Intelligent feedback-based precise current control method for LED lamp beads
By calculating the dynamic correlation coefficient matrix and predicting the optimal current value using a neural network model, and combining weighted averaging and adaptive closed-loop control, the hysteresis problem of LED lamp bead current control is solved, achieving precise and stable current control and improving the luminous performance and lifespan of the LED display screen.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENZHEN INFILED ELECTRONICS
- Filing Date
- 2026-04-10
- Publication Date
- 2026-05-29
Smart Images
Figure CN122116800A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of LED display technology, specifically to a method for precise current control of LED beads based on intelligent feedback. Background Technology
[0002] LED LED displays, with their advantages of high brightness, fast response speed, high contrast, wide color gamut, and long lifespan, have been widely used in various fields such as outdoor advertising, sports venues, stage backgrounds, monitoring and dispatching, and public information display, becoming one of the core carriers of modern display technology. As LED displays rapidly develop towards high definition, small pitch, high brightness, and low power consumption, higher requirements are placed on the current control accuracy of LED beads.
[0003] Current control of LED beads often uses constant current drive with fixed parameters or simple PID regulation. It only uses the drive current as a single feedback parameter and does not fully consider the dynamic changes of optical parameters such as forward voltage, junction temperature, ambient temperature, light intensity, and color coordinates of LED beads. This results in lag in current regulation, making it difficult to output the optimal current value under the current operating conditions. Ultimately, this leads to problems such as uneven brightness, color distortion, and poor gray uniformity.
[0004] Therefore, to meet current needs, a precise current control method for LED chips based on intelligent feedback is proposed. Summary of the Invention
[0005] The purpose of this invention is to provide a precise current control method for LED beads based on intelligent feedback. By calculating a dynamic correlation coefficient matrix, it accurately identifies the key correlations between variables in the working state of LED beads, enhances the model's sensitivity to key features, effectively ensures the quality of input data, and improves the accuracy and reliability of LED bead current control. It predicts the optimal current value through a neural network model and quantifies the prediction reliability, providing robust references when prediction confidence is low by combining historical data and a random tree model. The final current setpoint is generated using a weighted average method, coupled with a two-stage adaptive closed-loop control mechanism, to achieve precise and stable control of LED bead current, improve the luminous performance and lifespan of LED displays, and solve the problems mentioned in the background art.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] The method for precise current control of LED chips based on intelligent feedback includes the following steps:
[0008] The driving current, forward voltage, junction temperature or ambient temperature and optical parameters of the LED beads are collected in real time. The optical parameters include, but are not limited to, light intensity and color coordinates. The driving current, forward voltage, temperature and optical parameters constitute the real-time feature vector of the LED beads' working status.
[0009] The real-time feature vector and the preset optical target value are used as the input set and fed into a pre-trained neural network model for prediction, outputting the optimal current value under the current operating conditions.
[0010] The optimal current value is used as the setpoint of the PID controller, and the collected real-time current value is used as the process value of the PID controller. The error between the setpoint and the process value is calculated, and a control signal is generated based on proportional, integral, and derivative operations to drive and adjust the actual current applied to the LED beads so that it approaches the optimal current value.
[0011] After adjusting the current via the PID controller, the optical parameters are re-acquired and compared with the preset optical target value. If the error is within the allowable range, the adjustment is considered successful; otherwise, readjustment is triggered until the optical target value is reached.
[0012] Furthermore, the real-time feature vector of the LED chip's operating state, composed of driving current, forward voltage, temperature, and optical parameters, includes the following steps:
[0013] Calculate the dynamic correlation coefficient matrix between variables in the real-time feature vector to obtain the linear and nonlinear correlation strength between variables under the current working state of the LED beads;
[0014] Based on the correlation coefficient matrix, real-time feature variables that are strongly correlated with the preset optical target values are extracted and given high feature weights to enhance the sensitivity of the neural network model to key features.
[0015] If any parameter in the collected real-time data deviates from its relationship with other related variables in the correlation coefficient matrix, then the data may be an outlier.
[0016] A data correction mechanism is triggered, using estimates based on relevant variables as temporary replacements to ensure the data quality of the input set.
[0017] Furthermore, the dynamic correlation coefficient matrix between variables in the real-time feature vector is calculated to obtain the linear and nonlinear correlation strength between variables under the current operating state of the LED beads, including:
[0018] Calculate the linear correlation coefficient and nonlinear correlation coefficient between any two vectors in the real-time feature vector. The linear correlation coefficient is obtained based on the Pearson correlation coefficient, and the nonlinear correlation coefficient is obtained based on the distance correlation coefficient.
