Energy Efficiency Analysis Method for Lighting Systems Based on Multimodal Data

Abstract

This invention discloses a method for energy efficiency analysis of lighting systems based on multimodal data, belonging to the field of energy efficiency analysis technology. The method involves filtering the raw data from each single-modal sensor, retaining preliminary data, and then dynamically determining the fusion coefficients of each single-modal sensor based on reinforcement learning using a dual-scale queue. The aligned preliminary data is then fused early according to the corresponding fusion coefficients to form a lighting system data chain containing multiple time-synchronized data blocks. Through multi-level credibility verification, each time-synchronized data block in the lighting system data chain is region-filtered. The fusion coefficients are then retrained using the selected and retained regions from each time-synchronized data block to generate a lighting system task data chain. This task data chain is used to map the electrical variable data of the lighting system, achieving lighting energy efficiency assessment. By generating a high-precision lighting system task data chain, real-time energy efficiency assessment and optimization of the lighting system can be achieved.

Description

Technical Field

[0001] This invention relates to the field of energy efficiency analysis technology, specifically to a method for energy efficiency analysis of lighting systems based on multimodal data. Background Technology

[0002] With increasing global focus on energy issues, energy-saving lighting has become crucial. As a major energy consumer, lighting systems urgently require effective energy efficiency analysis methods to optimize energy use and reduce energy consumption.

[0003] Traditional lighting dimming methods, such as analog dimming, face challenges in terms of adaptability to dynamic scenes and energy efficiency optimization. The singularity of environmental perception dimensions makes it difficult to coordinate color temperature, color rendering index, and glare control, potentially leading to problems such as excessive flicker in office areas and spectral imbalance in medical lighting.

[0004] With the development of the Internet of Things (IoT), artificial intelligence (AI), and sensing technologies, the market demand for intelligent lighting systems is growing rapidly. Intelligent lighting systems need to automatically adjust the brightness and color temperature of light sources based on environmental conditions and user behavior to provide a comfortable and healthy lighting environment while achieving energy savings. This requires accurate analysis of the energy efficiency of lighting systems to better optimize control strategies. However, traditional energy efficiency analysis methods are insufficient to meet the needs of multimodal data fusion and complex scene analysis. Currently, there is a lack of effective methods for using multimodal data for lighting system energy efficiency analysis, resulting in problems such as inaccurate data processing and insufficiently optimized fusion algorithms, making it difficult to fully leverage the advantages of multimodal data.

[0005] Therefore, in order to address the above problems, there is an urgent need for a lighting system energy efficiency analysis method based on multimodal data. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for energy efficiency analysis of lighting systems based on multimodal data, which solves the problems of high data noise, rigid fusion strategies, and insufficient credibility verification in traditional lighting energy efficiency analysis.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a lighting system energy efficiency analysis method based on multimodal data, comprising the following steps: acquiring raw data from each single-modal sensor deployed around the lighting system; evaluating edge parameter thresholds based on the spatiotemporal correlation of neighboring single-modal sensors; filtering the raw data of each single-modal sensor using dynamic confidence determined by fuzzy logic, and retaining preliminary data from each single-modal sensor; aligning the retained preliminary data of each single-modal sensor according to timestamps; dynamically determining the fusion coefficients of each single-modal sensor for the lighting system based on reinforcement learning of a dual-scale queue; performing early fusion of the aligned preliminary data according to the corresponding fusion coefficients to form a lighting system data chain containing multiple time-synchronized data blocks; performing multi-level credibility verification on the lighting system data chain; filtering each time-synchronized data block in the lighting system data chain; retraining the fusion coefficients using the filtered and retained regions in each time-synchronized data block; re-fusioning the retained time-synchronized data block regions using the retrained fusion coefficients to generate a lighting system task data chain; and mapping the electrical variable data of the lighting system using the lighting system task data chain to achieve lighting energy efficiency assessment.

[0008] Furthermore, the specific analysis of the edge parameter threshold assessment based on the spatiotemporal correlation of neighboring single-modal sensors is as follows: based on the physical deployment location of each single-modal sensor around the lighting system, the neighboring single-modal sensors within a preset spatial range are determined, and the historical data sequences of these sensors are extracted; the changing trend of the neighboring single-modal sensor data is analyzed based on a time sliding window to identify its time correlation coefficient. If the time correlation coefficient exceeds the preset spatiotemporal correlation threshold, these sensors are determined to have spatiotemporal correlation; for single-modal sensors with spatiotemporal correlation, the data fluctuation range is statistically analyzed through a dynamic sliding window, and the median within the fluctuation range is taken as the edge parameter threshold.

[0009] Furthermore, the specific logic for determining the dynamic confidence level is as follows: the input variables of the fuzzy logic are defined as the stability of the sensor data, the consistency with the data of neighboring single-modal sensors, and the confidence score of historical data; the membership function of the fuzzy logic is set to divide the input variables into three fuzzy levels: low, medium, and high, and fuzzy inference rules for dynamic confidence level are generated based on the rule base; fuzzy inference is performed in combination with the fuzzy inference rules to generate the dynamic confidence level of each single-modal sensor data; if the dynamic confidence level is lower than the preset confidence level threshold, the data is marked as low-confidence data.

[0010] Furthermore, the specific analysis of filtering the raw data of each single-modal sensor is as follows: filtering out low-confidence data, marking raw data with dynamic confidence levels higher than or equal to a preset confidence threshold as candidate data; performing spatiotemporal integrity verification on the candidate data, filtering out data with missing timestamps or spatial coordinates exceeding a preset spatial range, then performing edge parameter threshold comparison on the remaining candidate data, filtering out candidate data exceeding the edge parameter threshold, and classifying and storing the remaining candidate data according to sensor type to form a preliminary data set for each single-modal sensor.

