LED lamp drive circuit control method and system for preventing stroboscopic effect

Abstract

The invention discloses an LED lamp drive circuit control method and system for preventing a stroboscopic effect, and the method comprises the following steps: 1, collecting data detected by a sensor, and generating a feature data set; 2, constructing a feature vector; step 3, training an iterative improved LightGBM model, inputting the feature vector into the trained improved LightGBM model, and outputting a multi-dimensional result at a leaf node; 4, executing a multi-target sorting rule according to a multi-dimensional result, and determining a control gear; 5, the driving parameter set of the control gear is issued to an LED lamp driving circuit, and operation of the driving circuit is controlled; and 6, calculating a stroboscopic evaluation index in the rolling time window, and switching a new control gear when the stroboscopic evaluation index exceeds a threshold value. The method can effectively reduce the stroboscopic effect, and is suitable for indoor and outdoor LED illumination system control scenes with high illumination quality requirements.

Description

Technical Field

[0001] This invention relates to the field of cross-application technology of intelligent sensing and machine learning modeling, and in particular to a method and system for controlling LED lighting drive circuits to prevent flicker effects. Background Technology

[0002] Existing LED lighting driver circuits generally employ pulse-width modulation (PWM)-based control, adjusting carrier frequency and duty cycle to achieve brightness regulation and energy efficiency optimization. In this process, flicker is a significant issue. Flicker manifests primarily as periodic fluctuations in the light output signal over time, causing perceptible flickering and stripe interference in images captured by cameras. This not only reduces visual comfort but can also lead to eye strain and even cause serious interference in applications such as industrial lighting and high-definition video recording. To reduce flicker, current technologies typically suppress it by fixing and increasing the carrier frequency or introducing simple filters. However, these methods often sacrifice energy efficiency and heat dissipation, failing to balance flicker control with the overall performance of the driver circuit.

[0003] In the control process of LED driver circuits, in addition to flicker effect, there are also many other issues involved, such as temperature rise, electromagnetic interference, and brightness abrupt changes. Most existing methods rely on a single optimization objective, such as setting drive parameters solely based on flicker prevention, while neglecting the temperature rise risk and electromagnetic interference issues of LED driver circuits under high power density. As application scenarios become increasingly complex, ambient illuminance, power supply fluctuations, and load conditions all dynamically affect the operating state of LED lamps. Existing fixed-parameter drive strategies struggle to maintain stable performance under different operating conditions, easily leading to incomplete flicker suppression or reduced energy efficiency. Furthermore, existing technologies lack constraints on the risk of brightness abrupt changes during level switching, which can easily cause sudden changes in light output, further exacerbating visual discomfort.

[0004] Current research attempts to introduce machine learning methods into LED driver control, such as parameter tuning based on traditional decision trees or simple regression models. However, because the model can only output a single control gear category and fails to consider the multidimensional costs associated with gear switching, problems such as unstable decision results and unbalanced parameter adjustments still exist in actual operation. This single-output structure leads to a disconnect between the model and the actual control logic of the drive circuit, requiring an additional decision layer for secondary processing, increasing system complexity, and reducing the real-time performance and reliability of control.

[0005] Existing LED lighting driver circuit control methods struggle to achieve unified optimization across flicker suppression, temperature rise constraints, electromagnetic interference control, and smooth brightness switching. In particular, they lack methods for jointly modeling multi-source sensor data with multiple objectives. How to introduce improved machine learning models into driver circuit control to directly output multi-dimensional results that satisfy flicker prevention requirements while also considering constraints such as temperature rise, electromagnetic interference, and brightness abrupt changes, and how to achieve closed-loop integration with the driver circuit control logic, has become a pressing technological challenge.

[0006] Therefore, how to provide a control method for LED lamp driver circuits to prevent flicker is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] The LED lamp driving circuit control method for preventing flickering effect according to an embodiment of the present invention includes the following steps: Step 1: Collect data detected by the sensors; Step 2: Perform time-domain and frequency-domain analysis on the sensor-detected data to construct feature vectors; Step 3: Train the iteratively improved LightGBM model. Input the feature vector into the trained improved LightGBM model and output multi-dimensional results at the leaf nodes, including the control gear number and the corresponding temperature rise cost, electromagnetic interference cost, and brightness jump cost. Step 4: Based on the multi-dimensional results, execute the multi-objective sorting rules, prioritize the selection of the level that meets the anti-flicker index, and determine the control level according to the order of temperature rise cost, electromagnetic interference cost, and brightness jump cost among the levels that meet the anti-flicker index. Step 5: Send the set of drive parameters for the control gear to the LED lighting driver circuit to control the operation of the driver circuit; Step 6: Calculate the flicker evaluation index within the rolling time window, compare it with the threshold, and if the threshold is exceeded, trigger the re-inference of the improved LightGBM model and switch to a new control level.

[0008] Optionally, the process of collecting data detected by the sensors to generate a feature dataset specifically involves: Collect the LED light output waveform signal from the photodiode sensor and extract the original light intensity sequence representing the brightness change; Collect the illuminance values ​​output by the ambient light sensor and extract the numerical illuminance index representing the current light intensity; Acquire image frame sequences output by the camera sensor, analyze the inter-frame brightness distribution, and calculate stripe interference criterion data; The bus voltage, current and temperature signals in the drive circuit are collected to obtain power fluctuation characteristics, load state characteristics and thermal state characteristics, respectively.

[0009] Optionally, the step of performing time-domain and frequency-domain analysis on the data detected by the sensor to construct a feature vector specifically involves: Perform a fast Fourier transform on the original light intensity sequence to extract the harmonic energy values ​​of the dominant frequency component; A sliding window process is performed on the original light intensity sequence to calculate the average brightness value and brightness change amplitude within each window, thus obtaining the brightness fluctuation amplitude sequence. Time series modeling is performed on the brightness fluctuation amplitude sequence to extract the phase shift and periodic fluctuation trend, thus obtaining the brightness fluctuation phase; The luminance fluctuation amplitude sequence, luminance fluctuation phase, harmonic energy value, numerical illuminance index and stripe interference criterion data are combined according to a unified feature structure to generate a feature vector.