[0019] Extract auxiliary feature parameters of real-time feature vector pairs, including linear fit goodness, data signal-to-noise ratio, dynamic change rate of operating conditions, historical correlation stability, and correction reliability;
[0020] Using the auxiliary feature parameters as the criterion layer, the optimal weight allocation of the auxiliary feature parameters as the target layer, and the pairwise importance comparison of the criterion layer as the scheme layer, the hierarchical analysis method is adopted to determine the optimal weight allocation for each auxiliary feature parameter, obtain the sum of the products of the auxiliary feature parameters and their corresponding optimal weight allocation, and normalize the sum to obtain the initial linear weights.
[0021] When the value of the real-time feature vector pair is within the preset linear interval, the interval factor is determined to be the sum of 1 and the single interval gain; otherwise, the interval factor is determined to be the difference between 1 and the single interval gain. The product between the initial linear weight and the interval factor is used as the adaptive linear weight, and 1 minus the adaptive linear weight is the adaptive nonlinear weight.
[0022] Based on adaptive linear weights and adaptive nonlinear weights, the linear correlation coefficients and nonlinear correlation coefficients of real-time feature vector pairs are weighted and fused to obtain the comprehensive correlation degree of real-time feature vector pairs.
[0023] A dynamic correlation coefficient matrix is constructed based on the comprehensive correlation of all real-time feature vector pairs.
[0024] Furthermore, outputting the optimal current value under the current operating conditions includes the following steps:
[0025] When the neural network model outputs the best current prediction value, it simultaneously outputs a confidence coefficient that quantifies the reliability of this prediction.
[0026] If the prediction confidence coefficient is higher than the set threshold, the predicted value will be used as the set value of the PID controller.
[0027] If the prediction confidence coefficient is lower than the set threshold, it indicates that the current working conditions are beyond the effective coverage of the neural network model training data and that there is uncertainty in the prediction results. In this case, the prediction will be re-made based on the historical database.
[0028] Furthermore, predictions are re-made based on historical databases, including the following steps:
[0029] Retrieve the K most similar historical points from the historical database that are most similar to the current feature vector, and assign high weights to dimensions related to the prediction confidence coefficient;
[0030] Based on the retrieved K similar historical points, the current values with verified good control effects are extracted from the K historical points to form an empirical candidate value set;
[0031] The current feature vector is input into a preset random tree model, and a robust reference current value is output.
[0032] The optimal current value predicted by the neural network model, the median of the empirical candidate value set, and the robust reference current value are used as a common basis for decision-making.
[0033] The weights of the three factors are dynamically allocated based on the prediction confidence coefficient. If the confidence coefficient is low, the weight of the best current value predicted by the neural network model is reduced, while the weights of the candidate value set based on historical experience and the robust reference value are increased.
[0034] The final optimal current setting value is generated by using a weighted average method.
[0035] Furthermore, the real-time feature vectors and preset optical target values are used as input sets and fed into a pre-trained neural network model for prediction, including the following steps:
[0036] A multi-dimensional dataset of LED beads of different models operating at different temperatures and aging stages was constructed. The dataset includes the correspondence between feature vectors and their corresponding optimal current values.
[0037] A neural network model is constructed, and the model is trained using a dataset. The current prediction error, optical output error, and energy efficiency index are comprehensively analyzed to capture the time-series dependency between the feature vectors in the dataset and the corresponding optimal current values.
[0038] Furthermore, a control signal is generated based on proportional, integral, and derivative operations to drive and adjust the actual current applied to the LED beads, including the following steps:
[0039] Based on the generated optimal current setting value, a two-stage adaptive closed-loop current tracking control mechanism is set up.
[0040] With the optimal current setting as the target and the real-time acquired drive current value as feedback, a current closed loop is formed;
[0041] The primary control signal is calculated using proportional, integral, and derivative control algorithms. Based on the real-time feature vector, the measurable disturbance is compensated by feedforward. The feedforward compensation is then superimposed on the primary control signal to generate a current loop control command.
[0042] Furthermore, generating control signals based on proportional, integral, and derivative operations to drive and adjust the actual current applied to the LED beads also includes the following steps:
[0043] Based on the preset optical target value and with real-time acquired optical parameters as feedback, an outer performance monitoring loop is formed;
[0044] Optical parameters are collected according to the outer ring sampling period and compared with the preset target. If the deviation of the optical parameters persists and exceeds the allowable range, the optimal current value is finely adjusted and corrected according to the long-term deviation of the optical parameters to compensate for the long-term drift of the system.
[0045] The current loop control command is dynamically smoothed and filtered, and an anti-integral saturation limit is applied to generate the final safe drive signal and adjust the actual current of the LED beads.
[0046] Furthermore, it also includes the following steps:
[0047] If multiple adjustments fail to bring the light intensity and color coordinates to the target values, an alarm mechanism will be triggered and relevant logs will be recorded.
[0048] After the PID controller is successfully adjusted, the timestamp, real-time feature vector, optimal current value, actual PID output current value, and final optical parameter results are continuously recorded as a historical operating data sequence.
