An office lighting lamp configuration method and system based on deep reinforcement learning and graph neural network

Abstract

This invention relates to the field of intelligent lighting control technology, and particularly to a method and system for configuring office lighting fixtures based on deep reinforcement learning and graph neural networks. The method includes: acquiring multimodal sensor data; constructing a lighting fixture map based on the spatial coordinates of the fixtures and personnel flow information; automatically partitioning and grouping control groups using node embedding representations learned through graph neural networks; dynamically adjusting the non-triggering lighting delay time of each control group using reinforcement learning, with the control groups and sensor data as input; predicting lighting demand using a time series model based on sensor data and user historical behavior, and adaptively adjusting the lighting and fading speed; dynamically adjusting the effective sensing distance of the sensors by fusing environmental parameters; and generating lighting control commands by combining the partitioning and grouping results, delay time, fading speed, and sensing distance. This invention solves the technical problems of complex manual configuration and difficulty in adapting static configuration to dynamic environmental changes in existing office lighting systems.

Description

Technical Field

[0001] This invention relates to the field of intelligent lighting control technology, and in particular to a method and system for configuring office lighting fixtures based on deep reinforcement learning and graph neural networks. Background Technology

[0002] With the widespread application of IoT technology and smart lighting systems, lighting systems in modern office environments have evolved from simple manual switch control to intelligent control systems capable of automatically sensing the environment and responding to personnel activities. However, current smart lighting solutions on the market still heavily rely on manual configuration during deployment and maintenance. In a typical office lighting system deployment, professional engineers need to physically calibrate the location of each lamp, divide control groups according to functional areas, set sensor trigger thresholds, and define dimming curves and delay parameters. This manual configuration method has the following prominent problems: First, the complexity of zoned and grouped configuration is high; grouping lamps in large office buildings can consume several days or even weeks of manual time. Second, key parameters such as triggerless lighting delay time, lighting and off gradient speed, and sensor sensing distance are difficult to balance the diverse needs of user comfort and energy efficiency. Third, static configuration is difficult to adapt to dynamic changes in the office environment; factors such as personnel turnover, workstation adjustments, and seasonal changes can cause the initial configuration to gradually deviate from the optimal state, requiring professional engineers to repeatedly conduct on-site surveys and parameter adjustments, resulting in continuous investment in human resources and the risk of business interruption. Therefore, there is an urgent need for a method and system for configuring office lighting fixtures that can automatically complete the zoning and grouping of lighting fixtures and adaptive optimization of key control parameters without human intervention. Summary of the Invention

[0003] The main objective of this invention is to propose a method and system for configuring office lighting fixtures based on deep reinforcement learning and graph neural networks, aiming to solve the technical problems of complex manual configuration and difficulty in adapting static configuration to dynamic environmental changes in existing office lighting systems.

[0004] To achieve the above objectives, the first aspect of this invention proposes a method for configuring office lighting fixtures based on deep reinforcement learning and graph neural networks, comprising: Step S100: Acquire multimodal sensor data deployed in the office environment, and construct a lighting fixture map based on the spatial coordinates of the lighting fixtures and the personnel flow information extracted from the multimodal sensor data; learn the embedding representation of each node in the lighting fixture map through a graph neural network to automatically partition and group the lighting fixtures to form multiple lighting fixture control groups; Step S200: Using the data from the lighting control group and the multimodal sensor as input, continuously learn and dynamically adjust the triggerless lighting delay time of each lighting control group through reinforcement learning algorithm; Step S300: Based on the multimodal sensor data and user historical behavior, use a time series model to predict user lighting needs, and adaptively adjust the lighting speed and the extinguishing speed of the lamps according to the prediction results. Step S400: Integrate the environmental parameters from the multimodal sensor data to dynamically adjust the effective sensing distance of each lamp or sensor in the control group; Step S500: At least by combining the results of the automatic zoning and grouping, the non-trigger lighting delay time, the lighting gradation speed and the lighting gradation speed of the lamps, and the effective sensing distance, a lighting control command is generated to control the switching on and off and brightness adjustment of the corresponding lamps.

[0005] Preferably, step S100 includes: Step S110: Calculate the Euclidean distance between the lamps based on their spatial coordinates, establish spatial proximity edges between two lamps whose distance is less than a preset spatial threshold, and determine the edge weights of the spatial proximity edges based on the Gaussian kernel function of the distance. Step S120: Based on the frequency of personnel movement between different lighting areas extracted from the multimodal sensor data, determine the functional association edge weights used to characterize the degree of personnel flow association between lighting fixtures; Step S130: The edge weights of the spatially adjacent edges and the functionally associated edge weights are weighted and fused to construct the adjacency matrix of the lighting graph; Step S140: Use a graph attention network to learn node representations for the lighting graph, and aggregate neighbor node information through a multi-head attention mechanism to obtain low-dimensional embedding representations of each lighting node; Step S150: Construct a similarity matrix based on the low-dimensional embedding representation, use spectral clustering algorithm to cluster the lamp nodes, and calculate the contour coefficient under different group numbers. Take the group number corresponding to the maximum contour coefficient as the optimal group number to generate the multiple lamp control groups.

[0006] Preferably, step S140 includes: Step S141: Input the original features of each lamp node into the first graph attention layer, and map the original features to the hidden space through a learnable weight matrix to obtain the first hidden representation; Step S142: Input the first hidden representation into the second graph attention layer, calculate the attention coefficient between each lamp node and its neighbor nodes through a multi-head attention mechanism, and aggregate the information of multi-hop neighbor nodes based on the attention coefficients to obtain the second hidden representation; Step S143: Input the second hidden representation into the third graph attention layer, and project the aggregated node features into the low-dimensional embedding space through a single-head attention mechanism to obtain the low-dimensional embedding representation of each lamp node; Step S144: During the training of the graph attention network, the reconstructing of the adjacency matrix using the low-dimensional embedding representation is used as a self-supervised pre-training task. This makes the light fixture nodes with close spatial proximity or personnel flow associations closer in the low-dimensional embedding space. The reconstruction loss adopts the form of cross-entropy. In the formula, This represents the reconstruction loss; This represents an edge in the lighting fixture diagram; This represents the edge weight between node i and node j in the adjacency matrix; Represents the low-dimensional embedding representation vector of the lighting node j; Represents the low-dimensional embedding representation vector of lamp node i; This represents the inner product of the embedding vectors of lamp node i and lamp node j; This represents the sigmoid activation function, used to map the inner product of the embedding vectors to the predicted probability that there is a connecting edge between two nodes; Let represent the set of edges in the luminaire graph.

[0007] Preferably, step S200 includes: Step S210: For each lighting control group, model the control problem of the delay time of the lighting without trigger as a Markov decision process, define a state vector that includes the environmental state, user state, lighting state and scene context, and use the adjustment amount of the delay time as the action vector. Step S220: Design a reward function to give a positive reward for the situation that there is no human activity in the area corresponding to the control group after the delay ends, and to impose a negative penalty for the behavior of manually turning on the lights due to the delay being too short. Step S230: Construct a policy network with the state vector as input and the action vector as output, calculate the advantage function using the reward function, and train the policy network based on the advantage function using a proximal policy optimization algorithm; during the training process, an experience replay mechanism and a course learning strategy are adopted, and the delay time of each control group is dynamically adjusted through incremental updates.

