Smart lighting energy-saving method and system based on multi-modal data fusion

Abstract

The invention relates to an intelligent lighting energy-saving method and system based on multi-modal data fusion. The method comprises the following steps: periodically obtaining a decision trajectory data set; combining the experience pool of the previous period with the decision trajectory data set newly collected in the current period according to a set proportion, and screening according to priority weight; training to obtain candidate strategies; evaluating the candidate strategies in a test scene set; adding the strategy comprehensive score into a strategy library when the strategy comprehensive score meets an access condition; when the strategy capacity in the strategy library reaches a preset value, eliminating the strategy with the lowest comprehensive strategy score based on the Pareto frontier; constructing a fitness function of the strategy and establishing a feedback-weight mapping model used for obtaining a prediction score; and based on the feedback-weight mapping model, establishing a loss function gradient, and based on the loss function gradient, performing iterated weight calculation to obtain iterated weights after the iterated calculation. According to the invention, the long-term energy-saving efficiency and the user experience of the lighting system are improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of smart cities and the Internet of Things, and in particular to a smart lighting energy-saving method and system based on multi-modal data fusion. Background Art

[0002] With the rapid development of smart city and Internet of Things technologies, the demand for intelligent, energy-saving, and dynamic regulation in urban lighting, park lighting, and commercial area lighting management is becoming increasingly urgent. At present, lighting systems have extensive application requirements in various application scenarios such as urban roads, parks, ecological parks, commercial complexes, pedestrian streets, residential areas, campuses, and special scenarios (such as temporary lighting during large events, emergency rescue, and disaster weather). Relevant management parties, including city managers, property companies, park management parties, commercial operators, etc., generally hope to achieve energy conservation and consumption reduction, improve lighting quality, enhance safety protection, and further promote the construction of green and low-carbon cities through more efficient and intelligent lighting means.

[0003] However, traditional lighting systems, especially those for urban roads, parks, and commercial areas, usually adopt a control mode based on simple sensors (such as photosensors) with timers. Although such systems achieve lighting automation to a certain extent, they generally have the following deficiencies: Rigid strategies and lack of flexibility; most systems rely on preset schedules or fixed sensor thresholds and cannot be adjusted according to real-time environment, dynamic population, and vehicle information. For example, even when there are no pedestrians or vehicles passing by in a certain area in the early morning, the street lights are still fully lit for a long time, wasting electricity.

[0004] Single data source and insufficient perception; traditional solutions often rely only on light sensors or infrared sensors and cannot obtain more comprehensive scene information, making it difficult to achieve precise control. For example, using only light intensity to judge cannot distinguish whether the brightness is insufficient due to rain or due to the need for lighting at night.

[0005] Low energy efficiency and lack of optimization; in terms of multi-objective balance (such as energy consumption, comfort, and safety), traditional systems lack optimization capabilities and are prone to problems such as "either too bright or too dark", which not only wastes energy but also reduces the user experience.

[0006] Currently, some research and products have tried to introduce multi-sensor (such as temperature and humidity, video surveillance) data for optimization, but these solutions have the following technical defects: Insufficient depth of data fusion; the data of multi-sensors are often processed independently, lacking the ability of cross-sensor collaborative analysis and decision-making.

[0007] Lack of evolution ability of control strategies; most systems use rule-based strategies and are difficult to adapt to changing environments, resulting in a decline in optimization effects after long-term operation.

[0008] Lack of modeling of the dynamic relevance between the environment and the population; factors such as peak flow of people, traffic congestion, and bad weather have a direct impact on lighting requirements, and existing solutions often do not introduce such relevance modeling and dynamic prediction.

[0009] Insufficient communication and collaboration; most traditional lighting control systems are centralized and decentralized solutions, lacking an efficient and low-latency communication mechanism between nodes, and it is difficult to achieve collaborative control and data sharing at the streetlight pole level.

[0010] Limited ability to handle faults and emergencies; existing systems usually can only passively accept central instructions, lacking self-healing ability and active compensation mechanism. When some nodes fail, it will directly affect the lighting effect of the entire area. Summary of the Invention

[0011] Therefore, the present invention provides a smart lighting energy-saving method and system based on multi-modal data fusion to improve the long-term energy-saving efficiency and user experience of the lighting system.

[0012] To solve the above technical problems, the present invention provides a smart lighting energy-saving method based on multi-modal data fusion, including: S1. Periodically obtain a decision trajectory data set; wherein, the decision trajectory data set includes a state space, an action space, and a reward. The state space includes multi-modal environmental data such as light intensity, traffic flow, population density, weather information, and positioning information. The action space includes lighting actions such as brightness adjustment and switch state. The reward includes energy-saving effect, comfort, and safety; S2. Add the decision trajectory data set of the first cycle to the experience pool as a sample; in subsequent cycles, merge the experience pool of the previous cycle with the newly collected decision trajectory data set of the current cycle according to a set ratio, and screen according to the priority weight; S3. Use the samples in the experience pool for policy training, establish an objective function for the best candidate policy, and train the candidate policy according to the objective function. A safety constraint coefficient is introduced into the objective optimization function, and a PPO loss function is introduced during the training process to prevent the policy parameter update from being too large through probability ratio clipping; S4. Evaluate the candidate policy in the test scenario set to obtain multi-dimensional evaluation indicators including energy efficiency, safety, and user subjective comfort. According to the multi-dimensional evaluation indicators and the corresponding sub-item weights, obtain the comprehensive policy score; S5. In response to the comprehensive policy score meeting the admission conditions, add it to the policy library; when the policy capacity in the policy library reaches a predetermined value, based on the Pareto frontier, compare the scores of the candidate policies that meet the admission conditions with the existing policies in the policy library, and eliminate the policy with the lowest comprehensive policy score; S6. Construct a fitness function for the strategy according to the weights of the scores of each index in the multi-dimensional evaluation index; establish a feedback-weight mapping model for obtaining the predicted score based on each sub-item weight and the corresponding feedback data; establish a loss function gradient based on the feedback-weight mapping model, and perform iterative calculation of the sub-item weights based on the loss function gradient to obtain the sub-item weights after iterative calculation. Based on the sub-item weights after iterative calculation, recalculate the corresponding comprehensive score of the strategy according to the fitness function, and compare it with the score of the current strategy. If it is better than the score of the current strategy, dynamically update the sub-item weights of the current strategy.

