A multi-modal perception based LED energy-saving lighting intelligent control system

Abstract

This invention discloses an intelligent control system for LED energy-saving lighting based on multimodal perception, comprising: a data acquisition and preprocessing module, which acquires multimodal perception data and equipment operation data, processes them to generate a perception window and an operation window; an evidence feature generation module, which generates an evidence feature set from the perception window and the operation window; a reflection evidence generation module, which generates a reflection evidence set from the operation window, power illuminance calibration mapping, and topological adjacency matrix; a DBM inference module, which takes the evidence feature set and the reflection evidence set as input and outputs a causal posterior and control semantic parameter set; a DBO optimization module, which initializes and filters the output decision vector based on the control semantic parameter set; and a dimming execution feedback module, which issues dimming commands based on the decision vector, collects feedback, and updates the window. This invention achieves a closed-loop energy-saving dimming inference optimization, suppresses false positives for reflection, and improves regional brightness consistency.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence control technology, and in particular to an intelligent control system for LED energy-saving lighting based on multimodal perception. Background Technology

[0002] Existing LED lighting control solutions mostly adopt timing strategies, fixed scene rules, or closed-loop adjustment methods based on illuminance feedback. Some solutions add power, status and other operating quantities for correction, while others introduce image, acoustic or waveform sensing data to assist in judgment. However, most of them are characterized by independent processing of single-source features before threshold judgment or rule linkage, lacking a mechanism to jointly infer multi-source evidence in a unified model. This leads to the control basis under complex environmental disturbances being prone to fluctuation and lacking generalization.

[0003] In real-world scenarios, changes in lighting demand and abnormal disturbances are often not directly measurable. Simply relying on illuminance or power feedback can easily misjudge illuminance deviations caused by reflections and glare as increased demand, leading to unnecessary brightening and increased energy consumption. Existing technologies typically lack a calculable evidence path to construct the source of the deviation between "theoretical illuminance and measured illuminance," making it difficult to establish a closed-loop judgment from operational data to the cause of the anomaly. On the other hand, multi-zone lighting exhibits spatial coupling characteristics, and excessive differences in dimming between adjacent zones can easily cause visual fragmentation and local overexposure. However, common optimization controls mainly rely on direct search of independent variables for each zone or simple linkage, lacking a process of updating candidate control solutions with consistency constraints based on topological adjacency relationships. This results in unstable feasibility of control solutions under spatial continuity constraints, and insufficient robustness of optimization results to dynamic disturbances.

[0004] Therefore, how to provide an intelligent control system for LED energy-saving lighting based on multimodal perception is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent control system for LED energy-saving lighting based on multimodal sensing. This invention integrates multi-source sensing data and equipment operation data, constructs counterfactual residual evidence, completes causal posterior inference in DBM, and combines DBO to optimize zonal dimming decisions under topological constraints. It has the advantages of good energy-saving effect, strong anti-reflective interference and high spatial consistency.

[0006] An intelligent control system for LED energy-saving lighting based on multimodal sensing according to an embodiment of the present invention includes:

[0007] The data acquisition and preprocessing module is used to acquire the multimodal sensing data set and the device operation dataset, merge them for preprocessing, and generate a sensing window set and an operation window set.

[0008] The evidence feature generation module is used to extract dynamic change feature sequences based on the perception window set and state change feature sequences based on the running window set to generate an evidence feature set.

[0009] The reflection evidence generation module is used to generate a counterfactual illuminance sequence and calculate a counterfactual residual sequence based on the running window set and the power illuminance calibration mapping, and to calculate the residual synchronization sequence based on the topological adjacency matrix to generate a reflection evidence set.

[0010] The DBM inference module is used to construct DBM and set the set of visible variables, latent variables and causal variables. It takes the set of evidence features and the set of reflected evidence as input to perform posterior inference and outputs the causal posterior set and the set of control semantic parameters.

[0011] The DBO optimization module is used to construct the DBO and define discrete and continuous decision vectors. It initializes the population based on the control semantic parameter set to obtain a candidate control solution set, iteratively updates the candidate control solution set based on the topological adjacency matrix, selects the optimal control solution based on the fitness function, and outputs the decision vector.

[0012] The dimming execution feedback module is used to generate and issue zone dimming commands based on decision vectors, collect execution feedback, and update the perception window set and the running window set.

[0013] Optionally, modules can be integrated using the following methods:

[0014] S1. Acquire the multimodal sensing data set and the device operation data set, perform preprocessing, and generate the sensing window set and the operation window set;

[0015] S2. The perception window set extracts the dynamic change feature sequence, the running window set extracts the state change feature sequence, and generates the evidence feature set.

[0016] S3. The set of running windows is mapped with the power illuminance calibration to generate a counterfactual illuminance sequence, the counterfactual residual sequence is calculated, the residual synchronization sequence is calculated based on the topological adjacency matrix, and a set of reflection evidence is generated.

[0017] S4. Construct DBM, set the set of visible variables, the set of latent variables and the set of causal variables, input the set of evidence features and the set of reflected evidence to perform posterior inference, and output the causal posterior set and the set of control semantic parameters.

[0018] S5. Construct a DBO, define discrete and continuous decision vectors, and initialize the population based on the set of control semantic parameters to obtain a set of candidate control solutions.