[0011] Furthermore, the specific analysis of dynamically determining the fusion coefficients of each single-modal sensor based on dual-scale queue reinforcement learning is as follows: A dual-scale queue is constructed, comprising a macro-scale queue and a micro-scale queue. The macro-scale queue stores statistical features of sensor data at minute-level time intervals, while the micro-scale queue stores the raw data stream at second-level time intervals. Based on the reinforcement learning model, macro-state features and micro-state features are extracted from the macro-scale queue and micro-scale queue of the preliminary data retained by each single-modal sensor, respectively. The macro-state features include data mean, variance, and trend slope, while the micro-state features include instantaneous volatility and noise ratio. Based on the macro-state features and micro-state features, unique macro-feature values ​​and unique micro-feature values ​​of the preliminary data retained by each single-modal sensor are determined, respectively. Based on the unique macro-feature values ​​and unique micro-feature values ​​of the preliminary data retained by each single-modal sensor, the reinforcement learning model is collaboratively trained to output the fusion coefficients of each single-modal sensor.

[0012] Furthermore, the generation method of the time synchronization data blocks in the lighting system data chain of the multiple time synchronization data blocks is as follows: the preliminary data of each single-modal sensor after early fusion is divided into data blocks according to a fixed duration, and an event synchronization tag and spatial area identifier are added to each data block to form a time synchronization data block. The multiple time synchronization data blocks are connected to each other based on task events, and the communication connection is cut off based on abnormal events. The task events include periodic performance evaluation report triggering, user active query instructions and energy-saving strategy switching commands. The abnormal events include sensor data interruption, communication delay exceeding limits and abnormal fluctuation of electrical variable data.

[0013] Furthermore, the multi-level trustworthiness verification specifically includes: a first-level sensor perception verification port: a first-level data processing port is established between each single-modal sensor and its neighboring single-modal sensors. The first-level data processing port periodically aggregates mutually perceptible data windows and compares the trend differences between the two sides with an adaptive threshold. When the difference exceeds the limit, a local distortion marker is immediately transmitted to the adjacent first-level data processing port; a second-level data block communication verification port: a second-level data processing port is arranged at any time-synchronized data block and early fusion node. The second-level data processing port performs integrity-temporal continuity double checks on each time-synchronized data block during transmission and fusion, and uses cross-block fingerprint comparison to verify the consistency of the fusion results of neighboring data blocks. If fingerprint misalignment or link packet loss occurs, a link untrustworthy marker is simultaneously sent to the upstream and downstream second-level data processing ports and all first-level data processing ports within their coverage area. Note: The third-level data chain decision verification port: A third-level data processing port is established in the lighting system. The third-level data processing port calls the physical model and historical energy consumption model to globally fit the electrical variables generated by the entire data chain, and compares them with the on-site meter readings in real time. If the fitting residual exceeds the set threshold, the third-level data processing port maps and marks the data segment exceeding the set threshold as a decision-suspect grid, and simultaneously sends it to the corresponding second-level data processing port. All first-level, second-level, and third-level data processing ports form a multi-directional mesh connection. The marked information captured by any data processing port is reported in the form of an event packet and broadcast within the network. The radius is automatically expanded according to the event coordinates, and surrounding data processing ports are scheduled to enter for re-verification. Cross-confirmation is conducted to check whether there are missed or false detections. After confirmation, abnormal grids are uniformly marked. Unmarked data is summarized and reorganized into data blocks according to the single-modal sensor type.

[0014] Furthermore, the specific analysis of retraining the fusion coefficients is as follows: training samples are extracted from the regions selected and retained from each time-synchronized data block; unique macroscopic feature values ​​and unique microscopic feature values ​​of each time-synchronized data block are extracted based on the training samples; then, the unique macroscopic feature values ​​and unique microscopic feature values ​​of each time-synchronized data block are retrained using a reinforcement learning model; updated fusion coefficients are generated by adjusting the parameters of the reinforcement learning model; and the updated fusion coefficients are applied to the corresponding time-synchronized data blocks for fusion to generate the lighting system task data chain.

[0015] Furthermore, the specific analysis for evaluating lighting energy efficiency by mapping the electrical variable data of the lighting system using the lighting system task data chain is as follows: the multimodal information after fusing the various time-synchronized data blocks in the lighting system task data chain is matched with the electrical characteristic model of the lighting system to map the electrical variable data of the lighting system. The electrical variable data is used to generate the current energy efficiency index of the lighting system. The energy efficiency index is compared with the preset benchmark energy efficiency to evaluate the current energy efficiency level of the lighting system.

[0016] The present invention has the following beneficial effects:

[0017] This energy efficiency analysis method for lighting systems based on multimodal data effectively identifies anomalous data ranges by assessing edge parameter thresholds based on the spatiotemporal correlation of neighboring single-modal sensors. Combined with dynamic confidence level filtering determined by fuzzy logic, it removes unreliable data and retains more accurate preliminary data, providing a reliable foundation for subsequent analysis. Multi-level confidence verification further ensures the reliability of data in the lighting system data chain, making the final data used for energy efficiency assessment more accurate. Reinforcement learning based on dual-scale queues dynamically determines the fusion coefficients, which can be adjusted in real time according to data characteristics and system status, enabling optimal fusion of data from each single-modal sensor. Retraining the fusion coefficients further optimizes them based on the filtered regional data from the lighting system data chain, improving fusion quality. This system utilizes multimodal data more comprehensively, improving the accuracy of energy efficiency analysis. Communication connections between time-synchronized data blocks are triggered by task events, and communication connection termination commands are triggered by abnormal events. This allows the system to flexibly adjust data transmission and processing flows according to actual conditions, promptly integrating data when needed and preventing the propagation of erroneous data in case of anomalies. This enhances the system's adaptability and reliability in different scenarios and unforeseen circumstances. By mapping the lighting system's electrical variable data using the lighting system's task data chain, multimodal sensor data can be correlated with the actual electrical variables of the lighting system. By comprehensively considering multiple factors, a precise assessment of lighting energy efficiency can be achieved, providing a more accurate basis for energy-saving optimization of the lighting system and helping to improve the energy utilization efficiency of the lighting system to achieve energy-saving goals.

[0018] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0019] Figure 1 This is a flowchart of the lighting system energy efficiency analysis method based on multimodal data according to the present invention.

[0020] Figure 2 This is a logical schematic diagram of the lighting system energy efficiency analysis method based on multimodal data according to the present invention.

[0021] Figure 3 This is a schematic flowchart illustrating the structure for evaluating lighting energy efficiency according to the present invention. Detailed Implementation

[0022] This application embodiment achieves the technical effects of high-precision dynamic energy efficiency assessment, adaptive multimodal data fusion, and end-to-end reliability assurance through a lighting system energy efficiency analysis method based on multimodal data.