[0013] In an embodiment of the present invention, in the subsequent cycle, the experience pool of the previous cycle and the decision trajectory data set newly collected in the current cycle are merged according to a set ratio, including: The merging is carried out by the following formula: , where, represents the data set collected in the first cycle; is the data set collected in the t-th cycle; a is the historical data retention rate, taking 0.7, indicating that 70% of the old data is retained and 30% of the new data is added; is the merging operation according to the priority; After the end of the t-th cycle, and is the merged experience pool.

[0014] In an embodiment of the present invention, the priority weight is expressed as: , where: is the trajectory length normalization factor, which normalizes trajectories of different lengths to ensure comparability between long trajectories and short trajectories; is the trajectory; is the trajectory length; k is the moment; is a constant used to adjust the weight and the decay rate; is the value prediction deviation, which measures the unexpected degree of the current value function prediction; is k the immediate reward obtained at the moment; is the discount factor, taking 0.99, which measures the importance of future rewards; is the policy value function, which represents the expected return that can be obtained starting from state s and following this policy; The average state feature of this week; is the state vector at time in the trajectory k .

[0015] In an implementation manner of the present invention, the objective function is expressed as: , wherein, , wherein, is the parameter when the parameter is , is the largest, and it is expected to obtain an optimal candidate policy, the parameter and the corresponding policy ; is an unknown parameter, which is substituted into all experience pools for training, and it is expected to find an optimal parameter; is the data set of the current experience pool storing historical trajectories; is the safety constraint coefficient; is the safety constraint term; is the difference between the maximum immediate reward and the minimum immediate reward; and are the immediate rewards at time k and time k - 1 respectively.

[0016] In an implementation manner of the present invention, the PPO loss function is expressed as: , wherein, , , wherein, is a constant; is a clipping function; a Python slicing function that slices the data within the range of in , and the data outside this range will be discarded; is the parameter of the policy; is the strategy when the parameter is ; is the strategy parameter of the last update; is the receiving state s output action a probability distribution; is the value prediction deviation; is the discount factor; is a constant.

[0017] In an embodiment of the present invention, for the candidate strategy, it is evaluated under a test scenario set to obtain multi-dimensional evaluation indicators including energy efficiency, safety, and user subjective comfort. According to the multi-dimensional evaluation indicators and the corresponding sub-item weights, a comprehensive strategy score is obtained, including: It is evaluated under the test scenario set C ={ rainy night, sunny day, foggy day, peak}; The comprehensive strategy score is expressed as: , where is the scenario weight, satisfying , and each scenario weight is given by the system according to the importance; is the set of sub-item weights for each scenario; is the sub-item weight of energy efficiency; is the sub-item weight of safety; is the sub-item weight of user subjective comfort; The energy efficiency is expressed as: , where is the total energy consumption under the scenario; c is a certain scenario in the test scenario set C ; is the total energy consumption of the baseline strategy, i.e., the maximum system power consumption, under the scenario; The safety is expressed as: , where is the number of control times under the scenario c ; is the scenario cThe number of times the lower brightness exceeds the safety threshold; The fewer the number of violations, the closer the safety score is to 1, that is, the best; The subjective comfort of the user is expressed as: , Among them, is the i th user rating in scenario c; the rating is from 1 to 5 points; is the number of N users who have rated in scenario c; reflects the average rating of users in scenario c and reflects the subjective comfort of users.

[0018] In an embodiment of the present invention, in response to the comprehensive policy score meeting the admission conditions, it is added to the policy library; when the policy capacity in the policy library reaches a predetermined value, based on the Pareto frontier, the scores of the candidate policies that meet the admission conditions are compared with the existing policies in the policy library, and the policy with the lowest comprehensive policy score is eliminated; including: The admission conditions are expressed as: , Among them, and are the mean score and standard deviation of the policy library respectively; is the set of policies ; If the admission conditions are met, add the policy: , Among them, is the policy when the parameter is ; Based on the Pareto frontier, eliminate the policy: , Among them, is the process of comparing the scores of the candidate policies that meet the admission conditions with the existing policies in the policy library and eliminating the policy with the lowest score; is the policy to be introduced; When it satisfies: , Among them, is the lowest score among the original policies. When the policy is lower than the lowest score, the policy is eliminated at this time.

[0019] In an embodiment of the present invention, the fitness function is expressed as: , , Among them, , ; is the user comfort index; is the total energy consumption when executing the current policy; is the minimum total energy consumption in the policy set; is the maximum total energy consumption in the policy set; is the number of violations of this policy; is the total number of violations; is the comfort score of the i-th person among N people when applying this policy.