[0010] Optionally, the training iteratively improved LightGBM model specifically includes: During the training iteration, the current cumulative control level logit vector and the current cumulative multidimensional regression value vector are initialized as constant vectors. The multidimensional regression value vector includes the temperature rise cost component, the electromagnetic interference cost component, and the brightness jump cost component. Set monitoring labels, which consist of control level category labels, temperature rise cost labels, electromagnetic interference cost labels, and brightness jump cost labels; The predicted category probability distribution is calculated by normalizing the current cumulative control level logit vector using the Softmax function, and the predicted cost during the training phase is output by each component of the current cumulative multidimensional regression value vector. The predicted cost during the training phase includes the cost of temperature rise, the cost of predicted electromagnetic interference, and the cost of predicted brightness jump. Define a joint loss function, which is a weighted combination of a classification loss function and three regression loss functions. The classification loss function is the cross-entropy loss between the control gear category label and the predicted category probability distribution. The three regression loss functions are the squared loss between the temperature rise cost label and the predicted temperature rise cost, the squared loss between the electromagnetic interference cost label and the predicted electromagnetic interference cost, and the squared loss between the brightness jump cost label and the predicted brightness jump cost. Based on the joint loss function, the feature vectors are input into the improved LightGBM model to generate the first decision tree; Repeat the training iteration steps until the stopping conditions of N decision trees and split gain threshold are met, forming an improved LightGBM model composed of N decision trees.

[0011] Optionally, the improved LightGBM model, which inputs the feature vectors into the trained model, outputs multi-dimensional results at the leaf nodes, including the control gear number and the corresponding temperature rise cost, electromagnetic interference cost, and brightness jump cost, specifically: Input the feature vectors into the trained improved LightGBM model; Within each trained decision tree, a path is selected layer by layer to the leaf node based on the splitting condition; Read the control level logit vector and multidimensional regression value vector stored in the leaf node; The control gear logit vector is obtained by accumulating the control gear logit increments output by the leaf nodes of each decision tree, and the multidimensional regression value vector is obtained by accumulating the cost increments output by the leaf nodes of each decision tree; the cost increments are the temperature rise cost increment, the electromagnetic interference cost increment, and the brightness jump cost increment. The control gear logit increment is determined by aggregating the first and second gradients calculated from the sample set falling into the leaf node under cross-entropy loss. The cost increment is determined by aggregating the first and second gradients calculated under squared loss for the sample set falling into the leaf node. The cumulative control gear logit vector is obtained by accumulating the control gear logit vector among all trained decision trees, and the cumulative predicted cost is obtained by accumulating the multidimensional regression value vector. Normalize the cumulative control gear logit vector using the Softmax function to obtain the control gear probability vector; The control gear number is determined based on the control gear probability vector, and the control gear number is associated with the cumulative predicted cost. The output consists of a multi-dimensional result composed of the control gear number and the cumulative predicted cost.

[0012] Optionally, the improved LightGBM model, which inputs the feature vectors into the trained model, outputs multi-dimensional results at the leaf nodes, including the control gear number and the corresponding temperature rise cost, electromagnetic interference cost, and brightness jump cost, specifically: Candidate control gears with probability values ​​exceeding a preset threshold are selected from the control gear probability vector to form a candidate control gear set; The anti-flicker index is checked one by one in the candidate control gear set. Candidate control gears that do not meet the anti-flicker index are removed, and the candidate control gear set that meets the anti-flicker index is obtained. Among the candidate control levels that meet the anti-flicker index, they are sorted according to the numerical values ​​of the predicted temperature rise cost, the predicted electromagnetic interference cost, and the predicted brightness jump cost. The candidate control position with the lowest overall cost is selected as the final control position number based on the sorting results. The comprehensive cost is composed of the predicted temperature rise cost, the predicted electromagnetic interference cost, and the predicted brightness jump cost. In the set of candidate control positions that meet the anti-flicker index, the values ​​of the predicted temperature rise cost, the predicted electromagnetic interference cost, and the predicted brightness jump cost are compared in turn, and the candidate control position with the smallest comprehensive cost is selected as the final control position number. The final control gear number and the corresponding cumulative multidimensional regression value vector are output for generating control parameters for the drive circuit.

[0013] Optionally, the step of sending the set of drive parameters for the control gear to the LED lighting driver circuit to control the operation of the driver circuit specifically involves: Based on the final control gear number, the corresponding set of driving parameters is read from the preset parameter library. The set of driving parameters includes carrier frequency parameters, model predictive control weight parameters, proportional-integral gain parameters, and notch filter center frequency parameters. A safety constraint verification is performed on the set of driving parameters, which includes threshold comparison of bus voltage ripple, current rise slope, device temperature and brightness step amplitude; the set of driving parameters is gradually adjusted within a preset gradual time window to generate a transition parameter sequence arranged in time sequence; and the transition parameter sequence is sequentially sent to the LED lighting driver circuit. Parameters are synchronously written at the beginning boundary of the pulse width modulation switching cycle; After the parameters take effect, a closed-loop verification is performed: the sensor and drive feedback signals are read, the anti-flicker index is calculated and compared with the threshold. If the threshold is exceeded, the control level is switched again. The drive feedback signal refers to the power fluctuation characteristics, load state characteristics, and thermal state characteristics obtained by feature extraction from the bus voltage signal, current signal, and temperature signal collected by the drive circuit. Record the time information and control gear number when the drive parameters take effect.

[0014] Optionally, the flicker evaluation index calculated within the rolling time window is compared with a threshold. If the threshold is exceeded, the improved LightGBM model is re-inferred and a new control level is switched. Specifically: Within a set length of rolling time window, the raw light intensity sequence output by the photodiode sensor and the image frame sequence output by the camera sensor are continuously acquired; The light intensity sequence is filtered and normalized, the difference between the maximum and minimum brightness between two adjacent periods is calculated, the peak-to-valley difference of brightness fluctuation is obtained, and the flicker percentage is calculated based on the peak-to-valley difference and the average brightness of the period. Frequency domain energy distribution analysis is performed on the brightness fluctuation amplitude sequence to calculate the variance of frequency component energy and the relative offset of the dominant frequency intensity. The scintillation index is then calculated based on the frequency domain perturbation characteristics. The image frame sequence acquired by the camera sensor is subjected to grayscale conversion, enhancement and noise suppression processing. The brightness distribution difference between adjacent frames is calculated, and the stripe intensity index is calculated based on the frequency domain energy concentration, brightness skewness and brightness modulation depth. The flicker percentage, flicker index, and stripe intensity index within the rolling time window are compared with preset thresholds. If any index exceeds the corresponding threshold, the trained improved LightGBM model is called to re-input the current feature vector, and steps three and four are executed. The control gear number is updated based on the inference results of the improved LightGBM model, and the corresponding drive parameter set is reread to trigger the switching and distribution of control parameters.

