Modular building multi-module collaborative light effect configuration method and system

Abstract

The application provides a modular building multi-module cooperative light effect configuration method and system, comprising: constructing a digital twin model containing module space topological relationship, interface light transmission attribute and shielding relationship; based on the digital twin model, fusing personnel behavior recognition data, environmental light intensity distribution data and task scene label in each module, extracting spatial dependence relationship of light effect demand between modules through a graph neural network, and generating a cross-module light effect demand graph; according to the light effect demand graph, using a multi-agent reinforcement learning algorithm, cooperatively solving a linkage configuration scheme of light color temperature, illumination and dynamic change rhythm of each module; distributing the linkage configuration scheme to each module intelligent control terminal for execution, collecting sensor feedback data after execution, driving online iterative update of the digital twin model and the configuration strategy, and realizing a closed-loop self-optimization of perception-decision-execution. The application improves the overall consistency of space light effect and reduces the overall energy consumption of the system.

Description

Technical Field

[0001] This invention relates to the field of electrical control technology, specifically to a method and system for configuring multi-module collaborative lighting effects in modular buildings. Background Technology

[0002] With the development of industrialized and intelligent construction, modular buildings have been widely used in residential, commercial, and temporary facilities due to their advantages such as fast construction speed, controllable quality, and easy expansion. Modular buildings are typically assembled on-site from multiple prefabricated functional modules, with each module integrating water supply and drainage, electrical, and HVAC systems. However, existing modular building lighting systems (such as indoor and outdoor lighting, ambient lighting, and wayfinding signage) generally suffer from isolated control, lack of coordination, and high energy consumption. Summary of the Invention

[0003] Based on the above problems, this invention proposes a method and system for configuring multi-module collaborative lighting effects in modular buildings. Through this invention, dynamic collaboration and adaptive optimization of cross-module lighting environments are achieved, which not only improves the overall consistency of spatial lighting effects but also effectively reduces the overall energy consumption of the system and has the ability to dynamically adapt to changes in module combinations.

[0004] In view of this, one aspect of the present invention proposes a method for configuring multi-module collaborative lighting effects in modular buildings, comprising: Based on BIM data and real-time illumination parameters of each module collected by IoT sensors, a digital twin model is constructed, which includes the spatial topology of the modules, the light transmission properties of the interfaces, and the occlusion relationship, forming a global light environment perception base. Based on the digital twin model, the system integrates personnel behavior recognition data, ambient light intensity distribution data, and task scene labels within each module, and extracts the spatial dependency relationship of light effect requirements between modules through graph neural networks to generate a cross-module light effect requirement map. Based on the light effect demand map, with the optimization objectives of light field continuity and energy consumption balance between modules, a multi-agent reinforcement learning algorithm is adopted to collaboratively solve the linkage configuration scheme of light color temperature, illuminance and dynamic change rhythm of each module. The linkage configuration scheme is distributed to the intelligent control terminals of each module through edge computing nodes for execution. At the same time, sensor feedback data after execution is continuously collected to drive the online iterative update of the digital twin model and configuration strategy, thereby realizing closed-loop self-optimization of perception-decision-execution.

[0005] Optionally, the step of constructing a digital twin model containing the spatial topological relationships of the modules, the light transmission properties of the interfaces, and the occlusion relationships based on BIM data and real-time illumination parameters collected by IoT sensors to form a global light environment perception base includes: The BIM structural data of each module is analyzed, and the geometric boundaries, interface material transmittance, window-to-wall ratio, and the location and size of the shared interface between adjacent modules are extracted to form a basic dataset of static optical properties for each module. Using each module as a node and the shared interface between modules as directed edges, a spatial topology graph is constructed based on the extracted location and size of the shared interface. The edge attributes record the transmittance and orientation information of the corresponding interface, forming a module spatial topology graph that describes the light transmission path. Based on the geometric boundaries and orientations of each module in the module space topology diagram, and combined with the real-time solar altitude angle and azimuth angle, the shadow coverage area of ​​each module at the current moment is calculated by the shading of adjacent modules or building components, and a dynamic shading relationship matrix is ​​generated that updates over time. The IoT light sensor nodes deployed in each module are mapped to the corresponding module nodes in the module spatial topology map, and an association index is established between the sensor number, spatial coordinates and the module node to which it belongs, providing an addressing basis for the targeted fusion of subsequent real-time data. The system continuously receives real-time illumination data from each sensor node via the IoT data channel. Based on the historical baseline of the sensors, it performs anomaly detection on the current readings, eliminates abnormal values ​​caused by sensor malfunctions or signal interference, and outputs real-time illumination data streams from each module that have undergone quality verification. The static optical attribute dataset, the module space topology map, the dynamic occlusion relationship matrix, and the real-time illumination data stream are fused and assembled in multiple layers to construct a unified modular digital twin model of building lighting environment. The model is continuously and synchronously updated with the sensor sampling cycle as the driving beat to form a global lighting environment perception base.

[0006] Optionally, the step of generating a cross-module light effect demand map by integrating personnel behavior recognition data, ambient light intensity distribution data, and task scene labels within each module based on the digital twin model includes: Based on the fusion data of image sensors and human infrared sensors deployed in each module, the current behavior type of personnel in the module is identified in real time, and the set of personnel behavior tags for each module is output, including personnel density, activity intensity and main behavior type. Using the set of personnel behavior tags, the current ambient light intensity distribution data of each module read from the digital twin model, and the pre-configured module task scene tags as input, the original feature vector of light effect requirements of each module is formed by splicing them together in units of modules, which serves as the initial input of the nodes of the subsequent graph neural network; The original feature vector of the light effect requirement is loaded into the corresponding node of the module space topology graph. The directed edge structure of the topology graph carries the light transmission path relationship between modules, forming an initial light effect requirement feature graph carrying multimodal requirement features. The initial light effect demand feature map is input into a graph neural network. Through multiple rounds of neighborhood message passing and aggregation operations, the feature vector of each module node is gradually integrated with the demand state information of adjacent modules. The dependency relationship of light effect demand of each module in spatial topology is extracted, and the enhanced feature vector of each module is output after incorporating neighborhood dependency information. Based on the enhanced feature vector, the target color temperature, target illuminance and dynamic change rhythm of light effect required by each module at the current moment are predicted by the demand decoder, and the quantitative light effect demand prediction results of each module are generated. The quantized light effect demand prediction results are written back to the corresponding nodes of the module space topology graph, and the spatial dependency strength between modules is written to the corresponding directed edges to form a cross-module light effect demand graph that simultaneously carries node-level light effect demand and edge-level module dependency relationship.

[0007] Optionally, the step of collaboratively solving the linkage configuration scheme of light color temperature, illuminance, and dynamic change rhythm of each module based on the light effect demand map, with the optimization objectives of light field continuity and energy consumption balance among modules, and employing a multi-agent reinforcement learning algorithm, includes: An independent reinforcement learning agent is instantiated for each module. The quantized light effect demand prediction results of the corresponding nodes in the cross-module light effect demand map and the dependency relationship of the edge modules are used as the initial environmental observation inputs of each agent. The action space reference of each agent is initialized with the actual configuration state of the current light color temperature, illuminance and dynamic change rhythm of each module. At each decision step, each module agent reads the current light field state, real-time power consumption, and personnel behavior tags of its module from the digital twin model. At the same time, it reads the current light field state and dependency edge weights of the adjacent module nodes directly connected to its module from the cross-module light effect demand map. The current light field state, real-time power consumption, personnel behavior tags, current light field state, and dependency edge weights are concatenated into the local observation state vector of the agent for the current decision step, ensuring that each agent's decision-making simultaneously perceives the module's demand and cross-module dependencies. The action space of each module agent is defined as a discretized joint space for the selection of light color temperature adjustment, illuminance adjustment and dynamic change rhythm mode of that module. Each agent independently samples the configuration adjustment action of the current decision step of its module based on the local observation state vector through the policy network with shared parameters. The shared parameter mechanism enables each agent to implicitly share cross-module policy experience when making decisions. After each agent performs the sampling action, the instantaneous reward obtained by each module agent in this decision step is calculated according to the preset cooperative reward function. The reward signal simultaneously reflects the light efficiency satisfaction of this module, the light field continuity deviation with adjacent modules, the real-time power consumption of this module, and the stability of configuration switching, ensuring that the local optimization goals of each agent are consistent with the global cooperative goals. The local observation state vectors, sampling actions, and immediate rewards of each agent are stored in a centralized experience replay buffer. A centralized training and distributed execution framework is used to periodically sample batch data from the buffer to update the parameters of the shared policy network and the independent value networks of each module. The absolute value of the difference between adjacent windows of the cumulative sliding mean of all agents within M consecutive decision steps is continuously lower than the preset convergence judgment threshold as the convergence judgment condition. After triggering convergence, the current policy network parameters are frozen. Here, M is the sliding window length, and both M and the convergence judgment threshold are hyperparameters that are preset before training according to the building module size and configuration accuracy requirements. The converged policy network performs a complete forward inference on the current cross-module light effect demand map, outputs the deterministic optimal action of each module agent under the current environmental observation, and assembles the color temperature target value, illuminance target value and dynamic change rhythm mode of each module into a structured linkage configuration scheme.