[0049] Calculate the comprehensive performance indicators within the specified period, including but not limited to the steady-state error of optical parameters, current fluctuation variance, and power consumption per unit light output;
[0050] When the comprehensive performance index is lower than the preset threshold for several consecutive periods or when a real-time feature vector is detected to show a trend deviation, the neural network model update process is triggered.
[0051] New training samples are extracted from the cached historical running data sequence, and incremental learning algorithms are used to fine-tune the parameters of the pre-trained neural network model.
[0052] Furthermore, when outputting the optimal current prediction value, the neural network model simultaneously outputs a confidence coefficient that quantifies the reliability of this prediction, including:
[0053] Based on historical prediction data of the neural network model, the prediction error, working condition distribution deviation and model output stability parameters of the neural network model are obtained, and the current feature parameters of the LED beads are obtained to calculate the confidence calibration coefficient.
[0054] Based on the confidence calibration coefficient and combined with the piecewise Sigmoid function, the confidence coefficient of the reliability of this prediction is calculated.
[0055] Compared with the prior art, the beneficial effects of the present invention are:
[0056] 1. In this invention, by calculating the dynamic correlation coefficient matrix, the key correlations between variables in the working state of LED beads are accurately identified, providing real-time feature variables with high feature weights for the neural network model, enhancing the model's sensitivity to key features, and thus improving prediction accuracy; at the same time, it effectively ensures the quality of input data, ensures the stable operation of the entire intelligent feedback control system, improves the accuracy and reliability of LED bead current control, and extends the service life of the LED display screen.
[0057] 2. In this invention, the optimal current value is predicted by a neural network model and the reliability of the prediction is quantified. Historical data and a random tree model are combined to provide a robust reference when the prediction confidence is low. The final current setpoint is generated by a weighted average method and combined with a two-level adaptive closed-loop control mechanism to achieve precise and stable control of the LED lamp current, effectively compensate for interference and long-term drift, and improve the luminous performance and service life of the LED display screen. Attached Figure Description
[0058] Figure 1 This is a flowchart of the LED lamp bead precision current control method based on intelligent feedback according to the present invention. Detailed Implementation
[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0060] To address the technical problem that existing technologies cannot accurately adjust the current in real time according to changes in the actual operating status of LED chips, resulting in low luminous efficiency, unstable light color, and shortened lifespan of LED displays, thus failing to maintain optimal luminous performance, please refer to [link to relevant documentation]. Figure 1 This embodiment provides the following technical solution:
[0061] The method for precise current control of LED chips based on intelligent feedback includes the following steps:
[0062] Real-time acquisition of LED chip drive current, forward voltage, junction temperature or ambient temperature, and optical parameters, including but not limited to luminous intensity and color coordinates, constitutes a real-time feature vector of the LED chip's operating state, comprising the following steps:
[0063] Calculate the dynamic correlation coefficient matrix between variables in the real-time feature vector to obtain the linear and nonlinear correlation strength between variables under the current working state of the LED beads; based on the correlation coefficient matrix, extract real-time feature variables that are strongly correlated with the preset optical target value and assign them high feature weights. For example, within a specific temperature range, the positive voltage and luminous efficacy show a significant negative correlation, enhancing the sensitivity of the neural network model to key features; if any parameter in the collected real-time data, such as light intensity, deviates significantly from the expected relationship between it and other highly correlated variables, such as driving current and temperature, in the correlation coefficient matrix, it is judged that the data may be an outlier. For example, if the light intensity decreases abnormally when the current increases, a data correction mechanism is triggered, and the estimated value based on the correlated variables is used for temporary replacement to ensure the data quality of the input set.
[0064] The beneficial effects achieved by the above are as follows: by calculating the dynamic correlation coefficient matrix, the key correlations between variables in the working state of LED beads are accurately identified, providing real-time feature variables with high feature weights for the neural network model, enhancing the model's sensitivity to key features, and thus improving prediction accuracy; at the same time, it effectively ensures the quality of input data, ensures the stable operation of the entire intelligent feedback control system, improves the accuracy and reliability of LED bead current control, and extends the service life of the LED display screen.
[0065] In one embodiment, the dynamic correlation coefficient matrix between variables in the real-time feature vector is calculated to obtain the linear and nonlinear correlation strength between variables under the current operating state of the LED beads, including:
[0066] Calculate the linear correlation coefficient and nonlinear correlation coefficient between any two vectors in the real-time feature vector. The linear correlation coefficient is obtained based on the Pearson correlation coefficient, and the nonlinear correlation coefficient is obtained based on the distance correlation coefficient.