[0020] In an embodiment of the present invention, based on each of the sub-item weights and the corresponding feedback data, a feedback-weight mapping model for obtaining a predicted score is established; based on the feedback-weight mapping model, a loss function gradient is established, and based on the loss function gradient, an iterative calculation of the sub-item weights is performed to obtain the sub-item weights after iterative calculation. Based on the sub-item weights after iterative calculation, according to the fitness function, the corresponding policy comprehensive score is recalculated and compared with the score of the current policy. If it is better than the score of the current policy, the sub-item weights of the current policy are dynamically updated, including: Establish a feedback-weight mapping model: , Among them, is the predicted score; is the sub-item weight data of this trajectory in the currently executed policy; is the energy efficiency sub-item weight; is the safety sub-item weight; is the user subjective comfort sub-item weight; is the feedback data; is the theoretical energy efficiency upper limit of this position period; is the safety risk coefficient based on the environment; is the comfort preference of the user group; Establish a loss function gradient: , wherein, , wherein, is the weight of each sub - index; is a constant; , indicates that this function is a gradient operator function; is the actual score feedback by the user; is the predicted score, calculated from a formula; is a constant, the minimum value of the sub - item weight set by the system; is the current sub - item weight.

[0021] The present invention also provides an intelligent lighting energy - saving system based on multi - modal data fusion, including: A data acquisition module, configured to periodically obtain a decision - making trajectory data set; wherein, the decision - making trajectory data set includes a state space, an action space, and a reward. The state space includes multi - modal environmental data such as light intensity, traffic flow, pedestrian density, weather information, and positioning information. The action space includes lighting actions such as brightness adjustment and switch state. The reward includes energy - saving effect, comfort, and safety; An experience pool update module, configured to add the decision - making trajectory data set of the first cycle to the experience pool as a sample; in subsequent cycles, merge the experience pool of the previous cycle with the newly acquired decision - making trajectory data set of the current cycle according to a set ratio, and screen according to the priority weight; A policy training module, configured to use the samples in the experience pool for policy training, establish an objective function for the best candidate policy, and train to obtain a candidate policy according to the objective function. A safety constraint coefficient is introduced into the objective optimization function, and a PPO loss function is introduced during the training process to prevent the policy parameter update from being too large through probability ratio clipping; A policy comprehensive score acquisition module, configured to evaluate the candidate policy in a test scenario set to obtain multi - dimensional evaluation indexes including energy efficiency, safety, and user subjective comfort, and obtain a policy comprehensive score according to the multi - dimensional evaluation indexes and the corresponding sub - item weights; A policy library update module, configured to add it to the policy library in response to the policy comprehensive score meeting the admission conditions; when the policy capacity in the policy library reaches a predetermined value, based on the Pareto front, compare the scores of the candidate policies meeting the admission conditions with the existing policies in the policy library, and eliminate the policy with the lowest policy comprehensive score; The sub-item weight update module is used to construct a fitness function of the strategy according to the weights of the scores of each index in the multi-dimensional evaluation index; establish a feedback-weight mapping model for obtaining predicted scores based on each sub-item weight and the corresponding feedback data; establish a loss function gradient based on the feedback-weight mapping model, and perform iterative calculation of the sub-item weights based on the loss function gradient to obtain the sub-item weights after iterative calculation. Based on the sub-item weights after iterative calculation, recalculate the comprehensive score of the corresponding strategy according to the fitness function, and compare it with the score of the current strategy. If it is better than the score of the current strategy, dynamically update the sub-item weights of the current strategy.

[0022] The above technical solution of the present invention has the following advantages compared with the prior art: The intelligent lighting energy-saving method and system based on multi-modal data fusion described in the present invention significantly improve the comprehensive perception ability of complex application scenarios by fusing multi-modal environmental data such as light intensity, traffic flow, pedestrian density, weather information, and positioning information; adopt a decision trajectory experience pool and a multi-dimensional priority weight screening mechanism to realize the dynamic optimization and adaptive update of the strategy library; introduce multi-objective constraints such as energy efficiency, comfort, and safety during the strategy optimization process to comprehensively balance various indicators. Innovatively introduce a safety constraint coefficient into the objective function and use the PPO loss function to prevent the drastic fluctuation of strategy parameters to ensure the safety and reliability of the lighting effect and the subjective comfort of users; introduce reinforcement learning and evolutionary algorithms, which can continuously evolve and iterate control strategies according to historical operation data and user feedback, so as to dynamically adapt to the changing environment and pedestrian conditions, and improve the long-term energy-saving efficiency and user experience of the lighting system. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to make the content of the present invention easier to be clearly understood, the present invention will be further described in detail below according to specific embodiments of the present invention in conjunction with the drawings.

[0024] Figure 1 It is a flowchart of an intelligent lighting energy-saving method based on multi-modal data fusion in Embodiment 1.

[0025] Figure 2 It is a schematic diagram of the architecture of an intelligent lighting energy-saving system based on multi-modal data fusion in Embodiment 2.

[0026] Figure 3 It is a schematic diagram of the communication of an intelligent lighting energy-saving system based on multi-modal data fusion in Embodiment 2. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it, but the embodiments cited do not limit the present invention.

[0028] Embodiment 1 Referring to Figure 1 As shown, this embodiment provides a smart lighting energy-saving method based on multi-modal data fusion, including the following steps: S1. Periodically obtain a decision trajectory data set; wherein, the decision trajectory data set includes a state space, an action space, and a reward. The state space includes multi-modal environmental data such as light intensity, traffic flow, pedestrian density, weather information, and positioning information. The action space includes lighting actions such as brightness adjustment and switch state. The reward includes energy-saving effect, comfort, and safety. The specific decision trajectory data set is represented as follows: , Where: is the decision trajectory data set (experience data set) collected in the t-th week; is the number of trajectories in the t-th period, which depends on the number of interactions between the agent and the environment in the t-th period; is a trajectory, identifying a complete sequence executed by the agent in the environment, starting from the state until the termination state ends; End identifier, at this time there is no and behind, and no further actions are executed; , s is the state, and S is the set of all states (light, traffic flow, pedestrian flow, weather, etc.); , normalized action space; A is the set of all actions; , reward function; identifying the reward function in the t-th period; are the state, action, and reward function in the 0-th period, respectively.