[0008] Optionally, the step of distributing the linkage configuration scheme to the intelligent control terminals of each module through edge computing nodes for execution, while continuously collecting sensor feedback data after execution, driving the online iterative update of the digital twin model and configuration strategy, and realizing closed-loop self-optimization of perception-decision-execution, includes: The linkage configuration scheme is divided according to module affiliation. The edge computing node compiles the color temperature target value, illuminance target value and dynamic change rhythm mode of each module into an instruction format that can be directly parsed by the corresponding intelligent control terminal. The instruction is then sent to the intelligent control terminal of each module in parallel through the local IoT channel. The edge computing node also records the timestamp of the instruction sent by each module as a timing benchmark for subsequent execution feedback comparison. After receiving the configuration instructions from the edge computing node, the intelligent control terminal of each module drives the lighting hardware of this module to complete the configuration switching according to the target color temperature, illuminance and rhythm mode. After the execution is completed, the actual execution color temperature, illuminance and switching completion timestamp of this module are reported to the edge computing node. After the edge computing node completes the configuration switch of each module, it triggers the real-time illumination data acquisition process, collects the actual illumination state data of each module sensor after the new configuration is stable, compares it with the color temperature target value and the illuminance target value module by module, calculates the execution feedback error of each module, and identifies the configuration deviation module caused by hardware response deviation, sensor drift or environmental change. The synchronous update mechanism of inputting the actual illumination state data of each module sensor and the execution feedback error into the digital twin model is used to perform online correction on the interface light transmission attribute estimation and occlusion relationship matrix that deviate from the measured values ​​in the digital twin model, so that the light environment mapping accuracy of the digital twin model continues to improve with the accumulation of execution feedback. The execution feedback error is used to calculate the current iteration learning rate according to the adaptive learning rate adjustment mechanism. This learning rate drives the online parameter update of the shared policy network and the independent value network of each module, so that the configuration policy is continuously adaptively corrected in the direction of reducing execution error and improving cross-module light field continuity. After updating the online parameters of the digital twin model, the shared policy network, and the independent value network of each module, the edge computing node determines whether the current execution feedback error of each module exceeds the preset re-optimization trigger threshold. If the threshold is exceeded, the updated digital twin model will be used as a new starting point to re-trigger the cross-module light effect demand map generation process and the linkage configuration scheme solution process, and start the next round of perception-decision-execution cycle. If the threshold is not exceeded, the current configuration is maintained and sensor data continues to be monitored, thus achieving closed-loop self-optimization driven on demand.

[0009] Optionally, when constructing the digital twin model, the following module's light environment state fusion function is used to dynamically fuse IoT sensor data and BIM static attributes:

[0010] in, Let m be the light environment fusion state vector of the m-th module at time t; The static light transmittance attribute vector of the m-th module extracted from the BIM model includes window-to-wall ratio and material transmittance. The set of sensors deployed for the m-th module; The measured light intensity value collected by sensor s at time t; The spatial location weighting coefficient of sensor s is determined by its proportion of the coverage area within the module. The dynamic balance coefficient between BIM static priors and real-time sensor data is adaptively adjusted according to the rate of change of natural daylight. The natural lighting dynamic correction matrix for the m-th module at time t is calculated by combining outdoor meteorological light intensity and solar altitude angle. It reflects the real-time correction of the static light transmission properties of BIM by building orientation and shading, so that the static BIM parameters have the ability to adapt to time-varying changes. The reliability coefficient of sensor s at time t is obtained by jointly evaluating the sensor's historical error rate and current self-test status. The readings of faulty or drifting sensors are automatically downweighted to prevent abnormal sensor data from contaminating the fusion results. Let be the optical environment temporal inertia coefficient of the m-th module, reflecting the physical hysteresis characteristics of the module's spatial heat capacity and optical field changes; This is the time difference term of the light environment fusion state vector from the previous moment, i.e. It captures the historical trend of light environment changes and together with the inertia coefficient, it constitutes the time-series prediction correction term, making the model's prediction of fast light-changing scenes smoother and more accurate.

[0011] Optionally, when extracting the spatial dependency relationship of light effect demand between modules through graph neural networks, the edge weights of the dependency strength between modules are calculated using the following spatial light effect coupling coefficient:

[0012] in, The coupling weights for the light efficiency requirements between module m and its adjacent module n; , These are the hidden layer feature vectors of the graph neural network for modules m and n, respectively, which encode the light demand state and personnel behavior characteristics of each module; This is a learnable edge attention weight matrix; Let m be the set of directly adjacent modules in the spatial topology graph; The physical transmittance of the interface between modules m and n is provided by the interface transmittance attribute parameters in the digital twin model and is used to embed physical optical constraints into the graph structure learning process. The dynamic light flux transfer bias term between module m and module n at time t is obtained by normalizing the product of the measured illuminance difference at the interface of the two modules and the interface area. The real-time physical light transfer is injected into the attention score calculation so that the edge weights are dynamically adjusted with the actual light field changes. The Sigmoid activation function compresses the linear scores within the parentheses to the (0,1) interval, serving as a global gating factor for the edge weights. For learnable side-gated projection vectors; Let m be the feature vector, n be the feature vector, and n be the feature vector of the module pair relationship. splicing; The spatial relationship feature vector between module m and module n encodes the relative position, interface orientation angle and functional partition label of the two modules, introduces structured spatial priors for edge weights, and avoids spatial blind spots caused by relying solely on node features.

[0013] Optionally, in a multi-agent reinforcement learning algorithm, the collaborative reward function of each module agent is defined as:

[0014] in, Let m be the instantaneous reward obtained by the agent corresponding to module m in a single-step decision-making process. , These are the light field state vectors currently configured for module m and its adjacent module n, respectively, which include color temperature and illuminance components; The Euclidean distance between adjacent modules is used to quantize the optical field continuity deviation at the module interface; This represents the real-time power consumption of module m at the current moment. This represents the current actual light effect status of module m. The target luminous efficacy state derived from the luminous efficacy demand map; The cosine similarity between the actual and target light effect distributions is calculated as the light effect satisfaction evaluation function. The weighting coefficients for each optimization objective are adaptively adjusted based on the building usage scenario labels; The interface importance weight between module m and adjacent module n is determined by the frequency of personnel line-of-sight crossing of the interface between the two modules and the interface area. A higher light field continuity penalty is given to the key interface that is frequently crossed by personnel, focusing on optimizing the module boundary that has the greatest impact on visual experience. The energy consumption dynamic discount factor of module m at time t is taken to be close to 1 during peak building electricity consumption to strengthen energy-saving penalties, and appropriately relaxed during off-peak periods to achieve time-series coordination with grid load. Weighting coefficients for rewards that improve visual comfort; The visual comfort increment of module m at time t relative to the previous time is calculated as a weighted change in uniform glare value and illuminance uniformity, providing a positive incentive for decisions to improve the configuration scheme towards greater comfort. To configure the weighting coefficient for frequent switching penalties The configuration switching cost of module m at time t is measured by the difference in the light effect configuration vector between the current time and the previous time. This suppresses high-frequency jittering behavior of the agent and ensures the stability of the light effect configuration and the visual comfort of the user.

