Incremental decision tree lighting energy saving control method and system with concept drift gating

Abstract

This invention relates to the field of intelligent lighting control technology, specifically to a concept drift-gated incremental decision tree lighting energy-saving control method and system. By collecting lighting zone operation data, maintaining short- and long-term statistical baselines, calculating the input distribution drift index and the control output distribution drift index based on the effective control output after constraints, and fusing them to obtain a joint drift level, the method employs gated incremental updates, shadow training, model takeover, frozen updates, and stable strategy switching. Simultaneously, hard constraints on lighting output are applied after model output and before protocol encoding, combined with online acceptance and a backoff evidence chain. This scheme can reduce the risks of drift mislearning, dimming oscillations, color temperature abrupt changes, and energy consumption rebound, improving the adaptability, stability, and traceability of lighting control.

Description

Technical Field

[0001] This invention relates to the field of intelligent lighting control, specifically to an incremental decision tree lighting energy-saving control method and system based on concept drift gating. Background Technology

[0002] Architectural lighting systems typically adjust luminaire brightness and color temperature based on data such as illuminance, occupant status, sunlight or ambient light, area usage, and manual intervention to reduce energy consumption and maintain visual comfort. With the increasing prevalence of lighting controllers, lighting gateways, and sensor acquisition modules, utilizing online learning models for continuous adaptation of lighting control strategies has become a feasible technical approach.

[0003] In real-world operation, lighting data distribution exhibits significant non-stationarity. Seasonal and solar radiation variations alter the contribution of natural daylight, changes in zoning use alter lighting demands, shifts in user behavior change switching and dimming preferences, and sensor drift alters the reliability of input data. If the lighting system relies solely on online incremental learning to directly update the decision tree model, the model may mislearn short-term anomalies, sensor drift, or local behavioral fluctuations as stable control patterns, leading to policy oscillations, frequent dimming flicker, abrupt color temperature changes, increased manual intervention, and energy consumption rebound.

[0004] Existing lighting energy-saving control solutions often involve solar compensation, human-perceived dimming, gradual dimming, or online model updates, but lack a unified risk encapsulation mechanism for the incremental decision tree lighting control chain. Especially between model updates, model takeover, effective control command issuance, and failure rollback, without clear conceptual drift gating, same-path simulation verification, online acceptance, and evidence chain recording, the lighting control system can hardly simultaneously meet the requirements of continuous adaptation, stable execution, and auditable rollback. Summary of the Invention

[0005] The purpose of this invention is to provide a technical solution to address the aforementioned problems existing in the prior art. Specifically, this invention is achieved through the following technical solution: The incremental decision tree lighting energy-saving control method with concept drift gating includes the following steps: Step 1: Collect illuminance, personnel status, sunlight or ambient light, area usage status, manual intervention, luminaire power or drive status feedback, historical effective control output, and the original model output corresponding to the historical effective control output for the lighting zones, and maintain short-window statistical baseline and long-window statistical baseline. Step 2: Calculate the input distribution drift index of key input features based on the short-window statistical baseline and the long-window statistical baseline, and calculate the control output distribution drift index based on the historical effective control output and its corresponding original model output. Then, merge the input distribution drift index and the control output distribution drift index into a joint drift level. Step 3: Perform external control based on the joint drift level. When there is a slight drift, perform restricted incremental updates and generate candidate models. When there is a moderate drift, only train the shadow model. When there is a severe drift, freeze the incremental updates and switch to a stable lighting control strategy. Then, determine the incremental decision tree model with takeover permission as the current takeover model. Step four: According to the control cycle, apply a predetermined lighting output hard constraint to the target brightness and target color temperature output by the current takeover model or the stable lighting control strategy in the current control cycle, after the model output and before the lighting control protocol encoding, so as to form an effective control command for the current control cycle; and execute the same lighting output hard constraint simulation path to the target brightness and target color temperature output by the candidate model or the shadow model in the current control cycle, so as to form a simulation effective control command. Step 5: Conduct a controlled trial run on the candidate model that has been verified through the lighting output hard constraint simulation path and formed the effective control command of the simulation, and count the energy saving rate, response delay, manual reverse intervention rate and comfort risk event rate; when the candidate model meets the takeover conditions, the joint drift level allows takeover and the key feedback link is effective, grant the candidate model takeover permission and upgrade it to a stable version; otherwise, revert to the previous stable version model and record the revert evidence chain.

[0006] Furthermore, the key input features include illuminance, personnel status, sunlight or ambient light, area usage status, and manual intervention events; the control output distribution drift index is obtained by comparing the differences between the effective luminance output distribution, effective color temperature output distribution, strategy oscillation rate, and constraint saturation rate in the short-window statistical baseline and the long-window statistical baseline; the constraint saturation rate is determined based on the proportion of the original model output rewritten by the lighting output hard constraints.

[0007] Furthermore, the long-term window statistical baseline adopts a reference distribution of the same time period, same day type, or regional function class that matches the current lighting zone; the joint drift level is obtained by weighted fusion of the normalized input distribution drift index and the normalized control output distribution drift index.

[0008] Furthermore, it also includes: if manual reverse intervention occurs within a preset protection time after the effective control command is actually issued to the lighting drive interface, the training samples corresponding to the actually issued effective control command will be downweighted, or the control quantity that is stably maintained after manual reverse intervention will be used as the correction target.

[0009] Furthermore, it also includes: Based on the out-of-bounds, stuck, spiked, missing, and command-perception residuals of each sensor used to collect at least one of the following data: illuminance, personnel status, sunlight or ambient light, area usage status, and luminaire power or drive status feedback, the health status of each sensor is calculated, and a feature mask is generated based on the health status of each sensor. When the health status of any sensor is lower than a predetermined soft mask threshold, the weight of the feature formed by the data collected or fed back by that sensor in incremental training is reduced, or the feature is prevented from entering incremental training. Herein, the key sensor is the sensor whose output participates in the joint drift level calculation, online verification, or key feedback link validity judgment. When the health status of any of the key sensors is lower than a predetermined hard mask threshold, the external door control logic corresponding to severe drift is directly applied, the incremental update is frozen, and the system switches to a stable lighting control strategy.

[0010] Furthermore, the shadow model receives the same lighting operation data as the current takeover model, and obtains the shadow effective control command through the same lighting output hard constraint simulation path as the current takeover model. The shadow effective control command is used for simulation verification and index pre-evaluation, and is not sent to the lighting drive interface.

[0011] Furthermore, the online verification window uses personnel occupancy time slices, light-on events, or control events of lighting zones as statistical units; after rollback, it enters a rollback cooling state, in which drift indicators are calculated and candidate models are prohibited from taking over repeatedly, until the joint drift level falls back to a stable operating state or a slight drift state, the key feedback link is effective, and the candidate model is re-verified through the same path lighting output hard constraint simulation path and meets the takeover conditions of the online verification window.

[0012] Furthermore, the rollback evidence chain includes a timestamp, lighting zone identifier, input distribution drift index at the rollback trigger time, control output distribution drift index, joint drift level, execution status of external control, current takeover model version, candidate model version, previous stable version, original model output, effective control command, drive status feedback, acceptance index, rollback reason, and rollback cooling status. The rollback evidence chain is stored using hash concatenation or an equivalent tamper-proof method.

