielding sensitive segment distribution information such as {number of segments: 2, average length: 1.5}. This statistical analysis reveals the clustering patterns of sensitive events, and combined with drift velocity, it can infer the rules governing interval movement.

[0067] In one possible implementation, when determining the direction of movement and the expected drift distance based on the distribution information of sensitive segments and the drift speed, if the distribution shows that the segment extends backward and the speed is positive, then the direction is positive, and the distance is calculated by multiplying the speed by the period duration. For example, if the speed is 0.2 units / period, the expected distance is 1.0 unit. This determination process logically connects statistical results with dynamic parameters, providing a basis for extrapolation.

[0068] For example, when using linear extrapolation to handle the direction of movement and the expected drift distance, the position of the next cycle can be corrected using a linear formula, such as: new upper bound = current upper bound + direction coefficient. The distance, where the direction coefficient is 1 or -1, yields a correction range such as [2.5, 3.6]. This method can effectively optimize monitoring strategies and improve the system's response to random drift in eddy current sensitive areas.

[0069] S104. Generate a basic modulation sequence from the asymmetric pulse width modulation chaotic mapping, and extract the main frequency bands of harmonic energy distribution of the basic modulation sequence.

[0070] A fundamental modulation sequence is generated through asymmetric pulse width modulation (PWM) chaotic mapping. The amplitude of each frequency component is extracted from the fundamental modulation sequence to obtain a harmonic energy sequence. The harmonic energy sequence is segmented and statistically analyzed to obtain the total energy of each frequency band. The three frequency bands with the highest energy are determined by comparing the total energy of each band. If the total energy of a frequency band exceeds twice that of its adjacent bands, it is marked as an energy-concentrated band. Based on the marking results, all energy-concentrated bands are output as the main distribution bands. The main distribution bands are sorted in ascending order of frequency to obtain the final frequency band list.

[0071] Specifically, the asymmetric pulse width modulation chaotic mapping is generated based on an improved logic of piecewise linear mapping (Tent Map). The process of generating the basic modulation sequence does not rely on complex exponential operations, but instead employs the following logic: An asymmetric control parameter K is set between 0 and 1; when the current state value is less than the control parameter K, the next state value is equal to the current state value divided by K; when the current state value is greater than or equal to the control parameter K, the next state value is equal to (1 minus the current state value) divided by (1 minus K). By iterating the above logic in each switching cycle, a sequence value with aperiodic and pseudo-random characteristics is generated, and this sequence value is linearly mapped to the duty cycle or period jitter of the PWM signal, thereby achieving asymmetric diffusion of spectral energy. In this invention, the asymmetric control parameter K of the chaotic mapping takes a value range of 0.2-0.8, preferably 0.5. This parameter is adjusted according to the power level of the LED power supply: 0.2-0.4 for low-power power supplies below 100W, 0.4-0.6 for medium-power power supplies between 100-500W, and 0.6-0.8 for high-power power supplies above 500W. The sequence value is linearly mapped to the PWM duty cycle in the range of 10%-90%, and mapped to the period jitter in the range of ±5% of the switching cycle.

[0072] For example, in the harmonic analysis of power systems, asymmetric pulse width modulation chaotic mapping is a technique used to generate complex signal sequences. It modulates the pulse width by introducing the nonlinear dynamic characteristics of chaotic mapping, thereby generating a basic modulation sequence.

[0073] Specifically, this mapping can be based on a Logistic chaotic model, where the parameters are set between 3.57 and 4 to ensure chaotic behavior, and the pulse width modulation is asymmetric, i.e., the positive and negative pulse widths are not equal, to simulate the unbalanced load in actual power transmission.

[0074] For example, in a motor system controlled by a frequency converter, the initial condition is set to 0.5, and 1000 sequence points are generated iteratively. Each point corresponds to a modulation width value, and finally a basic modulation sequence is formed for subsequent harmonic detection.

[0075] like Figure 5As shown, the left subplot (a) illustrates the geometric characteristics of the Tent Map asymmetric mapping function. The horizontal axis represents the current state value xn, and the vertical axis represents the next state value xn+1. The mapping function is tent-shaped, formed by the intersection of two piecewise linear functions at the asymmetric control parameter K, creating a peak. The figure shows two sets of mapping curves for K=0.3 and K=0.7, demonstrating the asymmetric characteristics of the mapping function at different K values. The diagonal y=x is used as a reference line, and the spider diagram shows the trajectory of state iteration evolution from the initial value, intuitively reflecting the ergodic characteristics of the chaotic mapping. The right subplot (b) shows the time evolution curves of two sets of chaotic sequences corresponding to different K values ​​during 100 iterations. The sequences exhibit non-periodic pseudo-random characteristics, with state values ​​densely distributed between 0 and 1. The gray area marks the PWM duty cycle mapping range from 10% to 90%. The state values ​​of the chaotic sequence, after linear mapping, are used to determine the PWM duty cycle for each switching cycle, achieving a broadened distribution of spectral energy.

[0076] In one possible implementation, the amplitude of each frequency component can be extracted from the basic modulation sequence by Fourier transform. First, the sequence is converted into a frequency domain signal, and then the amplitude value at each frequency point is calculated to obtain the harmonic energy sequence.

[0077] For example, in the aforementioned motor system, a fast Fourier transform is performed on the sequence to extract the amplitudes from the fundamental frequency to higher harmonics, such as a fundamental frequency amplitude of 10, a second harmonic amplitude of 5, a third harmonic amplitude of 3, etc. These amplitude sequences reflect the energy distribution and help identify potential harmonic interference sources.

[0078] Specifically, when performing segmented statistics on harmonic energy sequences, the frequency range can be divided into multiple segments, such as 0-100Hz, 100-200Hz, etc., and the sum of the squares of the amplitudes in each segment can be calculated to obtain the total energy of each frequency band.

[0079] For example, in motor harmonic analysis, the total energy in the 0-100Hz range is 150, and in the 100-200Hz range it is 80, which makes it easier to quantify the energy concentration area.

[0080] For example, after comparing the total energy of each frequency band, the top three highest values ​​can be selected, such as the segments corresponding to 150, 80, and 50, as the top three frequency bands with the highest energy. This helps to prioritize high-energy regions to optimize the system filtering design.

[0081] In one possible implementation, if the total energy of a certain frequency band exceeds twice that of the adjacent frequency band, for example, 150 exceeds twice that of the adjacent 80 (i.e., exceeds 160, but 150 < 160 is not marked), and another band, such as 300, exceeds twice that of the adjacent 100, it is marked as an energy concentration frequency band. This threshold judgment can highlight abnormal energy peaks and is used to warn of equipment failures in power quality monitoring.

[0082] Specifically, based on the marking results, all energy-concentrated frequency bands are output as the main distribution frequency bands. For example, two bands, 200-300Hz and 400-500Hz, are marked. These bands represent the main distribution areas of harmonic energy. This step can provide targeted guidance for subsequent filter design, thereby improving the stability and efficiency of the system.

[0083] For example, the main frequency bands can be sorted in ascending order to obtain the final frequency band list, such as 200-300Hz first and then 400-500Hz. This sorting facilitates the sequential application of compensation strategies in the business process, ensuring that interference is gradually eliminated from low frequency to high frequency, and achieving more accurate power harmonic management.

[0084] S105. Determine the degree of overlap between the main frequency band of the harmonic energy distribution and the expected location range. If the overlap ratio exceeds a preset threshold, activate the frequency avoidance mechanism.