[0008] Preferably, in step S230, the steps of constructing a policy network with the state vector as input and the action vector as output, calculating the advantage function using the reward function, and training the policy network based on the advantage function using a proximal policy optimization algorithm include: Step S231: Construct the policy network of the Actor-Critic architecture, with the state vector as input, wherein the Actor network outputs the mean and logarithmic variance of the action vector, and the Critic network outputs the state value estimate; Step S232: A self-attention mechanism is introduced after the input layer of the policy network. The state vector is mapped to a query matrix, a key matrix, and a value matrix through a learnable projection matrix. After scaling dot product attention calculation, the state vector is residually connected to the original state vector to obtain an enhanced state representation. Step S233: Input the enhanced state representation into the Actor network and the Critic network respectively, the Actor network outputs the mean and logarithmic variance of the action vector, and the Critic network outputs the state value estimate; Step S234: Based on the state value estimate output by the Critic network and the reward function, calculate the advantage function using generalized advantage estimation; use the advantage function to update the parameters of the policy network through the truncation mechanism of the near-end policy optimization algorithm.

[0009] Preferably, in step S230, the step of employing an experience playback mechanism and a course learning strategy during the training process, and dynamically adjusting the delay time of each control group through incremental updates, includes: Step S235: Before the formal training begins, a model-independent meta-learning framework is used to perform meta-training on multiple office scenario tasks to obtain the meta-initial parameters of the policy network, and the meta-initial parameters are used as the initial parameters of the policy network. Step S236: During the formal training process, the training of the policy network is iteratively executed in order from a single standardized scenario to a variety of scenarios with increasing complexity. Step S237: When updating the parameters of the policy network in each iteration, a priority experience replay mechanism is adopted to assign sampling priority to the experience samples according to the temporal difference error of each experience sample, and training samples are extracted from the experience replay pool according to the sampling priority. Step S238: After each update of the parameters of the policy network, the action vector is incrementally added to the current delay time, and the updated delay time is limited to a preset effective range by a truncation function.

[0010] Preferably, step S300 includes: Step S310: Based on the multimodal sensor data and the recorded user historical behavior, construct time series features, which include time encoding, personnel presence status, ambient light intensity change rate, and user manual adjustment records; Step S320: Input the time series features into a long short-term memory network to obtain hidden states, and predict the probability of user demand for lighting adjustment and the expected gradation speed in the future time period based on the hidden states; Step S330: When the demand probability exceeds a preset threshold, the lighting transition speed and the lighting extinguishing transition speed are calculated by combining the expected transition speed, the difference between ambient light and target brightness, and the current time period information. Step S340: Obtain the user's implicit feedback and explicit satisfaction rating on the lighting changes, generate a comfort reward signal based on this, and feed the comfort reward signal back to the reinforcement learning algorithm to continuously optimize the gradual speed adjustment strategy.

[0011] Preferably, step S400 includes: Step S410: Perform weighted fusion on the PIR sensor data, microwave radar sensor data and ambient light sensor data in the multimodal sensor data, calculate the comprehensive probability of the presence of personnel through learnable weights, and determine the presence of personnel when the comprehensive probability is higher than a preset threshold, so as to obtain the personnel presence status. Step S420: Obtain environmental parameters from the multimodal sensor data, establish a multivariate regression influence model between the sensor's nominal sensing distance and the environmental parameters, and obtain environmental correction coefficients; the environmental parameters include ambient temperature, humidity, airflow, and obstacle density; Step S430: Use a scene classifier to identify scene categories based on current time, personnel density, and historical activity patterns, and determine a baseline sensing distance based on the scene category; Step S440: Calculate the real-time trigger accuracy and false trigger rate based on the personnel presence status, and fine-tune the baseline sensing distance online to obtain the effective sensing distance.

[0012] Preferably, step S500 includes: Step S510: Construct a shared feature extraction network to extract shared representations from the multimodal sensor data to support the determination of target control parameters, the target control parameters including the results of the automatic partitioning and grouping, the dynamically adjusted delay time, the lighting gradation speed and the lighting gradation speed of the lamps, and the dynamically adjusted effective sensing distance; Step S520: Based on the shared representation, perform multi-task joint training on the target control parameters. During the training process, an uncertainty weighting method is used to automatically balance the loss weights of each task, and a penalty function method is introduced to encode the collaborative constraints between the control parameters. Step S530: Generate the lighting control command by combining the target control parameters output after the multi-task joint training.

[0013] A second aspect of this invention proposes an office lighting fixture configuration system based on deep reinforcement learning and graph neural networks, comprising: The lighting fixture partitioning and grouping module is used to acquire multimodal sensor data deployed in the office environment, and construct a lighting fixture map based on the spatial coordinates of the lighting fixtures and personnel flow information extracted from the multimodal sensor data; the module learns the embedding representation of each node in the lighting fixture map through a graph neural network to automatically partition and group the lighting fixtures to form multiple lighting fixture control groups; The delay adaptive optimization module is used to input the data from the lighting control group and the multimodal sensor, and continuously learn and dynamically adjust the triggerless lighting delay time of each lighting control group through reinforcement learning algorithm; The gradual speed adjustment module is used to predict the user's lighting needs based on the multimodal sensor data and the user's historical behavior using a time series model, and to adaptively adjust the lighting speed and the turning-off speed of the lamps according to the prediction results. The sensing distance adjustment module is used to integrate environmental parameters from the multimodal sensor data and dynamically adjust the effective sensing distance of each lamp or sensor in the control group. The joint optimization and control module is used to integrate the results of the automatic partitioning and grouping, the delay time of the non-trigger lighting, the lighting gradation speed and the lighting gradation speed of the lamps, and the effective sensing distance to generate lighting control commands to control the switching on and off and brightness adjustment of the corresponding lamps.

[0014] This invention provides an office lighting fixture configuration method and system based on deep reinforcement learning and graph neural networks. It automatically learns the spatial and functional relationships between lighting fixtures using graph neural networks and combines this with spectral clustering to achieve manual partitioning and grouping, improving deployment efficiency and adapting to dynamic changes in office space layout. Reinforcement learning transforms the latency from a fixed configuration to a context-aware dynamic strategy, reducing ineffective lighting energy consumption while ensuring user comfort. Time series models predict user lighting needs and adaptively adjust the gradation speed accordingly, improving visual comfort and user experience of lighting changes. Multi-sensor fusion and environmental impact modeling dynamically adjust the sensing distance, improving the accuracy and robustness of personnel detection and reducing false and missed triggers. Shared representation learning and multi-task joint optimization achieve end-to-end coordination of four parameters, enabling the system to automatically converge to the globally optimal configuration without manual intervention.