[0019] S6. During the DBO iteration process, update the candidate control solution set based on the topological adjacency matrix, select the optimal control solution based on the fitness function, and output the decision vector.

[0020] S7. Generate and issue zone dimming commands based on decision vectors, and collect execution feedback to update the perception window set and the running window set.

[0021] Optionally, S2 specifically includes:

[0022] S21. Read data fragments corresponding to the same window number from the perception window set and the running window set to form a window input set;

[0023] S22. Execute a preset feature extraction operator on the window input set to generate a dynamic change feature sequence and a state change feature sequence, and perform window aggregation on the dynamic change feature sequence and the state change feature sequence to obtain a window feature set;

[0024] S23. Concatenate the window feature set into an evidence feature vector according to a fixed field order, perform linear normalization on the evidence feature vector, and summarize to form an evidence feature set.

[0025] Optionally, S3 specifically includes:

[0026] S31. Read the power preprocessing sequence from the running window set, call the corresponding mapping relationship in the power illuminance calibration mapping data for each power sample value in the power preprocessing sequence according to the sampling sequence number, obtain the theoretical illuminance sample value that corresponds one-to-one with each power sample value, and generate the counterfactual illuminance sequence by arranging them according to the sampling sequence number.

[0027] S32. Perform differential operation point by point on the counterfactual illuminance sequence and the illuminance preprocessing sequence in the same running window according to the same sampling number to obtain the illuminance residual value corresponding to each sampling time, and arrange them according to the sampling number to generate the counterfactual residual sequence.

[0028] S33. Determine the adjacent partition combination based on the topological adjacency matrix. For any adjacent partition combination, read the corresponding counterfactual residual value at the same sampling sequence number position. Determine whether the change direction of the counterfactual residual value of the adjacent partition is consistent and whether the absolute value of the residual exceeds the preset threshold at the same time. When the consistency condition is met, record it as a synchronous change mark. When the consistency condition is not met, record it as a non-synchronous change mark. Arrange the residual synchronization sequence according to the sampling sequence number.

[0029] S34. Combine the counterfactual residual sequence and the residual synchronization sequence in a fixed field order to form a set of reflective evidence.

[0030] Optionally, the DBM structure in S4 specifically includes:

[0031] The visible variable layer is divided into a first visible variable sublayer and a second visible variable sublayer. The first visible variable sublayer receives the set of evidence features, and the second visible variable sublayer receives the set of reflected evidence.

[0032] An updatable connection parameter is set between the first visible variable sublayer and the causal variable layer, and a fixed connection parameter is set between the second visible variable sublayer and the causal variable layer. The fixed connection parameter is determined by the power illuminance calibration mapping data and the topological adjacency matrix.

[0033] Set connection parameters between the causal variable layer and the hidden variable layer;

[0034] The variables in the first visible variable sublayer consist of dynamic change feature sequences and state change feature sequences, while the variables in the second visible variable sublayer consist of counterfactual residual sequences and residual synchronization sequences.

[0035] DBM takes a set of evidence features and a set of reflected evidence as input, and outputs a set of causal posterior evidence.

[0036] Optionally, the posterior inference in S4 specifically includes:

[0037] Within each partition, based on the set of evidence features input from the first visible variable sublayer, and combined with the corresponding updatable connection parameters, the initial activation state of each causal variable in the causal variable layer is calculated to form the partition's initial causal state set.

[0038] Based on the topological adjacency matrix, a spatial propagation operation is performed on the initial cause state set of adjacent partitions, propagating the cause state of each partition to adjacent partitions to generate a spatially propagated cause state set.

[0039] Read the set of reflected evidence from the second visible variable sub-layer, and perform counterfactual consistency verification on each cause state in the set of causes of spatial propagation. Counterfactual consistency verification includes residual synchronization marker matching judgment.

[0040] Perform a retrieval operation on the cause states that fail the counterfactual consistency check. The retrieval operation includes removing the corresponding cause state from the spatial propagation cause state set and generating a retrieval set of cause states.

[0041] Based on the collected set of cause states, and combined with the connection parameters between the cause variable layer and the latent variable layer, the state distribution of the cause variable layer is updated;

[0042] Normalize the updated causal variable layer state distribution and output the causal posterior set.

[0043] Optionally, S5 specifically includes:

[0044] S51. Read the control semantic parameter set, generate a control variable set based on the partition identifier and the control semantic parameter set, and encode the control variable set into a DBO individual representation to form an individual parameter template;

[0045] S52. Set the population size, number of iterations and search boundary of DBO. Randomly generate the initial population according to the individual parameter template to obtain the candidate control solution set. Perform boundary truncation and legality verification on each candidate control solution set.

[0046] S53. Write the partition association constraint information into the candidate control solution set in the initial population according to the topological adjacency matrix, complete the constraint initialization of the initial population and output the initialized population.

[0047] Optionally, S6 specifically includes:

[0048] For each candidate control solution in the candidate control solution set, determine the adjacent partition pairs formed by any partition and the adjacent partitions based on the topological adjacency matrix, and read the control variable values ​​corresponding to the adjacent partition pairs in the candidate control solution.

[0049] For each adjacent partition pair, calculate the difference between the corresponding control variable values ​​of the two partitions. When the absolute value of the difference exceeds the preset adjacent consistency threshold, it is determined that the current adjacent partition pair violates the consistency constraint.