[0015] Optionally, during the online iterative update process, the following adaptive learning rate adjustment mechanism is used to control the update step size of the digital twin model and configuration strategy:

[0016] in, Let be the adaptive iterative learning rate at time t; The initial baseline learning rate; The mean value of the sensor feedback error of each module within the sliding window at time t reflects the overall deviation between the current configuration scheme and the actual light environment. It is the error amplitude attenuation coefficient, which controls the sensitivity of the learning rate as the overall error increases, and prevents policy oscillation during the large error phase. This is the gain coefficient for the rate of change of error; The derivative of the mean error with respect to time represents the dynamic trend of the light environment deviation. When the deviation increases rapidly, the update step size is automatically suppressed to maintain system stability. The periodic time modulation factor is defined as follows: ,in For the daily cycle length of construction, The diurnal modulation amplitude coefficient is used to embed the prior diurnal regularity of the building's lighting environment into the learning rate adjustment. During the morning and evening periods when the lighting environment changes frequently, the learning rate is appropriately increased to accelerate adaptation, while the learning rate is reduced during the steady-state phase at night to reduce ineffective updates. The error-aware activation threshold controls the activation level when the learning rate is extremely small. For numerically stable smoothing terms, to prevent When the denominator approaches zero, it becomes singular. As an error-adaptive activation term, when the overall feedback error is extremely small, this term approaches zero, automatically suppressing the learning rate to near the point of stopping updates, thus avoiding invalid perturbations after convergence; when the error is large, this term approaches 1, and the learning rate returns to the normal adjustment range, realizing the error-driven wake-up mechanism.

[0017] Another aspect of the present invention provides a modular building multi-module collaborative lighting effect configuration system for executing a modular building multi-module collaborative lighting effect configuration method, comprising: a cloud server and an intelligent control terminal; The cloud server is configured as follows: Based on BIM data and real-time illumination parameters of each module collected by IoT sensors, a digital twin model is constructed, which includes the spatial topology of the modules, the light transmission properties of the interfaces, and the occlusion relationship, forming a global light environment perception base. Based on the digital twin model, the system integrates personnel behavior recognition data, ambient light intensity distribution data, and task scene labels within each module, and extracts the spatial dependency relationship of light effect requirements between modules through graph neural networks to generate a cross-module light effect requirement map. Based on the light effect demand map, with the optimization objectives of light field continuity and energy consumption balance between modules, a multi-agent reinforcement learning algorithm is adopted to collaboratively solve the linkage configuration scheme of light color temperature, illuminance and dynamic change rhythm of each module. The linkage configuration scheme is distributed to the intelligent control terminals of each module through edge computing nodes for execution. At the same time, sensor feedback data after execution is continuously collected to drive the online iterative update of the digital twin model and configuration strategy, thereby realizing closed-loop self-optimization of perception-decision-execution.

[0018] The technical solution of this invention provides a modular building multi-module collaborative lighting configuration method, comprising: constructing a digital twin model based on BIM data and real-time illumination parameters of each module collected by IoT sensors, including the spatial topology of the modules, the light transmission properties of the interfaces, and the occlusion relationships, forming a global lighting environment perception base; based on the digital twin model, integrating personnel behavior recognition data, ambient light intensity distribution data, and task scene labels within each module, extracting the spatial dependency relationship of lighting effect requirements between modules through a graph neural network, and generating a cross-module lighting effect requirement map; according to the lighting effect requirement map, taking the continuity of the light field and the balance of energy consumption between modules as optimization objectives, employing a multi-agent reinforcement learning algorithm to collaboratively solve the linkage configuration scheme of the light color temperature, illuminance, and dynamic change rhythm of each module's lights; distributing the linkage configuration scheme to the intelligent control terminals of each module through edge computing nodes for execution, while continuously collecting sensor feedback data after execution, driving the online iterative update of the digital twin model and configuration strategy, and realizing a closed-loop self-optimization of perception-decision-execution. By deeply coupling digital twins with multi-agent collaborative decision-making, the problem of fragmented light fields between modules and inconsistent overall visual experience caused by the isolated and independent control of light effects in modular buildings is solved. Dynamic collaboration and adaptive optimization of cross-module light environment are realized, which not only improves the overall consistency of spatial light effect, but also effectively reduces the overall energy consumption of the system and has the ability to dynamically adapt to changes in module combination. Attached Figure Description

[0019] Figure 1 This is a flowchart of a modular building multi-module collaborative lighting effect configuration method provided in one embodiment of the present invention; Figure 2 This is a schematic block diagram of a modular building multi-module collaborative lighting effect configuration system provided in one embodiment of the present invention. Detailed Implementation

[0020] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0022] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0023] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0024] The following reference Figures 1 to 2 This invention describes a modular building multi-module collaborative lighting effect configuration method and system provided by some embodiments of the present invention.

[0025] like Figure 1 As shown, one embodiment of the present invention provides a method for configuring multi-module collaborative lighting effects in modular buildings, including: Based on BIM data and real-time illumination parameters of each module collected by IoT sensors, a digital twin model is constructed, which includes the spatial topology of the modules, the light transmission properties of the interfaces, and the occlusion relationship, forming a global light environment perception base. Based on the digital twin model, the system integrates personnel behavior recognition data, ambient light intensity distribution data, and task scene labels within each module, and extracts the spatial dependency relationship of light effect requirements between modules through graph neural networks to generate a cross-module light effect requirement map. Based on the light effect demand map, with the optimization objectives of light field continuity and energy consumption balance between modules, a multi-agent reinforcement learning algorithm is adopted to collaboratively solve the linkage configuration scheme of light color temperature, illuminance and dynamic change rhythm of each module. The linkage configuration scheme is distributed to the intelligent control terminals of each module through edge computing nodes for execution. At the same time, sensor feedback data after execution is continuously collected to drive the online iterative update of the digital twin model and configuration strategy, thereby realizing closed-loop self-optimization of perception-decision-execution.

[0026] The technical solution adopted in this embodiment solves the problem of fragmented light fields between modules and inconsistent overall visual experience caused by the isolated independent control of light effects of each module in modular buildings through the deep coupling of digital twin and multi-agent collaborative decision-making. It realizes dynamic collaboration and adaptive optimization of cross-module light environment, effectively reduces the overall energy consumption of the system while improving the overall consistency of spatial light effect, and has the ability to dynamically adapt to changes in module combination.

[0027] In some possible embodiments of the present invention, the step of constructing a digital twin model containing the spatial topological relationships of the modules, the light transmission properties of the interfaces, and the occlusion relationships based on BIM data and real-time illumination parameters of each module collected by IoT sensors, to form a global light environment perception base, includes: The BIM structural data of each module is analyzed, and the geometric boundaries, interface material transmittance, window-to-wall ratio, and the location and size of the shared interface between adjacent modules are extracted to form a basic dataset of static optical properties for each module. Using each module as a node and the shared interface between modules as directed edges, a spatial topology graph is constructed based on the extracted location and size of the shared interface. The edge attributes record the transmittance and orientation information of the corresponding interface, forming a module spatial topology graph that describes the light transmission path. Based on the geometric boundaries and orientations of each module in the module space topology diagram, and combined with the real-time solar altitude angle and azimuth angle, the shadow coverage area of ​​each module at the current moment is calculated by the shading of adjacent modules or building components, and a dynamic shading relationship matrix is ​​generated that updates over time. The IoT light sensor nodes deployed in each module are mapped to the corresponding module nodes in the module spatial topology map, and an association index is established between the sensor number, spatial coordinates and the module node to which it belongs, providing an addressing basis for the targeted fusion of subsequent real-time data. The system continuously receives real-time illumination data from each sensor node via the IoT data channel. Based on the historical baseline of the sensors, it performs anomaly detection on the current readings, eliminates abnormal values ​​caused by sensor malfunctions or signal interference, and outputs real-time illumination data streams from each module that have undergone quality verification. The static optical attribute dataset, the module space topology map, the dynamic occlusion relationship matrix, and the real-time illumination data stream are fused and assembled in multiple layers to construct a unified modular digital twin model of building lighting environment. The model is continuously and synchronously updated with the sensor sampling cycle as the driving beat to form a global lighting environment perception base.