[0029] S2. Add the decision trajectory data set of the first period to the experience pool as a sample; in subsequent periods, merge the experience pool of the previous period with the newly collected decision trajectory data set of the current period according to a set ratio, and screen according to the priority weight.

[0030] Specifically, the merging is performed through the following formula: , Where, represents the data set collected in the first period; The data set collected for the t-th period; a is the historical data retention rate, taking 0.7, indicating that 70% of the old data is retained and 30% of the new data is added; is the merging operation according to priority; After the end of the t-th period, and the merged experience pool.

[0031] Specifically, the priority weight is expressed as: , where: The trajectory length normalization factor is used to normalize trajectories of different lengths to ensure comparability between long and short trajectories; is the trajectory; is the trajectory length of; k is the time; is a constant used to adjust the weight and the decay rate; is the value prediction deviation, measuring the unexpectedness of the current value function prediction; is k the immediate reward obtained at time; is the discount factor, taking 0.99, measuring the importance of future rewards; is the policy value function, indicating the expected return that can be obtained starting from the state s and following the policy under the policy π; This week's average state characteristics; is the trajectory in k the state vector at time.

[0032] S3. Use the samples in the experience pool for policy training, establish the objective function of the best candidate policy, and train the candidate policy according to the objective function. A safety constraint coefficient is introduced into the objective optimization function, and the PPO loss function is introduced during the training process to prevent the policy parameter update from being too large by probability ratio clipping.

[0033] Specifically, the objective function is expressed as: , Among them, , Among them, is the parameter at which is the largest. It is expected to obtain an optimal candidate strategy, and the parameter and the corresponding strategy ; is an unknown parameter. Substitute it into all experience pools for training, and expect to find an optimal parameter; is the data set of the current experience pool storing historical trajectories; is the safety constraint coefficient; is the safety constraint term; is the difference between the maximum immediate reward and the minimum immediate reward; and are the immediate rewards at time k and time k - 1 respectively.

[0034] Specifically, the PPO loss function is expressed as: , Among them, , , Among them, is a constant; is a clipping function; a Python slicing function that slices the data within the range of in range, and the data outside this range will be discarded; is the parameter of the strategy; is the strategy when the parameter is ; is the strategy parameter of the previous update; is for receiving the state s output action a probability distribution; is the value prediction deviation; is the discount factor; is a constant.

[0035] S4. Evaluate the candidate strategies under the test scenario set to obtain multi-dimensional evaluation indicators including energy efficiency, security, and user subjective comfort. According to the multi-dimensional evaluation indicators and the corresponding sub-item weights, obtain the comprehensive strategy score.

[0036] Evaluate under the test scenario set C ={ rainy night, sunny day, foggy day, peak}; The comprehensive strategy score is expressed as: , where is the scenario weight, satisfying , and the system assigns each scenario weight according to the importance; is the set of sub-item weights for each scenario; is the energy efficiency sub-item weight; is the security sub-item weight; is the user subjective comfort sub-item weight; The energy efficiency is expressed as: , where is the total energy consumption under the scenario; c is the test scenario set C in a certain scenario; is the total energy consumption of the baseline strategy, i.e., when the system power consumption is the largest, under the scenario; The security is expressed as: , where is the number of control times under the scenario c ; is the number of times the brightness exceeds the safety threshold under the scenario c ; The fewer the number of violations of , the closer the security score is to 1, that is, the best; The user subjective comfort is expressed as: , where is the i th user score under the scenario c; it is from 1 to 5 points; is the number of N users who have rated under the c scenario; reflects the average user score under the c scenario and reflects the user subjective comfort.

[0037] S5. In response to the comprehensive policy score meeting the admission criteria, add it to the policy library; when the policy capacity in the policy library reaches the predetermined value, based on the Pareto front, compare the scores of the candidate policies that meet the admission criteria with the existing policies in the policy library, and eliminate the policy with the lowest comprehensive policy score.

[0038] Specifically, the admission criteria are expressed as: , where , are the mean score and standard deviation of the policy library respectively; is the set of policies ; If the admission criteria are met, add the policy: , where is the policy with parameter ; Based on the Pareto front, eliminate the policy: , where is the process of comparing the scores of the candidate policies that meet the admission criteria with the existing policies in the policy library and eliminating the policy with the lowest score; is the policy to be introduced; When it satisfies: , where is the lowest score among the original policies. When the policy is lower than the lowest score, the policy is eliminated at this time.

[0039] S6. According to the weights of the scores of each index in the multi-dimensional evaluation index, construct the fitness function of the policy; based on each sub-item weight and the corresponding feedback data, establish a feedback-weight mapping model for obtaining the predicted score; based on the feedback-weight mapping model, establish the gradient of the loss function, and based on the gradient of the loss function, perform iterative calculation of the sub-item weights to obtain the sub-item weights after iterative calculation. Based on the sub-item weights after iterative calculation, according to the fitness function, recalculate the corresponding comprehensive policy score and compare it with the score of the current policy. If it is better than the score of the current policy, dynamically update the sub-item weights of the current policy.

[0040] Specifically, the fitness function is expressed as: , , Among them, , ; is the user comfort index; is the total energy consumption when executing the current policy; is the minimum total energy consumption in the policy set; is the maximum total energy consumption in the policy set; is the number of violations of this policy; is the total number of violations; is the comfort score of the i-th person among N people when applying this policy.