[0085] Acquire the spectral data of the harmonic signal. Spectral analysis is used to obtain the distribution values ​​of harmonic energy at each frequency point. The main distribution frequency band of the harmonic energy is determined based on the distribution values. The expected location range of the eddy current sensitive area is obtained. The intersection length between the main distribution frequency band and the expected location range is calculated. The overlap ratio is obtained by dividing the intersection length by the total length of the main distribution frequency band. It is determined whether the overlap ratio exceeds a preset threshold. If the overlap ratio exceeds the preset threshold, frequency avoidance is required. A set of available backup frequencies is obtained. The frequency point with the lowest overlap ratio with the main distribution frequency band is selected from the set of available backup frequencies. The selected frequency point is used as the avoidance target frequency. The avoidance target frequency is sent to the execution unit to complete the frequency adjustment. In this invention, the set of available backup frequencies is a pre-calibrated frequency range without eddy current loss sensitivity, specifically a frequency band deviating more than 5kHz from the center frequency of the eddy current sensitive area, and this frequency band must ensure the spectral broadening characteristics of PWM modulation. The frequency points in the set are spaced 1kHz apart, and frequency points that do not overlap with the main distribution frequency band of the harmonic energy are preferentially selected.

[0086] In this invention, the "overlap ratio" refers to the percentage of the intersection frequency length between the main harmonic energy distribution frequency band and the expected location range of the eddy current sensitive zone, relative to the total frequency length of the main harmonic energy distribution frequency band. It is the core quantitative indicator for determining whether to activate the frequency avoidance mechanism. The preset threshold for the overlap ratio in this invention is 30%, meaning that when the overlap ratio between the main harmonic energy distribution frequency band and the expected location range of the eddy current sensitive zone exceeds 30%, the frequency avoidance mechanism is immediately activated. This threshold can be fine-tuned according to the loss requirements of the LED power supply; in scenarios with low loss requirements, it can be reduced to 20%.

[0087] like Figure 3As shown, the frequency avoidance adaptive control mechanism receives two input signals: the predicted location range of the eddy current sensitive region and the main frequency band of the harmonic energy distribution of the fundamental modulation sequence. The overlap ratio calculation module calculates the ratio of the intersection frequency length to the total length of the main frequency band to obtain the overlap ratio. The judgment module compares the overlap ratio with a 30% threshold. If the overlap ratio is less than or equal to 30%, the fundamental modulation sequence is bypassed; if the overlap ratio is greater than 30%, the frequency avoidance loop is entered. In the avoidance loop, the offset calculation module uses a piecewise linear mapping to determine the offset, where 30% to 50% corresponds to a 1kHz offset, 50% to 70% corresponds to a 2kHz offset, and above 70% corresponds to a 3kHz offset. The chaotic mapping parameter adjustment module adjusts the control parameter K accordingly, with a parameter change of ±0.1 corresponding to a 0.5kHz frequency offset. The offset modulation sequence generation module outputs the offset modulation sequence, and the spectrum comparison and verification module compares the offset sequence with the fundamental sequence to evaluate the degree of preservation of the broadening characteristics. If the retention rate is less than 80%, the offset is reduced by 0.5kHz and the process returns to the parameter adjustment module for iteration. If the retention rate is greater than or equal to 80%, the optimized modulation sequence is output for PWM drive signal generation, and ripple and temperature data are fed back to the prediction model for updating.

[0088] like Figure 4 As shown, the horizontal axis represents the frequency range of 0 to 300 kHz, and the vertical axis represents the normalized energy density. The gray shaded area represents the eddy current sensitive region, which expands outwards from its center frequency f0. The dashed boundary indicates the drift range of the sensitive region caused by changes in magnetic flux density and temperature. The solid curve represents the harmonic energy distribution of the basic modulation sequence after chaotic mapping. Multiple harmonic peaks are distributed across different frequency bands, with some peaks falling within the eddy current sensitive region, forming overlapping areas marked by cross-shading. The dashed curve represents the harmonic energy distribution of the modulation sequence after adaptive adjustment of the chaotic mapping control parameter K. The harmonic peaks are moved away from the sensitive region through frequency offset avoidance operations, with arrows indicating the direction and amount of frequency offset. After offsetting, the overlap between the main harmonic energy distribution frequency band and the eddy current sensitive region is significantly reduced, effectively suppressing core eddy current losses.

[0089] In one possible implementation, the spectral data of the harmonic signal can be obtained through Fourier transform.

[0090] For example, the acquired harmonic signal is input into a digital signal processor. First, the signal is sampled, with the sampling frequency set to more than twice the highest frequency of the signal to avoid aliasing. Then, the spectrum is calculated using the Fast Fourier Transform algorithm to obtain the amplitude value at each frequency point. The square of these amplitude values ​​can represent the harmonic energy distribution.

[0091] Specifically, in power systems, harmonic signals may originate from nonlinear loads generated by frequency converters. The processor converts the time-domain signal into the frequency domain and outputs an array, where each element corresponds to the energy value of a frequency point. For example, in the range from 0Hz to 1000Hz, there is a point every 1Hz. The energy value is calculated through the amplitude spectrum, thus forming a complete spectrum data foundation to support subsequent analysis.

[0092] For example, when determining the main frequency bands for harmonic energy distribution, a threshold screening can be applied to these energy values. First, the average energy value of all frequency points is calculated. Then, points exceeding twice the average value are grouped into a high-energy group. Subsequently, clustering methods such as the K-means algorithm are used to aggregate these points into continuous frequency bands. Assuming the clustering results show that the energy is mainly concentrated in the 200-300Hz and 500-600Hz bands, these bands are considered the main distribution frequency bands. This method ensures the continuity and representativeness of the frequency bands and avoids interference from isolated points.

[0093] In one possible implementation, obtaining the expected location range of the eddy current sensitive region involves querying the system configuration.

[0094] For example, in wireless communication services, the projected range might be a pre-allocated channel, such as a frequency band from 400-500Hz. This can be obtained by reading device configuration files or databases to ensure it matches the actual service. Calculating the intersection length between the main distributed frequency bands and the projected location range is relatively straightforward.

[0095] For example, if the main frequency bands are 200-300Hz and 500-600Hz, and the expected range is 250-550Hz, then the intersection is 250-300Hz (length 50Hz) and 500-550Hz (length 50Hz), with a total intersection length of 100Hz. The overlap ratio is obtained by dividing the intersection length by the total length of the main frequency bands.

[0096] For example, if the total length of the main frequency band is (300-200)+(600-500)=200Hz, and the intersection is 100Hz, then the ratio is 100 / 200=0.5 or 50%. Determine if the overlap ratio exceeds a preset threshold, such as 30%. If it does, frequency avoidance is required, which helps reduce interference and improve system stability.

[0097] In one possible implementation, the set of available spare frequencies can be obtained from a pre-defined frequency pool.

[0098] For example, the set includes unused frequency bands such as 700-800Hz and 900-1000Hz, and availability is ensured through system queries. The frequency point with the lowest overlap with the main distributed frequency bands is selected from the available spare frequency set.

[0099] For example, the overlap between each candidate point and the main frequency band is calculated. If the overlap in the 700-800Hz range is 0, then that point is selected. The selected frequency point is then sent to the execution unit as the avoidance target frequency.

[0100] For example, commands can be sent to the tuner via a network interface to complete adjustments, thereby optimizing service performance.

[0101] S106. Calculate the offset based on the overlap ratio, and adjust the control parameters of the chaotic mapping through an adaptive frequency offset mechanism to obtain the offset modulation sequence.

[0102] Obtain the overlap ratio between the current sequence and the reference sequence. Determine the offset value based on the overlap ratio using a preset mapping relationship. Input the offset value into the adaptive frequency offset mechanism. Adjust the control parameters of the chaotic mapping using the adaptive frequency offset mechanism to obtain the adjusted control parameters. Drive the chaotic mapping operation using the adjusted control parameters to generate an initial modulation sequence. Perform offset processing on the initial modulation sequence to obtain the offset modulation sequence. Output the offset modulation sequence as the base sequence for subsequent communication modulation. In this invention, the preset mapping relationship between the overlap ratio and the offset is a piecewise linear mapping: for an overlap ratio of 30%-50%, the offset is 1.0kHz; for 50%-70%, the offset is 2.0kHz; and for over 70%, the offset is 3.0kHz. The adaptive frequency offset mechanism achieves frequency offset by adjusting the control parameter K of the chaotic mapping. For every 0.1 adjustment of K, the corresponding frequency offset is 0.5kHz, and the adjusted K value remains within the range of 0.2-0.8.