[0028] In this embodiment, by progressively associating and ultimately unifying BIM static structure analysis, module topology mapping, dynamic occlusion calculation, sensor node mapping, and real-time data quality verification, the constructed digital twin model possesses both the static accuracy of the building structure and the real-time dynamics of the light environment. This effectively solves the problems in traditional light control systems, such as static configuration failing to respond to occlusion changes and abnormal sensor data directly contaminating control decisions. It provides a reliable and complete global perception foundation for subsequent cross-module light effect collaborative decision-making.

[0029] In some possible embodiments of the present invention, the step of generating a cross-module light effect demand map by fusing personnel behavior recognition data, ambient light intensity distribution data, and task scene labels within each module based on the digital twin model includes: Based on the fusion data of image sensors and human infrared sensors deployed in each module, the current behavior type of personnel in the module is identified in real time, and the set of personnel behavior tags for each module is output, including personnel density, activity intensity and main behavior type. Using the set of personnel behavior tags, the current ambient light intensity distribution data of each module read from the digital twin model, and the pre-configured module task scene tags (including scene types such as meeting, office, and leisure) as input, the original feature vector of light effect requirements of each module is formed by splicing them together by module, which serves as the initial input of the nodes of the subsequent graph neural network. The original feature vector of the light effect requirement is loaded into the corresponding node of the module space topology graph. The directed edge structure of the topology graph carries the light transmission path relationship between modules, forming an initial light effect requirement feature graph carrying multimodal requirement features. The initial light effect demand feature map is input into a graph neural network. Through multiple rounds of neighborhood message passing and aggregation operations, the feature vector of each module node is gradually integrated with the demand state information of adjacent modules. The dependency relationship of light effect demand of each module in spatial topology is extracted, and the enhanced feature vector of each module is output after incorporating neighborhood dependency information. Based on the enhanced feature vector, the target color temperature, target illuminance and dynamic change rhythm of light effect required by each module at the current moment are predicted by the demand decoder, and the quantitative light effect demand prediction results of each module are generated. The quantified light effect demand prediction results are written back to the corresponding nodes of the module space topology graph. At the same time, the spatial dependency strength between modules is written to the corresponding directed edges, forming a cross-module light effect demand graph that carries both node-level light effect demand and edge-level module dependency relationship, which can be directly called in the subsequent collaborative configuration scheme generation steps.

[0030] Understandably, the specific implementation steps of this process include: writing the predicted results of the target color temperature, target illuminance, and dynamic change rhythm of light effect for each module into the corresponding node attribute fields of the module spatial topology graph, one by one according to the module number index. After writing, each node attribute field retains its original personnel behavior label and task scenario label, forming a node-level multi-attribute light effect requirement record; extracting the attention weights corresponding to each directed edge from the final message passing round of the graph neural network as the spatial dependency strength value between the two modules connected by that edge. This dependency strength value reflects the degree of mutual influence of the light effect requirements of adjacent modules; and writing the extracted spatial dependency strength values ​​of each directed edge according to the source module and target module number indexes of the edge. The dependency strength attribute field of the directed edge corresponding to the module spatial topology graph, together with the original interface transmittance and orientation attributes of the edge, constitutes a complete edge-level attribute record. All nodes and directed edges of the module spatial topology graph are traversed to verify whether each node has completely written the target color temperature, target illuminance, and rhythm prediction results, and whether each directed edge has completely written the spatial dependency strength value. For nodes or edges with missing attributes, the corresponding module's demand prediction is recalculated to ensure that the graph data completely covers all modules. The verified module spatial topology graph, along with the current timestamp, is encapsulated to form a versioned cross-module light effect demand graph data object, which can be directly called by the multi-agent reinforcement learning collaborative configuration steps. Through the hierarchical writing and integrity verification of node demand attributes and edge-level dependency strength, the cross-module light effect demand graph data is ensured to be complete and structurally sound, providing reliable input for subsequent collaborative configuration decisions.

[0031] In this embodiment, by uniformly expressing the multimodal features of personnel behavior recognition, ambient light intensity perception, and task scene labeling on a graph structure, and by explicitly modeling the spatial light effect dependencies between modules using the neighborhood message passing mechanism of graph neural networks, the generated cross-module light effect demand map can perceive the impact of the demand status of adjacent modules on the configuration of this module compared to the independent demand analysis scheme of each module. This avoids demand conflicts and light field fragmentation caused by ignoring cross-module dependencies, and provides high-quality demand input with complete spatial structure and multimodal information fusion for subsequent multi-agent collaborative configuration decisions.

[0032] In some possible embodiments of the present invention, the step of using a multi-agent reinforcement learning algorithm to collaboratively solve the linkage configuration scheme of light color temperature, illuminance, and dynamic change rhythm of each module, based on the light effect demand map and with the optimization objectives of light field continuity and energy consumption balance among modules, includes: An independent reinforcement learning agent is instantiated for each module. The quantized light effect demand prediction results of the corresponding nodes in the cross-module light effect demand map and the dependency relationship of the edge modules are used as the initial environmental observation inputs of each agent. The action space reference of each agent is initialized with the actual configuration state of the current light color temperature, illuminance and dynamic change rhythm of each module. At each decision step, each module agent reads the current light field state, real-time power consumption, and personnel behavior tags of its module from the digital twin model. At the same time, it reads the current light field state and dependency edge weights of the adjacent module nodes directly connected to its module from the cross-module light effect demand map. The current light field state, real-time power consumption, personnel behavior tags, current light field state, and dependency edge weights are concatenated into the local observation state vector of the agent for the current decision step, ensuring that each agent's decision-making simultaneously perceives the module's demand and cross-module dependencies. The action space of each module agent is defined as a discretized joint space for the selection of light color temperature adjustment, illuminance adjustment and dynamic change rhythm mode of that module. Each agent independently samples the configuration adjustment action of the current decision step of its module based on the local observation state vector through the policy network with shared parameters. The shared parameter mechanism enables each agent to implicitly share cross-module policy experience when making decisions. It is understood that in this embodiment, the policy networks of each module agent share the same set of neural network weight parameters, rather than each module maintaining its own independent set of parameters. During each decision step, each module agent inputs its local observed state vector into this shared policy network. The network independently infers forward based on its different inputs, outputting the probability distribution of its respective module's configuration adjustment actions. Then, it samples this distribution to obtain the specific configuration adjustment action for the current decision step. The essence of this mechanism is "shared parameters, different inputs, and independent outputs." The decision-making processes of each agent do not interfere with each other, but the underlying policy function is the same. This allows the policy experience accumulated by any module to synchronously strengthen the decision-making capabilities of all modules through gradient backpropagation.

[0033] After each agent performs the sampling action, the instantaneous reward obtained by each module agent in this decision step is calculated according to the preset cooperative reward function. The reward signal simultaneously reflects the light efficiency satisfaction of this module, the light field continuity deviation with adjacent modules, the real-time power consumption of this module, and the stability of configuration switching, ensuring that the local optimization goals of each agent are consistent with the global cooperative goals. Understandably, the core idea of ​​the collaborative reward function is to simultaneously associate the immediate reward of a single module agent with signals from five dimensions: cross-module light field continuity penalty, which calculates the weighted distance between the current light field state vectors of the current module and each adjacent module. The higher the frequency of personnel crossing adjacent modules, the greater the penalty weight, and the greater the difference in light field, the more points are deducted; real-time power consumption penalty of the current module, which applies a negative reward to the current power consumption of the current module and multiplies it by a discount factor that dynamically changes with the peak and valley periods of the power grid, strengthening energy-saving constraints during peak electricity consumption periods; light efficiency demand satisfaction reward, which calculates the cosine similarity between the actual light efficiency state of the current module and the target light efficiency state in the cross-module light efficiency demand map. The higher the similarity, the greater the positive reward; visual comfort improvement reward, which provides a positive incentive for the improvement of the glare value and illuminance uniformity of the current module compared to the previous decision step; and configuration switching jitter penalty, which applies a negative penalty to the difference between the current and previous configuration vectors of the current module, suppressing high-frequency invalid switching. The five signals are summed by their respective weight coefficients, which are adaptively adjusted according to the current building usage scenario label, so that the optimization focus of the reward function dynamically shifts with the scenario.