[0067] Extract auxiliary feature parameters of real-time feature vector pairs, including linear fit goodness, data signal-to-noise ratio, dynamic change rate of operating conditions, historical correlation stability, and correction reliability;
[0068] Using the auxiliary feature parameters as the criterion layer, the optimal weight allocation of the auxiliary feature parameters as the target layer, and the pairwise importance comparison of the criterion layer as the scheme layer, the hierarchical analysis method is adopted to determine the optimal weight allocation for each auxiliary feature parameter, obtain the sum of the products of the auxiliary feature parameters and their corresponding optimal weight allocation, and normalize the sum to obtain the initial linear weights.
[0069] When the value of the real-time feature vector pair is within the preset linear interval, the interval factor is determined to be the sum of 1 and the single interval gain; otherwise, the interval factor is determined to be the difference between 1 and the single interval gain. The product between the initial linear weight and the interval factor is used as the adaptive linear weight, and 1 minus the adaptive linear weight is the adaptive nonlinear weight.
[0070] Based on adaptive linear weights and adaptive nonlinear weights, the linear correlation coefficients and nonlinear correlation coefficients of real-time feature vector pairs are weighted and fused to obtain the comprehensive correlation degree of real-time feature vector pairs.
[0071] A dynamic correlation coefficient matrix is constructed based on the comprehensive correlation of all real-time feature vector pairs.
[0072] In this embodiment, a larger linear fit goodness value indicates a more obvious linear relationship between the real-time feature vector pairs.
[0073] In this embodiment, the dynamic change rate of the operating conditions is the mean of the first difference of the variables within the window, which is the dynamic change rate of the real-time feature vector pairs.
[0074] In this embodiment, the specific steps of the hierarchical analysis method are to compare the importance of auxiliary feature parameters pairwise, verify whether the comparison results are reasonable (consistency check), and obtain the optimal weight allocation for each auxiliary feature parameter through the consistency check results.
[0075] In this embodiment, the single-interval gain is, for example, 0.1.
[0076] The beneficial effects of the above design scheme are as follows: By calculating the linear and nonlinear correlation coefficients between any two vectors in the real-time feature vector, where the linear correlation coefficient is based on the Pearson correlation coefficient and the nonlinear correlation coefficient is based on the distance correlation coefficient, the Pearson correlation coefficient accurately locks the linear association, and the distance correlation coefficient comprehensively covers the nonlinear dependence. The two complement each other to avoid the limitations of a single coefficient, fully restore the true association between variables, and extract auxiliary feature parameters of the real-time feature vector pairs, including linear fit goodness, data signal-to-noise ratio, dynamic change rate of working conditions, historical association stability, and correction credibility. This provides reference from multiple dimensions such as fitting effect and data quality, abandons the bias of single-dimensional judgment, and lays a solid objective foundation for weight allocation. The auxiliary feature parameters are used as the criterion layer, the optimal weight allocation of the auxiliary feature parameters is used as the target layer, and the pairwise importance comparison of the criterion layer is used as the scheme layer. The hierarchical analysis method is used to determine the optimal weight allocation for each auxiliary feature parameter, obtain the sum of the products of the auxiliary feature parameters and their corresponding optimal weight allocations, and normalize the sum to obtain the initial... Linear weights are assigned by systematically comparing feature importance using the analytic hierarchy process (AHP) to quantify weight allocation, avoid subjective assumptions, and ensure that the initial linear weights align with core control requirements. When the values of the real-time feature vector pairs are within a preset linear interval, the interval factor is determined to be the sum of 1 and the single interval gain; otherwise, the interval factor is determined to be the difference between 1 and the single interval gain. The product of the initial linear weights and the interval factor is used as the adaptive linear weights. 1 minus the adaptive linear weights yields the adaptive nonlinear weights. The weights are dynamically adjusted according to the variable value intervals, strengthening linear weights in linearly advantageous intervals and emphasizing nonlinear weights outside the intervals to adapt to different operating conditions. Based on the adaptive linear and adaptive nonlinear weights, the linear and nonlinear correlation coefficients of the real-time feature vector pairs are weighted and fused to obtain the comprehensive correlation degree of the real-time feature vector pairs. A dynamic correlation coefficient matrix is constructed based on the comprehensive correlation degree of all real-time feature vector pairs. The comprehensive correlation degree is obtained by weighted fusion, and the dynamic correlation coefficient matrix is constructed to output high-quality correlation data, providing a reliable prerequisite for subsequent precise LED current control.
[0077] The real-time feature vector and preset optical target values, such as target light intensity and target color coordinates, are used as input sets and fed into a pre-trained neural network model for prediction, outputting the optimal current value under the current operating conditions; including the following steps:
[0078] A multi-dimensional dataset of LED beads of different models operating at different temperatures and aging stages was constructed. The dataset included the correspondence between feature vectors and their corresponding optimal current values. A neural network model was constructed and trained using the dataset. The current prediction error, optical output error, and energy efficiency index were comprehensively analyzed to capture the time-series dependency between feature vectors and their corresponding optimal current values in the dataset.