[0041] Establish a feedback-weight mapping model: , Among them, is the predicted score; is the item weight data of this trajectory in the currently executed policy; is the energy efficiency item weight; is the safety item weight; is the user subjective comfort item weight; is the feedback data; is the theoretical energy efficiency upper limit of this location period; is the safety risk coefficient based on the environment; is the comfort preference of the user group; Establish the loss function gradient: , Among them, , Among them, is the weight of each sub-index; is a constant; , indicates that this function is a gradient operator function; is the actual score feedback by the user; is the predicted score, calculated from the formula; is a constant, the minimum value of the sub-item weight set by the system; is the current sub-item weight.

[0042] Based on the sub-item weights after iterative calculation , through the fitness function , recalculate the corresponding comprehensive policy score, and compare it with the score of the current policy. If it is better than the score of the current policy, dynamically update the sub-item weights of the current policy.

[0043] Embodiment 2 Based on the same inventive concept, this embodiment provides a smart lighting energy-saving system based on multi-modal data fusion. The principle of solving problems is similar to that of the method for smart lighting energy-saving based on multi-modal data fusion, and the repeated parts will not be elaborated here.

[0044] This embodiment provides a smart lighting energy-saving system based on multi-modal data fusion, including: A data acquisition module for periodically obtaining a decision trajectory data set; wherein, the decision trajectory data set includes a state space, an action space, and a reward. The state space includes multi-modal environmental data such as light intensity, traffic flow, pedestrian density, weather information, and location information. The action space includes lighting actions such as brightness adjustment and switch state. The reward includes energy-saving effect, comfort, and safety; An experience pool update module for adding the decision trajectory data set of the first cycle to the experience pool as a sample; in subsequent cycles, merging the experience pool of the previous cycle with the newly acquired decision trajectory data set of the current cycle according to a set ratio, and screening according to the priority weight; A policy training module for using the samples in the experience pool for policy training, establishing an objective function for the best candidate policy, and training to obtain a candidate policy according to the objective function. A safety constraint coefficient is introduced into the objective optimization function, and a PPO loss function is introduced during the training process to prevent the policy parameter update from being too large through probability ratio clipping; A comprehensive policy score acquisition module for evaluating the candidate policy in a test scenario set to obtain multi-dimensional evaluation indicators including energy efficiency, safety, and user subjective comfort, and obtaining a comprehensive policy score according to the multi-dimensional evaluation indicators and the corresponding sub-item weights; A policy library update module for adding it to the policy library in response to the comprehensive policy score meeting the admission conditions; when the policy capacity in the policy library reaches a predetermined value, based on the Pareto frontier, compare the scores of the candidate policies meeting the admission conditions with the existing policies in the policy library, and eliminate the policy with the lowest comprehensive policy score; The sub-item weight update module is used to construct a fitness function of the strategy according to the weights of the scores of each index in the multi-dimensional evaluation index; establish a feedback-weight mapping model for obtaining the predicted score based on each sub-item weight and the corresponding feedback data; establish a loss function gradient based on the feedback-weight mapping model, and perform iterative calculation of the sub-item weights based on the loss function gradient to obtain the sub-item weights after iterative calculation. Based on the sub-item weights after iterative calculation, recalculate the comprehensive score of the corresponding strategy according to the fitness function, and compare it with the score of the current strategy. If it is better than the score of the current strategy, dynamically update the sub-item weights of the current strategy.

[0045] It should be noted that, as shown in Figure 2 Real-time collection and transmission of multi-modal environmental data are realized through the data collection and transmission module (data collection module) installed on the lamp post, specifically including: light sensor, human presence sensor, environmental sensor, motion sensor, camera, Beidou positioning module and XingFlash module. The functions of various sensors are as follows: Light sensor: used to collect ambient brightness data; Human presence sensor and motion sensor: used to detect the relative distance, traveling direction and speed of the human body and the lamp post; Camera: used for area detection (such as pedestrian flow, crowd density) and providing monitoring data; Environmental sensor: used to collect environmental parameters such as temperature, humidity, and weather; Beidou positioning module: used to achieve precise positioning of the lighting lamp post and provide data support for cluster division and regional collaboration; XingFlash module: realizes short-range automatic networking communication and multi-node data sharing of smart lamp posts.

[0046] The environmental information collected in real time by the above sensors is uploaded to the control layer through the data transmission module, and after data processing processes such as data cleaning, normalization, and feature extraction, it is stored in the cloud platform. The system automatically adjusts and optimizes the brightness and energy-saving modes of the functional modules of the lamp posts in the area by using the modules integrated in the control layer (experience pool update module to sub-item weight update module) based on the preprocessed multi-modal data, so as to achieve the best lighting comfort and the optimal energy consumption.

[0047] In addition, as shown in Figure 3As shown, the system can divide the control area into multiple clusters (such as roads, parks, commercial areas, etc.) according to the geographical environment and functional characteristics of urban streets. The intelligent lighting terminals within each cluster are connected to the cloud platform, and each intelligent lighting terminal realizes self-organizing network communication through the StarFlash module. Through the StarFlash self-organizing network and data sharing between modules, sensor data and control states are communicated with each other to jointly decide on lighting strategies. The system supports brightness adjustment, switch control, low-power mode control, etc. of lighting modules to achieve high-efficiency energy saving and intelligent management.

[0048] Referring to Figure 2 As shown, the data processing and control process includes: Data acquisition and preprocessing: Clean, normalize, aggregate, and extract features from multi-source raw data, and store the processed results in the management cloud platform to achieve unified management of historical data. In addition, the system also includes: Control layer: The control layer includes a central control service and a local control service (edge computing gateway). The central control service issues control strategies through the Internet of Things protocol, and the local control service dynamically adjusts the lighting state and device operation parameters according to the real-time sensor data using the method in Embodiment 1 to achieve local and global collaborative control and optimization.