[0015] Furthermore, this invention also improves the alignment between partitioning and grouping and actual usage patterns by constructing a lighting graph through the fusion of spatial proximity and personnel flow edge weights; enhances the representation ability of complex spatial layouts by automatically learning neighbor node weights through a graph attention network; improves the flexibility and adaptability of system deployment by automatically determining the optimal number of groups through contour coefficients; improves the hierarchical representation ability of the embedding representation of the topology by aggregating multi-hop neighbor information layer by layer through a three-layer graph attention network; improves the discriminative power and grouping accuracy of the embedding representation by reconstructing the pre-training task through self-supervised graphs; improves the adaptability of the delay policy to dynamic environments by modeling delay control as a Markov decision process and designing a composite reward function; improves the training stability and representation ability of the policy network by enhancing state representation through an Actor-Critic architecture and self-attention mechanism; improves the accuracy and update stability of policy gradient estimation by updating parameters through generalized advantage estimation and PPO truncation mechanism; improves the cold start efficiency of new scenario deployment by obtaining meta-initial parameters through model-independent meta-learning; and improves the efficiency of cold start deployment in new scenarios by learning through courses and... Prioritize experience replay training to improve the generalization ability of the strategy and the learning efficiency of difficult samples; improve the smoothness and safety of the control strategy by adjusting the delay time through incremental updates and truncation functions; improve the predictability and accuracy of gradual speed adjustment by predicting user lighting needs through long short-term memory networks; improve the matching degree between lighting changes and user personalized preferences by optimizing the gradual strategy through multi-source comfort feedback closed loop; improve the accuracy and robustness of personnel detection by determining the presence status of personnel through multi-sensor weighted fusion; improve the adaptability of sensors in different physical environments and scenarios by dynamically adjusting the sensing distance through environmental multivariate regression models and scene classifiers; improve the long-term stability of sensing distance in dynamic environments by fine-tuning the real-time trigger accuracy and false trigger rate online; improve information sharing and parameter efficiency among control tasks through shared feature extraction and multi-task joint training; improve the stability and convergence efficiency of multi-task training by automatically balancing task losses through uncertainty weighting; and improve the engineering feasibility and control coordination of joint optimization results by encoding parameter collaborative constraints through penalty function method.

[0016] In summary, the present invention proposes an office lighting fixture configuration method and system based on deep reinforcement learning and graph neural networks, which solves the technical problems of complex manual configuration and difficulty in adapting static configuration to dynamic environmental changes in existing office lighting systems. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating an embodiment of the present invention for a method of configuring office lighting fixtures based on deep reinforcement learning and graph neural networks; Figure 2 A flowchart of an automatic zoning and grouping method for lighting fixtures provided in an embodiment of the present invention; Figure 3 A flowchart illustrating a graph attention network node representation learning method provided in an embodiment of the present invention; Figure 4 A flowchart of an embodiment of the present invention provides an adaptive optimization method for triggerless light-up delay time; Figure 5 A flowchart illustrating a policy network training method based on an Actor-Critic architecture provided in an embodiment of the present invention; Figure 6 A flowchart illustrating a strategy network training organization method based on meta-learning and curriculum learning, provided in an embodiment of the present invention; Figure 7 A flowchart of an embodiment of the present invention provides an adaptive adjustment method for the gradual speed of lighting on and off; Figure 8 A flowchart of a method for dynamically adjusting the effective sensing distance of a sensor provided in an embodiment of the present invention; Figure 9 A flowchart of a multi-task joint optimization and control instruction generation method provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of an office lighting fixture configuration system based on deep reinforcement learning and graph neural networks according to an embodiment of the present invention; Figure 11 This is a schematic diagram of the structure of a computer device for configuring office lighting fixtures based on deep reinforcement learning and graph neural networks, according to an embodiment of the present invention.

[0019] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] It should be noted that if the embodiments of the present invention involve directional indicators, such as up, down, left, right, front, back, etc., the directional indicators are only used to explain the relative positional relationship and movement of the components in a specific posture. If the specific posture changes, the directional indicators will also change accordingly.

[0022] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the meaning of "and / or" throughout the text includes three parallel solutions; for example, "A and / or B" includes solution A, solution B, or a solution where both A and B are satisfied simultaneously. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0023] like Figures 1 to 9 As shown, the first aspect of this invention proposes a method for configuring office lighting fixtures based on deep reinforcement learning and graph neural networks, comprising: Step S100: Acquire multimodal sensor data deployed in the office environment, and construct a lighting fixture map based on the spatial coordinates of the lighting fixtures and the personnel flow information extracted from the multimodal sensor data; learn the embedding representation of each node in the lighting fixture map through a graph neural network to automatically partition and group the lighting fixtures to form multiple lighting fixture control groups; Step S200: Using the data from the lighting control group and the multimodal sensor as input, continuously learn and dynamically adjust the triggerless lighting delay time of each lighting control group through reinforcement learning algorithm; Step S300: Based on the multimodal sensor data and user historical behavior, use a time series model to predict user lighting needs, and adaptively adjust the lighting speed and the extinguishing speed of the lamps according to the prediction results. Step S400: Integrate the environmental parameters from the multimodal sensor data to dynamically adjust the effective sensing distance of each lamp or sensor in the control group; Step S500: At least by combining the results of the automatic zoning and grouping, the delay time for the lamps without triggering, the lighting speed and the extinguishing speed of the lamps, and the effective sensing distance, a lighting control command is generated to control the switching on and off and the brightness adjustment of the corresponding lamps.

[0024] For details, see Figure 1 As shown, in a specific embodiment of the present invention, firstly, ambient light data, personnel activity data, and lighting fixture operating status data are continuously collected by multimodal sensors deployed in the office environment; then, based on the physical coordinates of each lighting fixture in the building floor plan, the Euclidean distance between the lighting fixtures is calculated and spatial proximity edges are established. Simultaneously, functional association edges are established based on the frequency of personnel movement between different lighting fixture coverage areas. The weights of these two types of edges are weighted and fused to construct a complete lighting fixture map; subsequently, a graph neural network is used to learn node representations on the lighting fixture map, so that lighting fixtures with spatially adjacent locations and frequent personnel movement are placed closer to each other in the embedded space. Then, a clustering algorithm is used to automatically divide the embedded representations, and the optimal grouping is automatically determined based on the clustering effect evaluation index. The system generates multiple lighting control groups. After grouping the lighting, the delay time control problem for each control group is modeled as a decision-making process involving continuous states and actions. State information includes the current time period, the density and duration of people staying in the area, the level of natural light, and the historical trigger frequency. Actions are defined as adjustments to the delay time. The system continuously interacts with the office area through a reinforcement learning algorithm: when a person enters, the lights are turned on; when the person leaves, the timer starts; if no one re-enters during the delay period, the lights are automatically turned off and a positive reward is given based on energy saving; if the delay is too short and people manually turn the lights back on, a penalty is imposed. Through continuous trial and error, the system gradually learns the activity patterns of people in the area and converges to the optimal delay strategy.

[0025] In the reinforcement learning framework of this embodiment, the agent observes the environment and acquires a state vector at each time step. The state vector is composed of the following four components: In the formula, This is a subvector representing the environmental state, containing the current hour, day of the week, month, natural light intensity, ambient temperature, and ambient humidity. This is a user state subvector, containing the number of people currently detected, the distribution of people in each area, the movement speed of people, and the duration of people's stay. This is a state subvector for each light fixture, containing the current brightness value, the most recent trigger time, and the cumulative runtime of each light fixture. It is a scenario context state sub-vector that encodes the current business scenario, including normal working hours, lunch break, overtime hours, and holidays.