[0013] Furthermore, the lighting output hard constraint includes: When the change in target brightness relative to the effective brightness of the previous control cycle exceeds the allowable brightness change range determined by the product of the current control cycle length and the predetermined upper limit of the brightness change rate, the effective brightness of the previous control cycle and the allowable brightness change range are used to limit the effective brightness of the current control cycle. When the interval between the current control time and the last dimming time is less than the predetermined minimum dimming interval, the effective brightness and effective color temperature of the previous control cycle are maintained to form a hold command. When the change in target color temperature relative to the effective color temperature of the previous control cycle exceeds the predetermined color temperature transition threshold, the effective color temperature of the previous control cycle and the color temperature transition threshold are used to limit the effective color temperature of the current control cycle.

[0014] A concept-drift-gated incremental decision tree lighting energy-saving control system, applying the concept-drift-gated incremental decision tree lighting energy-saving control method, includes: The module includes: data acquisition and baseline maintenance module, joint drift determination module, gating and model version management module, lighting output hard constraint module, protocol adaptation and feedback write-back module, online acceptance and rollback module, evidence chain recording module, communication module, and data processing module. The data acquisition and baseline maintenance module, joint drift determination module, gating and model version management module, lighting output hard constraint module, protocol adaptation and feedback write-back module, online acceptance and rollback module, evidence chain recording module, and communication module are all connected to the data processing module. The acquisition and baseline maintenance module is used to acquire illuminance, personnel status, sunlight or ambient light, area usage status, manual intervention, lamp power or drive status feedback, historical effective control output, and the original model output corresponding to the historical effective control output of the lighting zone, and to maintain short-window statistical baseline and long-window statistical baseline. The joint drift determination module is used to calculate the input distribution drift index based on the short-window statistical baseline and the long-window statistical baseline, and to calculate the control output distribution drift index based on the historical effective control output and the original model output corresponding to the historical effective control output, and output the joint drift level. The gating and model version management module is used to control the entry of training samples, restricted incremental updates, shadow model training, frozen incremental updates and stable lighting control strategy switching according to the joint drift level, and to manage the current takeover model, candidate model, shadow model and the previous stable version model, wherein the current takeover model is the incremental decision tree model that currently has takeover permission; The lighting output hard constraint module is used to apply a predetermined lighting output hard constraint to the target brightness and target color temperature output by the current takeover model or the stable lighting control strategy in the current control cycle, after the model output and before the lighting control protocol encoding, in order to form an effective control command for the current control cycle. It also executes the same lighting output hard constraint simulation path as the current takeover model on the target brightness and target color temperature output by the candidate model or shadow model in the current control cycle, in order to form an effective simulation control command. The protocol adaptation and feedback write-back module is used to convert the valid control command of the current control cycle into a lighting drive interface command and receive drive status feedback, power feedback or command confirmation status, and write the feedback back to the acquisition and baseline maintenance module, the online acceptance and rollback module and the evidence chain recording module. The online acceptance and rollback module is used to verify the candidate model that has formed the effective control command through the simulation path of the same path lighting output hard constraint. In the online verification window, the candidate model is run under controlled trial operation with the same lighting output hard constraint and protocol adaptation path as the current takeover model. The energy saving rate, response delay, manual reverse intervention rate and comfort risk event rate are statistically analyzed. The corresponding indicators of the current takeover model or the previous stable version model under the matching operation conditions are used as the acceptance benchmark. When the acceptance indicators meet the takeover conditions, the joint drift level allows takeover, and the key feedback link is effective, the candidate model is granted takeover permission and promoted to a stable version. When the acceptance indicators degrade, the joint drift level is upgraded, or the key feedback link fails, the current takeover model is rolled back to the previous stable version model. The evidence chain recording module is used to record the input distribution drift index, control output distribution drift index, joint drift level, execution status of external control, model version switching, original model output, effective control commands, drive status feedback, acceptance indicators and rollback reasons at the rollback trigger time, so as to form a rollback evidence chain.

[0015] Compared with the prior art, the present invention has the following advantages and beneficial effects: By combining input distribution drift and the effective control of output distribution after constraints as the basis for drift determination, the risk of misjudgment based solely on input changes or changes in the original model output is reduced. By using the constraint saturation rate to reflect the degree of conflict between the original model output and the lighting execution boundary, external gate control can perceive the constraints encountered by the model policy at the physical execution layer. By using graded gating corresponding to slight drift, moderate drift, and severe drift, we can distinguish between restricted incremental updates, shadow training, frozen updates, and stable policy switching. The candidate model is verified by using the same lighting output hard constraint simulation path as the current takeover model, and effective control commands are actually issued via the same path during the controlled trial operation, reducing the risk of officially taking over the site lighting without verification. Dimming flicker and color temperature abrupt changes are reduced by setting hard constraints on lighting output after model output and before protocol encoding. By manually reversing the bias and interlocking sensor health, the contamination of model updates by abnormal samples is suppressed. By recording drift indicators, model version switching, original model output, effective control commands, driving feedback, and rollback reasons through the rollback evidence chain, subsequent auditing and fault location are facilitated. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of the embodiments of the present invention and form part of this application, are not intended to limit the embodiments of the present invention.

[0017] In the attached diagram: Figure 1 A flowchart illustrating the incremental decision tree lighting energy-saving control method based on concept drift gating. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments and accompanying drawings. The illustrative embodiments and descriptions of this invention are for illustrative purposes only and are not intended to limit the invention. It should be noted that this invention is already in the actual research and development stage.

[0019] Example 1 like Figure 1 As shown, the incremental decision tree lighting energy-saving control method with concept drift gating includes the following steps: Step 1: Collect illuminance, personnel status, sunlight or ambient light, area usage status, manual intervention, luminaire power or drive status feedback, historical effective control output, and the original model output corresponding to the historical effective control output for the lighting zones, and maintain short-window statistical baseline and long-window statistical baseline. Step 2: Calculate the input distribution drift index of key input features based on the short-window statistical baseline and the long-window statistical baseline, and calculate the control output distribution drift index based on the historical effective control output and its corresponding original model output. Then, merge the input distribution drift index and the control output distribution drift index into a joint drift level. Step 3: Perform external control based on the joint drift level. When there is a slight drift, perform restricted incremental updates and generate candidate models. When there is a moderate drift, only train the shadow model. When there is a severe drift, freeze the incremental updates and switch to a stable lighting control strategy. Then, determine the incremental decision tree model with takeover permission as the current takeover model. Step four: According to the control cycle, apply a predetermined lighting output hard constraint to the target brightness and target color temperature output by the current takeover model or the stable lighting control strategy in the current control cycle, after the model output and before the lighting control protocol encoding, so as to form an effective control command for the current control cycle; and execute the same lighting output hard constraint simulation path to the target brightness and target color temperature output by the candidate model or the shadow model in the current control cycle, so as to form a simulation effective control command. Step 5: Conduct a controlled trial run on the candidate model that has been verified through the lighting output hard constraint simulation path and formed the effective control command of the simulation, and count the energy saving rate, response delay, manual reverse intervention rate and comfort risk event rate; when the candidate model meets the takeover conditions, the joint drift level allows takeover and the key feedback link is effective, grant the candidate model takeover permission and upgrade it to a stable version; otherwise, revert to the previous stable version model and record the revert evidence chain.