[0103] In one possible implementation, to obtain the overlap ratio between the current sequence and the reference sequence, the two sequences can first be aligned and compared. For example, in a wireless communication system, the current sequence may be based on a bit stream of historical modulated data, while the reference sequence originates from a base sequence defined by a standard protocol. The overlap ratio can be obtained by calculating the number of matching bits at the same position and dividing it by the total bit length.

[0104] Specifically, if the current sequence is "10101011" and the reference sequence is "10101101", there are 6 matching bits, a total length of 8 bits, and an overlap ratio of 0.75. This calculation process helps to assess the similarity of the sequences, thus providing a basis for subsequent adjustments and avoiding signal interference.

[0105] For example, the process of determining the offset value based on the overlap ratio using a preset mapping relationship can be achieved by setting up a mapping table where different offset values ​​correspond to different ratio ranges. For instance, the offset is 2 when the ratio is between 0.5 and 0.75, and 3 when it exceeds 0.75. In actual operations, this is similar to the offset adjustment of data frames in satellite communications to ensure transmission efficiency.

[0106] In one possible implementation, the offset value is input into an adaptive frequency offset mechanism, which then dynamically adjusts its parameters based on the input. The adaptive frequency offset mechanism is a feedback-based control system used to optimize signal frequency distribution. Its principle is to correct frequency deviations by monitoring the offset in real time, thereby maintaining system stability.

[0107] For example, in mobile communication networks, when the offset is 3, the mechanism automatically increases the frequency step size to reduce potential spectrum conflicts.

[0108] Specifically, when adjusting the control parameters of a chaotic map using an adaptive frequency offset mechanism, the chaotic map, a nonlinear dynamic system often used to generate pseudo-random sequences, typically has control parameters such as the r value in a Logistic map, usually between 3.57 and 4. The adjustment process involves multiplying the offset by a scaling factor and adding it to the original parameters to obtain the adjusted control parameters.

[0109] For example, if the original r is 3.8, the offset is 0.2, and the scaling factor is 0.5, then the new r is 3.9. This adjustment enhances the randomness of the sequence and is suitable for encrypted communication.

[0110] For example, the process of generating an initial modulation sequence by driving a chaotic map with adjusted control parameters can be calculated iteratively using the Logistic equation x_{n+1}=r x_n (1-x_n), with an initial value of 0.5, and run 100 times to obtain the sequence. This sequence serves as the modulation basis and can be used to modulate carrier signals in IoT device communication, improving anti-interference capabilities.

[0111] In one possible implementation, the initial modulation sequence is offset, and the resulting offset modulation sequence can be achieved through a cyclic shift operation, such as shifting the sequence left by the offset number of bits. If the initial sequence is "001101", an offset of 2 bits will result in "110100". This process helps optimize the autocorrelation property of the sequence.

[0112] Specifically, after the offset modulation sequence is output as the base sequence for subsequent communication modulation, it can be directly applied to QPSK modulation to achieve data encoding in 5G networks, thereby improving transmission reliability.

[0113] S107. Compare the spectrum of the offset modulation sequence with that of the basic modulation sequence to determine the degree of preservation of the spectrum broadening characteristics and the changes in energy distribution.

[0114] Step 1: Acquire the modulation signal data of the offset sequence and the base sequence. Perform preliminary sampling and digitization on both sequences using a pre-established signal processing module to obtain digitized modulation signals for both sequences. Step 2: For the digitized modulation signals, use a spectrum analysis tool to perform frequency domain transformation on both the offset sequence and the base sequence to determine their spectral data. Step 3: Extract the broadening characteristic parameters of the offset sequence and the base sequence from the spectral data. Analyze the differences in broadening characteristics between the two sequences in the frequency domain through comparison to determine the retention status. Step 4: Based on the results of the broadening characteristic differences, acquire the energy distribution data of both sequences and calculate the changes in energy distribution using the same spectrum analysis tool. Step 5: If the changes in energy distribution exceed a preset threshold, perform secondary processing on the modulation signal of the offset sequence to obtain adjusted spectral data. Step 6: By comparing the adjusted spectral data with the spectral data of the base sequence, analyze the specific changes in distribution differences to determine the final matching degree of the spectral broadening characteristics and energy distribution. In this invention, the preset threshold for energy distribution change is 10%. That is, when the rate of change of energy distribution of the offset sequence relative to the base sequence exceeds 10%, the offset sequence modulation signal is subjected to secondary processing. The secondary processing adopts an equalization filtering method to reduce the rate of change of energy distribution to within 10%.

[0115] The "degree of preservation of spectrum broadening characteristics" mentioned in this invention refers to the proportion of the offset modulation sequence relative to the basic modulation sequence in terms of 3dB bandwidth and spectrum uniformity, expressed as a percentage. The closer the value is to 100%, the more complete the broadening characteristics are preserved. At the same time, it is necessary to ensure that there is no local concentration of harmonic energy.

[0116] Specifically, when acquiring the modulation signal data of the offset sequence and the base sequence, it is first necessary to extract these signals from the communication system.

[0117] For example, in wireless communication networks, the offset sequence might originate from a parameter-tuned chaotic generator, while the base sequence serves as a standard reference to ensure signal stability. A pre-built signal processing module can be understood as an integrated hardware or software unit, including components such as analog-to-digital converters and filters, used to process analog signals. Specifically, the process begins with the radio frequency signal received by the antenna, converting the high-frequency signal to an intermediate frequency through down-conversion, and then using the sampling theorem for uniform sampling, for example, capturing signal details at twice the Nyquist frequency to avoid aliasing, thus obtaining a digitally modulated signal. These signals are stored in discrete numerical form for easy subsequent calculations.

[0118] In one embodiment, when performing frequency domain conversion on a digitally modulated signal, the spectrum analysis tool typically refers to a software tool that implements the Fourier transform algorithm, such as a Fast Fourier Transform module, which converts the time-domain signal into a frequency-domain representation.

[0119] For example, when processing offset sequences, a window function is first applied to the signal to reduce leakage effects, and then its spectral amplitude and phase are calculated. The same operation is performed on the base sequence to finally determine the spectral data, which includes parameters such as main lobe width and side lobe level, to help reveal the frequency component distribution of the signal.

[0120] For example, the process of extracting broadening characteristic parameters from spectral data involves identifying the bandwidth indicators of the spectrum. Broadening characteristic parameters may include 3dB bandwidth or effective bandwidth, etc. By comparing these parameters of the offset sequence and the base sequence, the differences such as the increase or decrease in bandwidth caused by the offset are analyzed to determine the preservation status, that is, to assess whether the offset has maintained the original spectral integrity without introducing too much distortion.

[0121] Specifically, when obtaining energy distribution data based on the results of the difference in broadening characteristics, energy distribution data refers to the power integral of each frequency component in the spectrum. The same spectrum analysis tools are used to calculate the changes, such as integrating a specific frequency band through the power spectral density function and comparing the energy shift of the offset sequence relative to the base sequence. This helps to quantify the impact of the shift on the signal energy.

[0122] In one embodiment, if the energy distribution changes beyond a preset threshold, such as a change rate exceeding 10%, the modulated signal of the offset sequence is subjected to secondary processing, for example, applying an equalization filter to adjust the amplitude response, to obtain adjusted spectral data. This processing aims to restore energy balance.

[0123] For example, by comparing the adjusted spectral data with the spectral data of the base sequence, the specific changes in distribution differences, such as the degree of main lobe alignment or side lobe suppression ratio, can be analyzed to determine the final spectral broadening characteristics and the degree of matching of energy distribution. This can bring about the technical effect of signal compatibility in services, ensuring the reliable transmission of the communication system.