[0034] The local observation state vectors, sampling actions, and immediate rewards of each agent are stored in a centralized experience replay buffer. A centralized training and distributed execution framework is used to periodically sample batch data from the buffer to update the parameters of the shared policy network and the independent value networks of each module. The absolute value of the difference between adjacent windows of the cumulative sliding mean of all agents within M consecutive decision steps is continuously lower than the preset convergence judgment threshold as the convergence judgment condition. After triggering convergence, the current policy network parameters are frozen. Here, M is the sliding window length, and both M and the convergence judgment threshold are hyperparameters that are preset before training according to the building module size and configuration accuracy requirements. The converged policy network performs a complete forward inference on the current cross-module light effect demand map, outputs the deterministic optimal action of each module agent under the current environmental observation, and assembles the color temperature target value, illuminance target value and dynamic change rhythm mode of each module into a structured linkage configuration scheme for subsequent configuration distribution steps to call.

[0035] In this embodiment, by directly embedding the node requirements and edge dependencies in the cross-module light effect requirement map into the local observation state of each agent, and coordinating the joint optimization of multi-agent strategies with a centralized training and distributed execution framework, the solved linkage configuration scheme can meet the independent light effect requirements of each module, explicitly constrain the continuity of the light field at the interface of adjacent modules, and take into account the global energy consumption balance. This overcomes the problems of interface light field abrupt change and local energy consumption overload caused by neglecting the coupling constraints between modules in the traditional module-by-module independent optimization scheme. Moreover, the shared parameter mechanism enables the policy network to have good scalability and adaptability when the number of modules increases.

[0036] In some possible embodiments of the present invention, the step of distributing the linkage configuration scheme to the intelligent control terminals of each module through edge computing nodes for execution, while continuously collecting sensor feedback data after execution, driving the online iterative update of the digital twin model and configuration strategy, and realizing the closed-loop self-optimization of perception-decision-execution includes: The linkage configuration scheme is divided according to module affiliation. The edge computing node compiles the color temperature target value, illuminance target value and dynamic change rhythm mode of each module into an instruction format that can be directly parsed by the corresponding intelligent control terminal. The instruction is then sent to the intelligent control terminal of each module in parallel through the local IoT channel. The edge computing node also records the timestamp of the instruction sent by each module as a timing benchmark for subsequent execution feedback comparison. After receiving the configuration instructions from the edge computing node, the intelligent control terminal of each module drives the lighting hardware of this module to complete the configuration switching according to the target color temperature, illuminance and rhythm mode. After the execution is completed, the actual execution color temperature, illuminance and switching completion timestamp of this module are reported to the edge computing node for subsequent feedback error calculation. After the edge computing node completes the configuration switch of each module, it triggers the real-time illumination data acquisition process, collects the actual illumination state data of each module sensor after the new configuration is stable, compares it with the color temperature target value and the illuminance target value module by module, calculates the execution feedback error of each module, and identifies the configuration deviation module caused by hardware response deviation, sensor drift or environmental change. The synchronous update mechanism of inputting the actual illumination state data of each module sensor and the execution feedback error into the digital twin model is used to perform online correction on the interface light transmission attribute estimation and occlusion relationship matrix that deviate from the measured values ​​in the digital twin model, so that the light environment mapping accuracy of the digital twin model continues to improve with the accumulation of execution feedback. The execution feedback error is used to calculate the current iteration learning rate according to the adaptive learning rate adjustment mechanism. This learning rate drives the online parameter update of the shared policy network and the independent value network of each module, so that the configuration policy is continuously adaptively corrected in the direction of reducing execution error and improving cross-module light field continuity. After updating the online parameters of the digital twin model, the shared policy network, and the independent value network of each module, the edge computing node determines whether the current execution feedback error of each module exceeds the preset re-optimization trigger threshold. If the threshold is exceeded, the updated digital twin model will be used as a new starting point to re-trigger the cross-module light effect demand map generation process and the linkage configuration scheme solution process, and start the next round of perception-decision-execution cycle. If the threshold is not exceeded, the current configuration is maintained and sensor data continues to be monitored, thus achieving closed-loop self-optimization driven on demand.

[0037] In this embodiment, by linking edge compilation and distribution of configuration instructions, terminal execution status reporting, sensor feedback error assessment, online correction of digital twin models, adaptive updating of configuration strategies, and closed-loop trigger judgment into a complete self-optimizing loop, the system can synchronously correct the dual deviations of the perception model and decision-making strategy after each round of execution, avoiding the problem of configuration errors accumulating and expanding over time in traditional open-loop control systems. The on-demand triggering mechanism reduces unnecessary computational overhead while ensuring the system's timely response to environmental changes, enabling the modular building lighting configuration system to have the ability to continuously self-calibrate and progressively improve performance during long-term operation.

[0038] In some possible embodiments of the present invention, when constructing a digital twin model, the following module optical environment state fusion function is used to dynamically fuse IoT sensor data and BIM static attributes:

[0039] in, Let m be the light environment fusion state vector of the m-th module at time t; This is the static light transmission attribute vector of the m-th module extracted from the BIM model, which includes window-to-wall ratio, material transmittance, etc. The dynamic balance coefficient represents the effect of fusing BIM static prior and real-time sensor data at time t on the balance of... The weight assigned to one side has a value range of (0,1); The set of sensors deployed for the m-th module; The measured light intensity value collected by sensor s at time t; The spatial location weighting coefficient of sensor s is determined by its proportion of the coverage area within the module. The dynamic balance coefficient between BIM static priors and real-time sensor data is adaptively adjusted according to the rate of change of natural daylight. The natural lighting dynamic correction matrix for the m-th module at time t is calculated by combining outdoor meteorological light intensity and solar altitude angle. It reflects the real-time correction of the static light transmission properties of BIM by building orientation and shading, so that the static BIM parameters have the ability to adapt to time-varying changes. The reliability coefficient of sensor s at time t is obtained by jointly evaluating the sensor's historical error rate and current self-test status. The readings of faulty or drifting sensors are automatically downweighted to prevent abnormal sensor data from contaminating the fusion results. Let be the optical environment temporal inertia coefficient of the m-th module, reflecting the physical hysteresis characteristics of the module's spatial heat capacity and optical field changes; This is the time difference term of the light environment fusion state vector from the previous moment, i.e. It captures the historical trend of light environment changes and together with the inertia coefficient, it constitutes the time-series prediction correction term, making the model's prediction of fast light-changing scenes smoother and more accurate.

[0040] In some possible embodiments of the present invention, when extracting the spatial dependency relationship of light effect demand between modules through a graph neural network, the edge weight of the dependency strength between modules is calculated using the following spatial light effect coupling coefficient:

[0041] in, The coupling weights for the light efficiency requirements between module m and its adjacent module n; , These are the hidden layer feature vectors of the graph neural network for modules m and n, respectively, which encode the light demand state and personnel behavior characteristics of each module; This is a learnable edge attention weight matrix; Let m be the set of directly adjacent modules in the spatial topology graph; The physical transmittance of the interface between modules m and n is provided by the interface transmittance attribute parameters in the digital twin model and is used to embed physical optical constraints into the graph structure learning process. The dynamic light flux transfer bias term between module m and module n at time t is obtained by normalizing the product of the measured illuminance difference at the interface of the two modules and the interface area. The real-time physical light transfer is injected into the attention score calculation so that the edge weights are dynamically adjusted with the actual light field changes. The Sigmoid activation function compresses the linear scores within the parentheses to the (0,1) interval, serving as a global gating factor for the edge weights. For learnable side-gated projection vectors; Let m be the feature vector, n be the feature vector, and n be the feature vector of the module pair relationship. splicing; The spatial relationship feature vector between module m and module n encodes the relative position, interface orientation angle and functional partition label of the two modules, introduces structured spatial priors for edge weights, and avoids spatial blind spots caused by relying solely on node features.