[0015] An LED lamp driving circuit control system for preventing flickering according to an embodiment of the present invention includes the following modules: The sensor data acquisition module is used to acquire data detected by the sensor. The time-frequency analysis and feature vector construction module is used to perform time-domain and frequency-domain analysis on the data detected by the acquired sensors and generate feature vectors in a unified format. The model training and inference module is used to train the iteratively improved LightGBM model. The control gear filtering and sorting module is used to determine the candidate control gear with the lowest overall cost as the final control gear number. The drive parameter generation and distribution module is used to control the operation of the LED lighting drive circuit according to the final control gear number; The closed-loop verification module is used to read the sensor and drive feedback signals after the drive parameters take effect, calculate the flicker index and compare it with the threshold. If the threshold is exceeded, the control gear is triggered to switch again. The flicker index calculation module is used to calculate the flicker percentage, flicker index, and stripe intensity index, and compare them with preset thresholds.

[0016] The beneficial effects of this invention are: This invention constructs a multi-objective driving circuit control method based on an improved LightGBM model, which prevents LED lamp flicker while taking into account multiple performance indicators such as temperature rise, electromagnetic interference, and brightness jump, achieving the best overall dynamic control effect.

[0017] By collecting multi-source feature data including light intensity sequences, image frames, illuminance values, voltage, current, and temperature from various sensors, and combining time-domain and frequency-domain analysis to extract key features, the modeling accuracy of flicker risk factors has been effectively improved.

[0018] An improved LightGBM model is used to output the control gear number and three-class cost values ​​at the leaf nodes. A joint loss function is introduced for model training, which enables the model to not only classify control gears, but also to regress and predict temperature rise, electromagnetic interference and brightness jumps, thereby improving the interpretability and generalization ability of the overall control strategy.

[0019] During the control level selection stage, by setting anti-flicker index thresholds and a comprehensive cost ranking mechanism, the selected level is ensured to have the best cost performance under the premise of meeting safety and stability requirements, effectively reducing the system risks caused by brightness fluctuations, electrical interference and device temperature rise.

[0020] By combining parameter gradual change processing with a closed-loop feedback verification mechanism, the system ensures that the adjustment process of drive parameters is smooth and controllable, and can quickly revert to a stable level or re-trigger model inference when there is a risk of flickering, thereby enhancing the system's dynamic adaptability and robustness under complex working conditions.

[0021] This invention improves the visual comfort and electrical stability of LED lighting fixtures while achieving synergistic optimization of multi-dimensional performance, demonstrating good practical value and promising prospects for promotion. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0023] Figure 1 This is an overall flowchart of the LED lamp driving circuit control method for preventing flicker effect proposed in this invention; Figure 2 This is a schematic diagram of the improved LightGBM model structure of the LED lamp driving circuit control method for preventing flicker effect proposed in this invention. Figure 3 This is a flowchart of the control level selection and sorting decision-making process for the LED lamp driver circuit control method for preventing flicker effect proposed in this invention. Figure 4 This is a schematic diagram of the LED lamp driving circuit control system proposed in this invention for preventing flickering. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0025] refer to Figure 1-3 A method for controlling LED lamp driver circuits to prevent flickering includes the following steps: Step 1: Collect data detected by the sensors; Step 2: Perform time-domain and frequency-domain analysis on the sensor-detected data to construct feature vectors; Step 3: Train the iteratively improved LightGBM model. Input the feature vector into the trained improved LightGBM model and output multi-dimensional results at the leaf nodes, including the control gear number and the corresponding temperature rise cost, electromagnetic interference cost, and brightness jump cost. Step 4: Based on the multi-dimensional results, execute the multi-objective sorting rules, prioritize the selection of the level that meets the anti-flicker index, and determine the control level according to the order of temperature rise cost, electromagnetic interference cost, and brightness jump cost among the levels that meet the anti-flicker index. Step 5: Send the set of drive parameters for the control gear to the LED lighting driver circuit to control the operation of the driver circuit; Step 6: Calculate the flicker evaluation index within the rolling time window, compare it with the threshold, and if the threshold is exceeded, trigger the re-inference of the improved LightGBM model and switch to a new control level.

[0026] This invention proposes a control method for LED lighting driver circuits to prevent flicker. First, multi-source sensor data is collected and analyzed in both the time and frequency domains to construct a feature vector containing brightness fluctuation characteristics, illuminance indices, and electrical parameters. Then, the feature vector is input into a trained improved LightGBM model, which outputs a control level number and corresponding multi-dimensional results such as temperature rise cost, electromagnetic interference cost, and brightness jump cost. Based on these results, levels that meet the anti-flicker indices are prioritized, and the optimal control level is determined according to a multi-objective ranking rule. Finally, the set of driving parameters corresponding to the control level is sent to the LED driver circuit, and flicker evaluation indices are continuously monitored within a rolling time window. When a threshold is exceeded, the model is automatically re-inferred and the control level is switched, achieving dynamic prediction and closed-loop optimization control of LED flicker effects.

[0027] In this embodiment, the process of acquiring data detected by the sensor and generating a feature dataset specifically involves: Collect the LED light output waveform signal from the photodiode sensor and extract the original light intensity sequence representing the brightness change; Collect the illuminance values ​​output by the ambient light sensor and extract the numerical illuminance index representing the current light intensity; The image frame sequence output by the camera sensor is acquired, the inter-frame brightness distribution is analyzed, and stripe interference criterion data is calculated. The stripe interference criterion data is extracted in the following way: grayscale and noise suppression processing is performed on the image frame sequence acquired by the camera sensor, the inter-frame brightness changes are extracted and frequency domain transformation analysis is performed, and stripe interference index is constructed based on the characteristics such as frequency domain energy concentration, local brightness fluctuation skew distribution and brightness modulation depth, which is used to determine the intensity of image stripes caused by LED flicker.

[0028] The bus voltage, current and temperature signals in the drive circuit are collected to obtain power fluctuation characteristics, load state characteristics and thermal state characteristics, respectively.

[0029] This invention achieves comprehensive perception and characterization of the operating status of LED lamps by constructing a multi-source heterogeneous feature dataset. Specifically, it includes: acquiring LED light output waveform signals from photodiode sensors and extracting the original light intensity sequence to reflect brightness change trends; acquiring illuminance values ​​from ambient light sensors to obtain numerical illuminance indices representing the current ambient light intensity; acquiring image frame sequences from camera sensors, extracting inter-frame brightness changes through grayscale and noise suppression processing, and constructing stripe interference criteria by combining features such as frequency domain energy concentration, brightness skewness, and modulation depth to assess image interference caused by flicker; and simultaneously acquiring bus voltage, current, and temperature signals from the drive circuit to extract power fluctuation characteristics, load state characteristics, and thermal state characteristics, providing accurate input for subsequent flicker risk modeling and intelligent selection of control levels.