[0124] S108. If the degree of preservation of the spectrum broadening characteristic is lower than the preset preservation range, then reduce the offset and re-execute the frequency avoidance mechanism.

[0125] Obtain the current spectral broadening characteristic retention value. Determine if the retention value is below a preset retention range. If the retention value is below the preset retention range, obtain the current offset value. Perform a reduction process on the current offset value to obtain an adjusted offset value. Replace the original offset value with the adjusted offset value. Re-execute the frequency avoidance mechanism based on the adjusted offset value. Obtain the spectral broadening characteristic retention value after re-executing the frequency avoidance mechanism. In this invention, the preset retention range for the spectral broadening characteristic is 80%-100%, meaning that the spectral broadening characteristic retention of the offset modulation sequence must be no less than 80%. The step size for offset reduction is 0.5kHz, and the frequency avoidance mechanism is re-executed after each reduction until the retention falls within the preset range.

[0126] In one possible implementation, when it is necessary to assess the degree to which the spectral broadening characteristics are preserved, relevant parameters are first extracted from the current modulated signal using signal analysis software. For example, the time-domain signal is converted into frequency-domain data using Fourier transform, thereby calculating a quantitative index of the broadening characteristics. This process involves integrating the power spectral density distribution of the signal to obtain a value representing the spectral width, such as a percentage of bandwidth.

[0127] Specifically, in wireless communication systems, if the spectrum broadening of the basic signal is a standard value, the degree of preservation of the current signal can be determined by comparing the similarity of the two spectra. For example, by using the correlation coefficient calculation method, the current spectrum curve is matched point-to-point with the reference curve to obtain a value between 0 and 1, representing the percentage of preservation.

[0128] For example, if the calculated retention value is 0.75, while the preset retention range is 0.8 to 1.0, it needs to be determined whether it is below the lower limit of this range. In this case, the system will automatically trigger an alarm mechanism and record the current status log for subsequent optimization.

[0129] In one possible implementation, once it's confirmed that the retention level is below a preset range, the current offset value is obtained. This offset typically refers to the displacement adjustment of the signal in the frequency domain, such as a fine-tuning value on the carrier frequency, for example, shifting from an initial 5kHz to the current 3kHz. This value can be directly read by querying the configuration parameters of the signal processing module, avoiding errors caused by manual intervention.

[0130] Specifically, this acquisition process may involve reading data from hardware registers or retrieving log records from a software database to ensure the accuracy of the offset, thereby providing a reliable basis for subsequent adjustments.

[0131] For example, when performing a reduction process on the current offset, the value can be subtracted by a fixed step size. For instance, if the original offset is 4kHz, the reduction process will result in an offset of 2kHz. This step size is set based on empirical values ​​to gradually approach the optimal position.

[0132] In one possible implementation, after replacing the original value with the adjusted offset, the system updates the configuration file, for example, by reloading the parameters in the digital signal processor, to ensure that the new offset takes effect immediately.

[0133] Specifically, this replacement process includes backing up the original configuration, writing the new value, and verifying consistency to avoid system interruption and thus maintain the stability of the communication link.

[0134] For example, when the frequency avoidance mechanism is re-executed based on the adjusted offset, an interference environment is simulated, such as in a multi-user wireless network, the occupancy of adjacent channels is detected, and then the signal transmission frequency is adjusted to avoid collisions.

[0135] In one possible implementation, a new value for the degree of preservation of spectral broadening characteristics is obtained after re-execution. By repeating the aforementioned frequency domain transformation and comparison steps, an updated index is obtained, for example, a new value of 0.85, which helps to verify the effectiveness of the adjustment.

[0136] Specifically, if the value stabilizes within the preset range after multiple iterations, reliable signal transmission can be ensured. In actual satellite communication services, this mechanism can reduce packet loss caused by interference and improve overall system performance.

[0137] S109. If the degree of preservation of the spectrum broadening characteristic meets the preset preservation range, the offset modulation sequence is output for the generation of the pulse width modulation drive signal in the next cycle.

[0138] Acquire the spectrum broadening data for the current period. Calculate the spectrum broadening preservation level using the spectrum analysis module. If the spectrum broadening preservation level is within a preset preservation range, acquire the offset-processed modulation sequence. Perform offset processing on the original modulation sequence using the sequence offset module to obtain the offset sequence. Generate the pulse width modulation parameters for the next period based on the offset sequence. Convert the pulse width modulation parameters into a drive signal waveform using the pulse width modulation generator. Output the drive signal waveform to the power conversion circuit.

[0139] For example, in a power conversion system, it is first necessary to obtain the spectrum broadening data for the current cycle, which involves real-time monitoring of the signal spectrum distribution in the circuit.

[0140] Specifically, this data can be extracted from the converter's output waveform using a digital signal processor. For example, in switching power supply applications, the acquired data might include broadening values ​​in the frequency range from 10kHz to 100kHz to ensure electromagnetic compatibility. Next, a spectrum analysis module calculates the degree of spectral broadening preservation. This module, typically based on the Fourier transform principle, performs frequency domain analysis on the acquired data to calculate the uniformity and stability of the broadening.

[0141] For example, if the preset hold range is 80% to 95%, the module will compare the actual broadening with the ideal broadening ratio. The process includes converting the time-domain signal to the frequency domain, applying a window function to reduce leakage, and quantifying the hold level as a percentage value to determine whether adjustment is needed.

[0142] In one possible implementation, if the degree of spectral broadening preservation is within a preset preservation range, the modulation sequence after offset processing is obtained. Here, offset processing aims to optimize the randomness of the signal to reduce noise peaks.

[0143] For example, in the inverter control of electric vehicles, the system reads a processed sequence from memory. This sequence is generated using a pre-defined algorithm to ensure that the broadening effect is not diminished. A sequence offset module is used to perform offset processing on the original modulation sequence to obtain the offset sequence. This module works by introducing a pseudo-random offset.

[0144] For example, each element of the original sequence is added with a value based on a noise generator. The process includes initializing the offset vector, adding element by element, and normalizing the result to produce a more uniform spectral distribution. In practical applications, this can be used in industrial frequency converters to avoid harmonic interference at specific frequencies.

[0145] For example, the pulse width modulation (PWM) parameters for the next cycle are generated based on the offset sequence. This step maps the offset sequence to the modulation ratio and duty cycle. For instance, in a photovoltaic inverter system, the sequence value determines the on-time of each switching cycle. A list of parameters, such as duty cycles from 0.5 to 0.8, is generated to ensure a smooth transition between consecutive cycles. The PWM parameters are then converted into a drive signal waveform using a PWM generator. This generator operates based on a comparator and a clock signal. The process includes loading parameters into a register, comparing a triangular wave with a reference value to generate a square wave, and adding a dead time to prevent short circuits. In wind power generation circuits, this enables efficient energy conversion and reduces electromagnetic interference.

[0146] In one possible implementation, the drive signal waveform is output to the power conversion circuit, which ensures that the signal directly drives devices such as IGBTs or MOSFETs. For example, in a high-speed rail traction system, the circuit amplifies the power after the waveform is output to achieve motor control. The entire process from data acquisition to output forms a closed loop, supporting the stable operation of the system.

[0147] S110. Feedback the ripple data and temperature increment collected during actual operation to the sensitive interval prediction model to update the internal state of the sensitive interval prediction model.