[0049] User display layer: Provides a visual management interface for users, supporting functions such as device distribution, area division, device control, energy-saving status information, real-time energy consumption monitoring, etc., facilitating management and decision-making.

[0050] Specifically, the system calculates the human walking direction and speed in real time through motion sensors and human presence sensors, and synchronizes data with surrounding lamp poles through the StarFlash communication channel. The StarFlash technology is a wireless communication technology for short distances, high speed, low latency, and high reliability, supporting automatic discovery, automatic routing, and self-organizing network communication between terminals, greatly improving the response speed and data collaboration ability of the intelligent lighting system.

[0051] In actual operation, the system can input relevant data such as the real-time collected human walking speed, pedestrian flow, and environmental brightness into the strategy operation engine, calculate and issue the lamp pole brightness coefficient and lighting time. For example, when it detects someone approaching, the system calculates the predicted time T for the pedestrian to reach the next lighting node through the distance S and speed V arrival : T arrival = S / V, and realizes regional collaborative dimming through environmental data. The light sensor triggers the basic lighting threshold (such as automatically turning on when the ambient light is below 50 Lux), the camera automatically identifies the pedestrian flow density and grades the lighting using the YOLO algorithm, and the Beidou module realizes the precise cluster division of lighting nodes.

[0052] The system also supports spatio-temporal collaborative control, for example: Phase synchronization: The brightness of adjacent lamp posts changes gradually to avoid sudden changes in illumination (e.g., from 100% to 20% is completed in 3 seconds). Fault self-healing: When a certain node fails, the brightness compensation of adjacent nodes is automatically increased (compensation formula): ΔB = B 失效 / N 相邻节点， ΔB, the brightness compensation value, is the brightness that the surrounding adjacent nodes need to increase. B 失效 , the original brightness of the failed node, is the brightness when the faulty node is working normally; N 相邻节点， The number of effective adjacent nodes, which is the number of nodes that can perform brightness compensation for the failed node; Environment adaptability: In high-temperature and high-humidity environments, the brightness is reduced to 80% to reduce equipment heat generation; in rainy and foggy weather, the brightness is increased to 120% to enhance road safety.

[0053] Therefore, when any intelligent lighting terminal in the cluster loses its network connection with the platform, as long as the device's star flash module and the cluster's self-organizing network function are normal, it can upload data through the data upload channels of other terminals; when the environment changes, it will automatically notify the nearby (next) node to execute the strategy to form a linkage.

[0054] In terms of star flash self-organizing network communication, the link quality assessment (LQ) is calculated by weighted normalization based on the received signal strength (RSSI), bit error rate (BER), and throughput (TP): , Definition of the normalization function: , x is the received signal strength (RSSI), bit error rate (BER), or throughput (TP); are the minimum and maximum values of each parameter.

[0055] The routing selection score S comprehensively considers the link quality and the remaining energy of the node: , Among them, E is the remaining energy of the node, and the total energy consumption of the path is calculated as: , , is the total sum of the energy consumed by the path; , is the energy required for node i to send data; , is the energy required for node i to receive data; Under the adaptive channel selection model, the interference level is calculated as follows: , wherein, is the total interference of channel c; N is the set of neighbor nodes; is the interference contribution of neighbor node i on channel c; is the weight of node i , which is inversely proportional to the distance.

[0056] Optimal channel selection function: , wherein, is the set of nodes with the minimum total interference of adjacent link nodes providing brightness compensation, and is the set of optimal path nodes.

[0057] Through the above star flash ad hoc network model, the system can automatically select the best link quality, optimal routing and channels, realize efficient and reliable data transmission and policy coordination, and provide a solid communication and perception foundation for the intelligent lighting energy-saving system.

[0058] In summary, the system first divides the area into several clusters according to the geographical environment and functional characteristics. The lighting nodes (such as lamp posts) within each cluster realize ad hoc network communication through star flash modules, so as to realize the efficient sharing and synchronization of sensor data and control states. During the data acquisition and processing process, the system not only integrates multi-source information such as light intensity, human presence, movement direction and speed, and the pedestrian flow density obtained by the camera, but also combines the accurate position information provided by the Beidou positioning module to provide comprehensive and accurate data input for the subsequent evolutionary algorithm.

[0059] In terms of generating control strategies, the system adopts a multi-objective optimization method, comprehensively weighs factors such as energy consumption, user comfort and environmental impact, establishes an algorithm for optimal control strategies, and effectively avoids problems such as over-illumination and blind area illumination. Compared with traditional fixed lighting strategies, the system of the present invention can dynamically respond to changes in the environment and pedestrian flow, has a significant energy-saving effect, and the energy-saving rate can reach 45% to 80%.

[0060] In addition, the system introduces a predictive lighting mechanism. By analyzing the movement direction and speed of pedestrians, it predicts the time when pedestrians reach the next lighting node, and activates the corresponding lamp post in advance to achieve a seamless lighting experience. The system also has a fault self-healing ability. When a certain node fails, adjacent nodes will automatically increase the brightness for compensation to ensure the continuity of road lighting and driving safety.

[0061] The SparkLink wireless communication technology not only enables efficient networking communication among lighting nodes, but also supports functions such as automatic node discovery, intelligent routing selection, link quality assessment, and adaptive channel selection. This short-distance, high-speed, and low-latency wireless communication method significantly improves the data synchronization efficiency and system response speed within the cluster.

[0062] Meanwhile, the system supports the continuous collection and analysis of historical data and user feedback (such as ratings within the APP, etc.). The policy library will be automatically updated and optimized weekly based on the actual operation data and user experience, achieving the continuous evolution and adaptive iteration of control strategies, and continuously improving the overall energy efficiency and intelligence level of the system.