[0042] In some possible embodiments of the present invention, the collaborative reward function of each module agent in the multi-agent reinforcement learning algorithm is defined as:

[0043] in, Let m be the instantaneous reward obtained by the agent corresponding to module m in a single-step decision-making process. , These are the light field state vectors currently configured for module m and its adjacent module n, respectively, which include color temperature and illuminance components; The Euclidean distance between adjacent modules is used to quantize the optical field continuity deviation at the module interface; This represents the real-time power consumption of module m at the current moment. This represents the current actual light effect status of module m. The target luminous efficacy state derived from the luminous efficacy demand map; The cosine similarity between the actual and target light effect distributions is calculated as the light effect satisfaction evaluation function. The weighting coefficients for each optimization objective are adaptively adjusted based on the building usage scenario labels; The interface importance weight between module m and adjacent module n is determined by the frequency of personnel line-of-sight crossing of the interface between the two modules and the interface area. A higher light field continuity penalty is given to the key interface that is frequently crossed by personnel, focusing on optimizing the module boundary that has the greatest impact on visual experience. The energy consumption dynamic discount factor of module m at time t is taken to be close to 1 during peak building electricity consumption to strengthen energy-saving penalties, and appropriately relaxed during off-peak periods to achieve time-series coordination with grid load. Weighting coefficients for rewards that improve visual comfort; The visual comfort increment of module m at time t relative to the previous time is calculated as a weighted change in uniform glare value and illuminance uniformity, providing a positive incentive for decisions to improve the configuration scheme towards greater comfort. To configure the weighting coefficient for frequent switching penalties The configuration switching cost of module m at time t is measured by the difference in the light effect configuration vector between the current time and the previous time. This suppresses high-frequency jittering behavior of the agent and ensures the stability of the light effect configuration and the visual comfort of the user.

[0044] In some possible embodiments of the present invention, during the online iterative update process, the following adaptive learning rate adjustment mechanism is used to control the update step size of the digital twin model and the configuration strategy:

[0045] in, Let be the adaptive iterative learning rate at time t; The initial baseline learning rate; The mean value of the sensor feedback error of each module within the sliding window at time t reflects the overall deviation between the current configuration scheme and the actual light environment. It is the error amplitude attenuation coefficient, which controls the sensitivity of the learning rate as the overall error increases, and prevents policy oscillation during the large error phase. This is the gain coefficient for the rate of change of error; The derivative of the mean error with respect to time represents the dynamic trend of the light environment deviation. When the deviation increases rapidly, the update step size is automatically suppressed to maintain system stability. The periodic time modulation factor is defined as follows: ,in For the daily cycle length of construction, The diurnal modulation amplitude coefficient is used to embed the prior diurnal regularity of the building's lighting environment into the learning rate adjustment. During the morning and evening periods when the lighting environment changes frequently, the learning rate is appropriately increased to accelerate adaptation, while the learning rate is reduced during the steady-state phase at night to reduce ineffective updates. The error-aware activation threshold controls the activation level when the learning rate is extremely small. For numerically stable smoothing terms, to prevent When the denominator approaches zero, it becomes singular. As an error-adaptive activation term, when the overall feedback error is extremely small (the system has converged), this term approaches zero, automatically suppressing the learning rate to near the point of stopping updates, thus avoiding invalid disturbances after convergence; when the error is large, this term approaches 1, and the learning rate returns to the normal adjustment range, realizing the error-driven wake-up mechanism.

[0046] In some possible embodiments of the present invention, for scenarios where the module combination structure changes (modules are added, removed, or reorganized), the following module optical efficiency migration similarity index is used to evaluate the reusability of the configuration between the old and new modules:

[0047] in, For the original module m to the new module Configuration migration similarity score; , These are the original module m and the new module respectively. The light environment fusion state vector, and Since the definition is consistent, the steady-state value is taken here; The cosine similarity of the optical environment states of the two modules measures the intrinsic similarity of their optical properties. This represents the shortest path distance between the old and new modules in the building space topology map. This represents the spatial distance attenuation coefficient, reflecting the physical law that the applicability of luminous efficacy migration decreases with increasing spatial distance. When If the preset threshold is exceeded, the original module configuration scheme is directly reused as the initial configuration of the new module; otherwise, the re-optimization process is triggered. For module pairs The functional area matching coefficient is determined by the similarity of the functional tags of the two modules—when the functions are the same (e.g., both are office areas). Amplify the cosine similarity score to encourage configuration reuse among similar modules; when functional differences are significant. Compressed scores to suppress irrational migration across functional areas; The interface connectivity correction factor is defined as the geometric mean of the ratio of the number of shared interfaces between two modules to the total number of interfaces on their respective perimeters. It reflects the impact of the actual physical connectivity between modules on the applicability of configuration migration. The higher the connectivity, the higher the migration similarity. The historical configuration stability entropy value of the original module m is calculated from the Shannon entropy of the configuration state sequence of the module within the historical time window. The lower the entropy value, the more stable and reliable the historical configuration is. For the new module The initial light environment uncertainty entropy value is calculated from the Shannon entropy of its initial sensor reading distribution; This is a historical stability reward gain coefficient; This is a historical stability enhancement; the more stable the original module's historical configuration ( The smaller the value, the higher the initial uncertainty of the new module. When the value is larger, this item approaches zero, indicating limited migration value; conversely, when the original module is highly stable, this item improves the migration score, prioritizing the reuse of well-proven stable configurations.

[0048] Please refer to Figure 2 Another embodiment of the present invention provides a modular building multi-module collaborative lighting effect configuration system for executing a modular building multi-module collaborative lighting effect configuration method, comprising: a cloud server and an intelligent control terminal; The cloud server is configured as follows: Based on BIM data and real-time illumination parameters of each module collected by IoT sensors, a digital twin model is constructed, which includes the spatial topology of the modules, the light transmission properties of the interfaces, and the occlusion relationship, forming a global light environment perception base. Based on the digital twin model, the system integrates personnel behavior recognition data, ambient light intensity distribution data, and task scene labels within each module, and extracts the spatial dependency relationship of light effect requirements between modules through graph neural networks to generate a cross-module light effect requirement map. Based on the light effect demand map, with the optimization objectives of light field continuity and energy consumption balance between modules, a multi-agent reinforcement learning algorithm is adopted to collaboratively solve the linkage configuration scheme of light color temperature, illuminance and dynamic change rhythm of each module. The linkage configuration scheme is distributed to the intelligent control terminals of each module through edge computing nodes for execution. At the same time, sensor feedback data after execution is continuously collected to drive the online iterative update of the digital twin model and configuration strategy, thereby realizing closed-loop self-optimization of perception-decision-execution.

[0049] It should be known that, Figure 2The block diagram of the modular building multi-module collaborative lighting effect configuration system shown is for illustrative purposes only, and the number of modules shown does not limit the scope of protection of this invention. The modular building multi-module collaborative lighting effect configuration system provided in this embodiment can be used to execute various embodiments of the corresponding modular building multi-module collaborative lighting effect configuration method. For specific implementation details, please refer to the descriptions of the respective method embodiments, which will not be repeated here.

[0050] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0051] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0052] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical or other forms.

[0053] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0054] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0055] If the integrated units described above are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0056] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0057] The embodiments of this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

[0058] While the present invention has been disclosed above, it is not limited thereto. Any person skilled in the art can easily conceive of variations or substitutions without departing from the spirit and scope of the present invention, and various modifications and alterations can be made, including combinations of the different functions and implementation steps described above, as well as software and hardware implementation methods, all of which are within the protection scope of the present invention.

Claims

1. A method for configuring multi-module collaborative lighting effects in modular buildings, characterized in that, include: Based on BIM data and real-time illumination parameters of each module collected by IoT sensors, a digital twin model is constructed, which includes the spatial topology of the modules, the light transmission properties of the interfaces, and the occlusion relationship, forming a global light environment perception base. Based on the digital twin model, the system integrates personnel behavior recognition data, ambient light intensity distribution data, and task scene labels within each module, and extracts the spatial dependency relationship of light effect requirements between modules through graph neural networks to generate a cross-module light effect requirement map. Based on the light effect demand map, with the optimization objectives of light field continuity and energy consumption balance between modules, a multi-agent reinforcement learning algorithm is adopted to collaboratively solve the linkage configuration scheme of light color temperature, illuminance and dynamic change rhythm of each module. The linkage configuration scheme is distributed to the intelligent control terminals of each module through edge computing nodes for execution. At the same time, sensor feedback data after execution is continuously collected to drive the online iterative update of the digital twin model and configuration strategy, thereby realizing closed-loop self-optimization of perception-decision-execution.