[0148] The process acquires ripple data and temperature increments generated during operation. The ripple data and temperature increments are synchronized and aligned using a data acquisition module to obtain a time-matched ripple-temperature sequence. This time-matched ripple-temperature sequence is input into the sensitive interval prediction model. Based on the current internal state of the sensitive interval prediction model, forward computation is performed on the input ripple-temperature sequence to obtain the current sensitive interval estimate. The deviation between the current sensitive interval estimate and the actual acquired ripple data is calculated. If the deviation exceeds a preset threshold, a model parameter adjustment process is triggered, updating the learnable parameters of the sensitive interval prediction model using a gradient descent method. The updated learnable parameters are written back to the sensitive interval prediction model, completing the internal state refresh. In this invention, the gradient descent method uses small-step iterative updates with a learning rate of 0.01-0.05 and 3-5 iterations. Each iteration only fine-tunes the learnable parameters of the sensitive interval prediction model to avoid parameter mutations that could lead to inaccurate model predictions. The loss function is constructed based on the deviation between the sensitive interval estimate and the actual ripple data, with minimizing the deviation as the optimization objective. In this invention, the preset threshold for the deviation of ripple data is 0.2V (peak-to-peak value), and the preset threshold for the deviation of temperature increment is 2℃. When either deviation exceeds the threshold, the model parameter adjustment process is triggered. The ripple data acquisition frequency is 10Hz, and the temperature increment acquisition frequency is 1Hz. The two are synchronized through timestamps.

[0149] The internal state of the sensitive interval prediction model described in this invention refers to the set of core parameters, historical data, and intermediate calculation results stored during model operation. Specifically, it includes the model's learnable parameters, historical center frequency and drift velocity data of the eddy current sensitive interval stored in the sliding window, a deviation correction coefficient table, and the currently calculated equivalent drift acceleration and weighted drift velocity. This internal state directly determines the calculation logic and output results of the model in predicting the expected location range of the eddy current sensitive interval. Real-time updating of the model's internal state is the core foundation for achieving accurate dynamic prediction of the eddy current sensitive interval.

[0150] Specifically, the model parameter adjustment process employs a "deviation correction lookup table method." When the actual collected ripple data exceeds a preset safety threshold, it indicates that the current frequency avoidance has failed, meaning there is a deviation in the predicted sensitive interval position. At this time, the system executes the following correction logic: if the current actual operating frequency is higher than the upper limit of the predicted interval and high ripple occurs, it indicates that the actual position of the sensitive interval is too high, and a positive offset is added to the correction coefficient table of the prediction model; if the current actual operating frequency is lower than the lower limit of the predicted interval and high ripple occurs, it indicates that the actual position of the sensitive interval is too low, and a negative offset is added to the correction coefficient table of the prediction model. Through this lookup table correction based on actual feedback, the baseline parameters of the prediction model are gradually calibrated, making it more adaptable to characteristic drift caused by core aging. The correction coefficient table of the deviation correction lookup method in this invention uses the core temperature increment (°C) and the change in magnetic flux density (T) as two-dimensional indexes. The positive / negative offset values ​​range from 0.5 to 2.0 kHz. For every 5°C increase in temperature increment and every 0.1T increase in magnetic flux density change, the offset increases by 0.5 kHz. The correction coefficient table is calibrated according to the core material type (ferrite / silicon steel) to adapt to different LED power supply cores.

[0151] In one possible implementation, the ripple data generated during operation can be acquired in real time using a high-precision current sensor.

[0152] For example, in a switching power supply system, a sensor records the fluctuation range of the output voltage every 0.5 seconds, while a thermistor monitors the temperature change of the circuit board, calculating the increment value from a reference temperature of 25 degrees Celsius. The data collected in this way includes peak-to-peak ripple values ​​such as 2.3 volts and temperature increments such as 3.7 degrees Celsius, forming a raw dataset for subsequent analysis.

[0153] Specifically, the data acquisition module uses a timestamp matching algorithm to synchronize and align ripple data and temperature increments.

[0154] For example, by aligning the acquisition time of ripple data with the acquisition time of temperature increment, samples with a time deviation of less than 0.1 seconds are considered as a matching pair, resulting in a time-matched ripple-temperature sequence. For instance, in the sequence, a ripple of 2.3 volts at time 1 corresponds to a temperature increment of 3.7 degrees Celsius, and a ripple of 1.8 volts at time 2 corresponds to a temperature increment of 4.2 degrees Celsius. This ensures the temporal consistency of the sequence and avoids analysis errors caused by asynchronous sampling.

[0155] For example, this time-matched ripple-temperature sequence can be input into a sensitive interval prediction model. This model can be a recurrent neural network-based structure containing hidden layers to capture sequence dependencies. Specifically, the sequence is taken as an input vector, converted into a feature representation by the input layer, and then the state is updated in the hidden layer. For example, the current hidden state is combined with the state at the previous time step and the current input to calculate the new state. Finally, the output layer generates a sensitive interval estimate, such as predicting the voltage sensitive interval as 1.5 to 2.5 volts.

[0156] In one possible implementation, when performing forward computation on the input ripple-temperature sequence based on the current internal state of the sensitive interval prediction model, the sequence is processed element by element.

[0157] For example, the model state is first initialized as a zero vector, then matrix multiplication and activation function operation are performed on the first ripple-temperature pair to obtain an intermediate output. Then, subsequent elements are recursively processed until the end of the sequence, and the estimated value of the sensitive interval at the current time, such as 2.0 volts, is output. This reflects the model's stepwise reasoning ability on time series data.

[0158] Specifically, when calculating the deviation using the current sensitive interval estimate and the actual collected ripple data, the absolute difference formula can be used.

[0159] For example, the deviation between the estimated value of 2.0 volts and the actual ripple of 2.3 volts is 0.3 volts. If this deviation exceeds a preset threshold such as 0.2 volts, it indicates that the model prediction is inaccurate and further intervention is needed.

[0160] For example, if the deviation exceeds a preset threshold, the model parameter adjustment process is triggered. The learnable parameters of the prediction model in the sensitive interval are updated by using the gradient descent method. Specifically, the gradient of the loss function is calculated, and the weight matrix is ​​adjusted by a small step size, such as 0.01. After several iterations, the parameters are updated from the initial 0.5 to 0.48, thereby improving the model's adaptability to new data. This adjustment helps maintain the accuracy of the prediction.

[0161] In one possible implementation, when the updated learnable parameters are written back to the sensitive interval prediction model, the original parameter values ​​will be directly overwritten.

[0162] For example, new weights can be injected through the model's loading interface, and then a validation sequence can be tested to confirm the refresh effect, thus completing the internal state refresh and ensuring that the model can better handle ripple-temperature data in the next cycle.

[0163] like Figure 9As shown, this invention conducted comparative tests on eddy current losses before and after optimization at five typical power levels: 50W, 100W, 200W, 300W, and 500W. The left vertical axis displays the absolute values ​​of eddy current losses before and after optimization in the form of a bar chart, where solid black bars represent eddy current losses before optimization, and gray diagonally filled bars represent eddy current losses after optimization. At the 50W power level, eddy current losses decreased from 0.6W to 0.3W; at the 100W power level, from 1.2W to 0.5W; at the 200W power level, from 2.5W to 1.2W; at the 300W power level, from 4.0W to 1.8W; and at the 500W power level, from 5.8W to 2.3W. The right vertical axis displays the percentage reduction in losses at each power level in the form of a line graph, which are 50%, 58%, 52%, 55%, and 60%, respectively. As can be seen from the figure, the eddy current loss optimization scheme of the present invention can achieve a loss reduction effect of more than 50% across the entire power range, and the optimization effect is more significant at higher power levels, reaching a maximum reduction of 60% at 500W.

[0164] like Figure 10 As shown, this invention compared the conversion efficiency of a traditional fixed-frequency PWM scheme, a simple chaotic modulation PWM scheme, and the frequency avoidance PWM scheme proposed in this invention within a load range of 10% to 100%. The three curves are distinguished by dashed lines with square markers, dashed lines with triangle markers, and solid lines with dots, respectively. Each curve is accompanied by a light gray error band representing a measurement uncertainty of ±0.5%. The test results show that the conversion efficiency of the traditional fixed-frequency PWM scheme reaches a peak of 88% at 80% load, and then begins to decline; the simple chaotic modulation PWM scheme reaches a peak of 90.2% at 80% load; while the frequency avoidance PWM scheme of this invention reaches a maximum conversion efficiency of 93% at 80% load, significantly outperforming the other two schemes. Under 100% full load conditions, the conversion efficiency of the proposed scheme is 92.5%, an improvement of approximately 5 percentage points compared to the traditional scheme's 87%, fully verifying the technical advantages of the frequency avoidance strategy in improving power conversion efficiency.