[0030] In this embodiment, the step of performing time-domain and frequency-domain analysis on the data detected by the sensor to construct a feature vector specifically involves: Perform a fast Fourier transform on the original light intensity sequence to extract the harmonic energy values ​​of the dominant frequency component; A sliding window process is performed on the original light intensity sequence to calculate the average brightness value and brightness change amplitude within each window, thus obtaining the brightness fluctuation amplitude sequence. Time series modeling is performed on the brightness fluctuation amplitude sequence to extract the phase shift and periodic fluctuation trend, thus obtaining the brightness fluctuation phase; The luminance fluctuation amplitude sequence, luminance fluctuation phase, harmonic energy value, numerical illuminance index and stripe interference criterion data are combined according to a unified feature structure to generate a feature vector.

[0031] This invention performs time-domain and frequency-domain analysis on sensor-detected data to construct a unified feature vector, comprehensively extracting key indicators reflecting the LED's operating status. Specifically, this includes: performing a Fast Fourier Transform on the original light intensity sequence to extract the harmonic energy value of the dominant frequency component; using a sliding window process to calculate the brightness fluctuation amplitude sequence to characterize the brightness change range; further performing time-series modeling on the brightness fluctuation amplitude sequence to extract phase shift and periodic fluctuation trends, obtaining the brightness fluctuation phase; and simultaneously integrating numerical illuminance indicators and stripe interference criteria data to construct a feature vector containing multi-dimensional flicker-related features, providing a structured input basis for subsequent model inference and control level decisions.

[0032] In this embodiment, the training of the iteratively improved LightGBM model specifically refers to: During the training iteration, the current cumulative control level logit vector and the current cumulative multidimensional regression value vector are initialized as constant vectors. The multidimensional regression value vector includes the temperature rise cost component, the electromagnetic interference cost component, and the brightness jump cost component. Set monitoring labels, which consist of control level category labels, temperature rise cost labels, electromagnetic interference cost labels, and brightness jump cost labels; The predicted category probability distribution is calculated by normalizing the current cumulative control level logit vector using the Softmax function, and the predicted cost during the training phase is output by each component of the current cumulative multidimensional regression value vector. The predicted cost during the training phase includes the cost of temperature rise, the cost of predicted electromagnetic interference, and the cost of predicted brightness jump. Define a joint loss function, which is a weighted combination of a classification loss function and three regression loss functions. The classification loss function is the cross-entropy loss between the control gear category label and the predicted category probability distribution. The three regression loss functions are the squared loss between the temperature rise cost label and the predicted temperature rise cost, the squared loss between the electromagnetic interference cost label and the predicted electromagnetic interference cost, and the squared loss between the brightness jump cost label and the predicted brightness jump cost. Based on the joint loss function, the feature vectors are input into the improved LightGBM model to generate the first decision tree; Repeat the training iteration steps until the stopping conditions of N decision trees and split gain threshold are met, forming an improved LightGBM model composed of N decision trees.

[0033] This invention achieves multi-objective intelligent decision-making for LED driver control settings by constructing an improved LightGBM model. During training, a joint modeling strategy is introduced, combining the control setting logit vector and multi-dimensional regression value vector. This corresponds to three cost regression tasks: control setting classification, temperature rise, electromagnetic interference, and brightness fluctuation. A joint loss function, constructed using cross-entropy loss and squared loss, is employed for collaborative optimization. In each iteration, the feature vectors are trained under supervision based on the joint loss function, generating a decision tree model that includes control strategies and multi-dimensional performance indicators. Ultimately, an improved LightGBM model consisting of N decision trees is formed, providing an efficient and reliable inference foundation for subsequent driver control parameter generation.

[0034] In this embodiment, the improved LightGBM model, which has been trained by inputting feature vectors, outputs multi-dimensional results at the leaf nodes, including control gear numbers and corresponding temperature rise costs, electromagnetic interference costs, and brightness jump costs. Specifically: Input the feature vectors into the trained improved LightGBM model; Within each trained decision tree, a path is selected layer by layer to the leaf node based on the splitting condition; The splitting condition consists of the splitting features and corresponding splitting thresholds stored in the internal nodes of the tree. The value of the feature vector on the splitting feature is compared with the splitting threshold. If the condition is met, the tree enters the left child node; otherwise, it enters the right child node, until the leaf node is reached.

[0035] Read the control level logit vector and multidimensional regression value vector stored in the leaf node; The control gear logit vector is obtained by accumulating the control gear logit increments output by the leaf nodes of each decision tree, and the multidimensional regression value vector is obtained by accumulating the cost increments output by the leaf nodes of each decision tree; the cost increments are the temperature rise cost increment, the electromagnetic interference cost increment, and the brightness jump cost increment. The control gear logit increment is determined by aggregating the first and second gradients calculated from the sample set falling into the leaf node under cross-entropy loss. The cost increment is determined by aggregating the first and second gradients calculated under squared loss for the sample set falling into the leaf node. The cumulative control gear logit vector is obtained by accumulating the control gear logit vector among all trained decision trees, and the cumulative predicted cost is obtained by accumulating the multidimensional regression value vector. Normalize the cumulative control gear logit vector using the Softmax function to obtain the control gear probability vector; The control gear number is determined based on the control gear probability vector, and the control gear number is associated with the cumulative predicted cost. The output consists of a multi-dimensional result composed of the control gear number and the cumulative predicted cost.

[0036] This invention achieves multi-dimensional inference and output of LED driving control strategies by inputting the constructed feature vectors into a trained improved LightGBM model. In each decision tree, the feature vectors are filtered layer by layer according to the splitting conditions of internal nodes, eventually reaching the leaf node and reading its stored control level logit increment and cost increment, including temperature rise cost, electromagnetic interference cost, and brightness jump cost. By accumulating the increment results of all trees, the cumulative logit vector and multi-dimensional predicted cost value of the control level are obtained. These are then normalized using the Softmax function to generate a control level probability vector, which determines the final control level number and the corresponding multi-dimensional cost output. This process achieves joint modeling and fine evaluation of the control strategy, providing a decision basis for subsequent driving parameter generation.