[0026] Each component in the state vector has been standardized using Z-score normalization. In the formula, This represents the eigenvalues ​​after standardization. Represents the original eigenvalues; This represents the mean of the feature across all samples; This represents the standard deviation of the feature across all samples; that is, subtracting the mean from each feature value and then dividing by its standard deviation to eliminate the influence of different units on the reinforcement learning training process.

[0027] Action space Defined as a composite action space, it consists of four subspaces: In the formula, The partitioning and grouping action subspace is defined on a discrete set of group labels; The time delay action subspace is defined as a continuous interval. The unit is seconds; The gradual velocity motion subspace is defined as a two-dimensional continuous space. The unit is seconds; The sensing distance action subspace is defined as a continuous interval. The unit is meters.

[0028] For a continuous action subspace, the policy network outputs the mean of a Gaussian distribution. and variance The actual actions performed are obtained through reparameterization sampling: In the formula, This represents the actual action vector executed. This represents the average action output by the policy network based on state s; This represents the variance of the action output by the policy network based on state s; A random noise vector sampled from a standard normal distribution; This represents element-wise multiplication. This parameterization method ensures the differentiability of the policy during training while maintaining necessary exploration capabilities by introducing random perturbations near the predicted mean.

[0029] The reward function is designed as a composite reward function containing four components. The total reward signal is obtained by weighted summation of the components. In the formula, This represents the total reward value obtained by the agent at time step t; The energy-saving reward is negatively correlated with the actual energy consumption of the lighting fixtures, encouraging the system to reduce energy consumption without sacrificing user experience. The comfort reward component is calculated based on the user's implicit and explicit feedback. Implicit feedback includes whether the user manually turns the lights back on after they are automatically turned off, and explicit feedback includes the user's satisfaction rating submitted through the terminal. This indicates a responsiveness bonus component, penalizing the delay in the entry of lighting personnel; This represents a stability reward component, which penalizes frequent and large fluctuations in control parameters and encourages smooth and stable control strategy adjustments. The weighting coefficient represents the energy-saving reward component and is used to adjust the contribution of energy-saving targets to the total reward. The weighting coefficient represents the comfort reward component, used to adjust the contribution of the user's comfort goal to the total reward; The weighting coefficients representing the responsiveness reward components are used to control the strength of the response delay penalty in the total reward; The weighting coefficients representing the stability reward components are used to control the intensity of parameter fluctuation penalties in the total reward; each weighting coefficient is dynamically adjusted according to business priority and satisfies... .

[0030] Simultaneously, the time-series data collected by sensors and user historical behavior records are input into a preset time-series model to predict the probability of user demand for lighting adjustments and the desired gradation speed at different times. When the demand probability exceeds a set threshold, the lighting gradation speed and the gradation speed of turning off the lights are adjusted in advance, and user satisfaction feedback on lighting changes is continuously collected to optimize the adjustment strategy. Then, environmental parameters such as ambient temperature, humidity, airflow, and obstacle density are extracted from multimodal sensor data, and the effective sensing distance of the sensors is dynamically determined in combination with scene classification results. The accuracy of personnel presence detection is improved through a multi-sensor fusion voting mechanism. Finally, the above-mentioned automatic partitioning and grouping results, dynamic delay parameters, adaptive gradation speed parameters, and dynamic sensing distance parameters are jointly optimized to generate a unified lighting control command and issue it to each lamp for execution.

[0031] Understandably, this embodiment uses graph neural networks to automatically learn the spatial and functional relationships between lighting fixtures and combines spectral clustering to achieve manual partitioning and grouping, improving deployment efficiency and adapting to dynamic changes in office space layout. Reinforcement learning transforms the latency from a fixed configuration to a context-aware dynamic strategy, reducing ineffective lighting energy consumption while ensuring user comfort. Time series models predict user lighting needs and adaptively adjust the gradation speed accordingly, improving visual comfort and user experience of lighting changes. Multi-sensor fusion and environmental impact modeling dynamically adjust the sensing distance, improving the accuracy and robustness of personnel detection and reducing false triggers and missed triggers. Shared representation learning and multi-task joint optimization achieve end-to-end coordination of the four parameters, enabling the system to automatically converge to the globally optimal configuration without manual intervention.

[0032] Based on the above technical solutions, those skilled in the art can make corresponding equivalent improvements according to the specific characteristics of the application scenario or system requirements. For example, the statistical method of personnel traffic frequency can be replaced with a moving average model based on time decay weighting to more sensitively reflect recent changes in personnel flow patterns; or the fixed penalty coefficient in the reward function of reinforcement learning can be replaced with an adaptive coefficient that is dynamically adjusted according to the frequency of accidental shutdown to balance energy saving and comfort preferences under different usage periods.

[0033] Preferably, step S100 includes: Step S110: Calculate the Euclidean distance between the lamps based on their spatial coordinates, establish spatial proximity edges between two lamps whose distance is less than a preset spatial threshold, and determine the edge weights of the spatial proximity edges based on the Gaussian kernel function of the distance. Step S120: Based on the frequency of personnel movement between different lighting areas extracted from the multimodal sensor data, determine the functional association edge weights used to characterize the degree of personnel flow association between lighting fixtures; Step S130: The edge weights of the spatially adjacent edges and the functionally associated edge weights are weighted and fused to construct the adjacency matrix of the lighting graph; Step S140: Use a graph attention network to learn node representations for the lighting graph, and aggregate neighbor node information through a multi-head attention mechanism to obtain low-dimensional embedding representations of each lighting node; Step S150: Construct a similarity matrix based on the low-dimensional embedding representation, use spectral clustering algorithm to cluster the lamp nodes, and calculate the contour coefficient under different group numbers. Take the group number corresponding to the maximum contour coefficient as the optimal group number to generate the multiple lamp control groups.

[0034] For details, see Figure 2 As shown, in a specific embodiment of the present invention, all the lighting fixtures in the office space are first modeled as an undirected weighted graph. , where the set of nodes All correspondences Each lamp, at each node eigenvectors Spatial coordinates of the lamp Physical attributes such as installation height, rated power, and beam angle are considered. The construction of edges takes into account both spatial proximity and functional association: for spatially adjacent edges, if the planar Euclidean distance between two luminaires is less than a preset threshold... An edge is then established between the two light fixtures, and the edge weight is calculated using the Gaussian kernel function of the distance: In the formula, Indicates lighting fixtures i With lighting fixtures j The spatial proximity edge weight between the two lamps is in the range of (0,1]. The larger the weight, the closer the two lamps are in spatial position. Indicates lighting fixtures i Plane coordinates; Represents the planar coordinates of lamp j; The kernel bandwidth parameter indicates that the closer the light fixtures are, the greater the edge weight they receive.

[0035] For functionally related edges, based on sensor data to statistically analyze the frequency of personnel movement between different lighting areas, the edge weight is proportional to the joint triggering frequency: In the formula, Indicates lighting fixtures i With lighting fixtures j The functional association edge weight between the two lamps has a value range of [0,1]. The larger the weight, the closer the personnel flow association between the two lamps. This represents the maximum frequency of all lamp pairs, used to normalize the edge weights to the [0,1] interval; Indicates personnel at the lighting fixtures i With lighting fixtures j The frequency of travel between regions, with the denominator being the maximum value of the frequency of travel between all lamp pairs, is used to normalize the edge weights. Interval.