[0020] Specifically, the lighting edge controller maintains data records at the lighting zone level. Each data record includes a timestamp, lighting zone identifier, illuminance value, personnel status, sunlight or ambient light value, area usage status, manual intervention event, current brightness output, current color temperature output, luminaire power or drive status feedback, original model output, constrained effective control output, and model version identifier. The acquisition cycle can be set according to the lighting controller's capabilities and sensor sampling capabilities, and the drift statistics cycle can be the aggregation cycle after time alignment of the acquired data. The original model output refers to the target brightness and target color temperature output by the current takeover model, candidate model, or shadow model before the application of lighting output hard constraints; the constrained effective control output refers to the control output formed after the target brightness and target color temperature have undergone lighting output hard constraint processing but before specific lighting control protocol encoding.

[0021] The lighting edge controller forms a short-term statistical baseline based on recent data and a long-term statistical baseline based on historical stable operating data or reference data for similar time periods. The short-term statistical baseline characterizes the distribution of recent lighting operating data, while the long-term statistical baseline characterizes the historical stable operating distribution, the reference distribution for similar time periods, the reference distribution for the same day, or the reference distribution for regional functions. The long-term statistical baseline can be limited to a reference distribution for the same time period, the same day, or the regional function that matches the current lighting zone. The same time period is used to mitigate the impact of the diurnal cycle, the same day pattern is used to distinguish between weekdays and non-weekdays, and the regional function is used to distinguish between office areas, meeting rooms, corridors, public areas, or other areas with different lighting needs.

[0022] For key input features, the lighting edge controller acquires short-term window samples within the statistical window corresponding to the short-term window statistical baseline, and long-term window samples within the statistical window corresponding to the long-term window statistical baseline, and calculates the input distribution drift index based on the short-term window samples and the long-term window samples. The key input features include illuminance, personnel status, sunlight or ambient light, area usage status, and manual intervention events. For continuous key input features, the lighting edge controller can construct short-term and long-term window distributions using equal-width binning, equal-frequency binning, empirical distribution functions, or equivalent statistical methods, respectively. For categorical key input features, the lighting edge controller can construct short-term and long-term window distributions based on the frequency of occurrence of each category, respectively. The lighting edge controller obtains the input distribution drift index for the same key input feature by comparing the short-term and long-term window distributions. Distribution difference measures can be performed using Jensen-Shannon divergence, population stability indicators, two-sample Kolmogorov-Smirnov statistics, or equivalent distribution difference indicators.

[0023] For control output, the lighting edge controller calculates the control output distribution drift index based on historical effective control output and the corresponding original model output. The historical effective control output includes historical effective luminance output and historical effective color temperature output. The original model output refers to the target luminance and target color temperature output of the current takeover model, candidate model, or shadow model before applying hard constraints to the lighting output; the historical effective control output refers to the control output formed after the original model output has undergone hard constraint processing and is used for actual issuance, maintenance, or online verification statistics. Effective simulation control commands used for pre-evaluation but not actually issued can be recorded separately; when a candidate model enters controlled trial operation and effectively issues control commands, the effective control commands during the controlled trial operation can be used as historical effective control output for subsequent statistics.

[0024] Specifically, the calculation process for the effective luminance output distribution includes: According to lighting zones and control cycles, effective luminance values ​​are extracted from historical effective control outputs to form an effective luminance output sequence. Based on the statistical window corresponding to the short-term window statistical baseline, effective luminance short-term window samples are extracted from the effective luminance output sequence, and an effective luminance output short-term window distribution is constructed. Based on the statistical window corresponding to the long-term window statistical baseline, effective luminance long-term window samples are extracted from the effective luminance output sequence, and an effective luminance output long-term window distribution is constructed. For continuous luminance values, the effective luminance output distribution can be constructed using equal-width binning, equal-frequency binning, empirical distribution functions, or equivalent statistical methods. For discrete luminance levels, the effective luminance output distribution can be constructed using the frequency of occurrence of each luminance level. Both the effective luminance output short-term window distribution and the effective luminance output long-term window distribution can be normalized to frequency distributions.

[0025] Specifically, the calculation process for the effective color temperature output distribution includes: According to the lighting zones and control cycles, effective color temperature values ​​are extracted from historical effective control outputs to form an effective color temperature output sequence. Based on the statistical window corresponding to the short-term window statistical baseline, effective color temperature short-term window samples are extracted from the effective color temperature output sequence, and an effective color temperature output short-term window distribution is constructed. Based on the statistical window corresponding to the long-term window statistical baseline, effective color temperature long-term window samples are extracted from the effective color temperature output sequence, and an effective color temperature output long-term window distribution is constructed. For continuous color temperature values, the effective color temperature output distribution can be constructed using equal-width binning, equal-frequency binning, empirical distribution functions, or equivalent statistical methods. For discrete color temperature levels, the effective color temperature output distribution can be constructed using the frequency of occurrence of each color temperature level. For lighting zones that do not support color temperature adjustment, the effective color temperature output distribution may not be calculated, or the control output distribution drift index may be calculated only based on the effective brightness output distribution, strategy oscillation rate, and constraint saturation rate.

[0026] When the minimum dimming interval constraint is triggered and a hold command is formed, the lighting edge controller writes the effective luminance and effective color temperature of the previous control cycle as the effective luminance and effective color temperature values ​​of the current control cycle into the historical effective control output. This allows the hold state to participate in the statistics of effective luminance output distribution, effective color temperature output distribution, strategy oscillation rate, and constraint saturation rate. Therefore, the effective luminance output distribution and the effective color temperature output distribution reflect the actual executable control results of the lighting system after hard constraint processing of the lighting output, rather than the original output of the unconstrained model.

[0027] The effective luminance output distribution difference can be obtained by comparing the short-term and long-term distributions of the effective luminance output; the effective color temperature output distribution difference can be obtained by comparing the short-term and long-term distributions of the effective color temperature output. Distribution difference measures can be taken using Jensen-Shannon divergence, population stability index, two-sample Kolmogorov-Smirnov statistic, or equivalent distribution difference index. The lighting edge controller normalizes and aggregates at least two of the effective luminance output distribution difference, the effective color temperature output distribution difference, the strategy oscillation rate, and the constraint saturation rate to obtain the control output distribution drift index.

[0028] The strategy oscillation rate is used to characterize the frequency of adjustment direction reversal of effective brightness or effective color temperature within a statistical window. The lighting edge controller compares the effective brightness values ​​of adjacent control cycles to determine the brightness adjustment direction; when two adjacent comparable adjustment directions change from increasing to decreasing, or from decreasing to increasing, it is recorded as one brightness direction reversal. The lighting edge controller can also compare the effective color temperature values ​​of adjacent control cycles in the same way to determine the color temperature adjustment direction and count the number of color temperature direction reversals. The strategy oscillation rate can be obtained as the ratio between the number of brightness direction reversals, the number of color temperature direction reversals, or a weighted sum of both, and the number of comparable adjustments within the statistical window. If the effective brightness or effective color temperature remains unchanged between two adjacent control cycles, this comparison can be excluded from the direction reversal count, or recorded separately as a hold-in state.