[0079] When the neural network model outputs the optimal current prediction value, it simultaneously outputs a confidence coefficient quantifying the reliability of this prediction. If the prediction confidence coefficient is higher than a set threshold, the predicted value is used as the setpoint for the PID controller. If the prediction confidence coefficient is lower than the set threshold, it indicates that the current operating condition exceeds the effective coverage of the neural network model's training data, such as extreme temperature or deep aging, and the prediction result has uncertainty. In this case, the prediction is re-performed based on the historical database. This includes the following steps:
[0080] The system retrieves the K most similar historical points from the historical database that are most similar to the current feature vector, assigning high weights to dimensions related to the prediction confidence coefficient, such as key anomalies that led to the low confidence level. Based on the retrieved K similar historical points, it extracts current values from these K historical points that have been verified to have good control effects, such as stable light and color and compliant efficiency, forming an empirical candidate value set. The current feature vector is then input into a pre-defined random tree model, which outputs a robust reference current value. The random tree model serves as an auxiliary model, using the same training data as the neural network model, to provide conservative but robust reference predictions under uncertain operating conditions. The optimal current value predicted by the neural network model, the median of the empirical candidate value set, and the robust reference current value are used as the common decision-making basis. The weights of the three are dynamically allocated according to the prediction confidence coefficient. If the confidence coefficient is low, the weight of the optimal current value predicted by the neural network model is reduced, while the weights of the candidate value set based on historical experience and the robust reference value are increased. A weighted average method is used to generate the final optimal current setpoint. This is to tilt the prediction results towards historical results that have been successfully controlled in historical data and conservative and robust predictions when the uncertainty of the prediction results is high, thereby avoiding unreliable outputs of a single model on out-of-distribution samples.
[0081] The optimal current value is used as the setpoint for the PID controller, and the collected real-time current value is used as the process value for the PID controller. The error between the setpoint and the process value is calculated, and a control signal is generated based on proportional, integral, and derivative operations to drive and adjust the actual current applied to the LED beads, making it approach the optimal current value. This includes the following steps:
[0082] Based on the generated optimal current setpoint, a two-stage adaptive closed-loop current tracking control mechanism is established. The optimal current setpoint is used as the target, and the real-time acquired drive current value is used as feedback to form a current closed loop. The primary control signal is calculated using proportional, integral, and derivative control algorithms. Based on the real-time feature vector, the measurable disturbance is fed forward to compensate, and the feedforward compensation is superimposed with the primary control signal to generate the current loop control command. This reduces the burden on the PID controller and improves the system response speed and stability.
[0083] Using a preset optical target value as a benchmark and real-time acquired optical parameters as feedback, an outer performance monitoring loop is constructed. Optical parameters are acquired according to the outer loop sampling period and compared with the preset target. If the deviation of the optical parameters persists and exceeds the allowable range, the optimal current value is fine-tuned based on the long-term deviation of the optical parameters to compensate for systemic long-term drift. Dynamic smoothing filtering is applied to the current loop control commands. When the current setpoint is stable, a lower cutoff frequency is used to effectively filter out high-frequency noise. When the current setpoint changes abruptly, the cutoff frequency is automatically increased to ensure rapid response and effectively avoid the impact of abrupt changes in the control signal on the drive components and LEDs. Anti-integral saturation limiting is applied to generate the final safe drive signal, driving the integrated circuit to adjust the actual current of the LED beads. When the PID controller output approaches the physical limit due to long-term errors, such as a duty cycle of 0% or 100%, the accumulation of the integral term is paused to prevent integral saturation and avoid huge overshoot when the system exits the saturation region.
[0084] After adjusting the current via the PID controller, the optical parameters are re-acquired and compared with the preset optical target value. If the error is within the allowable range, the adjustment is considered successful; otherwise, readjustment is triggered until the optical target value is reached.
[0085] If multiple adjustments fail to achieve the target values for light intensity and color coordinates, an alarm mechanism will be triggered, and relevant logs will be recorded for subsequent analysis of the problem, providing a basis for model optimization and system maintenance. After successful PID controller adjustment, timestamps, real-time feature vectors, optimal current values, actual PID output current values, and the final optical parameter results will be continuously recorded as historical operating data sequences. Comprehensive performance indicators within a specified period will be calculated, including but not limited to steady-state error of optical parameters, current fluctuation variance, and power consumption per unit light output. When the comprehensive performance indicators are below the preset threshold for several consecutive periods or when real-time feature vectors are detected, such as positive voltage changes under the same current or decreased luminous efficacy showing a trend deviation, the neural network model update process will be triggered. New training samples will be extracted from the cached historical operating data sequences, and incremental learning algorithms will be used to fine-tune the parameters of the pre-trained neural network model. During the fine-tuning process, the focus can be on enabling the neural network model to learn the aging trajectory and individual characteristics of the current specific LED bead, thereby generating a personalized adaptive model for the LED bead, which will be used for subsequent predictive control.