[0050] For adjacent partition pairs that violate consistency constraints, the control variable values ​​of the two partitions are synchronously updated by distributing the difference exceeding the consistency threshold equally to the corresponding control variables of the adjacent partitions, so as to obtain updated control variable values ​​that satisfy the adjacent consistency constraints.

[0051] The updated control variable values ​​are written back to the corresponding candidate control solutions, and the above consistency constraint update is performed sequentially on all adjacent partition pairs to complete the update of the candidate control solution set based on the topological adjacency matrix.

[0052] The beneficial effects of this invention are:

[0053] (1) This invention preprocesses and windowes the multimodal sensing data set and the equipment operation data set to form an evidence feature set. Furthermore, it generates a counterfactual illuminance sequence by power illuminance calibration mapping, generates a counterfactual residual sequence by difference between the measured illuminance sequence and the counterfactual illuminance sequence, and generates a residual synchronization sequence by topological adjacency matrix. This constructs a reflection evidence set, realizes the computable evidence expression of illuminance anomalies, reduces the risk of misjudgment caused by single illuminance feedback, and improves the control robustness under complex environmental disturbances.

[0054] (2) In the DBM, the visible variable layer is divided into the first visible variable sub-layer corresponding to the evidence feature set and the second visible variable sub-layer corresponding to the reflection evidence set. Updateable connection parameters and fixed connection parameters are set so that the two types of evidence enter the cause variable layer with different parameter mechanisms. Combined with the posterior inference process of spatial propagation and retrieval, the posterior set of causes and the set of control semantic parameters are output to realize the differentiation and inference of the causes of demand change and the causes of reflection interference, thereby improving the consistency and stability of cause determination.

[0055] (3) In the DBO iteration process, the present invention performs consistency constraint update on the candidate control solution set based on the topological adjacency matrix, adopts the update mechanism of adjacency consistency threshold and equal amortization correction to suppress the sudden change of dimming in adjacent partitions, and selects the optimal control solution to generate decision vector under the drive of fitness function, so as to achieve the coordinated satisfaction of energy saving target, illuminance deviation constraint and spatial continuity constraint, and improve the overall consistency of dimming effect and energy consumption control level of partition dimming. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 The flowchart is a process for an intelligent control system for LED energy-saving lighting based on multimodal perception proposed in this invention.

[0058] Figure 2 This is a schematic diagram illustrating the generation relationship between the evidence feature set and the reflection evidence set of an LED energy-saving lighting intelligent control system based on multimodal perception proposed in this invention.

[0059] Figure 3 This is a schematic diagram of topology consistency update in the DBO of an LED energy-saving lighting intelligent control system based on multimodal perception proposed in this invention. Detailed Implementation

[0060] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0061] refer to Figures 1-3 A smart control system for LED energy-saving lighting based on multimodal sensing, comprising:

[0062] The data acquisition and preprocessing module is used to acquire the multimodal sensing data set and the device operation dataset, merge them for preprocessing, and generate a sensing window set and an operation window set.

[0063] The evidence feature generation module is used to extract dynamic change feature sequences based on the perception window set and state change feature sequences based on the running window set to generate an evidence feature set.

[0064] The reflection evidence generation module is used to generate a counterfactual illuminance sequence and calculate a counterfactual residual sequence based on the running window set and the power illuminance calibration mapping, and to calculate the residual synchronization sequence based on the topological adjacency matrix to generate a reflection evidence set.

[0065] The DBM inference module is used to construct DBM and set the set of visible variables, latent variables and causal variables. It takes the set of evidence features and the set of reflected evidence as input to perform posterior inference and outputs the causal posterior set and the set of control semantic parameters.

[0066] The DBO optimization module is used to construct the DBO and define discrete and continuous decision vectors. It initializes the population based on the control semantic parameter set to obtain a candidate control solution set, iteratively updates the candidate control solution set based on the topological adjacency matrix, selects the optimal control solution based on the fitness function, and outputs the decision vector.

[0067] The dimming execution feedback module is used to generate and issue zone dimming commands based on decision vectors, collect execution feedback, and update the perception window set and the running window set.

[0068] In this embodiment, the modules are interconnected using the following method:

[0069] S1. Acquire the multimodal sensing data set and the device operation data set, perform preprocessing, and generate the sensing window set and the operation window set;

[0070] S2. The perception window set extracts the dynamic change feature sequence, the running window set extracts the state change feature sequence, and generates the evidence feature set.

[0071] S3. The set of running windows is mapped with the power illuminance calibration to generate a counterfactual illuminance sequence, the counterfactual residual sequence is calculated, the residual synchronization sequence is calculated based on the topological adjacency matrix, and a set of reflection evidence is generated.

[0072] S4. Construct DBM, set the set of visible variables, the set of latent variables and the set of causal variables, input the set of evidence features and the set of reflected evidence to perform posterior inference, and output the causal posterior set and the set of control semantic parameters.

[0073] S5. Construct a DBO, define discrete and continuous decision vectors, and initialize the population based on the set of control semantic parameters to obtain a set of candidate control solutions.

[0074] S6. During the DBO iteration process, update the candidate control solution set based on the topological adjacency matrix, select the optimal control solution based on the fitness function, and output the decision vector.

[0075] S7. Generate and issue zone dimming commands based on decision vectors, and collect execution feedback to update the perception window set and the running window set.