[0029] The constraint saturation rate is used to characterize the proportion of the original model output that has been rewritten by the hard constraints of the lighting output. The lighting edge controller compares the original model output with the corresponding historical effective control output within the same statistical window. When the target brightness is limited due to the upper limit of the brightness change rate, the target color temperature is limited due to the color temperature transition threshold, or the effective brightness and effective color temperature of the previous control cycle are maintained due to the minimum dimming interval constraint, it is recorded as a constraint trigger event. The constraint saturation rate can be obtained by the ratio between the number of constraint trigger events and the number of effective original model output records within the statistical window. If brightness limiting and color temperature limiting are triggered simultaneously in the same control cycle, they can be counted as a single comprehensive constraint trigger event; alternatively, the brightness constraint saturation rate and the color temperature constraint saturation rate can be recorded separately and then normalized and aggregated.

[0030] After calculating the input distribution drift index and the control output distribution drift index, the lighting edge controller normalizes these indices to allow comparisons of different types of drift indices on the same or dimensionless scale. This normalization can be implemented using a normalization function derived from stable operating period samples. When stable operating period samples are insufficient, the normalization function can be initialized using historical stable operating samples, reference distributions from similar time periods, reference distributions from the same day, or regional functional class reference distributions. When zero variance, outliers, or insufficient samples occur during normalization, quantile truncation, lower bound variance, robust z-scores, or equivalent robust normalization methods can be used. The normalization function can map different drift indices to the same or dimensionless interval, or to comparable robust standard scores.

[0031] In one implementation, the joint drift metric is calculated according to the following relationship: .

[0032] in, The joint drift index at time t, This represents the input distribution drift index of the j-th key input feature at time t. The control output distribution drift index represents the value at time t. Represents the normalization function. Let represent the weight of the j-th key input feature, and m represent the number of key input features. This represents the fusion weight between the input distribution drift and the control output distribution drift, and Weights of each key input feature The weights are non-negative and can be determined based on samples from stable operation periods, regional functional categories, sensor reliability, or maintenance rules; the weights of multiple key input features can be normalized to a weight sum equal to 1. The joint drift metric is used to determine the boundaries for incremental updates, shadow training, candidate model takeover, and rollback triggering.

[0033] The process for determining the slight drift threshold, moderate drift threshold, and severe drift threshold includes: extracting joint drift index samples from stable operation period samples, wherein the stable operation period samples include operation samples of the current takeover model or the previous stable version model when the critical feedback link is effective, not in a rollback cooling state, no serious sensor failure has occurred, and the acceptance indicators have not degraded; archiving the joint drift index samples according to clear partitions, similar time periods, same day types, or regional functional classes; determining the slight drift threshold, moderate drift threshold, and severe drift threshold based on the empirical distribution, quantiles, robust z-score, or equivalent statistical rules of the archived joint drift index samples, and ensuring that the slight drift threshold is less than the moderate drift threshold, and the moderate drift threshold is less than the severe drift threshold. When the stable operation period samples are insufficient to stably determine the above thresholds, the lighting edge controller can be initialized using historical stable operation samples, reference distributions of the same time period, reference distributions of the same day, reference distributions of regional functions, or maintenance rules; when the subsequent candidate model is upgraded to a stable version and the operating status meets the stable operation conditions, the corresponding operating samples can be included in the stable operation period sample set to update or calibrate the slight drift threshold, the moderate drift threshold, and the severe drift threshold.

[0034] The lighting edge controller compares a joint drift metric with the slight drift threshold, the moderate drift threshold, and the severe drift threshold to determine the joint drift level. If the joint drift metric is below the slight drift threshold, the external controller maintains a stable operating state; if the joint drift metric reaches the slight drift threshold but is below the moderate drift threshold, the joint drift level is slight drift; if the joint drift metric reaches the moderate drift threshold but is below the severe drift threshold, the joint drift level is moderate drift; if the joint drift metric reaches the severe drift threshold, the joint drift level is severe drift. When the joint drift level corresponds to a stable operating state or slight drift, the external controller allows candidate models to enter restricted incremental update, simulation verification, or online verification processes; when the joint drift level is moderate drift, only candidate models or shadow models are allowed to enter simulation path verification and metric pre-evaluation, and no actual takeover permission is granted; when the joint drift level is severe drift, the external controller freezes incremental updates, revokes candidate model takeover permission, and switches to a stable lighting control strategy.

[0035] In this embodiment, external control is a state control mechanism located outside the incremental decision tree model. This state control mechanism controls training sample entry, restricted incremental updates, candidate model generation, shadow model training, frozen incremental updates, stable lighting control strategy switching, candidate model takeover, and model rollback based on the joint drift level and sensor health status. The external control is not simply alarm logic; it directly alters model update permissions and lighting takeover permissions.

[0036] The current takeover model refers to the incremental decision tree model currently authorized to take over lighting control and used to generate effective field control commands. The previous stable version model refers to the incremental decision tree model that has passed online verification and is recorded as a rollbackable version. The candidate model refers to the incremental decision tree model formed after performing limited incremental updates based on the current takeover model, used to apply for subsequent takeover authorization; when the current takeover model is unavailable, in a rollback process, or has had its takeover authorization revoked, the candidate model is generated based on the previous stable version model. The shadow model refers to the incremental decision tree model generated based on the current takeover model and trained and verified in an isolated path; when the current takeover model is unavailable, in a rollback process, or has had its takeover authorization revoked, the shadow model is generated based on the previous stable version model. The shadow model does not directly take over lighting control, nor does it issue control commands to the lighting driver interface.

[0037] In a slightly drifted state, the input distribution and the effective control output distribution after constraints have shifted, but the shift remains within a controllable range, and the reliability of the currently running data is relatively high. Therefore, the gating and model version management module allows data that meets the feature mask and sample permission conditions to enter the incremental decision tree model, and performs restricted incremental updates based on the current takeover model or the previous stable version model to generate candidate models. The restricted incremental updates include limiting the range of samples entering training, reducing the weight of low-confidence samples, excluding features blocked by feature masks, retaining the previous stable version model, and requiring candidate models to pass the simulation path verification and online verification window acceptance through same-path illumination output hard constraint before takeover.

[0038] The candidate model generation process includes: copying the tree structure and model parameters of the current takeover model or the previous stable version model, recording the source model version identifier, and assigning a candidate model version identifier to the candidate model; selecting training samples that meet the sample permission conditions from the operational data corresponding to the short-term statistical baseline; using illuminance, personnel status, sunlight or ambient light, area usage status, and manual intervention events as model input features, and using the effective control quantity that has been processed by hard constraints on lighting output and passed feedback verification as the training objective; and performing incremental decision tree updates on the candidate model to obtain the updated candidate model. For data involving manual reverse intervention or abnormal sensor health status, the system reduces the weight of the corresponding training samples, delays their entry into training, or prevents them from entering training.

[0039] The role of the candidate model is to serve as a potential replacement for the current takeover model. After generation, the candidate model does not immediately obtain on-site lighting takeover permission. Instead, it first enters the same lighting output hard constraint simulation path as the current takeover model, generating valid simulation control commands. After passing simulation path verification, the candidate model enters the online verification window for controlled trial operation. Only when the candidate model meets the takeover conditions, the joint drift level allows takeover, and the critical feedback link is effective, does the gating and model version management module grant the candidate model takeover permission and promote it to a stable version.