[0063] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. 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. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer-usable program code.

[0064] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in Figure 1 one or more flows and / or Figure 1 blocks.

[0065] These computer program instructions can 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, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that implement the functions specified in Figure 1 one or more flows and / or Figure 1 blocks.

[0066] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide means for implementing the functions in theFigure 1 one process or multiple processes and / or boxes Figure 1 steps of functions specified in one box or multiple boxes.

[0067] Finally, it should be noted that the above specific implementation manners are only used to illustrate the technical solutions of the present invention rather than to limit them. Although the present invention has been described in detail with reference to the examples, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A smart lighting energy-saving method based on multi-modal data fusion, characterized in that Including: S1. Periodically obtain a decision trajectory dataset; wherein, the decision trajectory dataset includes a state space, an action space, and rewards. The state space includes multi-modal environmental data such as light intensity, traffic flow, pedestrian density, weather information, and positioning information. The action space includes lighting actions such as brightness adjustment and switch states. The rewards include energy-saving effects, comfort, and safety; S2. Add the decision trajectory dataset of the first cycle to the experience pool as a sample; in subsequent cycles, merge the experience pool of the previous cycle with the newly collected decision trajectory dataset of the current cycle according to a set ratio, and screen according to priority weights; S3. Use the samples in the experience pool for policy training, establish an objective function for the best candidate policy, and train a candidate policy according to the objective function. A safety constraint coefficient is introduced into the objective optimization function, and a PPO loss function is introduced during the training process. Probability ratio clipping is used to prevent the policy parameter update from being too large; S4. Evaluate the candidate policy under a test scenario set to obtain multi-dimensional evaluation indicators including energy efficiency, safety, and user subjective comfort. According to the multi-dimensional evaluation indicators and the corresponding sub-item weights, obtain a comprehensive policy score; S5. In response to the comprehensive policy score meeting the admission conditions, add it to the policy library; when the policy capacity in the policy library reaches a predetermined value, based on the Pareto frontier, compare the scores of the candidate policies that meet the admission conditions with the existing policies in the policy library, and eliminate the policy with the lowest comprehensive policy score; S6. Construct a fitness function for the policy according to the weights of the scores of each indicator in the multi-dimensional evaluation indicators; establish a feedback-weight mapping model for obtaining a predicted score based on each sub-item weight and the corresponding feedback data; based on the feedback-weight mapping model, establish a loss function gradient, and based on the loss function gradient, perform iterative calculation of the sub-item weights to obtain the iteratively calculated sub-item weights. Based on the iteratively calculated sub-item weights, according to the fitness function, recalculate the corresponding comprehensive policy score, and compare it with the score of the current policy. If it is better than the score of the current policy, dynamically update the sub-item weights of the current policy.

2. The intelligent lighting energy-saving method based on multi-modal data fusion according to claim 1, wherein, In subsequent cycles, merge the experience pool of the previous cycle with the newly collected decision trajectory dataset of the current cycle according to a set ratio, including: Merge through the following formula: ， Among them, indicates that the dataset is collected in the first cycle; The data set collected for the t-th period; a is the historical data retention rate, taking 0.7, indicating that 70% of the old data is retained and 30% of new data is added; is an operation to merge by priority; After the end of the t-th cycle, and the merged experience pool.

3. A smart lighting energy-saving method based on multi-modal data fusion according to claim 1, characterized in that, The priority weight is expressed as: ， Wherein: A trajectory length normalization factor, which normalizes trajectories of different lengths to ensure comparability between long trajectories and short trajectories; is a locus; is the trajectory length; k is the time; is a constant used to adjust the weight and decay rate; is the value prediction deviation, which measures the unexpectedness of the prediction by the current value function; For k instant rewards obtained at all times; is the discount factor, taking 0.99, which measures the importance of future rewards; is the policy value function, representing the expected return that can be obtained starting from state s and following the policy under policy π; Average state characteristics for this week; is the trajectory in k the state vector at the moment.

4. A smart lighting energy-saving method based on multimodal data fusion according to claim 1, characterized in that The objective function is expressed as: ， Wherein, ， Among them, when the parameter is , is the largest, and it is expected to obtain an optimal candidate strategy, the parameter and the corresponding strategy ; is an unknown parameter, which is substituted into all experience pools for training, and it is expected to find an optimal parameter; A data set of the current experience pool storing historical trajectories; is the safety constraint coefficient; is a safety constraint item; is the difference between the maximum immediate reward and the minimum immediate reward; and are the immediate rewards at time k and time k-1 respectively.

5. A smart lighting energy-saving method based on multi-modal data fusion according to claim 1, characterized in that, The PPO loss function is expressed as: ， Wherein, ， ， Among them, is a constant; is a clipping function; a Python slicing function that slices in the data within the range, and the data outside this range will be discarded; is a parameter of the strategy; is the strategy when the parameter is ; are the policy parameters of the last update; For the reception state s Output action a Probability distribution; is the value prediction deviation; is the discount factor; is a constant.