[0076] In this embodiment, S1 specifically includes:

[0077] S11. Acquire a multimodal sensing data set, which is divided into image data sequence, waveform data sequence, and acoustic data sequence. The image data sequence is recorded in consecutive frames, the waveform data sequence is recorded in consecutive sampling, and the acoustic data sequence is recorded in consecutive sampling. Acquire a device operation data set, which includes illuminance measurement sequence, power measurement sequence, dimming control quantity sequence, and drive status sequence.

[0078] S12. Obtain the topological adjacency matrix and power illuminance calibration mapping data. The topological adjacency matrix is ​​used to represent the adjacency relationship between partitions, and the power illuminance calibration mapping data is used to represent the correspondence between power and illuminance.

[0079] S13. Perform decoding and frame scaling on the image data sequence, and output the image preprocessing sequence; perform amplitude normalization and DC component removal on the waveform data sequence, and output the waveform preprocessing sequence; perform amplitude normalization on the acoustic data sequence, and output the acoustic preprocessing sequence; perform outlier removal and dimensional normalization on the illuminance measurement sequence and power measurement sequence, and output the illuminance preprocessing sequence and power preprocessing sequence; perform validity verification on the dimming control quantity sequence and drive state sequence, and output the dimming preprocessing sequence and state preprocessing sequence.

[0080] S14. Set the window length and window step size, perform sliding window segmentation based on the image preprocessing sequence, waveform preprocessing sequence, and acoustic preprocessing sequence to generate a set of perception windows; perform sliding window segmentation based on the illuminance preprocessing sequence, power preprocessing sequence, dimming preprocessing sequence, and state preprocessing sequence to generate a set of running windows.

[0081] In this embodiment, S2 specifically includes:

[0082] S21. Read the data segments corresponding to the same window number from the perception window set and the running window set to form a window input set; the window input set includes frame sequence segments of the image preprocessing sequence within the window range, sampling sequence segments of the waveform preprocessing sequence within the window range, sampling sequence segments of the acoustic preprocessing sequence within the window range, sampling sequence segments of the illuminance preprocessing sequence within the window range, sampling sequence segments of the power preprocessing sequence within the window range, sampling sequence segments of the dimming preprocessing sequence within the window range, and status code sequence segments of the status preprocessing sequence within the window range;

[0083] S22. Execute a preset feature extraction operator on the window input set to generate a dynamic change feature sequence and a state change feature sequence. Perform window aggregation on the dynamic change feature sequence and the state change feature sequence to obtain a window feature set. The dynamic change feature sequence consists of adjacent frame pixel difference sequences of image preprocessing sequence segments, adjacent frame difference sequences of spectral amplitude sequences obtained by time-frequency transformation of waveform preprocessing sequence segments, and adjacent frame difference sequences of frequency band energy sequences obtained by frequency domain transformation of acoustic preprocessing sequence segments. The state change feature sequence consists of adjacent sampling point difference sequences of illuminance preprocessing sequence segments, adjacent sampling point difference sequences of power preprocessing sequence segments, adjacent sampling point difference sequences of dimming preprocessing sequence segments, and adjacent sampling point status code difference sequences of state preprocessing sequence segments. Perform window aggregation on the dynamic change feature sequence and the state change feature sequence to obtain a window feature set. Window aggregation is used to calculate the sequence mean and sequence energy.

[0084] S23. Concatenate the window feature set into an evidence feature vector according to a fixed field order, perform linear normalization on the evidence feature vector, and summarize to form an evidence feature set.

[0085] In this embodiment, S3 specifically includes:

[0086] S31. Read the power preprocessing sequence from the running window set, call the corresponding mapping relationship in the power illuminance calibration mapping data for each power sample value in the power preprocessing sequence according to the sampling sequence number, obtain the theoretical illuminance sample value that corresponds one-to-one with each power sample value, and generate the counterfactual illuminance sequence according to the sampling sequence number; the power illuminance calibration mapping is a deterministic mapping relationship between the power input value and the corresponding theoretical illuminance output value.

[0087] S32. Perform differential operation point by point on the counterfactual illuminance sequence and the illuminance preprocessing sequence in the same running window according to the same sampling number to obtain the illuminance residual value corresponding to each sampling time, and arrange them according to the sampling number to generate the counterfactual residual sequence.

[0088] S33. Determine adjacent partition combinations based on the topological adjacency matrix. For any adjacent partition combination, read the corresponding counterfactual residual values ​​at the same sampling sequence number position. Determine whether the direction of change of the counterfactual residual values ​​of adjacent partitions is consistent and whether the absolute value of the residuals exceeds the preset threshold simultaneously. When the consistency condition is met, record it as a synchronous change mark; when the consistency condition is not met, record it as a asynchronous change mark. Generate a residual synchronization sequence by arranging the samples according to the sampling sequence number. The synchronous change mark is recorded in binary form. A synchronous change mark of 1 indicates that the residual changes of adjacent partitions are synchronized at the corresponding sampling time, and a mark of 0 indicates that they are asynchronous. The preset threshold is set to 0.15 to filter out small residual fluctuations.

[0089] S34. Combine the counterfactual residual sequence and the residual synchronization sequence in a fixed field order to form a set of reflective evidence.

[0090] In this embodiment, the DBM structure in S4 specifically includes:

[0091] The visible variable layer is divided into a first visible variable sublayer and a second visible variable sublayer. The first visible variable sublayer receives the set of evidence features, and the second visible variable sublayer receives the set of reflected evidence.