[0036] The final adjacency matrix is ​​obtained by weighted fusion of the two types of associated edge weights: In the formula, This represents the adjacency matrix of the merged lighting fixture graph; This represents an adjustable fusion coefficient, with a value range of [0,1], used to balance the relative importance of spatial layout constraints and personnel flow patterns in group decision-making; This represents the spatial proximity edge weight matrix, whose elements are... ; This represents the functional association edge weight matrix, whose elements are: .

[0037] By collecting personnel activity data over a period of time, the system can automatically construct a functional relationship graph reflecting actual usage patterns. The graph neural network outputs low-dimensional embedding representations of each lighting fixture node. Next, spectral clustering algorithm is used for partitioning and grouping. First, a similarity matrix is ​​constructed based on the Euclidean distance between the embedding vectors. ,in Then construct the degree matrix. ,in And calculate the standardized Laplace matrix: In the formula, Represents the normalized Laplace matrix; Represents the identity matrix; This represents a similarity matrix constructed based on the low-dimensional embedding vectors of the lighting fixture nodes; Representation degree matrix.

[0038] calculate The former k The eigenvectors corresponding to the smallest eigenvalues Construct the feature matrix .right Each row is standardized to obtain a matrix. ,in .Will Each row is considered a data point, and the K-means algorithm is used to divide it into... k Clusters. Optimal number of groups. k The silhouette coefficient is automatically determined, and it measures the combined effect of intra-cluster compactness and inter-cluster separation. In the formula, Indicates sample i The average distance to other samples in the same cluster; Indicates sample i The average distance to all samples in the nearest other cluster. The range of values ​​for is [ [1,1], where the closer to 1, the better the clustering effect. (In the preset...) Within the range, traverse different group numbers and select the one that maximizes the average profile coefficient. kThe optimal number of groups is set as the value and does not require manual specification. This partitioning and grouping scheme has good adaptability, automatically adjusting the groups according to actual personnel flow patterns. When adding new lighting fixtures, only the corresponding node needs to be added to the graph and the clustering process needs to be rerun, without reconstructing the entire system.

[0039] Understandably, this embodiment constructs a lighting graph by integrating spatial proximity relationships and personnel flow patterns, thereby improving the fit between zoning and grouping and actual usage patterns; it automatically learns the weights of neighboring nodes through a graph attention network instead of using fixed weights, thereby improving the ability to represent complex spatial layouts and personnel flow patterns; and it automatically determines the optimal number of groups through contour coefficients without manual specification, thereby improving the flexibility and adaptability of system deployment.

[0040] Based on the above technical solutions, those skilled in the art can make corresponding equivalent improvements according to the specific characteristics of the application scenario or system requirements. For example, the weight calculation method of spatial proximity edges can be replaced from the Gaussian kernel function to the inverse distance weighting function to adapt to different spatial scales; or the weight of functionally related edges can be replaced from being based on the frequency of travel to being based on the probability of simultaneous appearance of personnel to capture collaborative office mode; or the fusion method of spatial proximity edges and functionally related edges can be replaced from weighted summation to taking the maximum value to retain stronger unilateral correlation signals.

[0041] Preferably, step S140 includes: Step S141: Input the original features of each lamp node into the first graph attention layer, and map the original features to the hidden space through a learnable weight matrix to obtain the first hidden representation; Step S142: Input the first hidden representation into the second graph attention layer, calculate the attention coefficient between each lamp node and its neighbor nodes through a multi-head attention mechanism, and aggregate the information of multi-hop neighbor nodes based on the attention coefficients to obtain the second hidden representation; Step S143: Input the second hidden representation into the third graph attention layer, and project the aggregated node features into the low-dimensional embedding space through a single-head attention mechanism to obtain the low-dimensional embedding representation of each lamp node; Step S144: During the training of the graph attention network, the reconstructing of the adjacency matrix using the low-dimensional embedding representation is used as a self-supervised pre-training task. This makes the light fixture nodes with close spatial proximity or personnel flow associations closer in the low-dimensional embedding space. The reconstruction loss adopts the form of cross-entropy. In the formula, This represents the reconstruction loss; This represents an edge in the lighting fixture diagram; This represents the edge weight between node i and node j in the adjacency matrix; Represents the low-dimensional embedding representation vector of the lamp node j; Represents the low-dimensional embedding representation vector of lamp node i; This represents the inner product of the embedding vectors of lamp node i and lamp node j; This represents the sigmoid activation function, used to map the inner product of the embedding vectors to the predicted probability that there is a connecting edge between two nodes; Let represent the set of edges in the luminaire graph.

[0042] For details, see Figure 3 As shown, in a specific embodiment of the present invention, the graph attention network uses a three-layer architecture to learn node representations for the lighting graph, and then uses the node feature matrix... and adjacency matrix The input is a graph attention network, where n represents the total number of lights in the office environment, and d represents the original feature dimension of each light fixture node. The original features include physical attributes such as the light fixture's spatial coordinates, installation height, rated power, and beam angle. The first layer is a feature transformation layer, which maps the original physical attributes of each light fixture node to the hidden space using a learnable weight matrix, resulting in the first hidden representation. In the formula, Let represent the first hidden representation matrix, where each row corresponds to the feature vector of a lamp node in the hidden space; This represents the attention layer of the first graph, which aggregates the first-order neighbor information of each node through a self-attention mechanism. This represents the dimension of the hidden layer. The second layer is a deep aggregation layer that captures multi-hop neighbor information through a multi-head attention mechanism to obtain the second hidden representation: In the formula, This represents the second hidden representation matrix; The second layer represents the attention layer, which employs a multi-head attention mechanism to compute multiple attention scores in parallel and then concatenate or average them to capture the relationships between multi-hop neighbor nodes. The third layer is the output projection layer, which uses a single-head attention mechanism to project the aggregated node features into a low-dimensional embedding space to obtain the low-dimensional embedding representation of each lamp node. In the formula, This represents the node embedding matrix, where each row corresponds to a low-dimensional embedding vector of a lamp node. The third graph represents the attention layer, which uses a single-head attention mechanism to output the final embedded representation. Denotes the dimension of the low-dimensional embedding space, and This achieves dimensionality reduction and feature compression. During the training of the graph attention network, a self-supervised graph reconstruction task is used as the pre-training objective. The original adjacency matrix is ​​reconstructed through node embedding, ensuring that spatially adjacent or closely related lighting fixtures are close to each other in the embedding space. The reconstruction loss is expressed as cross-entropy. By minimizing this reconstruction loss The network is forced to learn an embedded representation that accurately reconstructs the relationships between lighting fixtures, making nodes that are connected or have strong functional relationships in the original graph closer in the low-dimensional embedding space, providing high-quality input features for subsequent spectral clustering and grouping.

[0043] Understandably, this embodiment aggregates neighbor node information layer by layer through a three-layer graph attention network. The first layer realizes the nonlinear mapping of the original features to the hidden space. The second layer captures the structural dependencies of multi-hop neighbors through a multi-head attention mechanism. The third layer compresses the aggregated features into a low-dimensional space, improving the ability of the embedding representation to represent the hierarchical relationships of complex office space topology. By reconstructing the pre-training task through self-supervised graphs, the consistency between the inner product of the embedding vectors and the original adjacency matrix is ​​constrained by cross-entropy loss, so that lamps that are spatially adjacent or closely related to personnel flow are naturally close in the embedding space, improving the discriminative power of the embedding representation and the accuracy of subsequent partitioning and grouping.