[0023] The overall approach of this application is as follows: through dynamic screening and fusion of multimodal data, combined with a dual-scale reinforcement learning adaptive optimization fusion strategy, and then through multi-level credibility verification, a high-precision lighting system task data chain is generated, ultimately realizing real-time energy efficiency assessment and optimization of the lighting system.

[0024] Please see Figure 1 This invention provides a technical solution: a method for energy efficiency analysis of lighting systems based on multimodal data, comprising the following steps: Step S1, acquiring the raw data of each single-modal sensor deployed around the lighting system, evaluating edge parameter thresholds based on the spatiotemporal correlation of neighboring single-modal sensors, and filtering the raw data of each single-modal sensor using dynamic confidence determined by fuzzy logic, retaining the preliminary data of each single-modal sensor; Step S2, aligning the preliminary data retained by each single-modal sensor according to timestamps, and dynamically determining the fusion efficiency of each single-modal sensor for the lighting system based on reinforcement learning of a dual-scale queue. The initial aligned data is fused according to the corresponding fusion coefficients to form a lighting system data chain containing multiple time-synchronized data blocks. Step S3 involves multi-level credibility verification of the lighting system data chain, region filtering of each time-synchronized data block in the lighting system data chain, retraining the fusion coefficients using the filtered and retained regions in each time-synchronized data block, and re-fusioning the retained time-synchronized data block regions using the retrained fusion coefficients to generate a lighting system task data chain. The lighting system task data chain is then used to map the electrical variable data of the lighting system to achieve lighting energy efficiency assessment.

[0025] Specifically, in step S1, the assessment of edge parameter thresholds based on the spatiotemporal correlation of neighboring single-modal sensors utilizes the spatial proximity of sensors and data change trends to determine the edge parameter thresholds for preliminary data screening. The specific analysis logic is as follows: Based on the physical deployment location of each single-modal sensor, a set of neighboring single-modal sensors within a preset spatial proximity range is determined. For example, among multiple illuminance sensors deployed in a room, sensors within a certain distance threshold can be considered neighboring. Similarly, a set of neighboring single-modal sensors can be established between temperature sensors or current sensors in the same area. Historical data sequences of each neighboring single-modal sensor are extracted, and a time sliding window is used to analyze the data of adjacent sensors. The change trends of the neighboring single-modal sensor data are analyzed based on the time sliding window, and their temporal correlation coefficients are calculated. If the temporal correlation coefficient of a pair of neighboring single-modal sensors exceeds a preset spatiotemporal correlation threshold, it is determined that this set of sensor data has a high synchronous change trend on that time scale, i.e., spatiotemporal correlation exists. Based on a sensor group with spatiotemporal correlation, the data fluctuation range is further statistically analyzed through a dynamic sliding window. For example, the maximum and minimum values ​​of sensor readings are recorded within a dynamic window period, and the fluctuation range of the sensor group readings is calculated accordingly. The median of the data distribution within this fluctuation range is taken as the edge parameter threshold for subsequent data filtering.

[0026] The specific analysis logic for dynamic confidence in step S1 is as follows: Three input variables are defined for the fuzzy logic system: the stability of the sensor data, the consistency between the sensor data and the data from neighboring single-modal sensors, and the confidence score of the sensor's historical data. These three aspects comprehensively consider whether the current data is stable, whether it matches the surrounding sensors, and the sensor's past performance reliability. A suitable membership function is designed for each input variable, and its value range is divided into three fuzzy levels: "low," "medium," and "high." For example, data stability can be represented by the variance or fluctuation frequency of readings over a short period; a small variance indicates "high" stability, while significant fluctuations indicate "low" stability. Data consistency can be measured by the difference between the current sensor reading and the readings of neighboring single-modal sensors; a small difference indicates high consistency. The historical confidence score is given based on the validity and failure rate of the sensor's past data. A fuzzy rule base is established, combining the three fuzzy input variables mentioned above into several inference rules according to empirical rules to infer the dynamic confidence level of the data. For example, the rule could be: "If the data stability is high, the data consistency is high, and the historical reliability is high, then the dynamic confidence level is high"; "If the data stability is low and the data consistency is low, then the dynamic confidence level is low". Fuzzy inference is performed using the fuzzy inference rules to generate the dynamic confidence level of each single-modal sensor data. The dynamic confidence level is compared with a preset confidence level threshold. If the dynamic confidence level of a data point is lower than the threshold, the data is marked as low-confidence data and will not proceed to the next step of analysis; conversely, the original data with a dynamic confidence level higher than or equal to the threshold are marked as candidate data and used as candidate objects for subsequent processing.

[0027] The specific analysis logic for filtering the raw data of each single-modal sensor in step S1 and retaining the preliminary data of each single-modal sensor is as follows: For candidate data filtered through dynamic confidence, further spatiotemporal integrity verification and edge parameter threshold comparison are performed to finally determine the preliminary data set of each single-modal sensor. Spatiotemporal integrity verification is to ensure that the data is valid in both time and space dimensions. By removing candidate data with missing timestamps or that does not conform to the time sequence, the time axis of the input data is ensured to be continuous and aligned. The spatial identifier of the candidate data is checked, and data that exceeds the preset spatial range is removed to ensure the reliability of the spatial attributes of the data. After spatiotemporal integrity verification, the candidate data is compared with the previously determined edge parameter thresholds. Candidate data with values ​​exceeding the corresponding sensor edge parameter threshold range are removed. After the above steps, the remaining candidate data is considered to be both reliable and within the normal range. This data is classified and stored according to the sensor type and number to form the preliminary data set of each single-modal sensor. The preliminary data will serve as the basic input for multimodal fusion.