[0086] The beneficial effects achieved by the above are as follows: the optimal current value is predicted by a neural network model and the reliability of the prediction is quantified. Combined with historical data and a random tree model, a robust reference is provided when the prediction confidence is low. The final current setpoint is generated by a weighted average method. With the help of a two-level adaptive closed-loop control mechanism, the current of the LED beads is accurately and stably controlled, effectively compensating for interference and long-term drift, and improving the luminous performance and lifespan of the LED display screen.
[0087] Working principle: By real-time monitoring of the LED chip's drive current, forward voltage, temperature, and optical parameters, a real-time feature vector is formed. A neural network model is used to analyze the feature vector against the preset optical target value, outputting the optimal current prediction value and confidence coefficient. If the confidence coefficient is high, the predicted value is directly used as the PID controller setpoint. If the confidence coefficient is low, historical data and a random tree model are combined to re-predict the current. The actual current is adjusted through the PID controller and a dual closed-loop mechanism, with feedforward compensation and anti-integral saturation processing to ensure accurate and stable current. Simultaneously, operating data is continuously recorded. When performance indicators decline or operating conditions deviate, model updates are triggered to achieve adaptive optimization, adapting to LED chip aging and environmental changes.
[0088] In one embodiment, when the neural network model outputs the optimal current prediction value, it simultaneously outputs a confidence coefficient that quantifies the reliability of this prediction, including:
[0089] Based on historical prediction data of the neural network model, the prediction error, working condition distribution deviation and model output stability parameters of the neural network model are obtained, and the current feature parameters of the LED beads are obtained to calculate the confidence calibration coefficient.
[0090] Confidence calibration coefficient The calculation formula is as follows:
[0091] ;
[0092] in, , and These represent the weighting coefficients of the neural network model's prediction error, deviation from the operating condition distribution, and model output stability parameter, respectively. , and These represent the normalized values of the prediction error, deviation from the operating condition distribution, and the model output stability parameter of the neural network model, respectively. represents the natural constant, with a value of 2.72. Indicates the time decay coefficient. This represents the cumulative decay coefficient at junction temperature. This represents the attenuation coefficient due to exceeding operating limits. This indicates the cumulative working time of the LED beads. This indicates the rated operating factor of the LED chip. This indicates the cumulative junction temperature of the LED chip. This indicates the cumulative rated junction temperature of the LED chip. This indicates the upper limit of the junction temperature of the LED chip under normal operating conditions. This indicates the lower limit of the junction temperature of the LED chip under normal operating conditions. This indicates the real-time junction temperature of the LED chip. This represents the average junction temperature of the LED chip under normal operating conditions. ;
[0093] Based on the confidence calibration coefficient and combined with the piecewise Sigmoid function, the confidence coefficient of the reliability of this prediction is calculated.
[0094] The formula for calculating the confidence coefficient H is as follows:
[0095]
[0096] in, This represents the gain coefficient in the low reliability range. Indicates the gain coefficient in the reliable interval. Indicates the gain coefficient in the high reliability range. This represents the interval offset coefficient.
[0097] In this embodiment, the gain coefficient for the low reliability interval is 15, the gain coefficient for the medium reliability interval is 20, and the gain coefficient for the high reliability interval is 15, thus controlling the confidence mapping sensitivity of each reliability interval to be reasonable. The interval offset coefficient is 5 to ensure the continuity of the piecewise function.
[0098] In this embodiment, the time decay coefficient is used to quantify the decay effect of the cumulative working time of the LED beads on the prediction reliability; the longer the time, the more obvious the decay.
[0099] In this embodiment, the junction temperature cumulative attenuation coefficient quantifies the attenuation effect of the junction temperature cumulative value on the prediction reliability; the more the junction temperature accumulates, the more significant the attenuation.
[0100] In this embodiment, the cumulative junction temperature value of the LED bead represents the integral value of the junction temperature from startup to the current moment, characterizing the degree of thermal aging of the bead.
[0101] In this embodiment, the rated cumulative junction temperature of the LED bead is the maximum allowable cumulative junction temperature value specified by the manufacturer, which serves as the benchmark value for thermal aging degradation.
[0102] In this embodiment, the operating condition out-of-bounds attenuation coefficient is used to quantify the impact of the junction temperature out-of-bounds degree on the correction factor and control the attenuation rate.
[0103] In this embodiment, the confidence calibration coefficient All calculated parameters have been normalized and their values are set between [0, 1].