[0165] like Figure 11As shown, this invention uses a bi-sub-plot heatmap to comprehensively display the changes in output ripple and core temperature rise over operating time at five power levels. The upper sub-plot displays output ripple data, and the lower sub-plot displays core temperature rise data. Each sub-plot is divided into two parts by a vertical line, corresponding to the test results before and after optimization, respectively. Regarding output ripple, taking the 500W power level as an example, after 8 hours of operation, the ripple value before optimization was 3.5V, which decreased to 2.0V after optimization, a reduction of 42.9%. At the 50W power level, the ripple was 2.2V after 8 hours of optimization, which decreased to 1.0V after optimization, a reduction of 54.5%. Regarding core temperature rise, at the 500W power level, after 8 hours of operation, the temperature rise before optimization was 45 degrees Celsius, which decreased to 30 degrees Celsius after optimization, a reduction of 33.3%. The color depth in the heatmap intuitively reflects the numerical values; the optimized area is predominantly light gray, indicating that both ripple and temperature rise are effectively controlled. As can be seen from the figure, the present invention can significantly reduce output ripple and core temperature rise across the entire power range and operating cycle, with particularly outstanding optimization effects under high-power, long-term operating conditions.

[0166] Summary of Supplementary Notes to the Instruction Manual I. Specific Implementation of the Sensitive Interval Prediction Model The "sensitive interval prediction model" described in this invention employs a "sliding window weighted trend extrapolation algorithm," which is a deterministic, non-black-box prediction model. The specific implementation steps are as follows: 1. Set a sliding window of length N (N ranges from 10 to 20 switching cycles, and is automatically adjusted according to the LED power supply switching frequency; N=15 at 50kHz, N=12 at 100kHz, and N=10 at 200kHz).

[0167] 2. In each switching cycle, store the center frequency value (unit: kHz) of the current eddy current sensitive range at the end of the sliding window, and remove the oldest data point at the front of the window to keep the window length constant.

[0168] 3. Calculate the difference between the center frequencies of adjacent periods within the window to obtain the frequency drift velocity sequence. For example, if the center frequencies of 5 consecutive periods in the window are [550, 552, 555, 558, 560], then the drift velocity sequence is [2, 3, 3, 2].

[0169] 4. Assign time decay weights to the drift velocity sequence. The weight allocation rules are as follows: the weights for the first 1-5 periods before the current period are 0.8-1.0, the weights for the 6-10 periods are 0.4-0.7, and the weights for the 11th period and above are 0.1-0.3. For example, for the drift velocity sequence [2, 3, 3, 2], if the current period is the 5th period, the weight allocation is [0.9, 0.8, 0.7, 0.6].

[0170] 5. Calculate the current equivalent drift acceleration by weighted summation, which is the average drift velocity after weighting. For example, the weighted summation is (2×0.9 + 3×0.8 + 3×0.7 + 2×0.6) / (0.9+0.8+0.7+0.6) =2.5.

[0171] 6. Based on the current center frequency, weighted drift velocity, and equivalent drift acceleration, linear extrapolation is performed using kinematic principles to calculate the predicted position range for the next few periods. The specific extrapolation formula is: Future center frequency = Current center frequency + Drift velocity × Number of predicted periods + Drift acceleration × (Number of predicted periods) 2 ) / 2. For example, if the current center frequency is 560kHz, the drift velocity is 2.5kHz / cycle, and the drift acceleration is 0.5kHz / cycle. 2 Predicting the next two cycles, the future center frequency = 560 + 2.5 × 2 + 0.5 × (2 2 ) / 2 = 566kHz.

[0172] 7. Using the calculated future center frequency as a reference, and combining it with the current bandwidth (in kHz), define the expected location range for several future cycles. For example, if the bandwidth is 150 kHz, then the expected location range is [566-75, 566+75] = [491, 641] kHz.

[0173] The sensitive interval prediction model of this invention does not rely on training of complex neural networks, but is based on deterministic calculations of historical data, which ensures the feasibility and reliability of the model, and has a small computational load, making it suitable for real-time operation of embedded controllers.

[0174] II. Specific Implementation of the Support Vector Machine Algorithm The "Support Vector Machine algorithm" described in this invention is used to determine the center frequency range of the eddy current sensitive region. The specific implementation steps are as follows: 1. Extract historical data from a pre-built database of magnetic core material properties, including magnetic flux density values ​​(unit: T), temperature values ​​(unit: °C), center frequency range (unit: kHz), etc., to form a training sample set.

[0175] 2. Combine the magnetic flux density value and temperature value of each sample into a two-dimensional feature vector. For example, the feature vector of a sample with a magnetic flux density value of 1.2T and a temperature value of 50℃ is [1.2, 50].

[0176] 3. The radial basis function (RBF kernel) is adopted as the kernel function of the support vector machine. This kernel function can effectively handle nonlinear classification problems and performs well in the application scenario of this invention.

[0177] 4. The parameter γ of the RBF core is 0.5-2.0, adjusted according to the core material type: 1.0 for ferrite cores and 1.5 for silicon steel cores. The penalty parameter C is 1.0-5.0, 2.0 for ferrite cores and 3.0 for silicon steel cores.

[0178] 5. Train the support vector machine model using historical data, determine the optimal parameter combination through cross-validation, and ensure that the model's classification accuracy on the test set is not less than 90%.

[0179] 6. For new magnetic flux density and temperature values, the model inputs the feature vector, calculates the classification result (center frequency range) through a kernel function, and then outputs the center frequency range. For example, with an input magnetic flux density of 1.2T and a temperature of 50℃, the model outputs a center frequency range of [500, 600]kHz.

[0180] 7. If the center frequency range exceeds the preset threshold (50-200kHz), the temperature data and magnetic flux density data will be corrected a second time, with the correction range being temperature ±2℃ and magnetic flux density ±0.1T, and the center frequency range will be recalculated.

[0181] The support vector machine algorithm in this invention ensures the accuracy and reliability of classification results by reasonably selecting kernel functions and parameters, and has a moderate computational load, making it suitable for real-time operation in embedded systems.

[0182] III. Specific Implementation of Asymmetric Pulse Width Modulation Chaotic Mapping The "asymmetric pulse width modulation chaotic mapping" described in this invention is based on an improved logic of piecewise linear mapping (Tent Map), and the specific implementation steps are as follows: 1. Set an asymmetric control parameter K between 0.2 and 0.8. The specific value should be adjusted according to the power level of the LED power supply: 0.2-0.4 for low power supply below 100W, 0.4-0.6 for medium power supply between 100-500W, and 0.6-0.8 for high power supply above 500W.

[0183] 2. The initial state value is set to 0.5, which is a typical starting point for chaotic mapping.

[0184] 3. In each switching cycle, calculate the next state value based on the relationship between the current state value and the control parameter K: If the current state value is less than the control parameter K, then the next state value = current state value / K If the current state value is greater than or equal to the control parameter K, then the next state value = (1 - current state value) / (1 - K) 4. Through the above iterative logic, a sequence of values ​​with aperiodic and pseudo-random characteristics is generated. For example, when K=0.5, the iteration process is as follows: The initial value is 0.5. Since it equals K, the next state value = (1-0.5) / (1-0.5) = 1.0 Since the state value is 1.0, which is greater than K, the next state value = (1-1.0) / (1-0.5) = 0.0 The state value is 0.0. Since it is less than K, the next state value = 0.0 / 0.5 = 0.0 The state value is 0.0. Since it is less than K, the next state value = 0.0 / 0.5 = 0.0 The state value is 0.0. Since it is less than K, the next state value = 0.0 / 0.5 = 0.0 ...(In practical applications, more sequence points will be generated through iterative processes.) 5. Linearly map the generated chaotic sequence values ​​to the duty cycle or period jitter of the PWM signal, with a mapping range of 10%-90% duty cycle or ±5% switching period jitter.