[0028] In this implementation scheme, the specific logical analysis of analyzing the changing trends of neighboring single-modal sensor data and calculating their temporal correlation coefficients based on time-sliding window analysis is as follows: An identical time-sliding window is set for each neighboring single-modal sensor. This window has a preset time length and sliding step size. The time length is determined based on the regular cycle of data changes in the lighting system, and the sliding step size is set based on the data acquisition frequency. Historical data sequences of two neighboring single-modal sensors are loaded into their corresponding time-sliding windows in chronological order, ensuring that the data in the two windows correspond to the same time period. Within the current time-sliding window, the data change sequence of the first sensor is extracted, reflecting the increase or decrease of data from that sensor over time within the window period. Simultaneously, the data change sequence of the second sensor within the same window period is extracted. The overall trends of the two data change sequences are compared. If both sequences show an increase or decrease at the same time point, and the order of change is consistent and the relative relationship of the magnitude of change is stable, then the two trends are considered similar; otherwise, the trends are considered significantly different. Based on the trend comparison results above, the temporal correlation between two adjacent single-modal sensors within the current time sliding window is determined. When the trends are highly consistent, the temporal correlation coefficient is high; when the trends are partially consistent, the temporal correlation coefficient is medium; and when the trends differ significantly, the temporal correlation coefficient is low. The time sliding window is moved according to the set sliding step size, and the above steps are repeated until the complete historical data sequence of the two adjacent single-modal sensors is traversed, obtaining the temporal correlation coefficients for different time periods.

[0029] The edge parameter threshold is a benchmark value determined based on a sensor group with spatiotemporal correlation, after statistically analyzing the data fluctuation range through a dynamic sliding window. Specifically, it is the median of the data distribution within the reading fluctuation range of the sensor group during the dynamic window period. It reflects the reasonable fluctuation boundary of the data collected by the sensor group under normal operating conditions and serves as a quantitative standard for judging whether the raw data of a single-mode sensor is within the normal range.

[0030] Fuzzy inference is performed by combining fuzzy inference rules to generate dynamic confidence scores for each single-modality sensor data. Specifically, the fuzzy output of the dynamic confidence scores is converted into specific numerical values ​​through a defuzzification method. These numerical values ​​are the dynamic confidence scores of the corresponding original data and are used for subsequent screening and judgment of the original data.

[0031] The edge parameter thresholds set based on the spatiotemporal correlation of neighboring single-modal sensors can adapt to the normal variation range of the local environment, filtering out abnormal data exceeding this range. This avoids misjudgments caused by isolated anomalies from a single sensor, improving the reliability of initial data screening. The dynamic confidence assessment mechanism utilizes fuzzy logic to fuse multiple factors to determine data reliability. Compared to a single threshold or simple rules, it can more comprehensively and flexibly identify abnormal or low-quality data. After filtering out low-confidence data, the remaining candidate data has a higher overall quality, improving the accuracy of multimodal data fusion and analysis from the source. By verifying the spatiotemporal integrity and threshold range of the candidate data, the validity and correctness of the initial data from each sensor are ensured.

[0032] Specifically, the logical analysis of aligning the preliminary data retained by each single-modality sensor according to timestamps in step S2 is as follows: Preliminary data from different sensors are aligned and integrated according to timestamps. Considering that the sampling frequencies of different sensors may vary, a combination of interpolation and buffer queues is used to achieve time synchronization. Specifically, the data from high-speed sampling sensors is downsampled or aggregated to the same time granularity as that from low-speed sampling sensors. Data that is too slow is filled with estimated values ​​at intermediate times through interpolation, ensuring that data from all types of sensors are available at each time point under a unified time reference. The result of timestamp alignment is the generation of multiple time-synchronized data point sequences, ensuring that sensor readings from various modalities can be matched at each time point, creating conditions for subsequent data fusion.

[0033] The specific logical analysis of step S2, which uses reinforcement learning based on a dual-scale queue to dynamically determine the fusion coefficients of each single-modal sensor, is as follows: Based on time-aligned data, a reinforcement learning model is introduced to dynamically determine the fusion coefficients of each single-modal sensor data, thereby achieving adaptive weighted fusion of multimodal data. Specifically, a dual-scale queue structure is employed to extract different time-scale features of the sensor data: a macro-scale queue and a micro-scale queue. The macro-scale queue stores statistical features of the sensor data at minute-level time intervals, such as the mean, variance, and trend slope of the data over the most recent few minutes, reflecting the long-term trend and stability of the sensor data. The micro-scale queue stores the raw data stream or high-frequency detail data of the sensor at second-level time intervals to capture short-term instantaneous fluctuations and noise levels. The reinforcement learning model extracts state features from the macro-scale queue and micro-scale queue of each sensor: macro-state features include the moving average, recent variance, and trend slope of the sensor data, representing the sensor's behavioral characteristics over a longer period; micro-state features include real-time fluctuation amplitude and noise ratio, representing the sensor's data characteristics over a short period. By combining the aforementioned macroscopic and microscopic state characteristics, a unique state representation is provided for the current data of each sensor, containing unique macroscopic and microscopic feature values. Based on this state representation, a reinforcement learning model is collaboratively trained using the unique macroscopic and microscopic feature values ​​of the preliminary data retained by each single-modal sensor. The model outputs the fusion coefficients of each single-modal sensor. Specifically, the reinforcement learning model takes the state characteristics of each sensor as input and, after training, can provide the optimal weight allocation for each modality at the current moment. A collaborative training mechanism is employed during training to continuously adjust the fusion coefficients to maximize the preset reward function. By introducing dual-scale state characteristics, both long-term stability and short-term fluctuations can be considered during training, thus assigning appropriate weights to different sensors.

[0034] The specific logical analysis of the lighting system data chain with multiple time-synchronized data blocks in step S2 is as follows: Once the fusion coefficients of each sensor at the current moment are obtained, the preliminary single-modal data after alignment are weighted and combined according to these coefficients, thereby achieving early fusion of multimodal data. Early fusion refers to combining multi-source data into a comprehensive data in the early stage of data analysis to reduce the impact of single sensor errors on subsequent analysis. The fusion output data can be understood as a result vector or data block that comprehensively considers all sensor information at each time point. As time progresses, such fused data blocks are continuously generated, and these continuous fusion results are organized into a lighting system data chain ordered by time, with each node corresponding to a time-synchronized data block. In addition to containing the fused multimodal values, each time-synchronized data block also includes an event synchronization tag and a spatial area identifier for that moment. The event synchronization tag is used to record whether a specific task event has been triggered, such as routine energy efficiency report generation, user-initiated query requests, or energy-saving strategy switching execution; the spatial area identifier indicates the physical area to which the data block belongs, such as which building and which floor's lighting area it is. Multiple time-synchronized data blocks are linked together via communication triggered by task events: when a task event occurs, the system sends the corresponding time-segment data block to the central server or relevant modules for priority processing; when an abnormal event is detected, the communication link of the relevant data block is cut off or the data segment is marked as abnormal to prevent abnormal data from affecting the overall analysis. Task events include periodic performance evaluation report triggers, user-initiated query commands, and energy-saving strategy switching commands; abnormal events include sensor data interruption, communication latency exceeding limits, and abnormal fluctuations in electrical variable data.