[0040] Under moderate drift conditions, the deviation of the effective control output distribution after input distribution or constraint is higher than under slight drift conditions, increasing the risk of directly using incremental update results as candidate models to take over the on-site lighting. Therefore, the gating and model version management module does not grant actual takeover permission to newly generated candidate models or existing candidate models, and generates a shadow model based on the current takeover model or the previous stable version model. The shadow model receives the same lighting operation data as the current takeover model, is trained according to the same input feature construction method as the current takeover model, and outputs target brightness and target color temperature; however, the output of the shadow model only enters the same lighting output hard constraint simulation path as the current takeover model, and is used to generate shadow effective control commands and index pre-evaluation results, without being written to the lighting drive interface or changing the on-site lighting fixture status.

[0041] The shadow model generation process includes: copying the tree structure and model parameters of the current takeover model or the previous stable version model, recording the source model version identifier, and assigning a shadow model version identifier to the shadow model; inputting the lighting operation data collected under moderate drift conditions into the shadow model; performing incremental training on the shadow model without affecting the control output of the current takeover model, and performing simulation verification and index pre-evaluation in the isolated path; inputting the target brightness and target color temperature output by the shadow model into the same path lighting output hard constraint simulation path to obtain the shadow effective control command; and statistically analyzing the strategy oscillation rate, constraint saturation rate, comfort risk event rate, and energy-saving related indicators based on the shadow effective control command.

[0042] The role of the shadow model is to isolate the evaluation model's adaptation direction and control risks for moderate drift data. Compared to the candidate model, the shadow model does not participate in actual on-site deployment, does not assume current lighting control responsibilities, and does not trigger changes in the status of on-site lighting fixtures; the candidate model, on the other hand, is used to apply for takeover permission after a slight drift state or the rollback cooling-off period, and may be upgraded to a stable version. Therefore, the candidate model corresponds to an update path that can be verified before deployment, while the shadow model corresponds to an isolated evaluation path that cannot be directly deployed. The training results and pre-evaluation results of the shadow model can be used to adjust subsequent sample permission conditions, feature weights, or candidate model generation conditions.

[0043] In a severe drift state, the joint drift level indicates that the current input distribution, control output distribution, or key feedback loop has significantly deviated from the stable operating state. Continued incremental updates may lead to the model mislearning abnormal data, and the training results of candidate models or shadow models may lack a reliable basis. Therefore, the gating and model version management module freezes incremental updates, prevents new samples from entering the current incremental decision tree model, revokes the takeover permission for newly generated candidate models and existing candidate models, and switches to a stable lighting control strategy. The stable lighting control strategy prioritizes the previous stable version model; if the previous stable version model is unavailable, a rule-based baseline protection strategy is adopted. The rule-based baseline protection strategy maintains at least the minimum illuminance target, maximum luminance change rate limit, minimum dimming interval limit, and manual intervention protection under the occupied state to ensure the continuity of lighting control.

[0044] Through the above-described hierarchical processing, data changes under slight drift are considered controllable adaptive changes, and the system allows the generation of candidate models and entry into a restricted verification process; data changes under moderate drift are considered to have a high risk of mislearning, and the system only trains a shadow model and evaluates its control effect in an isolated path; data changes under severe drift are considered unsuitable for continued learning and takeover, and the system freezes incremental updates and switches to a stable lighting control strategy. Thus, the online adaptation, on-site takeover, and failure rollback of the incremental decision tree model are uniformly incorporated into the external control logic.

[0045] According to the control cycle, for the target luminance and target color temperature output by the current takeover model or stable lighting control strategy in the current control cycle, the lighting output hard constraint module executes constraints after the model output and before the lighting control protocol encoding. The pre-determined lighting output hard constraints include a pre-determined upper limit for the rate of change of luminance, a pre-determined minimum dimming interval, and a pre-determined color temperature transition threshold. The upper limit for the rate of change of luminance, the minimum dimming interval, and the color temperature transition threshold can be determined based on luminaire specifications, lighting driver interface capabilities, regional usage requirements, maintenance rules, or control statistics during stable operation. Let the target luminance output by the model be... The effective brightness of the previous control cycle was The effective brightness of the current control cycle is The upper limit of the brightness change rate is The current control cycle length is Then the constraint on the rate of change of brightness can be expressed as: ; in, This represents a limiting function that restricts x to a lower bound of l and an upper bound of h. This represents the allowable range of brightness change determined by the product of the upper limit of the brightness change rate and the current control cycle length. This represents the effective brightness of the current control cycle after being constrained by the rate of change of brightness. Therefore, the effective brightness is not determined solely by the upper limit of the rate of change of brightness, but rather by a combination of the target brightness, the effective brightness of the previous control cycle, the upper limit of the rate of change of brightness, and the length of the current control cycle. The current control cycle length can be a pre-configured fixed control cycle length, or it can be an actual control cycle length determined based on the time interval between the current control moment and the previous control moment.

[0046] Let the current control time be t, and the last dimming time be t. The minimum dimming interval is .when At this time, the lighting output hard constraint module maintains the effective brightness and effective color temperature of the previous control cycle to form a hold command for the current control cycle; the hold command, after protocol adaptation, is sent to the lighting driver interface to maintain the effective brightness and effective color temperature of the previous control cycle. At this time, the lighting output hard constraint module allows the generation of new effective control commands based on the luminance change rate constraint and color temperature transition constraint.

[0047] Let the target color temperature output by the model be The effective color temperature of the previous control cycle was The effective color temperature of the current control cycle is The color temperature transition threshold is Then the color temperature transition constraint can be expressed as: ; in, Represents a symbolic function. This represents the original target color temperature of the model. Indicates the effective color temperature of the previous control cycle. This indicates the magnitude of change in the target color temperature relative to the effective color temperature of the previous control cycle. This indicates the maximum allowable color temperature change range in a single control cycle. This represents the effective color temperature of the current control cycle after the color temperature transition constraint. Therefore, the effective color temperature is not determined solely by the color temperature transition threshold, but rather by the target color temperature, the effective color temperature of the previous control cycle, and the color temperature transition threshold, all of which are combined and limited.

[0048] The effective brightness and effective color temperature of the current control cycle form the effective control command for the current control cycle. This effective control command is converted into a lighting driver interface command by the protocol adaptation and feedback write-back module and then sent to the dimming driver. The lighting driver interface can use DALI, 0-10V, PWM, DMX, or an equivalent lighting dimming control interface. Before entering the online verification window, the candidate model or shadow model generates a simulation effective control command through the same lighting output hard constraint simulation path as the current takeover model. A simulation effective control command refers to the control command obtained after the candidate model or shadow model output has been processed through the same lighting output hard constraint simulation path as the current takeover model, used for pre-takeover verification or pre-evaluation of indicators. Simulation path verification means that the target brightness and target color temperature output by the candidate model or shadow model can form a simulation effective control command through the lighting output hard constraint simulation path, and can generate simulation records for statistical strategy oscillation rate, constraint saturation rate, and comfort risk event rate.

[0049] The candidate model verified through the simulation path continues to generate and issue actual effective control commands within the online verification window, following the same lighting output hard constraints and protocol adaptation path as the current takeover model, to form a controlled trial run. Controlled trial run refers to the process of generating and issuing actual effective control commands within the online verification window, following the same lighting output hard constraints and protocol adaptation path as the current takeover model, to statistically analyze acceptance indicators. The protocol adaptation and feedback write-back module records command confirmation status, drive status feedback, and power feedback, and writes the feedback into the short-time window statistical baseline, the online verification window, and the rollback evidence chain.