[0092] An updatable connection parameter is set between the first visible variable sublayer and the causal variable layer, and a fixed connection parameter is set between the second visible variable sublayer and the causal variable layer. The fixed connection parameter is determined by the power illuminance calibration mapping data and the topological adjacency matrix. The fixed connection parameter is a set of parameters calculated based on the power illuminance calibration mapping data and the topological adjacency matrix during the system initialization phase and remains unchanged during operation.

[0093] Set connection parameters between the causal variable layer and the hidden variable layer;

[0094] The variables in the first visible variable sublayer consist of dynamic change feature sequences and state change feature sequences, while the variables in the second visible variable sublayer consist of counterfactual residual sequences and residual synchronization sequences.

[0095] DBM takes a set of evidence features and a set of reflected evidence as input, and outputs a set of causal posterior evidence.

[0096] In this embodiment, the posterior inference in S4 specifically includes:

[0097] Within each partition, based on the set of evidence features input from the first visible variable sublayer, and combined with the corresponding updatable connection parameters, the initial activation state of each causal variable in the causal variable layer is calculated to form the partition's initial causal state set. The initial activation state is obtained by performing a linear mapping between the set of evidence features in the first visible variable sublayer and the updatable connection parameters and applying a nonlinear activation function. The initial activation state is a real value between 0 and 1.

[0098] Based on the topological adjacency matrix, a spatial propagation operation is performed on the initial cause state sets of adjacent partitions, propagating the cause state of each partition to adjacent partitions to generate a spatial propagation cause state set; the initial cause state set of the current partition is merged with the initial cause state sets of each adjacent partition; a weighted average operation is performed on the merged cause states according to the same cause variable number to obtain a spatial propagation cause state set containing the influence of the current partition and adjacent partitions.

[0099] Read the set of reflected evidence from the second visible variable sub-layer, and perform counterfactual consistency verification on each cause state in the set of causes of spatial propagation. Counterfactual consistency verification includes residual synchronization marker matching judgment.

[0100] A retrieval operation is performed on cause states that fail the counterfactual consistency check. The retrieval operation includes removing the corresponding cause state from the spatial propagation cause state set and generating a retrieval set of cause states. The retrieval operation includes replacing the current cause state with the initial cause state of the corresponding partition in the spatial propagation cause state set when a cause state fails the counterfactual consistency check in the current window; and removing the current cause state from the spatial propagation cause state set when a cause state fails the counterfactual consistency check at several consecutive sampling positions.

[0101] Based on the collected set of causal states, and combined with the connection parameters between the causal variable layer and the latent variable layer, the state distribution of the causal variable layer is updated. The collected set of causal states is concatenated with the current state of the latent variable layer to form a joint state vector. Matrix multiplication is performed on the joint state vector and the connection weight matrix between the causal variable layer and the latent variable layer, and corresponding bias terms are added to obtain the updated response value of the causal variable layer. The Sigmoid activation function is then applied to each updated response value to obtain the updated state value of the causal variable layer. All updated state values ​​constitute the state distribution of the causal variable layer.

[0102] Normalize the updated causal variable layer state distribution and output the causal posterior set.

[0103] In this embodiment, controlling the generation of the semantic parameter set includes:

[0104] The posterior probability values ​​of each causal variable in the posterior set of causes are sorted by cause number, and causal variables with posterior probability values ​​greater than a preset judgment threshold are selected.

[0105] For the selected causal variables, the corresponding control semantic parameter items are read according to the causal variable number and the control semantic parameter mapping table. The control semantic parameter mapping table is pre-configured during the system initialization phase and records the correspondence between the causal variable number and the corresponding control semantic parameter item.

[0106] The control semantic parameters include dimming direction indicator, dimming amplitude coefficient, and adjustment priority indicator;

[0107] The control semantic parameter items corresponding to all selected causal variables within the same partition are summarized to form a control semantic parameter set.

[0108] In this embodiment, S5 specifically includes:

[0109] S51. Read the control semantic parameter set, generate a control variable set based on the partition identifier and the control semantic parameter set, and encode the control variable set into a DBO individual representation to form an individual parameter template. The control variable set consists of control variable vectors that correspond one-to-one with the partition. Each control variable vector contains dimming control parameter items corresponding to the partition. The DBO individual representation is stored in vector form, and each element in the vector corresponds to the control variables of each partition in the partition order.

[0110] S52. Set the population size, number of iterations, and search boundary of DBO. Randomly generate an initial population according to the individual parameter template to obtain a candidate control solution set. Perform boundary truncation and legality verification on each candidate control solution set. The search boundary is jointly determined by the adjustment range parameter in the control semantic parameter set and the maximum allowable dimming range preset by the system. The legality verification includes judging whether the control variable exceeds the search boundary and whether it violates the basic operating constraints of the interval. When it violates the constraints, the corresponding control variable is truncated.

[0111] S53. Write the partition association constraint information into the candidate control solution set in the initial population according to the topological adjacency matrix, complete the constraint initialization of the initial population and output the initialized population.

[0112] In this embodiment, S6 specifically includes:

[0113] For each candidate control solution in the candidate control solution set, determine the adjacent partition pairs formed by any partition and the adjacent partitions based on the topological adjacency matrix, and read the control variable values ​​corresponding to the adjacent partition pairs in the candidate control solution.