[0035] In this implementation scheme, the unique macroscopic feature value is a comprehensive quantitative representation of the sensor's data behavior characteristics over a longer time period. It uniquely identifies the long-term trend and stability of the sensor's data at a macroscopic scale, serving as a crucial basis for the reinforcement learning model's state judgment and reflecting the overall pattern of sensor data on a minute-level time scale. The acquisition method involves: extracting statistical features of the sensor within a preset minute-level time interval from the macroscopic scale queue, including data mean, variance, and slope of change trends; normalizing these statistical features to eliminate dimensional differences between different features; and using a feature fusion algorithm to integrate the normalized statistical features into a unique numerical value, which is the unique macroscopic feature value.

[0036] A unique micro-feature value is a comprehensive quantitative representation of the instantaneous characteristics of sensor data within a short time period. It uniquely identifies the short-term fluctuations and noise levels of the sensor data at a microscale, providing reinforcement learning models with short-term data state information and reflecting the detailed changes in sensor data on a second-level time scale. The acquisition method involves: extracting the original data stream or high-frequency detail data of the sensor within a preset second-level time interval from the micro-scale queue, calculating its real-time fluctuation amplitude and noise ratio; standardizing these instantaneous features to ensure different features are within the same numerical range; and merging the standardized instantaneous features into a single unique value using a feature integration method. This value is the unique micro-feature value.

[0037] The specific logical analysis of the reinforcement learning model collaborative training based on the unique macroscopic and microscopic feature values ​​of the preliminary data retained by each single-modal sensor, and the output of the fusion coefficients of each single-modal sensor, is as follows: The reinforcement learning agent collects the unique macroscopic and microscopic feature values ​​of each single-modal sensor and combines them to form the complete state features of each sensor; the policy network receives the complete state features of all sensors as input, performs hierarchical processing on the input state features, extracts features from the macroscopic and microscopic features of each sensor respectively, and performs cross-sensor correlation analysis on the extracted features; based on the feature processing results, the policy network generates an initial fusion coefficient candidate value for each sensor based on the preset network parameters and the trained decision logic. This candidate value reflects the preliminary weight ratio of the sensor data in the current state; the generated initial fusion coefficient candidate value is constrained and adjusted to ensure that the sum of the fusion coefficients of all sensors is 1, and the value of each fusion coefficient is within a preset reasonable range to avoid extreme weight allocation; the adjusted fusion coefficient is output as the weight of the fusion of the single-modal sensor data at the current moment, and the output result is fed back to the reinforcement learning model for subsequent reward calculation and policy network parameter update.

[0038] By aligning with timestamps, data from different modalities can be compared and fused under the same time reference, avoiding information misalignment or association errors caused by asynchronous sensor sampling, and ensuring the consistency and correctness of the fused input. The reinforcement learning method based on dual-scale queues enables the fusion coefficients to be dynamically adjusted over time. When a sensor experiences abnormal fluctuations in the short term, its micro-features guide the reinforcement learning agent to reduce the weight of that sensor's data; conversely, when a sensor performs stably and reliably over a long period, its macro-features increase its base weight. Compared to fixed weights or simple weighted averaging, dynamic fusion coefficients can better adapt to complex environmental changes, improving the accuracy and robustness of multimodal data fusion results. Data blocks segmented by fixed duration are accompanied by clear time and spatial labels, allowing subsequent analysis to be conducted on specific time periods and regions. The task event triggering mechanism ensures that relevant data is extracted and processed in a timely manner at critical moments, while the abnormal event cutoff mechanism prevents faulty data from interfering with the entire lighting system data chain, structurally improving the system's robustness and security under abnormal conditions.

[0039] Specifically, the logical analysis of the multi-level credibility verification in step S3 is as follows:

[0040] The first-level sensor sensing and verification port: A first-level data processing port is set up between each single-modal sensor and its neighboring single-modal sensors. For example, this port is set up between the illuminance sensor and the nearby temperature sensor in a room. The first-level data processing port will periodically aggregate mutually perceptible data windows. The period length is determined according to the update frequency of the sensor data, such as once every 5 seconds. The data window covers all the raw data of the two sensors within the period. The first-level data processing port automatically generates an adaptive threshold based on the fluctuation range of historical data. This threshold can be dynamically adjusted according to the sensor's operating status. For example, the threshold is smaller when the sensor is in a stable environment and increases accordingly when the environment fluctuates greatly. The first-level data processing port compares the trend difference between the two sensors within the data window. If the difference exceeds the adaptive threshold, it immediately transmits a local distortion mark to the adjacent first-level data processing port. The mark contains the timestamp of the anomaly and the sensor identifier.

[0041] The second-level data block communication verification port: A secondary data processing port is deployed at any time-synchronized data block and early fusion node. For example, a secondary data processing port is set up at the fusion node responsible for integrating illuminance and human body sensing data in a lighting system. The secondary data processing port performs a dual check of integrity and temporal continuity for the time-synchronized data block. The integrity check verifies whether the number of preset fields of the data block is complete, such as whether it contains necessary information such as sensor ID, timestamp, and measurement value. The temporal continuity check compares the timestamp order of the data blocks to ensure that they are transmitted sequentially according to the acquisition time without any jumps or repetitions. The secondary data processing port uses a cross-block fingerprint comparison method to verify the consistency of the fusion results of neighboring data blocks. That is, it extracts the feature information of each data block to generate a unique identifier, uses the unique identifier as a fingerprint, and then compares the fingerprint differences of neighboring data blocks. If fingerprint misalignment occurs, such as the feature identifiers of adjacent data blocks do not match, or if there is packet loss in the link, such as missing data blocks during transmission, an untrusted link marker is sent synchronously to the upstream and downstream secondary data processing ports and all primary data processing ports within its coverage area. The marker contains the location information and anomaly type of the abnormal data block.