[0104] The beneficial effects of the above design scheme are as follows: By combining the three core dimensions of prediction error, operating condition distribution deviation, and model output stability with weighted calculations using weighted coefficients, it avoids the one-sidedness of a single indicator and comprehensively covers key influencing factors such as model performance and operating condition adaptability. This makes the confidence assessment more closely aligned with actual control scenarios. The introduction of time decay and junction temperature cumulative decay terms quantifies the impact of LED chip usage time and thermal aging on prediction reliability, aligning with the performance drift pattern of LED chips from new to aged. Furthermore, for situations where the real-time junction temperature deviates from the normal range, the calibration coefficient is adjusted using an operating condition out-of-range decay coefficient. To avoid misjudging high reliability under extreme operating conditions and improve the adaptability of the evaluation under different operating conditions, different mapping logics are set according to the calibration coefficient range. The middle range (0.2, 0.8) has higher gain and stronger sensitivity, while the two ends of the range transition smoothly. At the same time, the confidence level is limited to 0.05 to avoid evaluation failure caused by extreme values. It takes into account both sensitivity and robustness. Finally, through multi-dimensional fusion, full life cycle calibration and segmented mapping, the confidence coefficient is accurately and dynamically quantified, which provides a reliable basis for subsequent multi-source fusion prediction under low confidence, and ensures the stability and accuracy of current control.
[0105] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or high-voltage switchgear that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or high-voltage switchgear.
[0106] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for precise current control of LED beads based on intelligent feedback, characterized in that, Includes the following steps: The driving current, forward voltage, junction temperature or ambient temperature and optical parameters of the LED beads are collected in real time. The optical parameters include, but are not limited to, light intensity and color coordinates. The driving current, forward voltage, temperature and optical parameters constitute the real-time feature vector of the LED beads' working status. The real-time feature vector and the preset optical target value are used as the input set and fed into a pre-trained neural network model for prediction, outputting the optimal current value under the current operating conditions. The optimal current value is used as the setpoint of the PID controller, and the collected real-time current value is used as the process value of the PID controller. The error between the setpoint and the process value is calculated, and a control signal is generated based on proportional, integral, and derivative operations to drive and adjust the actual current applied to the LED beads so that it approaches the optimal current value. After adjusting the current via the PID controller, the optical parameters are re-acquired and compared with the preset optical target value. If the error is within the allowable range, the adjustment is considered successful; otherwise, readjustment is triggered until the optical target value is reached.
2. The method for precise current control of LED beads based on intelligent feedback according to claim 1, characterized in that, The real-time feature vector of LED chip operating status, composed of driving current, forward voltage, temperature, and optical parameters, includes the following steps: Calculate the dynamic correlation coefficient matrix between variables in the real-time feature vector to obtain the linear and nonlinear correlation strength between variables under the current working state of the LED beads; Based on the correlation coefficient matrix, real-time feature variables that are strongly correlated with the preset optical target values are extracted and given high feature weights to enhance the sensitivity of the neural network model to key features. If any parameter in the collected real-time data deviates from its relationship with the related variable in the correlation coefficient matrix, then the data may be an outlier. It also triggers a data correction mechanism, using estimates based on relevant variables as temporary replacements to ensure the data quality of the input set.
3. The method for precise current control of LED beads based on intelligent feedback according to claim 2, characterized in that, Calculate the dynamic correlation coefficient matrix among variables in the real-time feature vector to obtain the linear and nonlinear correlation strength among variables under the current operating state of the LED beads, including: Calculate the linear correlation coefficient and nonlinear correlation coefficient between any two vectors in the real-time feature vector. The linear correlation coefficient is obtained based on the Pearson correlation coefficient, and the nonlinear correlation coefficient is obtained based on the distance correlation coefficient. Extract auxiliary feature parameters of real-time feature vector pairs, including linear fit goodness, data signal-to-noise ratio, dynamic change rate of operating conditions, historical correlation stability, and correction reliability; Using the auxiliary feature parameters as the criterion layer, the optimal weight allocation of the auxiliary feature parameters as the target layer, and the pairwise importance comparison of the criterion layer as the scheme layer, the hierarchical analysis method is adopted to determine the optimal weight allocation for each auxiliary feature parameter, obtain the sum of the products of the auxiliary feature parameters and their corresponding optimal weight allocation, and normalize the sum to obtain the initial linear weights. When the value of the real-time feature vector pair is within the preset linear interval, the interval factor is determined to be the sum of 1 and the single interval gain; otherwise, the interval factor is determined to be the difference between 1 and the single interval gain. The product between the initial linear weight and the interval factor is used as the adaptive linear weight, and 1 minus the adaptive linear weight is the adaptive nonlinear weight. Based on adaptive linear weights and adaptive nonlinear weights, the linear correlation coefficients and nonlinear correlation coefficients of real-time feature vector pairs are weighted and fused to obtain the comprehensive correlation degree of real-time feature vector pairs. A dynamic correlation coefficient matrix is constructed based on the comprehensive correlation of all real-time feature vector pairs.