[0050] The online verification window is used for online acceptance testing of candidate models that have passed the simulation path verification. The statistical units in the online verification window can use personnel occupancy time slices, light-on events, or control events within the lighting zone. The personnel occupancy time slice refers to a statistical time segment where the occupancy status of personnel within the lighting zone remains consistent, or a statistical time segment divided by changes in occupancy status caused by personnel entering or leaving. The energy saving rate is based on the energy consumption of the luminaire power feedback under matched occupancy status, area usage status, and sunlight conditions; when luminaire power feedback is unavailable, the energy saving rate can be estimated based on drive status feedback, command confirmation status, and a preset rated power or calibrated power mapping. Response latency is based on the trigger event occurrence time, the effective control command issuance time, and the drive confirmation time; the manual reverse intervention rate is based on the manual reverse intervention event after the effective control command is issued. The comfort risk event rate can be obtained based on the statistics of comfort risk events. These comfort risk events include events where directly issuing the original model output would exceed the upper limit of the brightness change rate, events where directly issuing the original model output would violate the minimum dimming interval, events where directly issuing the original model output would exceed the color temperature transition threshold, and manual reverse intervention events after the effective control command is issued. For control events that have been corrected by hard constraints on lighting output, they can be counted as hard constraint trigger events in the comfort risk event count, without considering the corrected effective control command as an actual violation. The comfort risk event rate is the ratio of the number of comfort risk events to the number of statistical units within the online verification window. The statistical unit is at least one of personnel occupancy time slices, lighting events, or control events. The acceptance benchmark refers to the statistical indicators of the current takeover model or the previous stable version model under matched operating conditions. These matched operating conditions include at least the same lighting zone, the same or similar personnel occupancy time slices, the same or similar area usage status, and the same or similar sunlight or ambient light conditions.

[0051] If a candidate model meets the following conditions within the online verification window, the gating and model version management module will upgrade the candidate model to a stable version: the candidate model's energy saving rate is not lower than the reference energy saving rate, the response latency is not higher than the reference response latency, the manual reverse intervention rate is not higher than the reference manual reverse intervention rate, the comfort risk event rate is not higher than the reference comfort risk event rate, and the joint drift level allows takeover and the critical feedback link is effective. In this embodiment, the joint drift levels that allow takeover include stable operation and slight drift; when the joint drift level is moderate, only the candidate model or shadow model is allowed to enter the simulation path verification and index pre-evaluation, and no actual takeover permission is granted; when the joint drift level is severe, the candidate model takeover permission is revoked and the system switches to a stable lighting control strategy. Effective critical feedback link means that the command confirmation status, drive status feedback, power feedback, or selected alternative feedback within the online verification window can support the statistics of energy saving rate, response latency, manual reverse intervention rate, and comfort risk event rate. If the luminaire power feedback is unavailable, energy consumption can be estimated based on the drive status feedback, command confirmation status, and preset rated power or calibrated power mapping, and the estimation source can be written into the fallback evidence chain.

[0052] If energy efficiency degrades, response latency degrades, manual reverse intervention rate increases, comfort risk event rate increases, joint drift level upgrades, or critical feedback link fails, the online acceptance and rollback module will revoke the candidate model takeover permission, roll back the currently managed model to the previous stable version, and enter a rollback cooling state. The rollback cooling state refers to a state where, after a model rollback occurs, candidate models are temporarily prohibited from repeated takeovers, but data acquisition and drift calculations continue. In the rollback cooling state, only candidate models are allowed to enter simulation verification or index pre-evaluation. The conditions for lifting the rollback cooling state include: the joint drift level falling back to a stable operating state or a slight drift state, the critical feedback link being effective, and the candidate model re-passing the simulation path verification of the same path lighting output hard constraint and the takeover conditions of the online verification window. Once all the above conditions are met, the gating and model version management module will lift the rollback cooling state.

[0053] The rollback evidence chain is recorded append-only. Each external control state switch, candidate model takeover, acceptance failure, severe drift, rollback, and cooling state change forms an evidence chain record. The evidence chain record includes at least a timestamp, lighting zone identifier, input distribution drift index at the rollback trigger time, control output distribution drift index, joint drift level, external control execution status, current takeover model version, candidate model version, previous stable version, original model output, effective control commands, drive status feedback, energy saving rate, response latency, manual reverse intervention rate, comfort risk event rate, rollback reason, and rollback cooling state. The rollback reason can be represented by text identifiers, enumerated identifiers, or equivalent fields. The evidence chain record can be stored using hash concatenation or an equivalent tamper-proof method.

[0054] Example 2 In an improved embodiment, a preset protection time is set after the effective control command is issued. If manual reverse intervention occurs within the preset protection time, the system does not directly regard the control result as a stable preference, but instead marks the training sample corresponding to the effective control command as a low-confidence sample. The low-confidence samples can be processed by reducing their training weights, delaying their entry into incremental training, or removing them from the current batch of incremental updates.

[0055] If a new stable brightness or stable color temperature is formed after manual reverse intervention, and no further intervention occurs within the preset holding time, the system will write the stable control quantity back to the training sample label as the correction target, so that the incremental decision tree model can absorb the effective control information reflected by the user's comfort preferences in subsequent updates.

[0056] Example 3 The system calculates the sensor health status before external control judgment. The sensor health status is based at least on sensor out-of-range, jammed, spiked, missing, and command-sensor residuals. Command-sensor residuals refer to the difference between the expected change in sensitivity, power, or drive state of an effective control command and the actual feedback change.

[0057] Sensors used to calculate sensor health status include sensors that collect at least one of the following data: illuminance, personnel status, sunlight or ambient light, area usage status, luminaire power, or drive status feedback. Critical sensors are those that collect at least one of the following data: illuminance, personnel status, sunlight or ambient light, luminaire power, drive status feedback, or command confirmation status, and whose output participates in joint drift level calculation, online verification, or critical feedback link validity assessment.

[0058] Let the health status of the j-th sensor at time t be... The out-of-bounds component is The stuck component is The peak component is The residual components are The missing component is The corresponding weights are respectively , , , and The health status of the sensor can then be expressed as: .

[0059] in, This represents the health status of the j-th sensor at time t. This represents a limiting function that restricts x to the range of 0 to 1. , , , and These represent the degree of anomaly corresponding to out-of-bounds, stuck, spike, residual, and missing, respectively, and can be normalized to 0 to 1. , , , and These represent the non-negative weights of the corresponding abnormal components. The weights corresponding to each abnormal component can be determined based on maintenance rules, sensor specifications, historical fault statistics, or stable operation data, and can be normalized to a weight sum equal to 1.