6. The intelligent lighting energy-saving method based on multi-modal data fusion according to claim 1, characterized in that, Evaluate the candidate policy under a test scenario set to obtain multi-dimensional evaluation indicators including energy efficiency, safety, and user subjective comfort. According to the multi-dimensional evaluation indicators and the corresponding sub-item weights, obtain a comprehensive policy score, including: Evaluate on the test scenario set C ={ rainy night, sunny day, foggy day, peak period}; The comprehensive policy score is expressed as: ， Among them, is the scene weight, satisfying , and the system assigns weights to each scene according to the importance; is the set of sub-item weights under each scene; is the weight of the energy efficiency sub-item; is the weight of the safety sub-item; is the weight of the user's subjective comfort sub-item; Energy efficiency is expressed as: ， Among them, is the total energy consumption under the scenario; c For a test scenario set C in a certain scenario; In the scenario, the baseline strategy is the total energy consumption when the system power consumption is the largest; Safety is expressed as: ， Among them, is the number of controls c under the scenario; The number of times the brightness exceeds the safety threshold in the scenario c ; The fewer the number of violations, the closer the safety score is to 1, that is, the best; User subjective comfort is expressed as: ， Among them, is the i th user rating under scenario c; it ranges from 1 to 5 points; There are N users who have given ratings in scenario c; The average score of users in reaction c scenario, reflecting the subjective comfort of users.

7. A method for energy-saving of intelligent lighting based on multi-modal data fusion according to claim 6, characterized in that, In response to the satisfaction of the access condition by the comprehensive policy score, add it to the policy library; when the policy capacity in the policy library reaches a predetermined value, based on the Pareto front, compare the scores of the candidate policies that meet the access condition with the existing policies in the policy library, and eliminate the policy with the lowest comprehensive policy score; including: The access condition is expressed as: ， Among them, , are the mean score and standard deviation of the policy library respectively; For the policy of the set; If the access condition is satisfied, add the policy: ， Among them, is the strategy when the parameter is ; is a set of; Based on the Pareto front, eliminate the policy: ， Among them, is a process of comparing the scores of candidate policies for access conditions with the existing policies in the policy library and eliminating the policy with the lowest score. is the policy to be introduced; When the following is satisfied: ， Among them, is the lowest score in the original strategy. When the strategy is lower than the lowest score, the strategy is eliminated at this time.

8. A smart lighting energy-saving method based on multi-modal data fusion according to claim 1, characterized in that The fitness function is expressed as: ， ， Among them, , ; is the user comfort index; Total energy consumption when executing the current policy; is the minimum total energy consumption in the strategy set; is the maximum total energy consumption in the strategy set; This is the number of violations of this policy; is the total number of violations; The comfort score of the i-th person among N people when applying this strategy.

9. A smart lighting energy-saving method based on multi-modal data fusion according to claim 1, characterized in that, Based on each sub-item weight and the corresponding feedback data, establish a feedback-weight mapping model for obtaining the predicted score; based on the feedback-weight mapping model, establish the gradient of the loss function, based on the gradient of the loss function, perform iterative calculation of the sub-item weights to obtain the sub-item weights after iterative calculation, based on the sub-item weights after iterative calculation, according to the fitness function, recalculate the corresponding comprehensive policy score, and compare it with the score of the current policy. If it is better than the score of the current policy, dynamically update the sub-item weights of the current policy, including: Establish a feedback-weight mapping model: ， Among them, is the predicted score; It is the itemized weight data for this trajectory in the currently executed policy; is the weight of the energy efficiency sub-item; is the sub-item weight for safety; is the sub - weight of the user's subjective comfort; For feedback data; is the theoretical upper limit of energy efficiency for this position period; is the environment-based security risk coefficient; For the comfort preferences of the user group; Establish the gradient of the loss function: ， Wherein, ， Among them, is the weight of each sub-index; is a constant; , indicating that this function is a gradient operator function; is the actual score feedback by the user; It is a predicted score, calculated from a formula; is a constant, the minimum value of the sub-item weight set by the system; is the current sub-item weight.

10. A smart lighting energy-saving system based on multi-modal data fusion, characterized in that, Including: A data acquisition module for periodically acquiring a decision trajectory data set; wherein, the decision trajectory data set includes a state space, an action space, and a reward. The state space includes multi-modal environmental data such as light intensity, traffic flow, pedestrian density, weather information, and positioning information. The action space includes lighting actions such as brightness adjustment and switch state. The reward includes energy-saving effect, comfort, and safety; An experience pool update module for adding the decision trajectory data set of the first cycle to the experience pool as a sample; in subsequent cycles, merge the experience pool of the previous cycle with the newly acquired decision trajectory data set of the current cycle according to a set ratio, and screen according to the priority weight; A policy training module for using the samples in the experience pool to perform policy training, establish an objective function for the best candidate policy, and train the candidate policy according to the objective function. A safety constraint coefficient is introduced into the objective optimization function, and a PPO loss function is introduced during the training process to prevent the policy parameter update from being too large; A comprehensive policy score acquisition module for evaluating the candidate policy in a test scenario set to obtain multi-dimensional evaluation indicators including energy efficiency, safety, and user subjective comfort, and obtaining the comprehensive policy score according to the multi-dimensional evaluation indicators and the corresponding sub-item weights; A policy library update module for adding it to the policy library in response to the satisfaction of the access condition by the comprehensive policy score; when the policy capacity in the policy library reaches a predetermined value, based on the Pareto front, compare the scores of the candidate policies that meet the access condition with the existing policies in the policy library, and eliminate the policy with the lowest comprehensive policy score; The sub-item weight update module is used to construct a fitness function of the strategy according to the weights of the scores of each index in the multi-dimensional evaluation index; establish a feedback-weight mapping model for obtaining predicted scores based on each sub-item weight and the corresponding feedback data; establish a loss function gradient based on the feedback-weight mapping model, perform iterative calculation of the sub-item weights based on the loss function gradient to obtain the sub-item weights after iterative calculation, and based on the sub-item weights after iterative calculation, recalculate the corresponding comprehensive score of the strategy according to the fitness function and compare it with the score of the current strategy. If it is better than the score of the current strategy, dynamically update the sub-item weights of the current strategy.

Images