[0037] In this embodiment, the improved LightGBM model, which has been trained by inputting feature vectors, outputs multi-dimensional results at the leaf nodes, including control gear numbers and corresponding temperature rise costs, electromagnetic interference costs, and brightness jump costs. Specifically: Candidate control gears with probability values ​​exceeding a preset threshold are selected from the control gear probability vector to form a candidate control gear set; The anti-flicker index is checked one by one in the candidate control gear set. Candidate control gears that do not meet the anti-flicker index are removed, and the candidate control gear set that meets the anti-flicker index is obtained. Among the candidate control levels that meet the anti-flicker index, they are sorted according to the numerical values ​​of the predicted temperature rise cost, the predicted electromagnetic interference cost, and the predicted brightness jump cost. The predicted temperature rise cost is output by the temperature rise cost component of the current cumulative multidimensional regression value vector during the training iteration of the improved LightGBM model. The temperature rise cost component is obtained by accumulating the temperature rise cost increments output by each leaf node of the decision tree.

[0038] The predicted electromagnetic interference cost is output by the electromagnetic interference cost component of the current accumulated multidimensional regression value vector during the training iteration of the improved LightGBM model. The electromagnetic interference cost component is obtained by accumulating the electromagnetic interference cost increments output by each leaf node of the decision tree.

[0039] The predicted brightness jump cost is output by the brightness jump cost component of the current accumulated multidimensional regression value vector in the training iteration of the improved LightGBM model. The brightness jump cost component is obtained by accumulating the brightness jump cost increments output by each leaf node of the decision tree.

[0040] The candidate control position with the lowest overall cost is selected as the final control position number based on the sorting results. The comprehensive cost is composed of the predicted temperature rise cost, the predicted electromagnetic interference cost, and the predicted brightness jump cost. In the set of candidate control positions that meet the anti-flicker index, the values ​​of the predicted temperature rise cost, the predicted electromagnetic interference cost, and the predicted brightness jump cost are compared in turn, and the candidate control position with the smallest comprehensive cost is selected as the final control position number. The final control gear number and the corresponding cumulative multidimensional regression value vector are output for generating control parameters for the drive circuit.

[0041] This invention inputs feature vectors into a trained improved LightGBM model to obtain inference results including the probability distribution of control gears and multidimensional cost values. Candidate control gears with probability values ​​exceeding a preset threshold are then selected from the control gear probability vectors. Subsequently, anti-flicker index detection is performed on the candidate control gears sequentially, eliminating options that do not meet the conditions and retaining the set of control gears that meet the anti-flicker index. Based on this, the temperature rise cost increment, electromagnetic interference cost increment, and brightness jump cost increment output from each decision tree leaf node are extracted, accumulated tree by tree to form a predicted cost value component, and sorted according to this component. The control gear with the lowest overall cost is selected as the final control result. Finally, the control gear number and the corresponding cumulative multidimensional regression value vector are output together as the basis for generating the drive circuit control parameters, realizing multi-index joint optimization and dynamic control of the LED driving process.

[0042] In this embodiment, the step of sending the set of drive parameters for the control gear to the LED lighting driver circuit to control the operation of the driver circuit specifically involves: Based on the final control gear number, the corresponding set of driving parameters is read from the preset parameter library. The set of driving parameters includes carrier frequency parameters, model predictive control weight parameters, proportional-integral gain parameters, and notch filter center frequency parameters. A safety constraint verification is performed on the driving parameter set, which includes threshold comparison of bus voltage ripple, current rise slope, device temperature and brightness step amplitude; the driving parameter set is gradually adjusted within a preset gradual time window to generate a transition parameter sequence arranged in time sequence; the transition parameter sequence is sequentially sent to the LED lighting drive circuit: the carrier frequency setting of the PWM drive module, the weight configuration of the model prediction control module, the proportional coefficient and integral coefficient setting of the proportional-integral control module, and the center frequency setting of the notch filter module are written respectively; Parameters are synchronously written at the beginning boundary of the pulse width modulation switching cycle; After the parameters take effect, a closed-loop verification is performed: the sensor and drive feedback signals are read, the anti-flicker index is calculated and compared with the threshold. If the threshold is exceeded, the control level is switched again. The drive feedback signal refers to the power fluctuation characteristics, load state characteristics, and thermal state characteristics obtained by feature extraction from the bus voltage signal, current signal, and temperature signal collected by the drive circuit. Record the time information and control gear number when the drive parameters take effect.

[0043] This invention reads the set of drive parameters corresponding to the final control level from a preset parameter library and performs safety constraint verification on the carrier frequency parameters, model predictive control weight parameters, proportional-integral gain parameters, and notch filter center frequency parameters to ensure that they meet the safety threshold requirements for bus voltage ripple, current slope, and device temperature and brightness variation. During parameter distribution, a preset gradual time window is used for step-by-step adjustment, generating a transition parameter sequence and writing it sequentially into each control module to ensure smooth parameter switching. After synchronously writing parameters at the starting boundary of the pulse width modulation switching cycle, the system performs closed-loop verification, reads sensor and drive feedback signals in real time to calculate the anti-flicker index, and compares it with the threshold. Once an anomaly is detected, a control level re-switching mechanism is triggered. In addition, the system records the time information of the parameter's effective date and the control level number for subsequent traceability and scheduling optimization, realizing dynamic control and stable operation of the LED drive system.

[0044] In this embodiment, the flicker evaluation index calculated within the rolling time window is compared with a threshold. If the threshold is exceeded, the improved LightGBM model is re-inferred and a new control level is switched. Specifically: Within a set length of rolling time window, the raw light intensity sequence output by the photodiode sensor and the image frame sequence output by the camera sensor are continuously acquired; The light intensity sequence is filtered and normalized, the difference between the maximum and minimum brightness between two adjacent periods is calculated, the peak-to-valley difference of brightness fluctuation is obtained, and the flicker percentage is calculated based on the peak-to-valley difference and the average brightness of the period. Frequency domain energy distribution analysis is performed on the brightness fluctuation amplitude sequence to calculate the variance of frequency component energy and the relative offset of the dominant frequency intensity. The scintillation index is then calculated based on the frequency domain perturbation characteristics. The image frame sequence acquired by the camera sensor is subjected to grayscale conversion, enhancement and noise suppression processing. The brightness distribution difference between adjacent frames is calculated, and the stripe intensity index is calculated based on the frequency domain energy concentration, brightness skewness and brightness modulation depth. The flicker percentage, flicker index, and stripe intensity index within the scrolling time window are compared with preset thresholds, respectively. If any indicator exceeds the corresponding threshold, the trained improved LightGBM model is called to re-input the current feature vector, and steps three and four are executed. The control gear number is updated based on the inference results of the improved LightGBM model, and the corresponding drive parameter set is reread to trigger the switching and distribution of control parameters.