[0060] The out-of-bounds component can be determined based on the degree to which the sensor sampled value exceeds the sensor specification range or operation and maintenance range; the stuck component can be determined based on the duration for which the sensor sampled value remains unchanged or changes less than a threshold within a continuous sampling window; the spike component can be determined based on the abrupt change magnitude or frequency of adjacent sampled values; the residual component can be determined based on the difference between the expected change and the actual feedback change of the effective control command's control level, power, or drive state; the missing component can be determined based on the proportion of no returned data or invalid data within the sampling window. Predetermined soft mask thresholds and predetermined hard mask thresholds can be determined based on anomaly statistics during stable operation, sensor specification boundaries, maintenance rules, or a combination thereof. When the health status of any sensor is lower than the predetermined soft mask threshold, the system reduces the weight of features formed by the data collected or fed back by that sensor in incremental training, or prevents those features from entering incremental training; when the health status of any critical sensor is lower than the predetermined hard mask threshold, the system directly processes the external control logic corresponding to severe drift, freezes incremental updates, and switches to a stable lighting control strategy. Sensor health status, feature mask, and mask cause are written into the fallback evidence chain.

[0061] Example 4 The incremental decision tree lighting energy-saving control system with concept drift gating is deployed on the lighting edge controller or lighting gateway. The system includes a data acquisition and baseline maintenance module, a joint drift determination module, a gating and model version management module, a lighting output hard constraint module, a protocol adaptation and feedback write-back module, an online acceptance and rollback module, and an evidence chain recording module.

[0062] The data acquisition and baseline maintenance module connects to the illuminance acquisition interface, personnel detection interface, sunlight or ambient light acquisition interface, manual intervention acquisition interface, luminaire power or drive status feedback interface, and lighting drive interface feedback channel. This module performs time alignment, outlier marking, maintenance status filtering, and lighting zone archiving on multi-source data. It records the original model output, constrained effective control output, and model version identifier, and maintains short-window and long-window statistical baselines. The joint drift determination module reads the short-window and long-window statistical baselines, historical effective control outputs, and corresponding original model outputs from the acquisition and baseline maintenance module. It calculates the input distribution drift index and the control output distribution drift index and outputs the joint drift level.

[0063] The gating and model version management module includes a current takeover model register, a candidate model register, a shadow model register, a previous stable version register, a stable lighting control strategy register, and a gating state machine. The gating state machine generates training sample entry permissions, restricted incremental update permissions, candidate model generation instructions, shadow model training permissions, freeze incremental update instructions, candidate model takeover permissions, stable lighting control strategy switching instructions, and rollback instructions based on the joint drift level, online verification window results, and sensor health status.

[0064] In a slightly drifting state, the gating and model version management module prioritizes generating candidate models based on the currently takenover model. When the current takenover model is unavailable, in a rollback process, or has had its takeover permission revoked, the gating and model version management module generates candidate models based on the previous stable version model. The gating and model version management module assigns a candidate model version identifier to each candidate model and records the source model version identifier. The candidate model receives training samples that meet the sample permission conditions, performs restricted incremental updates, and after completing the update, enters the same-path lighting output hard-constraint simulation path and online verification window. After simulation path verification, online verification window acceptance, and takeover permission judgment, the candidate model can be promoted to a stable version.

[0065] Under moderate drift conditions, the gating and model version management module does not grant actual takeover permission to newly generated candidate models or existing candidate models, and prioritizes generating shadow models based on the currently takenover model. When the currently takenover model is unavailable, in a rollback process, or has had its takeover permission revoked, the gating and model version management module generates a shadow model based on the previous stable version model. The gating and model version management module assigns a shadow model version identifier to the shadow model and records the source model version identifier. The shadow model receives the same lighting operation data as the currently takenover model and performs incremental training, simulation verification, and index pre-evaluation in an isolated path. The shadow model is not granted lighting control takeover permission, does not issue valid control commands to the lighting driver interface, and does not directly change the status of on-site lighting fixtures.

[0066] In a severe drift state, the gating and model version management module stops generating new candidate models and shadow models, freezes incremental updates, revokes the takeover permission for newly generated candidate models and existing candidate models, and invokes the stable lighting control strategy. The stable lighting control strategy prioritizes calling the previous stable version model; when the previous stable version model is unavailable, it invokes the rule-based backup strategy to maintain lighting control continuity.

[0067] The protocol adaptation and feedback write-back module converts effective control commands into lighting drive interface commands and receives drive status feedback, power feedback, or command confirmation status. It then writes this feedback back to the acquisition and baseline maintenance module, the online acceptance and rollback module, and the evidence chain recording module. The online acceptance and rollback module verifies candidate models that form effective control commands through simulation paths with hard constraints on lighting output. These models undergo controlled trial operation in the online verification window using the same hard constraints and protocol adaptation path as the current takeover model. The module calculates energy savings, response latency, manual reverse intervention rate, and comfort risk event rate. When acceptance indicators meet takeover conditions, the joint drift level allows takeover, and the critical feedback link is effective, the online acceptance and rollback module sends takeover confirmation information to the gating and model version management module, which then upgrades the candidate model to a stable version. When acceptance indicators degrade, the joint drift level upgrades, or the critical feedback link fails, the online acceptance and rollback module triggers a rollback to the previous stable version model.

[0068] The evidence chain recording module receives data from the joint drift determination module, the gating and model version management module, the lighting output hard constraint module, the protocol adaptation and feedback write-back module, and the online acceptance and rollback module, forming a rollback evidence chain that includes the input distribution drift index at the rollback trigger time, the control output distribution drift index, the joint drift level, the execution status of the external gate control, the model version switch, the original model output, the effective control command, the drive status feedback, the acceptance index, and the rollback reason.

[0069] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. Incremental decision tree lighting energy saving control method with concept drift gating characterized in that, Includes the following steps: Step 1: Collect illuminance, personnel status, sunlight or ambient light, area usage status, manual intervention, luminaire power or drive status feedback, historical effective control output, and the original model output corresponding to the historical effective control output for the lighting zones, and maintain short-window statistical baseline and long-window statistical baseline. Step 2: Calculate the input distribution drift index of key input features based on the short-window statistical baseline and the long-window statistical baseline, and calculate the control output distribution drift index based on the historical effective control output and its corresponding original model output. Then, merge the input distribution drift index and the control output distribution drift index into a joint drift level. Step 3: Perform external control based on the joint drift level. When there is a slight drift, perform restricted incremental updates and generate candidate models. When there is a moderate drift, only train the shadow model. When there is a severe drift, freeze the incremental updates and switch to a stable lighting control strategy. Then, determine the incremental decision tree model with takeover permission as the current takeover model. Step four: According to the control cycle, apply a predetermined lighting output hard constraint to the target brightness and target color temperature output by the current takeover model or the stable lighting control strategy in the current control cycle, after the model output and before the lighting control protocol encoding, so as to form an effective control command for the current control cycle; and execute the same lighting output hard constraint simulation path to the target brightness and target color temperature output by the candidate model or the shadow model in the current control cycle, so as to form a simulation effective control command. Step 5: Conduct a controlled trial run on the candidate model that has been verified through the lighting output hard constraint simulation path and formed the effective control command of the simulation, and count the energy saving rate, response delay, manual reverse intervention rate and comfort risk event rate; when the candidate model meets the takeover conditions, the joint drift level allows takeover and the key feedback link is effective, grant the candidate model takeover permission and upgrade it to a stable version; otherwise, revert to the previous stable version model and record the revert evidence chain.