[0185] 6. By controlling the setting of parameter K, asymmetric characteristics can be achieved, that is, the positive and negative pulse widths are not equal. This helps to optimize the avoidance effect of specific frequencies while maintaining spectrum broadening.

[0186] The asymmetric pulse width modulation chaotic mapping of the present invention is implemented through simple iterative logic, with low computational load, easy to implement in embedded systems, and can effectively broaden the spectrum and reduce electromagnetic interference peaks.

[0187] IV. Basis for Determining the Threshold The threshold parameters involved in this invention have been determined through extensive experimental verification and engineering practice, and the specific determination criteria are as follows: 1. Preset Threshold for Overlap Ratio (30%): Statistical analysis of extensive experimental data shows that when the overlap ratio between the main harmonic energy distribution frequency band and the expected location range of the eddy current sensitive area exceeds 30%, eddy current losses will increase significantly, leading to increased ripple and temperature. In tests of 50-500W high-power LED power supplies, 30% was found to be the critical point where eddy current losses begin to increase significantly.

[0188] 2. Preset retention range of spectrum broadening characteristics (80%-100%): Through experimental measurement of the retention degree of spectrum broadening characteristics under different offsets, it was found that when the retention degree is below 80%, the spectrum broadening characteristics decrease significantly and cannot effectively suppress electromagnetic interference. 80% is the minimum acceptable value for maintaining spectrum broadening characteristics.

[0189] 3. Preset threshold for short-term random fluctuation range (0.08T): Experiments on ferrite cores show that when the magnetic flux density fluctuation range exceeds 0.08T, the sensitive range for eddy current loss drifts significantly, requiring intervention. The fluctuation threshold for silicon steel cores is adjusted to 0.10T.

[0190] 4. Energy distribution change preset threshold (10%): By comparing the energy distribution of the offset sequence and the base sequence, it was found that when the rate of change of energy distribution exceeds 10%, it will lead to uneven spectral energy distribution, which may cause new interference problems.

[0191] 5. Preset threshold for deviation (ripple peak-to-peak value 0.2V, temperature increment 2℃): In actual operation, when the ripple peak-to-peak value deviation exceeds 0.2V or the temperature increment deviation exceeds 2℃, it indicates that there is a significant deviation in the sensitive interval prediction model, and the model parameters need to be adjusted.

[0192] 6. Step size for offset reduction processing (0.5kHz): Through multiple experimental tests, it was determined that a step size of 0.5kHz can effectively avoid over-adjustment and achieve fast convergence while ensuring the spectral broadening characteristics.

[0193] 7. Learning rate of gradient descent (0.01-0.05): Experimental tests show that when the learning rate is in the range of 0.01-0.05, the model can converge stably and will not cause parameter oscillation due to excessive step size, nor will it cause slow convergence speed due to excessive step size.

[0194] 8. Number of iterations for gradient descent (3-5 times): Experiments have shown that 3-5 iterations are sufficient to stabilize the model parameters, while avoiding overfitting caused by overtraining.

[0195] V. Explanation of Other Technical Terms 1. Eddy Current Sensitive Range: This refers to the frequency range within which, under specific temperature and magnetic flux density, the nonlinear permeability of the magnetic core changes, leading to a change in the inductance of the transformer windings. This, in turn, causes a drift in the resonant frequency of the LC resonant circuit composed of the winding inductance and parasitic capacitance. When the harmonic frequency of the PWM modulation signal falls within this drifted resonant frequency range, it excites high-frequency oscillations, resulting in a sharp increase in the rate of change of magnetic flux (dB / dt), thereby triggering abnormal peak values ​​of eddy current losses.

[0196] 2. Spectrum broadening characteristics: refers to the width of the spectrum energy distribution of the PWM modulation signal on the frequency axis. The better the broadening characteristics, the more uniformly the spectrum energy is distributed, and the lower the peak value of electromagnetic interference.

[0197] 3. Harmonic energy main distribution frequency bands: These refer to the frequency bands in the PWM modulation signal where the energy is most concentrated. The concentration of harmonic energy in these frequency bands will lead to an increase in eddy current loss at specific frequencies.

[0198] 4. Short-term random drift trend: refers to the drift direction and trend of the eddy current sensitive range in a short period of time (such as within 10 switching cycles), which is affected by factors such as core temperature changes and operating point fluctuations.

[0199] 5. Drift speed: refers to the speed at which the eddy current sensitive region changes over time, with the unit being kHz / cycle, representing the frequency range of the eddy current sensitive region's movement per cycle.

[0200] 6. Overlap ratio: This refers to the percentage of the intersection frequency length between the main harmonic energy distribution frequency band and the expected location range of the eddy current sensitive area, which accounts for the total frequency length of the main harmonic energy distribution frequency band. It is the core quantitative indicator for determining whether to activate the frequency avoidance mechanism.

[0201] 7. Degree of Preservation of Spectral Broadening Characteristics: This refers to the proportion of the offset modulation sequence that is retained relative to the basic modulation sequence in terms of 3dB bandwidth and spectral uniformity. It is expressed as a percentage, and the closer the value is to 100%, the more complete the broadening characteristics are preserved.

[0202] 8. Deviation Correction Lookup Table Method: When the actual collected ripple data exceeds the preset safety threshold, the system looks up the corresponding offset from the pre-calibrated correction coefficient table according to the relative position of the actual operating frequency and the prediction interval, and corrects it, gradually calibrating the baseline parameters of the prediction model.

[0203] 9. Core Material Property Database: Contains parameters of commonly used core materials for LED power supplies, such as ferrite and silicon steel. Specific data items include magnetic flux density (0.5-1.5T), operating temperature (20-80℃), permeability, saturation magnetic flux density, temperature coefficient, resistivity, etc. K-nearest neighbor matching is used, with k value set to 3 and similarity judgment threshold of 90%.

[0204] 10. K-Nearest Neighbor Method: A distance-based classification algorithm that calculates the distance between the sample to be classified and each sample in the database, selects the K nearest samples, and determines the category of the sample to be classified based on the categories of these samples. In this invention, K is set to 3, and the similarity threshold is 90%. When the similarity is below 90%, a sensor data anomaly alarm is triggered.

[0205] Through the above detailed description, this invention ensures that those skilled in the art can implement the invention based on the description in the specification, thus avoiding review comments that the technical means are "vague".

[0206] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within this application. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A PWM control method for avoiding eddy current losses in LED power supply cores, characterized in that, The method includes: The magnetic flux density and temperature values ​​under the current working conditions are obtained and mapped through a pre-calibrated magnetic core material property database to obtain the center frequency and width of the eddy current sensitive region at the current moment. Collect actual operating point data within the most recent multiple switching cycles, and calculate the short-term random drift trend and drift velocity of the eddy current sensitive region based on the actual operating point data; The short-term random drift trend and the drift velocity are input into a preset sensitive interval prediction model to determine the expected location range of the eddy current sensitive interval in the future several periods. A basic modulation sequence is generated from an asymmetric pulse width modulation chaotic map, and the main frequency bands of harmonic energy distribution of the basic modulation sequence are extracted. Determine the degree of overlap between the main frequency band of the harmonic energy distribution and the expected location range. If the overlap ratio exceeds a preset threshold, activate the frequency avoidance mechanism. The offset is calculated based on the overlap ratio, and the control parameters of the chaotic mapping are adjusted through an adaptive frequency offset mechanism to obtain the offset modulation sequence. The spectrum of the offset modulation sequence is compared with that of the basic modulation sequence to determine the degree of preservation of the spectrum broadening characteristics and the changes in energy distribution. If the degree of preservation of the spectrum broadening characteristic is lower than the preset preservation range, then the offset is reduced and the frequency avoidance mechanism is re-executed; If the degree of preservation of the spectrum broadening characteristic meets the preset preservation range, the offset modulation sequence is output for the generation of the pulse width modulation drive signal in the next cycle. The ripple data and temperature increments collected during actual operation are fed back to the sensitive interval prediction model to update the internal state of the sensitive interval prediction model.