[0114] For each adjacent partition pair, calculate the difference between the corresponding control variable values ​​of the two partitions. When the absolute value of the difference exceeds the preset adjacent consistency threshold of 0.2, it is determined that the current adjacent partition pair violates the consistency constraint. The control variable value is a normalized dimming control quantity in scalar form, and the difference between the corresponding control variable values ​​of adjacent partitions is the absolute value of the difference between the two scalars.

[0115] For adjacent partition pairs that violate consistency constraints, the control variable values ​​of the two partitions are synchronously updated by equally distributing the difference exceeding the consistency threshold to the corresponding control variables of the adjacent partitions, thus obtaining updated control variable values ​​that satisfy the adjacent consistency constraints. When the absolute value of the difference between the control variable values ​​of adjacent partitions exceeds the adjacent consistency threshold, the excess part is calculated as the difference between the current absolute value of the difference and the adjacent consistency threshold. Half of the excess part is deducted from the larger control variable value, and half of the excess part is added to the smaller control variable value. This process is called equal distribution. The update of adjacent partition pairs is performed sequentially according to the fixed order of the partition numbers in the topological adjacency matrix. After completing one round of adjacent partition pair traversal in the same candidate control solution, the updated candidate control solution is obtained.

[0116] The updated control variable values ​​are written back to the corresponding candidate control solutions, and the above consistency constraint update is performed sequentially on all adjacent partition pairs to complete the update of the candidate control solution set based on the topological adjacency matrix.

[0117] In this embodiment, the evaluation criteria for the fitness function include:

[0118] The power consumption level under the control variables corresponding to the candidate control solution, the degree of illuminance deviation under the control variables corresponding to the candidate control solution, and the degree of consistency between the candidate control solution and the adjustment priority constraints in the control semantic parameter set;

[0119] The power consumption level is obtained by statistical analysis of the power preprocessing sequence in the running window set. The degree of illuminance deviation is evaluated by the degree of deviation between the illuminance preprocessing sequence in the running window set and the target illuminance range. The consistency of the adjustment priority constraint is determined by the matching relationship between the control variables of each partition in the candidate control solution and the corresponding adjustment priority identifier in the control semantic parameter set.

[0120] The function value is obtained by weighted summation. The weights of power consumption level, illuminance deviation under the control variable corresponding to the candidate control solution, and consistency between the candidate control solution and the adjustment priority constraint in the control semantic parameter set are 0.4, 0.4, and 0.2, respectively.

[0121] In this embodiment, the feedback update process specifically includes:

[0122] After the zonal dimming command is issued and executed, the corresponding illuminance measurement sequence, power measurement sequence, dimming control quantity sequence, and drive status sequence are collected in real time. The newly collected equipment operation data is then appended to the equipment operation data set. The image data sequence, waveform data sequence, and acoustic data sequence continuously collected by the multimodal sensing side are synchronously refreshed. The above feedback data undergoes a preprocessing process consistent with the previous steps, including outlier removal, dimension normalization, and validity verification. Then, the sliding window is re-executed according to the predetermined window length and window step size to form an updated sensing window set and operating window set. This is essentially a closed-loop control feedback and sliding window algorithm.

[0123] Example 1: To verify the feasibility of this invention in practice, it was applied to a type of public lighting scenario characterized by multi-zone lighting, complex environmental changes, and susceptibility to external interference. Such scenarios typically involve frequent human and vehicle activity, continuous distribution of lighting areas, and unstable environmental reflection conditions. Traditional closed-loop illuminance control methods are prone to misjudgment and over-dimming in these environments, leading to increased energy consumption and visual discomfort.

[0124] In this scenario, the system continuously collects multimodal sensing data sets and equipment operation data sets. The multimodal sensing data reflects the dynamic changes in the environment, while the equipment operation data reflects the working status of the lighting system itself. After preprocessing, the collected data forms a sensing window set and an operation window set. Based on these, dynamic change feature sequences and state change feature sequences are extracted to construct an evidence feature set. Simultaneously, based on the power data and power-illuminance calibration mapping relationship in the operation window set, the system generates a counterfactual illuminance sequence and differs it from the measured illuminance sequence to obtain a counterfactual residual sequence. This residual synchronization sequence is then calculated using a topological adjacency matrix, thereby constructing a reflection evidence set.

[0125] During the control decision-making phase, the system simultaneously inputs the evidence feature set and the reflection evidence set into the DBM inference model. By setting visible variable sub-layers with different connection mechanisms, environmental dynamic evidence and counterfactual evidence participate in the causal inference process in a differentiated manner. Through spatial propagation and retrieval mechanisms, the system can comprehensively analyze the consistency of abnormal changes in adjacent intervals, thereby distinguishing between changes in actual lighting demand and illuminance anomalies caused by external factors such as reflection and glare. Based on the inferred posterior set of causes, the system further generates a set of control semantic parameters to constrain the subsequent optimization process.

[0126] During the optimization phase, the system initializes the candidate control solution set based on the DBO algorithm and updates the candidate control solutions with consistency constraints according to the topological adjacency matrix in each iteration. When the difference in dimming control variables between adjacent partitions exceeds a preset threshold, the system synchronously adjusts the control variables of adjacent partitions through an equal-amount correction mechanism, effectively suppressing local dimming abrupt changes. Guided by the fitness function, the system comprehensively evaluates the consistency of energy consumption level, illuminance deviation, and adjustment priority to screen candidate control solutions, and finally outputs a stable and reasonable dimming decision vector.