2. The concept drift gated incremental decision tree lighting energy saving control method according to claim 1, characterized in that, The key input features include illuminance, personnel status, sunlight or ambient light, area usage status, and manual intervention events; the control output distribution drift index is obtained by comparing the differences between the effective luminance output distribution, effective color temperature output distribution, strategy oscillation rate, and constraint saturation rate in the short-window statistical baseline and the long-window statistical baseline; the constraint saturation rate is determined based on the proportion of the original model output rewritten by the lighting output hard constraints.

3. The incremental decision tree lighting energy-saving control method with concept drift gating according to claim 1, characterized in that, The long-term window statistical baseline adopts a reference distribution of the same time period, same day type or regional function class that matches the current lighting zone; the joint drift level is obtained by weighted fusion of the normalized input distribution drift index and the normalized control output distribution drift index.

4. The concept of drift gated incremental decision tree lighting energy saving control method according to claim 1, characterized in that, Also includes: If a manual reverse intervention occurs within a preset protection time after the effective control command is actually issued to the lighting drive interface, the training sample corresponding to the actually issued effective control command will be downweighted, or the control quantity that is stably maintained after the manual reverse intervention will be used as the correction target.

5. The Concept Drift Gated Incremental Decision Tree Lighting Energy Saving Control method according to claim 1, characterized in that, Also includes: Based on the out-of-bounds, stuck, spiked, missing, and command-perception residuals of each sensor used to collect at least one of the following data: illuminance, personnel status, sunlight or ambient light, area usage status, and luminaire power or drive status feedback, the health status of each sensor is calculated, and a feature mask is generated based on the health status of each sensor. When the health status of any sensor is lower than a predetermined soft mask threshold, the weight of the feature formed by the data collected or fed back by that sensor in incremental training is reduced, or the feature is prevented from entering incremental training. Herein, the key sensor is the sensor whose output participates in the joint drift level calculation, online verification, or key feedback link validity judgment. When the health status of any of the key sensors is lower than a predetermined hard mask threshold, the external door control logic corresponding to severe drift is directly applied, the incremental update is frozen, and the system switches to a stable lighting control strategy.

6. The concept drift gated incremental decision tree lighting energy saving control method of claim 1, wherein, The shadow model receives the same lighting operation data as the current takeover model, and obtains the shadow effective control command through the same lighting output hard constraint simulation path as the current takeover model. The shadow effective control command is used for simulation verification and index pre-evaluation, and is not sent to the lighting driver interface.

7. The concept drift gated incremental decision tree lighting energy saving control method of claim 1, wherein, The online verification window uses personnel occupancy time slices, light-on events, or control events in the lighting zone as statistical units. After rollback, it enters a rollback cooling state, in which drift indexes are calculated and candidate models are prohibited from taking over repeatedly until the joint drift level falls back to a stable operating state or a slight drift state, the key feedback link is effective, and the candidate model is re-verified through the same path lighting output hard constraint simulation path and meets the takeover conditions of the online verification window.

8. The concept drift gated incremental decision tree lighting energy saving control method of claim 1, wherein, The rollback evidence chain includes a timestamp, lighting zone identifier, input distribution drift index at the rollback trigger time, control output distribution drift index, joint drift level, execution status of external control, current takeover model version, candidate model version, previous stable version, original model output, effective control command, drive status feedback, acceptance index, rollback reason, and rollback cooling status. The rollback evidence chain is stored using hash concatenation or an equivalent tamper-proof method.

9. The concept drift gated incremental decision tree lighting energy saving control method of claim 1, wherein, The hard constraint on lighting output includes: When the change in target brightness relative to the effective brightness of the previous control cycle exceeds the allowable brightness change range determined by the product of the current control cycle length and the predetermined upper limit of the brightness change rate, the effective brightness of the previous control cycle and the allowable brightness change range are used to limit the effective brightness of the current control cycle. When the interval between the current control time and the last dimming time is less than the predetermined minimum dimming interval, the effective brightness and effective color temperature of the previous control cycle are maintained to form a hold command. When the change in target color temperature relative to the effective color temperature of the previous control cycle exceeds the predetermined color temperature transition threshold, the effective color temperature of the previous control cycle and the color temperature transition threshold are used to limit the effective color temperature of the current control cycle.

10. A concept drift gated incremental decision tree lighting energy saving control system characterized in that, The incremental decision tree lighting energy-saving control method using the concept of drift gating as described in any one of claims 1-9 includes: The module includes: data acquisition and baseline maintenance module, joint drift determination module, gating and model version management module, lighting output hard constraint module, protocol adaptation and feedback write-back module, online acceptance and rollback module, evidence chain recording module, communication module, and data processing module. The data acquisition and baseline maintenance module, joint drift determination module, gating and model version management module, lighting output hard constraint module, protocol adaptation and feedback write-back module, online acceptance and rollback module, evidence chain recording module, and communication module are all connected to the data processing module. The acquisition and baseline maintenance module is used to acquire illuminance, personnel status, sunlight or ambient light, area usage status, manual intervention, lamp power or drive status feedback, historical effective control output, and the original model output corresponding to the historical effective control output of the lighting zone, and to maintain short-window statistical baseline and long-window statistical baseline. The joint drift determination module is used to calculate the input distribution drift index based on the short-window statistical baseline and the long-window statistical baseline, and to calculate the control output distribution drift index based on the historical effective control output and the original model output corresponding to the historical effective control output, and output the joint drift level. The gating and model version management module is used to control the entry of training samples, restricted incremental updates, shadow model training, frozen incremental updates and stable lighting control strategy switching according to the joint drift level, and to manage the current takeover model, candidate model, shadow model and the previous stable version model, wherein the current takeover model is the incremental decision tree model that currently has takeover permission; The lighting output hard constraint module is used to apply a predetermined lighting output hard constraint to the target brightness and target color temperature output by the current takeover model or the stable lighting control strategy in the current control cycle, after the model output and before the lighting control protocol encoding, in order to form an effective control command for the current control cycle. It also executes the same lighting output hard constraint simulation path as the current takeover model on the target brightness and target color temperature output by the candidate model or shadow model in the current control cycle, in order to form an effective simulation control command. The protocol adaptation and feedback write-back module is used to convert the valid control command of the current control cycle into a lighting drive interface command and receive drive status feedback, power feedback or command confirmation status, and write the feedback back to the acquisition and baseline maintenance module, the online acceptance and rollback module and the evidence chain recording module. The online acceptance and rollback module is used to verify the candidate model that has formed the effective control command through the simulation path of the same path lighting output hard constraint. In the online verification window, the candidate model is run under controlled trial operation with the same lighting output hard constraint and protocol adaptation path as the current takeover model. The energy saving rate, response delay, manual reverse intervention rate and comfort risk event rate are statistically analyzed. The corresponding indicators of the current takeover model or the previous stable version model under the matching operation conditions are used as the acceptance benchmark. When the acceptance indicators meet the takeover conditions, the joint drift level allows takeover, and the key feedback link is effective, the candidate model is granted takeover permission and promoted to a stable version. When the acceptance indicators degrade, the joint drift level is upgraded, or the key feedback link fails, the current takeover model is rolled back to the previous stable version model. The evidence chain recording module is used to record the input distribution drift index, control output distribution drift index, joint drift level, execution status of external control, model version switching, original model output, effective control commands, drive status feedback, acceptance indicators and rollback reasons at the rollback trigger time, so as to form a rollback evidence chain.

Images