2. The modular building multi-module collaborative lighting effect configuration method according to claim 1, characterized in that, The steps of constructing a digital twin model containing the spatial topology of modules, interface light transmission properties, and occlusion relationships based on BIM data and real-time illumination parameters collected by IoT sensors to form a global light environment perception base include: The BIM structural data of each module is analyzed, and the geometric boundaries, interface material transmittance, window-to-wall ratio, and the location and size of the shared interface between adjacent modules are extracted to form a basic dataset of static optical properties for each module. Using each module as a node and the shared interface between modules as directed edges, a spatial topology graph is constructed based on the extracted location and size of the shared interface. The edge attributes record the transmittance and orientation information of the corresponding interface, forming a module spatial topology graph that describes the light transmission path. Based on the geometric boundaries and orientations of each module in the module space topology diagram, and combined with the real-time solar altitude angle and azimuth angle, the shadow coverage area of ​​each module at the current moment is calculated by the shading of adjacent modules or building components, and a dynamic shading relationship matrix is ​​generated that updates over time. The IoT light sensor nodes deployed in each module are mapped to the corresponding module nodes in the module spatial topology map, and an association index is established between the sensor number, spatial coordinates and the module node to which it belongs, providing an addressing basis for the targeted fusion of subsequent real-time data. The system continuously receives real-time illumination data from each sensor node via the IoT data channel. Based on the historical baseline of the sensors, it performs anomaly detection on the current readings, eliminates abnormal values ​​caused by sensor malfunctions or signal interference, and outputs real-time illumination data streams from each module that have undergone quality verification. The static optical attribute dataset, the module space topology map, the dynamic occlusion relationship matrix, and the real-time illumination data stream are fused and assembled in multiple layers to construct a unified modular digital twin model of building lighting environment. The model is continuously and synchronously updated with the sensor sampling cycle as the driving beat to form a global lighting environment perception base.

3. The modular building multi-module collaborative lighting effect configuration method according to claim 2, characterized in that, The step of generating a cross-module light effect demand map by integrating personnel behavior recognition data, ambient light intensity distribution data, and task scene labels within each module based on the digital twin model includes: Based on the fusion data of image sensors and human infrared sensors deployed in each module, the current behavior type of personnel in the module is identified in real time, and the set of personnel behavior tags for each module is output, including personnel density, activity intensity and main behavior type. Using the set of personnel behavior tags, the current ambient light intensity distribution data of each module read from the digital twin model, and the pre-configured module task scene tags as input, the original feature vector of light effect requirements of each module is formed by splicing them together in units of modules, which serves as the initial input of the nodes of the subsequent graph neural network; The original feature vector of the light effect requirement is loaded into the corresponding node of the module space topology graph. The directed edge structure of the topology graph carries the light transmission path relationship between modules, forming an initial light effect requirement feature graph carrying multimodal requirement features. The initial light effect demand feature map is input into a graph neural network. Through multiple rounds of neighborhood message passing and aggregation operations, the feature vector of each module node is gradually integrated with the demand state information of adjacent modules. The dependency relationship of light effect demand of each module in spatial topology is extracted, and the enhanced feature vector of each module is output after incorporating neighborhood dependency information. Based on the enhanced feature vector, the target color temperature, target illuminance and dynamic change rhythm of light effect required by each module at the current moment are predicted by the demand decoder, and the quantitative light effect demand prediction results of each module are generated. The quantized light effect demand prediction results are written back to the corresponding nodes of the module space topology graph, and the spatial dependency strength between modules is written to the corresponding directed edges to form a cross-module light effect demand graph that simultaneously carries node-level light effect demand and edge-level module dependency relationship.

4. The modular building multi-module collaborative lighting effect configuration method according to claim 3, characterized in that, The steps of using a multi-agent reinforcement learning algorithm to collaboratively solve the linkage configuration scheme of light color temperature, illuminance, and dynamic change rhythm of each module, based on the light effect demand map and with the optimization objectives of light field continuity and energy consumption balance among modules, include: An independent reinforcement learning agent is instantiated for each module. The quantized light effect demand prediction results of the corresponding nodes in the cross-module light effect demand map and the dependency relationship of the edge modules are used as the initial environmental observation inputs of each agent. The action space reference of each agent is initialized with the actual configuration state of the current light color temperature, illuminance and dynamic change rhythm of each module. At each decision step, each module agent reads the current light field state, real-time power consumption, and personnel behavior tags of its module from the digital twin model. At the same time, it reads the current light field state and dependency edge weights of the adjacent module nodes directly connected to its module from the cross-module light effect demand map. The current light field state, real-time power consumption, personnel behavior tags, current light field state, and dependency edge weights are concatenated into the local observation state vector of the agent for the current decision step, ensuring that each agent's decision-making simultaneously perceives the module's demand and cross-module dependencies. The action space of each module agent is defined as a discretized joint space for the selection of light color temperature adjustment, illuminance adjustment and dynamic change rhythm mode of that module. Each agent independently samples the configuration adjustment action of the current decision step of its module based on the local observation state vector through the policy network with shared parameters. The shared parameter mechanism enables each agent to implicitly share cross-module policy experience when making decisions. After each agent performs the sampling action, the instantaneous reward obtained by each module agent in this decision step is calculated according to the preset cooperative reward function. The reward signal simultaneously reflects the light efficiency satisfaction of this module, the light field continuity deviation with adjacent modules, the real-time power consumption of this module, and the stability of configuration switching, ensuring that the local optimization goals of each agent are consistent with the global cooperative goals. The local observation state vectors, sampling actions, and immediate rewards of each agent are stored in a centralized experience replay buffer. A centralized training and distributed execution framework is used to periodically sample batch data from the buffer to update the parameters of the shared policy network and the independent value networks of each module. The absolute value of the difference between adjacent windows of the cumulative sliding mean of all agents within M consecutive decision steps is continuously lower than the preset convergence judgment threshold as the convergence judgment condition. After triggering convergence, the current policy network parameters are frozen. Here, M is the sliding window length, and both M and the convergence judgment threshold are hyperparameters that are preset before training according to the building module size and configuration accuracy requirements. The converged policy network performs a complete forward inference on the current cross-module light effect demand map, outputs the deterministic optimal action of each module agent under the current environmental observation, and assembles the color temperature target value, illuminance target value and dynamic change rhythm mode of each module into a structured linkage configuration scheme.

5. The modular building multi-module collaborative lighting effect configuration method according to claim 4, characterized in that, The steps of distributing the linkage configuration scheme to the intelligent control terminals of each module through edge computing nodes for execution, while continuously collecting sensor feedback data after execution, and driving the online iterative update of the digital twin model and configuration strategy to achieve closed-loop self-optimization of perception-decision-execution, include: The linkage configuration scheme is divided according to module affiliation. The edge computing node compiles the color temperature target value, illuminance target value and dynamic change rhythm mode of each module into an instruction format that can be directly parsed by the corresponding intelligent control terminal. The instruction is then sent to the intelligent control terminal of each module in parallel through the local IoT channel. The edge computing node also records the timestamp of the instruction sent by each module as a timing benchmark for subsequent execution feedback comparison. After receiving the configuration instructions from the edge computing node, the intelligent control terminal of each module drives the lighting hardware of this module to complete the configuration switching according to the target color temperature, illuminance and rhythm mode. After the execution is completed, the actual execution color temperature, illuminance and switching completion timestamp of this module are reported to the edge computing node. After the edge computing node completes the configuration switch of each module, it triggers the real-time illumination data acquisition process, collects the actual illumination state data of each module sensor after the new configuration is stable, compares it with the color temperature target value and the illuminance target value module by module, calculates the execution feedback error of each module, and identifies the configuration deviation module caused by hardware response deviation, sensor drift or environmental change. The synchronous update mechanism of inputting the actual illumination state data of each module sensor and the execution feedback error into the digital twin model is used to perform online correction on the interface light transmission attribute estimation and occlusion relationship matrix that deviate from the measured values ​​in the digital twin model, so that the light environment mapping accuracy of the digital twin model continues to improve with the accumulation of execution feedback. The execution feedback error is used to calculate the current iteration learning rate according to the adaptive learning rate adjustment mechanism. This learning rate drives the online parameter update of the shared policy network and the independent value network of each module, so that the configuration policy is continuously adaptively corrected in the direction of reducing execution error and improving cross-module light field continuity. After updating the online parameters of the digital twin model, the shared policy network, and the independent value network of each module, the edge computing node determines whether the current execution feedback error of each module exceeds the preset re-optimization trigger threshold. If the threshold is exceeded, the updated digital twin model will be used as a new starting point to re-trigger the cross-module light effect demand map generation process and the linkage configuration scheme solution process, and start the next round of perception-decision-execution cycle. If the threshold is not exceeded, the current configuration is maintained and sensor data continues to be monitored, thus achieving closed-loop self-optimization driven on demand.