[0127] In practical operation, traditional methods often incorrectly increase brightness levels due to abnormal measured illuminance when high reflectivity interference occurs in the environment. However, the method of this invention can identify the source of such anomalies through counterfactual residuals and spatial synchronization features, avoiding unnecessary brightening and, in some cases, even proactively reducing brightness, effectively suppressing glare and reducing energy consumption. Under multi-zone continuous lighting conditions, this invention reduces abrupt brightness changes between adjacent zones through a topological consistency constraint update mechanism, resulting in a smoother overall brightness distribution and improved visual consistency and comfort.

[0128] Table 1: Comparison of Illumination Control Performance

[0129] Comparison indicators Traditional illuminance closed-loop control Multimodal sensing control only No counterfactual inference scheme was introduced. Method of the present invention Average energy consumption level 0.82 0.74 0.71 0.58 Illuminance overshoot incidence 21.6% 15.2% 12.4% 4.1% Reflective interference misjudgment rate 18.9% 11.3% 9.6% 2.8% Regional brightness consistency index 0.63 0.71 0.76 0.91 Number of dimming abrupt changes in adjacent zones 27 19 14 5 Control stability score 0.68 0.74 0.79 0.93

[0130] As shown in Table 1, the method of this invention has a lower average energy consumption than the control scheme, indicating that the system can avoid over-dimming caused by misjudgment through counterfactual inference and topological consistency constraints. Regarding the false judgment rate of reflective interference, the method of this invention reduces the false judgment ratio, demonstrating that counterfactual residual and synchronicity analysis can effectively identify illuminance anomalies caused by reflection, such as in snowy conditions. This allows the method to proactively reduce brightness rather than increase it under specific circumstances, solving the problem that traditional methods struggle to handle similar environments.

[0131] Regarding regional brightness consistency, the index value of the method in this invention is higher than that of the control scheme, and the number of dimming abrupt changes between adjacent partitions is reduced. This indicates that the equal-amount correction and update mechanism based on the topological adjacency matrix can continuously constrain spatial consistency during the optimization process and avoid drastic changes in local brightness. The comprehensive control stability score results show that the present invention has higher control stability and continuous energy-saving capability under complex environmental change conditions.

[0132] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart control system for LED energy-saving lighting based on multimodal sensing, characterized in that, include: The data acquisition and preprocessing module is used to acquire the multimodal sensing data set and the device operation dataset, merge them for preprocessing, and generate a sensing window set and an operation window set. The evidence feature generation module is used to extract dynamic change feature sequences based on the perception window set and state change feature sequences based on the running window set to generate an evidence feature set. The reflection evidence generation module is used to generate a counterfactual illuminance sequence and calculate a counterfactual residual sequence based on the running window set and the power illuminance calibration mapping, and to calculate the residual synchronization sequence based on the topological adjacency matrix to generate a reflection evidence set. The DBM inference module is used to construct DBM and set the set of visible variables, latent variables and causal variables. It takes the set of evidence features and the set of reflected evidence as input to perform posterior inference and outputs the causal posterior set and the set of control semantic parameters. The DBO optimization module is used to construct the DBO and define discrete and continuous decision vectors. It initializes the population based on the control semantic parameter set to obtain a candidate control solution set, iteratively updates the candidate control solution set based on the topological adjacency matrix, selects the optimal control solution based on the fitness function, and outputs the decision vector. The dimming execution feedback module is used to generate and issue zone dimming commands based on decision vectors, collect execution feedback, and update the perception window set and the running window set.

2. The LED energy-saving lighting intelligent control system based on multimodal perception according to claim 1, characterized in that, The modules are connected in the following way: S1. Acquire the multimodal sensing data set and the device operation data set, perform preprocessing, and generate the sensing window set and the operation window set; S2. The perception window set extracts the dynamic change feature sequence, the running window set extracts the state change feature sequence, and generates the evidence feature set; S3. The set of running windows is mapped with the power illuminance calibration to generate a counterfactual illuminance sequence, the counterfactual residual sequence is calculated, the residual synchronization sequence is calculated based on the topological adjacency matrix, and a set of reflection evidence is generated. S4. Construct DBM, set the set of visible variables, the set of latent variables and the set of causal variables, input the set of evidence features and the set of reflected evidence to perform posterior inference, and output the causal posterior set and the set of control semantic parameters. S5. Construct a DBO, define discrete and continuous decision vectors, and initialize the population based on the set of control semantic parameters to obtain a set of candidate control solutions. S6. During the DBO iteration process, update the candidate control solution set based on the topological adjacency matrix, select the optimal control solution based on the fitness function, and output the decision vector. S7. Generate and issue zone dimming commands based on decision vectors, and collect execution feedback to update the perception window set and the running window set.

3. The LED energy-saving lighting intelligent control system based on multimodal perception according to claim 2, characterized in that, S2 specifically includes: S21. Read data segments corresponding to the same window number from the perception window set and the running window set to form a window input set; S22. Execute a preset feature extraction operator on the window input set to generate a dynamic change feature sequence and a state change feature sequence, and perform window aggregation on the dynamic change feature sequence and the state change feature sequence to obtain a window feature set; S23. Concatenate the window feature set into an evidence feature vector according to a fixed field order, perform linear normalization on the evidence feature vector, and summarize to form an evidence feature set.