2. The method according to claim 1, characterized in that, The process of obtaining the magnetic flux density and temperature values ​​under the current operating conditions, and mapping them through a pre-calibrated magnetic core material property database to obtain the center frequency and width of the eddy current sensitive region at the current moment, includes: Acquire magnetic flux density and temperature data in the current working environment, and collect and record them in real time through sensors; By matching the collected magnetic flux density data and temperature data through a pre-built magnetic core material property database, the corresponding material performance parameters are obtained. Based on the matched material performance parameters and the current environmental conditions, the preliminary characteristic values ​​of the eddy current sensitive region are calculated. The support vector machine algorithm is used to classify the preliminary feature values ​​and determine the center frequency range of the eddy current sensitive region. If the center frequency range exceeds the preset threshold range, the temperature data and magnetic flux density data are corrected a second time to obtain the adjusted characteristic value. For the adjusted eigenvalues, calculate the width of the eddy current sensitive region and determine the final region parameters; Based on the center frequency and width determined above, the eddy current sensitive region description data for the current moment is generated.

3. The method according to claim 1, characterized in that, The process of collecting actual operating point data over the most recent multiple switching cycles and calculating the short-term random drift trend and drift velocity of the eddy current sensitive region based on the actual operating point data includes: An initial dataset is constructed by acquiring actual operating point data from multiple switching cycles. The data is then cleaned and organized using automated tools to obtain structured operating point records. Based on the structured working point records, the eddy current sensitive interval is segmented and processed to extract key data points within each cycle range, thereby determining the range of short-term random fluctuations. If the short-term random fluctuation range exceeds the preset threshold, the key data points are weighted to calculate the corresponding drift trend and obtain the preliminary trend direction. By conducting time series analysis on the initial trend direction and combining it with actual data within the period range, the dynamic changes in drift speed can be determined. If the dynamic changes in drift velocity exhibit nonlinear characteristics, a linear regression algorithm is used to fit the changes to obtain a stable velocity estimate. Based on the stable velocity estimate, a secondary analysis is performed on the stochastic characteristics within the eddy current sensitive range to determine the final short-term stochastic drift trend. By combining the final short-term random drift trend with the results of interval analysis, a corresponding drift characteristic description is generated, thus completing a comprehensive assessment of the eddy current sensitive interval.

4. The method according to claim 1, characterized in that, The step of inputting the short-term random drift trend and the drift velocity into a preset sensitive interval prediction model to determine the expected location range of the eddy current sensitive interval within a few future periods includes: To obtain short-term random drift trends and drift speeds; By processing short-term random drift trends and drift speeds using a pre-set sensitive interval prediction model, the predicted location range of the eddy current sensitive interval within several future periods can be obtained. For the expected location range, the interval boundary extraction method is used to obtain the upper and lower bounds of the eddy current sensitive interval; Based on the upper and lower bound positions, determine whether the current position is within the eddy current sensitive range. If it is within the eddy current sensitive range, mark the current period as a sensitive period. From the marked sensitive period sequence, the number of consecutive sensitive period segments and the length of each segment are counted to obtain the distribution information of sensitive segments; Based on the distribution information of sensitive segments and drift speed, determine the direction of movement and the expected drift distance of the sensitive section; The linear extrapolation method is used to process the direction of movement and the expected drift distance of the sensitive region, so as to obtain the corrected position range of the eddy current sensitive region in the next cycle.

5. The method according to claim 1, characterized in that, The process of generating a basic modulation sequence from an asymmetric pulse width modulation chaotic map and extracting the main frequency bands of harmonic energy distribution of the basic modulation sequence includes: The basic modulation sequence is generated by asymmetric pulse width modulation chaotic mapping; The harmonic energy sequence is obtained by extracting the amplitude of each frequency component from the basic modulation sequence; The total energy of each frequency band is obtained by segmenting and statistically analyzing the harmonic energy sequence. By comparing the total energy of each frequency band, the three frequency bands with the highest energy are determined. If the total energy of a certain frequency band exceeds twice that of the adjacent frequency bands, it is marked as an energy-concentrated frequency band. Based on the labeling results, all energy-concentrated frequency bands are output as the main distribution frequency bands; The final list of frequency bands is obtained by sorting the main frequency bands in ascending order of frequency.

6. The method according to claim 1, characterized in that, The determination of the degree of overlap between the main frequency band of the harmonic energy distribution and the expected location range, and if the overlap ratio exceeds a preset threshold, then the frequency avoidance mechanism is activated, including: Obtain the spectral data of the harmonic signal; The distribution of harmonic energy at each frequency point is obtained through spectrum analysis; Determine the main frequency bands of harmonic energy distribution based on the distribution values; Obtain the estimated location range occupied by the current business; Calculate the intersection length between the main distribution frequency band and the expected location range; The overlap ratio is obtained by dividing the intersection length by the total length of the main distribution frequency bands. Determine whether the overlap ratio exceeds a preset threshold; If the overlap ratio exceeds a preset threshold, then frequency avoidance needs to be performed. Obtain the set of available backup frequencies; Select the frequency point with the lowest overlap with the main distribution frequency band from the set of available spare frequencies; The selected frequency point will be used as the target frequency to avoid; The frequency to avoid the target is sent to the execution unit to complete the frequency adjustment.

7. The method according to claim 1, characterized in that, The step of calculating the offset based on the overlap ratio and adjusting the control parameters of the chaotic mapping through an adaptive frequency offset mechanism to obtain the offset modulation sequence includes: Obtain the overlap ratio between the current sequence and the reference sequence; The offset value is determined based on the overlap ratio using a preset mapping relationship; Input the offset value into the adaptive frequency offset mechanism; The control parameters of the chaotic mapping are adjusted by an adaptive frequency offset mechanism to obtain the adjusted control parameters; The chaotic mapping is driven by the adjusted control parameters to generate an initial modulation sequence. The initial modulation sequence is offset to obtain the offset modulation sequence. The offset modulation sequence is output as the base sequence for subsequent communication modulation.

8. The method according to claim 1, characterized in that, The step of comparing the spectrum of the offset modulation sequence with that of the basic modulation sequence to determine the degree of preservation of spectral broadening characteristics and the changes in energy distribution includes: Step 1: Acquire the modulation signal data of the offset sequence and the base sequence. Perform preliminary sampling and digitization processing on the two sequences through a pre-established signal processing module to obtain the digital modulation signals of the two sequences. Step 2: For the digitally modulated signal, use a spectrum analysis tool to perform frequency domain transformation on the offset sequence and the fundamental sequence respectively to determine the spectrum data of the two sequences; Step 3: Extract the broadening characteristic parameters of the offset sequence and the base sequence from the spectral data, analyze the difference in broadening characteristics of the two sequences in the frequency domain through comparison, and determine the retention status; Step 4: Based on the results of the differences in broadening characteristics, obtain the energy distribution data of the two sequences, and use the same spectral analysis tool to calculate the changes in energy distribution; Step 5: If the change in energy distribution exceeds the preset threshold, the modulated signal of the offset sequence is processed a second time to obtain the adjusted spectrum data. Step Six: By comparing the adjusted spectral data with the spectral data of the base sequence, analyze the specific changes in distribution differences and determine the degree of matching between the final spectral broadening characteristics and energy distribution.

Images