[0045] This invention continuously acquires the raw light intensity sequence output by a photodiode sensor and the image frame sequence output by a camera sensor within a set rolling time window, and calculates the flicker percentage, flicker index, and stripe intensity index respectively to dynamically assess the flicker risk of an LED driving system. Specifically, the flicker percentage is calculated by analyzing the peak-to-valley difference of brightness fluctuations and the periodic average brightness; the flicker index is calculated by analyzing the frequency domain energy distribution characteristics; and the stripe intensity index is extracted by analyzing the brightness difference between image frames and the frequency domain perturbation characteristics. These three factors together constitute the flicker evaluation index system. When any index exceeds a set threshold, the system immediately calls the trained improved LightGBM model, re-executes the inference process based on the current feature vector, generates a new control level number, and reads the corresponding driving parameter set to trigger a parameter update operation. This enables rapid response and dynamic adjustment to flicker anomalies, improving system stability and visual comfort.

[0046] refer to Figure 4 An LED lighting driver circuit control system for preventing flickering includes the following modules: The sensor data acquisition module is used to acquire data detected by the sensor. The time-frequency analysis and feature vector construction module is used to perform time-domain and frequency-domain analysis on the data detected by the acquired sensors and generate feature vectors in a unified format. The model training and inference module is used to train the iteratively improved LightGBM model. The control gear filtering and sorting module is used to determine the candidate control gear with the lowest overall cost as the final control gear number. The drive parameter generation and distribution module is used to control the operation of the LED lighting drive circuit according to the final control gear number; The closed-loop verification module is used to read the sensor and drive feedback signals after the drive parameters take effect, calculate the flicker index and compare it with the threshold. If the threshold is exceeded, the control gear is triggered to switch again. The flicker index calculation module is used to calculate the flicker percentage, flicker index, and stripe intensity index, and compare them with preset thresholds.

[0047] This invention constructs a modular system structure for flicker-proof control of LED lighting fixtures, including a sensor data acquisition module, a time-frequency analysis and feature vector construction module, a model training and inference module, a control level selection and sorting module, a drive parameter generation and distribution module, a closed-loop verification module, and a flicker index calculation module. The system first acquires multi-source signals such as light intensity, voltage, current, and temperature through the sensor data acquisition module. After time-domain and frequency-domain analysis, a unified format feature vector is constructed and input into an improved LightGBM model for training and inference, outputting the control level number and its multi-dimensional cost. In the control level selection and sorting module, the optimal control level is selected based on a comprehensive cost, and the drive parameter generation and distribution module allocates the corresponding parameters to drive the circuit. After the drive parameters take effect, the closed-loop verification module continuously reads the feedback signal and, in conjunction with the flicker index calculation module, monitors the flicker percentage, flicker index, and stripe intensity. If any index exceeds the limit, the system will trigger re-inference and level switching, thereby achieving dynamic closed-loop adjustment of flicker suppression and drive control.

[0048] In this embodiment, the improved LightGBM model has undergone the following structural optimizations to address the multi-objective decision-making requirements in LED lighting anti-flicker control: The improved LightGBM model employs joint supervision labels during the training phase. These labels include control level category labels, temperature rise cost labels, electromagnetic interference cost labels, and brightness jump cost labels, used to simultaneously perform classification and regression tasks. During training, the current cumulative control level logit vector and the current cumulative multidimensional regression value vector are initialized. The multidimensional regression value vector includes temperature rise cost components, electromagnetic interference cost components, and brightness jump cost components.

[0049] The model defines a joint loss function, which is a weighted combination of a classification loss function consisting of the cross-entropy loss between the control gear category label and the predicted category probability distribution, and three regression loss functions consisting of the squared loss between the temperature rise cost label and the predicted temperature rise cost, the squared loss between the electromagnetic interference cost label and the predicted electromagnetic interference cost, and the squared loss between the brightness jump cost label and the predicted brightness jump cost.

[0050] In each training iteration, the feature vectors are sequentially input into each decision tree to be trained, and gradient information is calculated based on the joint loss function. The control gear logit increment is determined by aggregating the first and second gradients calculated under cross-entropy loss for the sample set falling into the leaf node; the temperature rise cost increment, electromagnetic interference cost increment, and brightness jump cost increment are determined by aggregating the first and second gradients under squared loss, respectively.

[0051] The model constructs multiple subtrees during training, with each subtree storing the control level logit increment and multidimensional cost increment at its leaf nodes. In all trained decision trees, the control level logit increment is accumulated tree by tree to generate a cumulative control level logit vector; the temperature rise cost increment, electromagnetic interference cost increment, and brightness jump cost increment are accumulated tree by tree to generate a cumulative multidimensional regression value vector.

[0052] During the inference phase, the model normalizes the cumulative control level logit vector using the Softmax function to obtain a control level probability vector. From this probability vector, a set of candidate control levels with probability values ​​exceeding a preset threshold is selected. This set is then sorted based on the predicted temperature rise cost, predicted electromagnetic interference cost, and predicted brightness jump cost to determine the candidate control level number with the lowest overall cost, which is used to drive parameter generation.

[0053] This invention proposes an improved LightGBM model training and inference method for flicker prevention control of LED lighting fixtures. By introducing joint supervision labels, the classification task of control levels and the multi-objective regression task of temperature rise, electromagnetic interference, and brightness jump are unified into a single model. During the training phase, the logit vector of the control level and the multi-dimensional regression value vector are initialized, and multiple decision trees are iteratively constructed based on the joint loss function. Each decision tree outputs the logit increment of the control level and the multi-dimensional cost increment at its leaf nodes. During training, the learning paths of classification and regression directions are updated by gradient aggregation. During the inference phase, the logit increments of all subtrees are normalized using Softmax to select a set of candidate control levels that meet the probability threshold. Then, the control levels are sorted according to the cost values ​​of temperature rise, electromagnetic interference, and brightness jump output by the model to determine the control level number with the minimum comprehensive cost, providing multi-objective optimization results for the generation of driving parameters.

[0054] Example 1: To verify the feasibility of this invention in practice, it was applied to a lighting system renovation project in a large indoor industrial work area. This work area previously used a traditional constant-gear drive method to control LED lights, which frequently resulted in problems such as fluctuations in light source brightness, strong electromagnetic interference, and stripe interference in camera images. These issues severely impacted operator efficiency and the accuracy of the intelligent monitoring system, especially during periods of intensive equipment operation and frequent power fluctuations, when the flickering problem was particularly pronounced.