[0042] The third-level data chain decision verification port: A third-level data processing port is established in the lighting system. This port is connected to the physical model library and historical energy consumption model library of the lighting system. The third-level data processing port calls the physical model, such as the model reflecting the relationship between the power of lighting equipment and light intensity, and the historical energy consumption model, such as the energy consumption data model of the same period in the past year, to perform global fitting on the electrical variables generated by the entire data chain. The fitting process is achieved by comparing the consistency of the trend between the data chain output and the model prediction. The third-level data processing port compares the actual readings of the field meters in real time. If the fitting residual exceeds the set threshold, the data segment exceeding the threshold is mapped and marked as a decision-suspect grid. The threshold is determined based on the historical fitting accuracy. The grid contains the time range of abnormal data and the corresponding data block number, and is sent to the corresponding second-level data processing port.

[0043] All primary, secondary, and tertiary data processing ports form a multi-directional mesh connection. The tagging information captured by any port is reported in the form of event packets and propagated throughout the network. The event packet contains event coordinates, such as sensor location or data block number. The verification radius is automatically expanded based on the event coordinates. For example, when a primary port reports an anomaly, the expanded radius covers the three surrounding primary ports.

[0044] The surrounding data processing ports are then re-verified. For example, primary ports re-compare data trends, and secondary ports re-check data block integrity. Cross-verification is used to determine if there are any missed or false detections. Once an anomaly is confirmed, the anomaly grid is uniformly marked. Unmarked data is aggregated and reassembled into data blocks according to single-modal sensor type to provide reliable data for subsequent fusion processing.

[0045] Sensor perception verification: This process utilizes data from neighboring single-modal sensors to cross-verify the reliability of their readings. Specifically, it checks the consistency and physical plausibility of outputs from different types of sensors in the same or adjacent areas. For example, if the fused results in a time-synchronized data block show a sharp increase in lighting power in a certain area, but a neighboring illuminance sensor does not detect a corresponding increase in brightness, this inconsistency will be flagged. By allowing neighboring single-modal sensors to "cross-verify" with each other, it can identify individual sensor readings that are abnormal and not completely filtered out by previous steps. Once an anomaly is found in a data block related to a particular sensor, the system will remove the relevant data in subsequent processing. Sensor perception verification provides the first line of defense, further verifying the authenticity of the data through redundant information from neighboring sensors, filtering out residual anomalies, and monitoring and improving the reliability of sensor-level data in real time, thus laying the foundation for the reliability of the entire lighting system's data chain.

[0046] Data block communication verification: After passing the first-level verification, the reliability of each time-synchronized data block during transmission and fusion is further verified. This includes identifying the consistency of the data block fusion results and the integrity during transmission. Specifically, integrity checks are performed on each received data block, such as verifying data length and checksum, to ensure that the data block has not been tampered with and that the information is complete during network transmission. The numerical changes between adjacent time-synchronized data blocks are compared to see if they are continuous and smooth, with a focus on whether there are any abrupt changes in a single data block. If a data block shows a drastic change compared to the preceding and following data blocks that cannot be explained by the normal behavior of the lighting system, it means that an error has occurred in the fusion calculation or transmission and storage of that data block. In this case, the data block is marked as abnormal and isolated or removed in the lighting system data chain processing. Data block communication verification adds a layer of network transmission and sequence logic protection to the lighting system data chain, enabling timely detection of abnormal data blocks caused by communication failures or calculation errors. This prevents incomplete or distorted data from flowing into the final energy efficiency assessment, thereby maintaining the temporal consistency and reliability of the lighting system data chain.

[0047] Lighting System Data Link Decision Verification: The credibility of the final energy efficiency assessment results output by the lighting system data link, which contains multiple time-synchronized data blocks, is verified. Based on the physical model of the lighting system and historical energy consumption patterns, the results are reviewed for reasonableness. Specifically, the electrical variable data mapped from the lighting system data link is verified using the physical model of the lighting system's electrical characteristics. Given the current on / off state and brightness adjustment level of the luminaires, the model calculates the theoretical power consumption range. If the lighting system data link results exceed this reasonable range, the results are deemed problematic. Simultaneously, the energy efficiency indicators output by the lighting system data link are compared with actual energy consumption measurements. If a significant deviation is found between the lighting system data link results and actual measurements, it indicates that the fusion results of the lighting system data link may be inaccurate or incomplete. Through a dual method of physical model verification and actual measurement comparison, the final output energy efficiency assessment results are ensured to conform to the objective laws and actual conditions of the lighting system. Once an anomaly is detected in the lighting system data link decision verification, the aforementioned levels and algorithm modules are traced back to locate the cause of the problem. If necessary, the credibility level of that section of the result is lowered in the assessment report. The decision-making level verification takes a holistic approach, starting from the principles of lighting systems and macro-level energy consumption trends, to ensure the credibility and reliability of the final energy efficiency assessment results. This level of verification not only prevents errors multiple times during data processing but also identifies anomalies in the final decision-making stage, minimizing the impact of potential biases on actual judgments and ensuring the accuracy of energy efficiency assessment conclusions.

[0048] The specific logic analysis for retraining the fusion coefficients in step S3 is as follows: For each time-synchronized data block in the lighting system data chain, a region selection process is performed. Data blocks not marked as abnormal, along with high-quality segments within them, are selected as the basis for retraining. Each selected and retained time-synchronized data block region is extracted as a training sample. For these training sample data, their macroscopic and microscopic feature values ​​are recalculated. Based on the unique macroscopic and microscopic feature values ​​extracted for each sample, the reinforcement learning model is trained again, adjusting its internal parameters to generate updated fusion coefficients for each single-modal sensor. Since this round of training only uses high-quality data block regions, the model can more accurately learn the relationship between sensor data and energy consumption, thereby correcting any previously potentially deviated weight settings. After training, the system applies the updated fusion coefficients to the corresponding time-synchronized data blocks, recalculates the fusion of these data blocks, replaces the corresponding parts in the original lighting system data chain, and generates an optimized lighting system task data chain.