[0044] Based on the above technical solutions, those skilled in the art can make corresponding equivalent improvements according to the specific characteristics of the application scenario or system requirements. For example, the three-layer graph attention network can be adjusted to two or four layers to achieve a balance between computational overhead and representational capability; or the number of attention heads in the multi-head attention mechanism can be increased or decreased according to the size of the lighting network to adapt to office space layouts of different densities; or the cross-entropy loss function can be replaced with the mean squared error loss function to simplify the computational complexity of the training target.

[0045] Preferably, step S200 includes: Step S210: For each lighting control group, model the control problem of the delay time of the lighting without trigger as a Markov decision process, define a state vector that includes the environmental state, user state, lighting state and scene context, and use the adjustment amount of the delay time as the action vector. Step S220: Design a reward function to give a positive reward for the situation that there is no human activity in the area corresponding to the control group after the delay ends, and to impose a negative penalty for the behavior of manually turning on the lights due to the delay being too short. Step S230: Construct a policy network with the state vector as input and the action vector as output, calculate the advantage function using the reward function, and train the policy network based on the advantage function using a proximal policy optimization algorithm; during the training process, an experience replay mechanism and a course learning strategy are adopted, and the delay time of each control group is dynamically adjusted through incremental updates.

[0046] For details, see Figure 4 As shown, in a specific embodiment of the present invention, an independent delay time optimization model is established for each lighting control group that has been partitioned and grouped. The no-trigger lighting delay time refers to the length of time a lighting fixture remains lit after the sensor no longer detects human activity. This parameter directly affects two mutually constraining objectives: user comfort and energy consumption. Excessive delay leads to unnecessary energy waste, while insufficient delay may cause the lights to turn off prematurely while people are still present. To achieve adaptive optimization of this parameter, the delay control problem is modeled as a Markov decision process involving continuous states and continuous actions. The state space includes the following characteristics: the current time period... t Normalized to the [0,1] interval; the average length of stay of people in this area Population density in the area Current level of natural light The lighting control group label for this area ; and a histogram of sensor trigger frequencies in the area over the past 24 hours. The state vector is embedded through a linear layer into In a continuous dimensional space, nonlinear correlations between different state characteristics are captured; the action space is defined as a one-dimensional continuous space, representing the adjustment amount of the delay time. Its value range is ,in This indicates the maximum step size for a single adjustment. Each decision is made within the current delay time. Based on this, incremental adjustments are made, and the adjusted delay time is limited to a preset effective range using a truncation function: In the formula, Indicates the updated delay time; Indicates the current delay time; This indicates the amount of time delay adjustment for the output of this decision; This indicates the preset minimum allowable delay time; This indicates the preset maximum allowable delay time; This represents a truncation function that restricts the input value to a specific range. Within the range. An incremental update strategy is used to avoid drastic changes in delay values, ensuring smooth control.

[0047] The reward function comprehensively considers two indicators: false shutdown rate and energy efficiency, guiding the agent to learn a delay strategy that balances comfort and energy saving. The reward value is calculated using the following formula: In the formula, This represents the reward value obtained by the delay control task within the current delay period; This indicates that the indicator function has not been retried. If no human activity is detected again in the area after the delay, the value is 1; otherwise, it is 0. This indicates the actual delay duration of the current delay period; This represents the preset maximum allowable delay time, used to normalize the delay duration so that the energy-saving reward component has a value range of [0,1]. This represents the penalty coefficient for accidental shutdown, used to balance comfort and energy-saving goals; This indicates the function to be retriggered. If the light is manually turned back on during the delay period, the value is 1; otherwise, it is 0.

[0048] When the delay is set reasonably and there is no human activity in the area, the agent receives a positive reward proportional to the delay duration. The shorter the delay, the lower the normalized reward value. When the delay is too short and people manually turn the lights back on, the agent is penalized with a fixed coefficient to suppress accidental light-off behavior.

[0049] Then, a policy network is constructed with state vectors as input and action vectors as outputs. This network is trained using a proximal policy optimization algorithm. The advantage function is calculated using a reward function, and the network parameters are updated based on this advantage function. During training, an experience replay mechanism is used to store historical interaction samples and randomly sample from them to break data correlation. A learning strategy is employed to gradually transition from simple to complex scenarios to improve policy generalization ability. Through continuous interaction and trial and error with the office environment, the policy network gradually learns the optimal delay strategy that adapts to the activity patterns of personnel in each control group, causing the delay time to gradually converge from the initial default value to the optimal value that balances comfort and energy saving.

[0050] Understandably, this embodiment models the delay time control problem as a Markov decision process, transforming the traditional fixed delay configuration into an adaptive decision capable of sensing the context, thereby improving the adaptability of the delay strategy to the activity patterns of people in different areas. By designing a composite reward function that comprehensively considers the penalties for accidental shutdown and the incentives for energy saving, the reinforcement learning agent is guided to automatically find the optimal balance between user comfort and energy efficiency, improving the rationality of the system's decision-making in complex dynamic environments. By adjusting the delay time through incremental updates and combining it with a truncation function to limit the adjustment range, the smoothness and stability of the control strategy are improved, avoiding drastic changes in delay parameters from interfering with the user experience.

[0051] Based on the above technical solutions, those skilled in the art can make corresponding equivalent improvements according to the specific characteristics of the application scenario or system requirements. For example, the fixed penalty coefficient in the reward function can be replaced with an adaptive coefficient that is dynamically adjusted according to the frequency of accidental shutdown to flexibly balance energy saving and comfort preferences under different usage periods; or the experience playback mechanism can be replaced with a priority experience playback mechanism to accelerate the learning efficiency of key samples.

[0052] Preferably, in step S230, the steps of constructing a policy network with the state vector as input and the action vector as output, calculating the advantage function using the reward function, and training the policy network based on the advantage function using a proximal policy optimization algorithm include: Step S231: Construct the policy network of the Actor-Critic architecture, with the state vector as input, wherein the Actor network outputs the mean and logarithmic variance of the action vector, and the Critic network outputs the state value estimate; Step S232: A self-attention mechanism is introduced after the input layer of the policy network. The state vector is mapped to a query matrix, a key matrix, and a value matrix through a learnable projection matrix. After scaling dot product attention calculation, the state vector is residually connected to the original state vector to obtain an enhanced state representation. Step S233: Input the enhanced state representation into the Actor network and the Critic network respectively, the Actor network outputs the mean and logarithmic variance of the action vector, and the Critic network outputs the state value estimate; Step S234: Based on the state value estimate output by the Critic network and the reward function, calculate the advantage function using generalized advantage estimation; use the advantage function to update the parameters of the policy network through the truncation mechanism of the near-end policy optimization algorithm.