[0055] In this application scenario, the goal of upgrading the LED lighting system is not only to solve the visible flicker problem, but also to ensure the stability of the light source brightness, keep the impact of electromagnetic interference on the camera equipment within an acceptable range, and maintain the long-term operation of the lighting fixtures under low heat load. A data acquisition module equipped with photodiodes and ambient light sensors, along with a data acquisition path connecting to the feedback signal of the drive circuit, was deployed on-site. A camera acquisition system was also integrated to assist in flicker and image stripe detection. By deploying the control method of this invention, multi-dimensional input features, including light intensity fluctuations, power supply status, and temperature changes, were collected and uniformly formatted into feature vectors, which were then input into the trained improved LightGBM model.

[0056] The model, based on a joint supervision label and joint loss function mechanism, outputs not only the currently recommended control level number during each inference round, but also the predicted cost value of that level in terms of temperature rise, electromagnetic interference, and brightness fluctuation. The system automatically filters candidate levels above the probability threshold based on flicker evaluation results and performs cost ranking to select the optimal solution. Finally, the control parameters, after safety verification, are written into the drive circuit to achieve dynamic adjustment of core parameters such as PWM modulation frequency, filter frequency, and proportional gain.

[0057] After the system starts running, it automatically performs rolling flicker index analysis every 10 minutes. If the flicker percentage, stripe intensity, or brightness fluctuation exceeds the preset threshold, the model inference is immediately re-initiated and the control level is switched to complete the dynamic closed-loop optimization of the drive control. The experiment lasted for a whole week, covering different load states during day and night, and the control effect indicators during typical time periods were recorded. Specific experimental data are shown in Table 1:

[0058] Table 1. Evaluation of LED luminaire control effect based on the improved LightGBM model.

[0059] Comparison data in Table 1 shows that before the upgrade, the flicker percentage of LED lights in the area reached a maximum of 26.8% during high-load operation, the stripe intensity of the camera images was close to 0.38, brightness fluctuated frequently, and the surface temperature of the equipment reached over 90℃. After adopting the method of this invention, the average flicker percentage decreased to 10.9%, the stripe intensity was controlled below 0.13, and the heat load of the lights remained stable below 75℃. The false judgment rate of the image recognition system decreased from 12.6% to 1.7%, and the number of interference alarms caused by power fluctuations decreased by more than 85%. The entire system can operate stably under multi-objective constraints and effectively adapt to environmental changes, demonstrating excellent flicker control capabilities and multi-factor balancing capabilities.

[0060] This embodiment starts with a real and complex environment, demonstrating the effective application of the present invention under multiple constraints such as high fluctuating load, electrical interference and flicker suppression. It not only significantly improves the stability and intelligence of LED driving systems, but also provides a feasible technical path for lighting flicker control in industrial scenarios.

[0061] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for controlling LED lamp driving circuits to prevent flickering, characterized in that, Includes the following steps: Step 1: Collect data detected by the sensors and generate a feature dataset; Step 2: Perform time-domain and frequency-domain analysis on the sensor-detected data to construct feature vectors; Step 3: Train the iteratively improved LightGBM model. Input the feature vector into the trained improved LightGBM model and output multi-dimensional results at the leaf nodes, including the control gear number and the corresponding temperature rise cost, electromagnetic interference cost, and brightness jump cost. Step 4: Based on the multi-dimensional results, execute the multi-objective sorting rules, prioritize the selection of the level that meets the anti-flicker index, and determine the control level according to the order of temperature rise cost, electromagnetic interference cost, and brightness jump cost among the levels that meet the anti-flicker index. Step 5: Send the set of drive parameters for the control gear to the LED lighting driver circuit to control the operation of the driver circuit; Step 6: Calculate the flicker evaluation index within the rolling time window, compare it with the threshold, and if the threshold is exceeded, trigger the re-inference of the improved LightGBM model and switch to a new control level.

2. The LED lamp driving circuit control method for preventing flickering effect according to claim 1, characterized in that, The data detected by the acquisition sensor is used to generate a feature dataset, specifically as follows: Collect the LED light output waveform signal from the photodiode sensor and extract the original light intensity sequence representing the brightness change; Collect the illuminance values ​​output by the ambient light sensor and extract the numerical illuminance index representing the current light intensity; Acquire image frame sequences output by the camera sensor, analyze the inter-frame brightness distribution, and calculate stripe interference criterion data; The bus voltage, current and temperature signals in the drive circuit are collected to obtain power fluctuation characteristics, load state characteristics and thermal state characteristics, respectively.

3. The LED lamp driving circuit control method for preventing flickering effect according to claim 1, characterized in that, The process of performing time-domain and frequency-domain analysis on the data detected by the sensor to construct a feature vector is as follows: Perform a fast Fourier transform on the original light intensity sequence to extract the harmonic energy values ​​of the dominant frequency component; A sliding window process is performed on the original light intensity sequence to calculate the average brightness value and brightness change amplitude within each window, thus obtaining the brightness fluctuation amplitude sequence. Time series modeling is performed on the brightness fluctuation amplitude sequence to extract the phase shift and periodic fluctuation trend, thus obtaining the brightness fluctuation phase; The luminance fluctuation amplitude sequence, luminance fluctuation phase, harmonic energy value, numerical illuminance index and stripe interference criterion data are combined according to a unified feature structure to generate a feature vector.

4. The LED lamp driving circuit control method for preventing flickering effect according to claim 1, characterized in that, The trained iteratively improved LightGBM model is specifically as follows: During the training iteration, the current cumulative control level logit vector and the current cumulative multidimensional regression value vector are initialized as constant vectors. The multidimensional regression value vector includes the temperature rise cost component, the electromagnetic interference cost component, and the brightness jump cost component. Set monitoring labels, which consist of control level category labels, temperature rise cost labels, electromagnetic interference cost labels, and brightness jump cost labels; The predicted category probability distribution is calculated by normalizing the current cumulative control level logit vector using the Softmax function, and the predicted cost during the training phase is output by each component of the current cumulative multidimensional regression value vector. The predicted cost during the training phase includes the cost of temperature rise, the cost of predicted electromagnetic interference, and the cost of predicted brightness jump. Define a joint loss function, which is a weighted combination of a classification loss function and three regression loss functions. The classification loss function is the cross-entropy loss between the control gear category label and the predicted category probability distribution. The three regression loss functions are the squared loss between the temperature rise cost label and the predicted temperature rise cost, the squared loss between the electromagnetic interference cost label and the predicted electromagnetic interference cost, and the squared loss between the brightness jump cost label and the predicted brightness jump cost. Based on the joint loss function, the feature vectors are input into the improved LightGBM model to generate the first decision tree; Repeat the training iteration steps until the stopping conditions of the number of N decision trees and the split gain threshold are met, forming an improved LightGBM model composed of N decision trees.