4. The LED energy-saving lighting intelligent control system based on multimodal perception according to claim 3, characterized in that, S3 specifically includes: S31. Read the power preprocessing sequence from the running window set, call the corresponding mapping relationship in the power illuminance calibration mapping data for each power sample value in the power preprocessing sequence according to the sampling sequence number, obtain the theoretical illuminance sample value that corresponds one-to-one with each power sample value, and generate the counterfactual illuminance sequence by arranging them according to the sampling sequence number. S32. Perform differential operation point by point on the counterfactual illuminance sequence and the illuminance preprocessing sequence in the same running window according to the same sampling number to obtain the illuminance residual value corresponding to each sampling time, and arrange them according to the sampling number to generate the counterfactual residual sequence. S33. Determine the adjacent partition combination based on the topological adjacency matrix. For any adjacent partition combination, read the corresponding counterfactual residual value at the same sampling sequence number position. Determine whether the change direction of the counterfactual residual value of the adjacent partition is consistent and whether the absolute value of the residual exceeds the preset threshold at the same time. When the consistency condition is met, record it as a synchronous change mark. When the consistency condition is not met, record it as a non-synchronous change mark. Arrange the residual synchronization sequence according to the sampling sequence number. S34. Combine the counterfactual residual sequence and the residual synchronization sequence in a fixed field order to form a set of reflective evidence.

5. The LED energy-saving lighting intelligent control system based on multimodal perception according to claim 4, characterized in that, The DBM structure in S4 specifically includes: The visible variable layer is divided into a first visible variable sublayer and a second visible variable sublayer. The first visible variable sublayer receives the set of evidence features, and the second visible variable sublayer receives the set of reflected evidence. An updatable connection parameter is set between the first visible variable sublayer and the causal variable layer, and a fixed connection parameter is set between the second visible variable sublayer and the causal variable layer. The fixed connection parameter is determined by the power illuminance calibration mapping data and the topological adjacency matrix. Set connection parameters between the causal variable layer and the hidden variable layer; The variables in the first visible variable sublayer consist of dynamic change feature sequences and state change feature sequences, while the variables in the second visible variable sublayer consist of counterfactual residual sequences and residual synchronization sequences. DBM takes a set of evidence features and a set of reflected evidence as input, and outputs a set of causal posterior evidence.

6. The LED energy-saving lighting intelligent control system based on multimodal perception according to claim 5, characterized in that, The posterior inference in S4 specifically includes: Within each partition, based on the set of evidence features input from the first visible variable sublayer, and combined with the corresponding updatable connection parameters, the initial activation state of each causal variable in the causal variable layer is calculated to form the partition's initial causal state set. Based on the topological adjacency matrix, a spatial propagation operation is performed on the initial cause state set of adjacent partitions, propagating the cause state of each partition to adjacent partitions to generate a spatially propagated cause state set. Read the set of reflected evidence from the second visible variable sub-layer, and perform counterfactual consistency verification on each cause state in the set of causes of spatial propagation. Counterfactual consistency verification includes residual synchronization marker matching judgment. Perform a retrieval operation on the cause states that fail the counterfactual consistency check. The retrieval operation includes removing the corresponding cause state from the spatial propagation cause state set and generating a retrieval set of cause states. Based on the collected set of cause states, and combined with the connection parameters between the cause variable layer and the latent variable layer, the state distribution of the cause variable layer is updated; Normalize the updated causal variable layer state distribution and output the causal posterior set.

7. The LED energy-saving lighting intelligent control system based on multimodal perception according to claim 6, characterized in that, S5 specifically includes: S51. Read the control semantic parameter set, generate a control variable set based on the partition identifier and the control semantic parameter set, and encode the control variable set into a DBO individual representation to form an individual parameter template; S52. Set the population size, number of iterations and search boundary of DBO. Randomly generate the initial population according to the individual parameter template to obtain the candidate control solution set. Perform boundary truncation and legality verification on each candidate control solution set. S53. Write the partition association constraint information into the candidate control solution set in the initial population according to the topological adjacency matrix, complete the constraint initialization of the initial population and output the initialized population.

8. The LED energy-saving lighting intelligent control system based on multimodal perception according to claim 7, characterized in that, S6 specifically includes: For each candidate control solution in the candidate control solution set, determine the adjacent partition pairs formed by any partition and the adjacent partitions based on the topological adjacency matrix, and read the control variable values ​​corresponding to the adjacent partition pairs in the candidate control solution. For each adjacent partition pair, calculate the difference between the corresponding control variable values ​​of the two partitions. When the absolute value of the difference exceeds the preset adjacent consistency threshold, the current adjacent partition pair is determined to violate the consistency constraint. For adjacent partition pairs that violate consistency constraints, the control variable values ​​of the two partitions are synchronously updated by distributing the difference exceeding the consistency threshold equally to the corresponding control variables of the adjacent partitions, so as to obtain updated control variable values ​​that satisfy the adjacent consistency constraints. The updated control variable values ​​are written back to the corresponding candidate control solutions, and the above consistency constraint update is performed sequentially on all adjacent partition pairs to complete the update of the candidate control solution set based on the topological adjacency matrix.

Images