[0053] For details, see Figure 5As shown, in a specific embodiment of the present invention, the policy network adopts an Actor-Critic architecture. The Actor network takes a state vector as input and outputs the mean and log-variance of the action vector; the Critic network takes the same state vector as input and outputs a single state value estimate. The Actor network structure includes an input layer, two fully connected hidden layers, and an output layer. The hidden layers use the ReLU activation function, and the output layer contains two units, outputting the mean and log-variance of the action vector, respectively. The Critic network shares a similar hidden layer structure with the Actor network, but its output layer contains only one unit, outputting a value estimate for the current state. To enhance the expressive power of the state representation, a self-attention mechanism is introduced after the input layer of the policy network. A learnable projection matrix maps the state vector to a query matrix, a key matrix, and a value matrix. After scaling dot product attention calculation, a residual connection is made with the original state vector to obtain the enhanced state representation. In the formula, This represents the state vector of the input policy network; This represents the enhanced state after self-attention enhancement; This represents the scaled dot product attention function; This represents the learnable query projection matrix. Represents the learnable key projection matrix. The projection matrix represents the learnable values, with all dimensions being equal to 1. ,in Let be the dimension of the state vector. For the dimension of attention head.

[0054] During the parameter update phase, the advantage function is calculated using generalized advantage estimation based on the state value estimate and reward function output by the Critic network. For each time step, the temporal difference error is calculated: In the formula, The time difference error at time step t, The reward value obtained at the current moment. As a discount factor, For the Critic network's estimation of the value of the next state. This is the Critic network's value estimate for the current state. Then, an exponentially weighted sum of the temporal difference errors is performed to obtain the generalized advantage estimate: In the formula, Indicates time step t The generalized advantage estimate is used to measure the degree of advantage of the current action relative to the average action; The decay parameter represents the generalized advantage estimation, used to achieve a balance between bias and variance, and its value ranges from [0,1]. This represents the number of steps to look forward. After obtaining the advantage function, the parameters of the policy network are updated using the truncation mechanism of the near-end policy optimization algorithm. The optimization objective function is: In the formula, This represents the objective function for optimizing the policy after truncation. Represents the parameters of the policy network; The importance sampling ratio represents the difference between the current strategy and the old strategy in terms of action. The ratio of the probabilities; Indicates the value of the dominant function; This represents the cutoff range parameter, used to limit the magnitude of policy updates; This represents a truncation function that limits the importance sampling ratio to a certain value. Within the specified interval. This truncation mechanism effectively limits the magnitude of policy changes during each parameter update, improving policy performance while ensuring training stability.

[0055] Understandably, this embodiment separates policy learning from value estimation through the Actor-Critic architecture, improving the stability and convergence speed of reinforcement learning training; it uses a self-attention mechanism to enable adaptive interaction between different features in the state vector, improving the policy network's ability to represent complex environmental states; it uses residual connections to enhance features while preserving original state information, improving the robustness and training stability of the policy network; and it uses generalized advantage estimation to calculate the advantage function, achieving a balance between bias and variance, thus improving the accuracy of policy gradient estimation.

[0056] Based on the above technical solutions, those skilled in the art can make corresponding equivalent improvements according to the specific characteristics of the application scenario or system requirements. For example, the number of fully connected layers in the Actor network and Critic network can be adjusted to three or four layers to cope with a larger state space; or the parameters in the generalized advantage estimation can be adjusted to balance bias and variance; or the ReLU activation function can be replaced with GELU or LeakyReLU activation function to improve gradient flow.

[0057] Preferably, in step S230, the step of employing an experience playback mechanism and a course learning strategy during the training process, and dynamically adjusting the delay time of each control group through incremental updates, includes: Step S235: Before the formal training begins, a model-independent meta-learning framework is used to perform meta-training on multiple office scenario tasks to obtain the meta-initial parameters of the policy network, and the meta-initial parameters are used as the initial parameters of the policy network. Step S236: During the formal training process, the training of the policy network is iteratively executed in order from a single standardized scenario to a variety of scenarios with increasing complexity. Step S237: When updating the parameters of the policy network in each iteration, a priority experience replay mechanism is adopted to assign sampling priority to the experience samples according to the temporal difference error of each experience sample, and training samples are extracted from the experience replay pool according to the sampling priority. Step S238: After each update of the parameters of the policy network, the action vector is incrementally added to the current delay time, and the updated delay time is limited to a preset effective range by a truncation function.

[0058] For details, see Figure 6 As shown, in a specific embodiment of the present invention, during the construction and training of the policy network, a model-independent meta-learning framework is first employed to perform meta-training on multiple office scenario tasks with different spatial layouts and personnel activity patterns, in order to find an initial policy parameter that can quickly adapt to new tasks. The inner loop of the meta-training performs meta-training on each task. Perform one step of gradient descent: In the formula, Indicates task The parameters obtained after one step of gradient descent are as follows. Indicates the current meta-parameter. This represents the learning rate of the inner loop. Indicates task loss function on, The parameter is The policy network. The meta-objective is to minimize the expected loss of the adapted parameters over all tasks: In the formula, This indicates the distribution of tasks in an office setting.

[0059] Meta-learning pre-training allows the policy network to acquire good initial parameters, enabling it to quickly adapt to new office areas with only a small amount of interaction data, thus shortening the system's cold start time. During formal training, a course-based learning strategy is adopted, training in sequence from a single standardized scenario to a variety of scenarios with increasing complexity. Initially, training is conducted only in standard office areas, normal working hours, and scenarios with medium personnel density. Subsequently, complex scenarios such as mixed functional areas, irregular workstation layouts, and high-flow areas are gradually introduced to effectively prevent the policy from getting stuck in local optima in the initial stage.

[0060] Simultaneously, a priority experience replay mechanism is adopted. In the experience replay pool, each experience sample is updated from its current state. Execution of actions Receive rewards and the next state Composition, and assigning priority scores to each sample: In the formula, This represents the priority score of the empirical sample; Indicates an immediate reward; Indicates the discount factor; and These represent the Critic network's value estimates for the next state and the current state, respectively. This is the temporal difference error, reflecting the learning value of the sample; This serves as a priority index, controlling the degree of influence of priorities. This is a preset small constant to prevent samples with priority of zero from being left unsampled. The sampling probability is proportional to the priority. By sampling empirical samples with larger time-series differential errors more frequently, and prioritizing experience replay, the training focuses on state transitions that the current policy has not yet fully grasped, such as scenarios where personnel are still present but sensors occasionally miss detections, thereby accelerating policy convergence. Through continuous interaction and trial and error with the office environment, the policy network gradually learns the optimal delay strategy that adapts to the activity patterns of personnel in each control group, causing the delay time to gradually converge from the initial default value to the optimal value that balances comfort and energy saving.

[0061] Understandably, this embodiment uses a model-independent meta-learning framework to perform meta-training on multiple office scenario tasks to obtain meta-initial parameters, enabling the policy network to quickly adapt when deployed in a new office area without having to train from scratch, thus improving the system's cold start efficiency. By organizing the training of the policy learning strategy in a sequence from a single standardized scenario to a variety of scenarios with increasing complexity, the policy network is prevented from getting stuck in local optima in the initial stage, improving the generalization ability and robustness of the final policy in a changing office environment. Through a priority experience replay mechanism, sampling priority is allocated according to the temporal difference error, so that training resources are concentrated on state transitions that the current policy has not yet fully mastered, improving the learning efficiency and convergence speed of the policy network for difficult samples.