[0049] Step S3, which uses the lighting system task data chain to map the electrical variable data of the lighting system to achieve lighting energy efficiency assessment, involves the following specific analysis: After obtaining the final optimized lighting system task data chain, this series of highly reliable fused data is used for lighting system energy efficiency assessment, i.e., electrical variable mapping and energy efficiency index calculation. The specific steps are as follows: An electrical characteristic model of the lighting system is pre-established. This model defines the correspondence between inputs and outputs based on the specifications and topology of the lighting equipment. During the electrical variable mapping stage, the multimodal fusion information contained in each time-synchronized data block in the lighting system task data chain is converted and mapped into relevant electrical parameters of the lighting system. For example, a data block corresponds to the state of a certain area at a certain time, where the fused information includes ambient brightness, personnel activity status, and lamp dimming levels. Based on the electrical model, the actual power consumption and current load of the lighting circuit in that area can be calculated. The lighting system task data chain provides comprehensive state information, which the electrical characteristic model then transforms into energy consumption-related variables of the lighting system. Based on the electrical variable data obtained from these mappings, the energy efficiency index of the lighting system is calculated. For example, the energy efficiency index can be set as the power consumption per unit of light output, or the energy utilization efficiency score for meeting lighting needs within a certain time period. The currently calculated energy efficiency index is compared with a preset benchmark energy efficiency, which is determined based on industry standards, historical best records, or pre-set energy-saving targets. Through comparative analysis, if the current energy efficiency index is close to or better than the benchmark value, it indicates that the lighting system is operating in a highly efficient state; if it is significantly lower than the benchmark value, it indicates that the current system energy efficiency is low, there is room for optimization, and the control strategy needs to be adjusted or related equipment needs to be repaired. The evaluation results can be output in an intuitive report format, including the energy efficiency index value, the percentage difference relative to the benchmark, and the energy efficiency level rating, such as excellent, good, average, or poor.

[0050] In this implementation scheme, a three-tiered port-based layered verification system comprehensively checks for data anomalies from sensor perception and data block communication to data link decision-making, significantly reducing the probability of erroneous data entering subsequent processes. Multi-directional mesh connectivity enables the propagation and re-verification of labeled information, effectively avoiding missed or false detections. For example, if a port misjudges a case, re-verification of surrounding ports can correct it promptly. The dynamic adjustment mechanism of each port allows the verification process to adapt to different operating states, ensuring stable operation even in complex environments and providing a high-quality data foundation for lighting system energy efficiency assessment. Retraining the fusion coefficients enables self-correction and optimization. When the environment changes or the initial model weights are not accurate enough, secondary learning using reliable data feedback effectively avoids the continued use of outdated fusion strategies. The updated fusion coefficients make the lighting system data link more consistent with the current actual state, improving the accuracy and reliability of data in the lighting system task data link. The online adaptive optimization mechanism ensures that the energy efficiency analysis method remains efficient and reliable over time, reducing the need for manual intervention. Applying the results of the aforementioned data processing to actual energy management decisions achieves a closed loop from data to energy efficiency insights. Based on accurately fused data, energy efficiency indicators are calculated and compared with benchmark levels. Managers can clearly understand the current performance level of the lighting system. Compared with methods that rely solely on single sensor data or rough estimations, this provides a more accurate and targeted energy efficiency assessment, enabling timely detection of energy efficiency anomalies and identification of problem areas. This provides a scientific basis for subsequent energy-saving measures and achieves intelligent management of the energy efficiency of the lighting system.

[0051] In summary, this application has at least the following effects:

[0052] By employing spatiotemporal correlation analysis and dynamic confidence screening, noisy data is effectively eliminated, retaining high-quality data and making subsequent analysis more accurate. Multi-level confidence verification ensures the reliability of data from the sensor level to the decision level, reducing the risk of misjudgment. Dual-scale queue reinforcement learning dynamically adjusts the fusion coefficients to adapt to real-time scenarios such as changes in illumination and human activities, avoiding the rigidity problem of traditional fixed-weight fusion. The mechanism continuously optimizes the data fusion strategy, improving the stability of long-term energy efficiency analysis. Multimodal information is integrated to accurately map electrical variable data, achieving high-precision energy efficiency calculation. Communication resource allocation is optimized to reduce redundant data transmission and improve system response speed. A rapid fault-tolerance mechanism triggered by abnormal events ensures continuous system operation.

[0053] Those skilled in the art will understand that embodiments of the present invention can be provided as methods. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0054] This invention is described with reference to a flowchart of a method according to embodiments of the invention. It should be understood that the combination of each step in the flowchart can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process. Figure 1 A device for a function specified in one or more processes.

[0055] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 The function specified in one or more processes.

[0056] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 Steps of a specified function in one or more processes.

[0057] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0058] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for energy efficiency analysis of lighting systems based on multimodal data, characterized in that, Includes the following steps: The system acquires raw data from each single-modal sensor deployed around the lighting system, evaluates edge parameter thresholds based on the spatiotemporal correlation of neighboring single-modal sensors, and filters the raw data of each single-modal sensor using dynamic confidence determined by fuzzy logic, retaining the preliminary data of each single-modal sensor. The initial data retained by each single-modal sensor are aligned according to the timestamp. The fusion coefficient of each single-modal sensor for the lighting system is dynamically determined based on the reinforcement learning of the dual-scale queue. The aligned initial data is then fused early according to the corresponding fusion coefficient to form a lighting system data chain containing multiple time-synchronized data blocks. The specific analysis of the reinforcement learning dynamic determination of the fusion coefficients of each single-modal sensor based on dual-scale queues is as follows: A dual-scale queue is constructed, comprising a macro-scale queue and a micro-scale queue. The macro-scale queue is used to store statistical features of sensor data at minute-level time intervals, and the micro-scale queue is used to store raw data streams at second-level time intervals. Based on the reinforcement learning model, macro-state features and micro-state features are extracted from the macro-scale queue and micro-scale queue of the preliminary data retained by each single-modal sensor. The macro-state features include data mean, variance and trend slope, and the micro-state features include instantaneous volatility and noise ratio. Based on macroscopic and microscopic state characteristics, unique macroscopic and unique microscopic feature values ​​are determined for the preliminary data retained by each single-mode sensor. The reinforcement learning model is trained collaboratively based on the unique macroscopic and microscopic feature values ​​of the preliminary data retained by each single-modal sensor, and the fusion coefficients of each single-modal sensor are output. Multi-level credibility verification is performed on the lighting system data chain. Regions are filtered for each time synchronization data block in the lighting system data chain. The fusion coefficients are retrained using the filtered and retained regions in each time synchronization data block. The retrained fusion coefficients are used to re-fuse the retained time synchronization data block regions to generate the lighting system task data chain. The lighting system task data chain is used to map the electrical variable data of the lighting system to achieve the evaluation of lighting energy efficiency.