5. The LED lamp driving circuit control method for preventing flickering effect according to claim 1, characterized in that, The improved LightGBM model, which inputs feature vectors into its training, outputs multi-dimensional results at its leaf nodes, including control gear numbers and corresponding temperature rise costs, electromagnetic interference costs, and brightness jump costs. Specifically: Input the feature vectors into the trained improved LightGBM model; Within each trained decision tree, a path is selected layer by layer to the leaf node based on the splitting condition; Read the control level logit vector and multidimensional regression value vector stored in the leaf node; The control gear logit vector is obtained by accumulating the control gear logit increments output by the leaf nodes of each decision tree, and the multidimensional regression value vector is obtained by accumulating the cost increments output by the leaf nodes of each decision tree; the cost increments are the temperature rise cost increment, the electromagnetic interference cost increment, and the brightness jump cost increment. The control gear logit increment is determined by aggregating the first and second gradients calculated from the sample set falling into the leaf node under cross-entropy loss. The cost increment is determined by aggregating the first and second gradients calculated under squared loss for the sample set falling into the leaf node. The cumulative control gear logit vector is obtained by accumulating the control gear logit vector among all trained decision trees, and the cumulative predicted cost is obtained by accumulating the multidimensional regression value vector. Normalize the cumulative control gear logit vector using the Softmax function to obtain the control gear probability vector; The control gear number is determined based on the control gear probability vector, and the control gear number is associated with the cumulative predicted cost. The output consists of a multi-dimensional result composed of the control gear number and the cumulative predicted cost.

6. The LED lamp driving circuit control method for preventing flickering effect according to claim 1, characterized in that, The multi-objective ranking rule based on multi-dimensional results is specifically as follows: Candidate control gears with probability values ​​exceeding a preset threshold are selected from the control gear probability vector to form a candidate control gear set; The anti-flicker index is checked one by one in the candidate control gear set. Candidate control gears that do not meet the anti-flicker index are removed, and the candidate control gear set that meets the anti-flicker index is obtained. Among the candidate control levels that meet the anti-flicker index, they are sorted according to the numerical values ​​of the predicted temperature rise cost, the predicted electromagnetic interference cost, and the predicted brightness jump cost. The candidate control position with the lowest overall cost is selected as the final control position number based on the sorting results. The comprehensive cost is composed of the predicted temperature rise cost, the predicted electromagnetic interference cost, and the predicted brightness jump cost. In the set of candidate control positions that meet the anti-flicker index, the values ​​of the predicted temperature rise cost, the predicted electromagnetic interference cost, and the predicted brightness jump cost are compared in turn, and the candidate control position with the smallest comprehensive cost is selected as the final control position number. Output the final control gear number and the corresponding cumulative multidimensional regression value vector.

7. The LED lamp driving circuit control method for preventing flickering effect according to claim 1, characterized in that, The step of sending the set of drive parameters for the control gear to the LED lighting driver circuit to control the operation of the driver circuit specifically involves: Based on the final control gear number, the corresponding set of driving parameters is read from the preset parameter library. The set of driving parameters includes carrier frequency parameters, model predictive control weight parameters, proportional-integral gain parameters, and notch filter center frequency parameters. A safety constraint verification is performed on the set of driving parameters, which includes threshold comparison of bus voltage ripple, current rise slope, device temperature and brightness step amplitude; the set of driving parameters is gradually adjusted within a preset gradual time window to generate a transition parameter sequence arranged in time sequence; and the transition parameter sequence is sequentially sent to the LED lighting driver circuit. Parameters are synchronously written at the beginning boundary of the pulse width modulation switching cycle; After the parameters take effect, a closed-loop verification is performed: the sensor and drive feedback signals are read, the anti-flicker index is calculated and compared with the threshold. If the threshold is exceeded, the control level is switched again. The drive feedback signal refers to the power fluctuation characteristics, load state characteristics, and thermal state characteristics obtained by feature extraction from the bus voltage signal, current signal, and temperature signal collected by the drive circuit. Record the time information and control gear number when the drive parameters take effect.

8. The LED lamp driving circuit control method for preventing flickering effect according to claim 1, characterized in that, The calculated flicker evaluation index within the rolling time window is compared with a threshold. If the threshold is exceeded, the improved LightGBM model is re-inferred and a new control level is switched. Specifically: Within a set length of rolling time window, the raw light intensity sequence output by the photodiode sensor and the image frame sequence output by the camera sensor are continuously acquired; The light intensity sequence is filtered and normalized, the difference between the maximum and minimum brightness between two adjacent periods is calculated, the peak-to-valley difference of brightness fluctuation is obtained, and the flicker percentage is calculated based on the peak-to-valley difference and the average brightness of the period. Frequency domain energy distribution analysis is performed on the brightness fluctuation amplitude sequence to calculate the variance of frequency component energy and the relative offset of the dominant frequency intensity. The scintillation index is then calculated based on the frequency domain perturbation characteristics. The image frame sequence acquired by the camera sensor is subjected to grayscale conversion, enhancement and noise suppression processing. The brightness distribution difference between adjacent frames is calculated. The stripe intensity index is calculated based on the frequency domain energy concentration, brightness skewness and brightness modulation depth. The flicker evaluation indicators are flicker percentage, flicker index, and stripe intensity index; The flicker evaluation index within the rolling time window is compared with the preset threshold. If any index exceeds the corresponding threshold, the trained improved LightGBM model is called to re-input the current feature vector, and the processes of steps three and four are executed. The control gear number is updated based on the inference results of the improved LightGBM model, and the corresponding drive parameter set is reread to trigger the switching and distribution of control parameters.

9. An LED lighting fixture driver circuit control system for preventing flicker, comprising the LED lighting fixture driver circuit control method for preventing flicker as described in any one of claims 1 to 8, characterized in that, Includes the following modules: The sensor data acquisition module is used to acquire data detected by the sensor. The time-frequency analysis and feature vector construction module is used to perform time-domain and frequency-domain analysis on the data detected by the acquired sensors and generate feature vectors in a unified format. The model training and inference module is used to train the iteratively improved LightGBM model. The control gear filtering and sorting module is used to determine the candidate control gear with the lowest overall cost as the final control gear number. The drive parameter generation and distribution module is used to control the operation of the LED lighting drive circuit according to the final control gear number; The closed-loop verification module is used to read the sensor and drive feedback signals after the drive parameters take effect, calculate the flicker index and compare it with the threshold. If the threshold is exceeded, the control gear is triggered to switch again. The flicker index calculation module is used to calculate the flicker percentage, flicker index, and stripe intensity index, and compare them with preset thresholds.

Images