6. The modular building multi-module collaborative lighting effect configuration method according to claim 5, characterized in that, When constructing a digital twin model, the following module's light environment state fusion function is used to dynamically fuse IoT sensor data and BIM static attributes: in, Let m be the light environment fusion state vector of the m-th module at time t; The static light transmittance attribute vector of the m-th module extracted from the BIM model includes window-to-wall ratio and material transmittance. The set of sensors deployed for the m-th module; The measured light intensity value collected by sensor s at time t; The spatial location weighting coefficient of sensor s is determined by its proportion of the coverage area within the module. The dynamic balance coefficient between BIM static priors and real-time sensor data is adaptively adjusted according to the rate of change of natural daylight. The natural lighting dynamic correction matrix for the m-th module at time t is calculated by combining outdoor meteorological light intensity and solar altitude angle. It reflects the real-time correction of the static light transmission properties of BIM by building orientation and shading, so that the static BIM parameters have the ability to adapt to time-varying changes. The reliability coefficient of sensor s at time t is obtained by jointly evaluating the sensor's historical error rate and current self-test status. The readings of faulty or drifting sensors are automatically downweighted to prevent abnormal sensor data from contaminating the fusion results. Let be the optical environment temporal inertia coefficient of the m-th module, reflecting the physical hysteresis characteristics of the module's spatial heat capacity and optical field changes; This is the time difference term of the light environment fusion state vector from the previous moment, i.e. It captures the historical trend of light environment changes and together with the inertia coefficient, it constitutes the time-series prediction correction term, making the model's prediction of fast light-changing scenes smoother and more accurate.

7. The modular building multi-module collaborative lighting effect configuration method according to claim 6, characterized in that, When extracting the spatial dependency relationship of light effect demand between modules using a graph neural network, the edge weights of the dependency strength between modules are calculated using the following spatial light effect coupling coefficient: in, The coupling weights for the light efficiency requirements between module m and its adjacent module n; , These are the hidden layer feature vectors of the graph neural network for modules m and n, respectively, which encode the light demand state and personnel behavior characteristics of each module; This is a learnable edge attention weight matrix; Let m be the set of directly adjacent modules in the spatial topology graph; The physical transmittance of the interface between modules m and n is provided by the interface transmittance attribute parameters in the digital twin model and is used to embed physical optical constraints into the graph structure learning process. The dynamic light flux transfer bias term between module m and module n at time t is obtained by normalizing the product of the measured illuminance difference at the interface of the two modules and the interface area. The real-time physical light transfer is injected into the attention score calculation so that the edge weights are dynamically adjusted with the actual light field changes. The Sigmoid activation function compresses the linear scores within the parentheses to the (0,1) interval, serving as a global gating factor for the edge weights. For learnable side-gated projection vectors; Let m be the feature vector, n be the feature vector, and n be the feature vector of the module pair relationship. splicing; The spatial relationship feature vector between module m and module n encodes the relative position, interface orientation angle and functional partition label of the two modules, introduces structured spatial priors for edge weights, and avoids spatial blind spots caused by relying solely on node features.

8. The modular building multi-module collaborative lighting effect configuration method according to claim 7, characterized in that, In multi-agent reinforcement learning algorithms, the collaborative reward function of each agent module is defined as: in, Let m be the instantaneous reward obtained by the agent corresponding to module m in a single-step decision-making process. , These are the light field state vectors currently configured for module m and its adjacent module n, respectively, which include color temperature and illuminance components; The Euclidean distance between adjacent modules is used to quantize the optical field continuity deviation at the module interface; This represents the real-time power consumption of module m at the current moment. This represents the current actual light effect status of module m. The target luminous efficacy state derived from the luminous efficacy demand map; The cosine similarity between the actual and target light effect distributions is calculated as the light effect satisfaction evaluation function. The weighting coefficients for each optimization objective are adaptively adjusted based on the building usage scenario labels; The interface importance weight between module m and adjacent module n is determined by the frequency of personnel line-of-sight crossing of the interface between the two modules and the interface area. A higher light field continuity penalty is given to the key interface that is frequently crossed by personnel, focusing on optimizing the module boundary that has the greatest impact on visual experience. The energy consumption dynamic discount factor of module m at time t is taken to be close to 1 during peak building electricity consumption to strengthen energy-saving penalties, and appropriately relaxed during off-peak periods to achieve time-series coordination with grid load. Weighting coefficients for rewards that improve visual comfort; The visual comfort increment of module m at time t relative to the previous time is calculated as a weighted change in uniform glare value and illuminance uniformity, providing a positive incentive for decisions to improve the configuration scheme towards greater comfort. To configure the weighting coefficient for frequent switching penalties The configuration switching cost of module m at time t is measured by the difference in the light effect configuration vector between the current time and the previous time. This suppresses high-frequency jittering behavior of the agent and ensures the stability of the light effect configuration and the visual comfort of the user.

9. The modular building multi-module collaborative lighting effect configuration method according to claim 8, characterized in that, During online iterative updates, the following adaptive learning rate adjustment mechanism is used to control the update step size of the digital twin model and configuration strategy: in, Let be the adaptive iterative learning rate at time t; The initial baseline learning rate; The mean value of the sensor feedback error of each module within the sliding window at time t reflects the overall deviation between the current configuration scheme and the actual light environment. It is the error amplitude attenuation coefficient, which controls the sensitivity of the learning rate as the overall error increases, and prevents policy oscillation during the large error phase. This is the gain coefficient for the rate of change of error; The derivative of the mean error with respect to time represents the dynamic trend of the light environment deviation. When the deviation increases rapidly, the update step size is automatically suppressed to maintain system stability. The periodic time modulation factor is defined as follows: ,in For the daily cycle length of construction, The diurnal modulation amplitude coefficient is used to embed the prior diurnal regularity of the building's lighting environment into the learning rate adjustment. During the morning and evening periods when the lighting environment changes frequently, the learning rate is appropriately increased to accelerate adaptation, while the learning rate is reduced during the steady-state phase at night to reduce ineffective updates. The error-aware activation threshold controls the activation level when the learning rate is extremely small. For numerically stable smoothing terms, to prevent When the denominator approaches zero, it becomes singular. As an error-adaptive activation term, when the overall feedback error is extremely small, this term approaches zero, automatically suppressing the learning rate to near the point of stopping updates, thus avoiding invalid perturbations after convergence; when the error is large, this term approaches 1, and the learning rate returns to the normal adjustment range, realizing the error-driven wake-up mechanism.

10. A modular building multi-module collaborative lighting effect configuration system, used to execute the modular building multi-module collaborative lighting effect configuration method as described in any one of claims 1 to 9, characterized in that, include: Cloud servers and intelligent control terminals; The cloud server is configured as follows: Based on BIM data and real-time illumination parameters of each module collected by IoT sensors, a digital twin model is constructed, which includes the spatial topology of the modules, the light transmission properties of the interfaces, and the occlusion relationship, forming a global light environment perception base. Based on the digital twin model, the system integrates personnel behavior recognition data, ambient light intensity distribution data, and task scene labels within each module, and extracts the spatial dependency relationship of light effect requirements between modules through graph neural networks to generate a cross-module light effect requirement map. Based on the light effect demand map, with the optimization objectives of light field continuity and energy consumption balance between modules, a multi-agent reinforcement learning algorithm is adopted to collaboratively solve the linkage configuration scheme of light color temperature, illuminance and dynamic change rhythm of each module. The linkage configuration scheme is distributed to the intelligent control terminals of each module through edge computing nodes for execution. At the same time, sensor feedback data after execution is continuously collected to drive the online iterative update of the digital twin model and configuration strategy, thereby realizing closed-loop self-optimization of perception-decision-execution.

Images