2. The method for energy efficiency analysis of lighting systems based on multimodal data according to claim 1, characterized in that, The specific analysis of the edge parameter threshold evaluation based on the spatiotemporal correlation of neighboring single-modal sensors is as follows: Based on the physical deployment location of each single-mode sensor around the lighting system, determine the neighboring single-mode sensors within a preset spatial range and extract the historical data sequences of these sensors. Based on the time sliding window analysis, the changing trend of neighboring single-modal sensor data is identified, and their time correlation coefficient is determined. If the time correlation coefficient exceeds the preset spatiotemporal correlation threshold, these sensors are determined to have spatiotemporal correlation. For single-mode sensors with spatiotemporal correlation, the data fluctuation range is statistically analyzed using a dynamic sliding window, and the median within the fluctuation range is taken as the threshold for edge parameters.

3. The method for energy efficiency analysis of lighting systems based on multimodal data according to claim 1, characterized in that, The specific logic for determining the dynamic confidence level is as follows: The input variables for fuzzy logic are defined as the stability of sensor data, the consistency with data from neighboring single-modal sensors, and the reliability score of historical data. Define the membership function of fuzzy logic, divide the input variables into three fuzzy levels: low, medium, and high, and generate fuzzy inference rules with dynamic confidence based on the rule base; Fuzzy inference is performed by combining fuzzy inference rules to generate dynamic confidence scores for each single-modal sensor data. If the dynamic confidence score is lower than the preset confidence score threshold, the data is marked as low-confidence data.

4. The method for energy efficiency analysis of lighting systems based on multimodal data according to claim 3, characterized in that, The specific analysis of filtering the raw data from each single-mode sensor is as follows: Low-confidence data is filtered out, and original data with dynamic confidence levels higher than or equal to a preset confidence threshold are marked as candidate data; Spatiotemporal integrity verification is performed on the candidate data to filter out data with missing timestamps or spatial coordinates exceeding the preset spatial range. Then, edge parameter threshold comparison is performed on the retained candidate data to filter out candidate data exceeding the edge parameter threshold. The remaining candidate data is classified and stored according to sensor type to form a preliminary data set for each single-modal sensor.

5. The method for energy efficiency analysis of lighting systems based on multimodal data according to claim 1, characterized in that, The time synchronization data blocks in the lighting system data chain are generated as follows: the preliminary data of each single-modal sensor after early fusion are divided into data blocks according to a fixed duration, and an event synchronization tag and spatial area identifier are added to each data block to form a time synchronization data block. The multiple time synchronization data blocks are connected to each other based on task events, and the communication connection is cut off based on abnormal events. The task events include periodic performance evaluation report triggering, user active query instructions, and energy-saving strategy switching commands. The abnormal events include sensor data interruption, communication delay exceeding limits, and abnormal fluctuation of electrical variable data.

6. The method for energy efficiency analysis of lighting systems based on multimodal data according to claim 1, characterized in that, The multi-level credibility verification specifically includes: First-level sensor sensing and verification port: A first-level data processing port is set up between each single-mode sensor and its neighboring single-mode sensor. The first-level data processing port periodically aggregates mutually perceptible data windows and compares the trend differences on both sides with an adaptive threshold. When the difference exceeds the limit, a local distortion mark is immediately transmitted to the adjacent first-level data processing port. Second-level data block communication verification port: A secondary data processing port is deployed at any time-synchronized data block and early fusion node. The secondary data processing port performs integrity-timing continuity double checks on each time-synchronized data block during transmission and fusion, and uses cross-block fingerprint comparison to verify the consistency of fusion results of neighboring data blocks. If fingerprint misalignment or link packet loss occurs, an untrusted link marker is sent synchronously to the upstream and downstream secondary data processing ports and all primary data processing ports within its coverage area. The third-level data chain decision verification port: A third-level data processing port is established in the lighting system. The third-level data processing port calls the physical model and historical energy consumption model to perform global fitting of the electrical variables generated by the entire data chain, and compares the on-site electricity meter readings in real time. If the fitting residual exceeds the set threshold, the third-level data processing port maps and marks the data segment exceeding the set threshold as a decision-suspect grid, and simultaneously sends it to the corresponding second-level data processing port. All primary, secondary, and tertiary data processing ports form a multi-directional mesh connection. The tagging information captured by any data processing port is reported in the form of an event packet and broadcast within the network. The radius is automatically expanded according to the event coordinates, and surrounding data processing ports are scheduled to enter for re-verification. Cross-checking is performed to confirm whether there are any missed or false detections. Once confirmed, the abnormal grid is marked uniformly. The unmarked data is summarized and reorganized into data blocks according to the single-modal sensor type.

7. The method for energy efficiency analysis of lighting systems based on multimodal data according to claim 6, characterized in that, The specific analysis of retraining the fusion coefficients is as follows: training samples are extracted from the recombined data blocks, and unique macroscopic and microscopic feature values ​​of each time-synchronized data block are extracted based on the training samples. Then, the unique macroscopic and microscopic feature values ​​of each time-synchronized data block are retrained using a reinforcement learning model. The parameters of the reinforcement learning model are adjusted to generate updated fusion coefficients. The updated fusion coefficients are applied to the corresponding time-synchronized data blocks for fusion to generate the lighting system task data chain.

8. The method for energy efficiency analysis of lighting systems based on multimodal data according to claim 1, characterized in that, The specific analysis for evaluating lighting energy efficiency by mapping the electrical variable data of the lighting system using the lighting system task data chain is as follows: the multimodal information after fusing the various time-synchronized data blocks in the lighting system task data chain is matched with the electrical characteristic model of the lighting system to map the electrical variable data of the lighting system. The electrical variable data is used to generate the current energy efficiency index of the lighting system. The energy efficiency index is compared with the preset benchmark energy efficiency to evaluate the current energy efficiency level of the lighting system.

Images