4. The method for precise current control of LED beads based on intelligent feedback according to claim 1, characterized in that, Output the optimal current value under the current operating conditions, including the following steps: When the neural network model outputs the best current prediction value, it simultaneously outputs a confidence coefficient that quantifies the reliability of this prediction. If the prediction confidence coefficient is higher than the set threshold, the predicted value will be used as the set value of the PID controller. If the prediction confidence coefficient is lower than the set threshold, it indicates that the current working conditions are beyond the effective coverage of the neural network model training data and that there is uncertainty in the prediction results. In this case, the prediction will be re-made based on the historical database.
5. The method for precise current control of LED beads based on intelligent feedback according to claim 4, characterized in that, Re-predicting based on historical databases includes the following steps: Retrieve the K most similar historical points from the historical database that are most similar to the current feature vector, and assign high weights to dimensions related to the prediction confidence coefficient; Based on the retrieved K similar historical points, the current values with verified good control effects are extracted from the K historical points to form an empirical candidate value set; The current feature vector is input into a preset random tree model, and a robust reference current value is output. The optimal current value predicted by the neural network model, the median of the empirical candidate value set, and the robust reference current value are used as a common basis for decision-making. The weights of the three factors are dynamically allocated based on the prediction confidence coefficient. If the confidence coefficient is low, the weight of the best current value predicted by the neural network model is reduced, while the weights of the candidate value set based on historical experience and the robust reference value are increased. The final optimal current setting value is generated by using a weighted average method.
6. The method for precise current control of LED beads based on intelligent feedback according to claim 1, characterized in that, The real-time feature vector and the preset optical target values are used as input sets and fed into a pre-trained neural network model for prediction, including the following steps: A multi-dimensional dataset of LED beads of different models operating at different temperatures and aging stages was constructed. The dataset includes the correspondence between feature vectors and their corresponding optimal current values. A neural network model is constructed, and the model is trained using a dataset. The current prediction error, optical output error, and energy efficiency index are comprehensively analyzed to capture the time-series dependency between the feature vectors in the dataset and the corresponding optimal current values.
7. The method for precise current control of LED beads based on intelligent feedback according to claim 1, characterized in that, The control signal is generated based on proportional, integral, and derivative operations to drive and adjust the actual current applied to the LED beads, including the following steps: Based on the generated optimal current setting value, a two-stage adaptive closed-loop current tracking control mechanism is set up. With the optimal current setting as the target and the real-time acquired drive current value as feedback, a current closed loop is formed; The primary control signal is calculated using proportional, integral, and derivative control algorithms. Based on the real-time feature vector, the measurable disturbance is compensated by feedforward. The feedforward compensation is then superimposed on the primary control signal to generate the current loop control command.
8. The method for precise current control of LED beads based on intelligent feedback according to claim 6, characterized in that, The control signal is generated based on proportional, integral, and derivative operations to drive and adjust the actual current applied to the LED beads. The process also includes the following steps: Based on the preset optical target value and with real-time acquired optical parameters as feedback, an outer performance monitoring loop is formed; Optical parameters are collected according to the outer ring sampling period and compared with the preset target. If the deviation of the optical parameters persists and exceeds the allowable range, the optimal current value is finely adjusted and corrected according to the long-term deviation of the optical parameters to compensate for the long-term drift of the system. The current loop control command is dynamically smoothed and filtered, and an anti-integral saturation limit is applied to generate the final safe drive signal and adjust the actual current of the LED beads.
9. The method for precise current control of LED beads based on intelligent feedback according to claim 1, characterized in that, It also includes the following steps: If multiple adjustments fail to bring the light intensity and color coordinates to the target values, an alarm mechanism will be triggered and relevant logs will be recorded. After the PID controller is successfully adjusted, the timestamp, real-time feature vector, optimal current value, actual PID output current value, and final optical parameter results are continuously recorded as a historical operating data sequence. Calculate the comprehensive performance indicators within the specified period, including but not limited to the steady-state error of optical parameters, current fluctuation variance, and power consumption per unit light output; When the comprehensive performance index is lower than the preset threshold for several consecutive periods or when a real-time feature vector is detected to show a trend deviation, the neural network model update process is triggered. New training samples are extracted from the cached historical running data sequence, and incremental learning algorithms are used to fine-tune the parameters of the pre-trained neural network model.
10. The method for precise current control of LED beads based on intelligent feedback according to claim 4, characterized in that, When outputting the optimal current prediction, the neural network model simultaneously outputs a confidence coefficient that quantifies the reliability of this prediction, including: Based on historical prediction data of the neural network model, the prediction error, working condition distribution deviation and model output stability parameters of the neural network model are obtained, and the current feature parameters of the LED beads are obtained to calculate the confidence calibration coefficient. Based on the confidence calibration coefficient and combined with the piecewise Sigmoid function, the confidence coefficient of the reliability of this prediction is calculated.