[0062] Based on the above technical solutions, those skilled in the art can make corresponding equivalent improvements according to the specific characteristics of the application scenario or system requirements. For example, the scenario complexity progression method in the course learning strategy can be replaced by dividing according to scenario type to dividing according to personnel density or activity frequency range to adapt to different training data organization needs; or the priority index in the priority experience replay mechanism can be adjusted to control the balance between priority sampling and uniform sampling; or priority experience replay can be replaced by ranking-based experience replay to simplify priority calculation; or the number of office scenario tasks used in the meta-training stage can be increased or decreased according to the richness of pre-training data to adapt to different pre-training resource conditions.

[0063] Preferably, step S300 includes: Step S310: Based on the multimodal sensor data and the recorded user historical behavior, construct time series features, which include time encoding, personnel presence status, ambient light intensity change rate, and user manual adjustment records; Step S320: Input the time series features into a long short-term memory network to obtain hidden states, and predict the probability of user demand for lighting adjustment and the expected gradation speed in the future time period based on the hidden states; Step S330: When the demand probability exceeds a preset threshold, the lighting transition speed and the lighting extinguishing transition speed are calculated by combining the expected transition speed, the difference between ambient light and target brightness, and the current time period information. Step S340: Obtain the user's implicit feedback and explicit satisfaction rating on the lighting changes, generate a comfort reward signal based on this, and feed the comfort reward signal back to the reinforcement learning algorithm to continuously optimize the gradual speed adjustment strategy.

[0064] For details, see Figure 7 As shown, in a specific embodiment of the present invention, the light transition speed includes the lighting speed and the turning-off speed, which are key parameters affecting user visual comfort. Too rapid a change in brightness can cause glare and visual fatigue, while too slow a change may cause users to wait too long when they need lighting. To achieve adaptive adjustment of the transition speed, a time-series feature is first constructed based on multimodal sensor data and recorded user historical behavior. The feature sequence input into the Long Short-Term Memory network is... , including the past T Environmental observations and user behavior records at each time step. Feature vector for each time step. Includes: the sine code of the current hour Sum and cosine encoding Based on the periodicity of capture time, the unique thermal encoding of the day of the week, the presence status of people in the area, and the rate of change of ambient light intensity. The system also records user manual adjustments to the lights over the past hour. Long Short-Term Memory (LSTM) networks encode temporal features into fixed-dimensional hidden state vectors. Then, the probability of user demand for lighting adjustment and the expected gradation speed are predicted in the near future through a fully connected layer: In the formula, This represents the hidden state vector output by the Long Short-Term Memory network; Represents the forward propagation function of a Long Short-Term Memory (LSTM) network; This represents the time series features of the input; This represents the predicted probability of a user's need for light adjustment, with a value range of [0,1]. This represents the sigmoid activation function, used to map the output to a probability interval; and These represent the weight matrix and bias vector of the demand probability prediction layer, respectively. This represents the predicted user expectation rate of change, with a value range of []. 1,1]; Represents the hyperbolic tangent activation function; and These represent the weight matrix and bias vector of the expected gradual velocity prediction layer, respectively.

[0065] Based on the predicted output of a Long Short-Term Memory (LSTM) network, a two-stage decision-making process is employed to determine the gradual change rate. The first stage is demand forecasting, where the demand probability... Exceeding the preset threshold When the system anticipates that the user will need to adjust the lighting, it enters a pre-ready state. The second stage involves speed calculation, taking into account the user's desired gradual change speed. The difference between ambient light and target brightness Based on the current time period information, calculate the light-up transition speed: In the formula, This represents the calculated speed at which the lights gradually change color; Indicates the basic gradient time; The weighting coefficients represent the user's desired gradation speed; The weighting coefficient represents the brightness difference. This represents the weighting coefficient for the morning period; This represents the difference between ambient light and target brightness. This represents a morning time indicator function, which takes a value of 1 during the morning time and a value of 0 during other times. This indicates the amount of compensation accelerated in the morning; and These represent the minimum and maximum allowable values ​​for the gradual change speed, respectively. This represents a truncation function that restricts the calculation result to... Within the interval, the calculation of the lamp-off speed uses a formula that is symmetrical to but independently parameterized as the lamp-on speed.

[0066] The adaptive optimization of the gradient speed relies on a quantitative assessment of user comfort. First, multi-source comfort feedback signals are collected, including implicit and explicit feedback. Implicit feedback indirectly infers comfort by analyzing the user's manual adjustment behavior after the automatic lighting change. If the user manually adjusts the light shortly after the automatic change, it indicates that the current gradient speed may be inappropriate; if the user never intervenes in the automatic dimming process, it indicates that the current setting is acceptable. Explicit feedback collects user-submitted satisfaction ratings through terminal applications or workstation panels. Let the comfort rating be... The comfort bonus is calculated using the following formula: In the formula, This represents the comfort bonus component, with a value range of []. 1,1]; This represents the user's satisfaction rating. A positive reward is given when the rating is above the median, and a negative reward is given when the rating is below the median. The comfort reward signal is fed back to the reinforcement learning algorithm to drive the continuous optimization of the gradual speed strategy.

[0067] Understandably, this embodiment uses a long short-term memory network to perform temporal modeling of users' lighting demand patterns, capturing the temporal regularity and periodicity of user behavior, and improving the accuracy of predicting future lighting adjustment needs. A two-stage decision-making process decouples the demand probability from the calculation of the gradual change speed, triggering speed adjustment only when the predicted demand exceeds a threshold, improving system response efficiency and avoiding unnecessary parameter fluctuations. The optimal gradual change speed is calculated by comprehensively considering the user's desired gradual change speed, the difference in ambient brightness, and time-of-day characteristics, improving the matching degree between lighting changes and user visual comfort. A multi-source comfort feedback closed-loop mechanism transforms implicit user behavior and explicit evaluations into reward signals, improving the adaptability of the gradual change speed strategy to users' personalized preferences.

[0068] Based on the above technical solutions, those skilled in the art can make corresponding equivalent improvements according to the specific characteristics of the application scenario or system requirements. For example, the preset threshold of the demand probability can be adjusted in a personalized manner according to the sensitivity of user feedback; or the calculation method of comfort reward can be changed from linear mapping to nonlinear mapping to adapt to the deviation of different user rating distributions; or the judgment time window of implicit feedback can be adjusted to adapt to the difference in reaction speed of different users to changes in light.

[0069] Preferably, step S400 includes: Step S410: Perform weighted fusion on the PIR sensor data, microwave radar sensor data and ambient light sensor data in the multimodal sensor data, calculate the comprehensive probability of the presence of personnel through learnable weights, and determine the presence of personnel when the comprehensive probability is higher than a preset threshold, so as to obtain the personnel presence status. Step S420: Obtain environmental parameters from the multimodal sensor data, establish a multivariate regression influence model between the sensor's nominal sensing distance and the environmental parameters, and obtain environmental correction coefficients; the environmental parameters include ambient temperature, humidity, airflow, and obstacle density; Step S430: Use a scene classifier to identify scene categories based on current time, personnel density, and historical activity patterns, and determine a baseline sensing distance based on the scene category; Step S440: Calculate the real-time trigger accuracy and false trigger rate based on the personnel presence status, and fine-tune the baseline sensing distance online to obtain the effective sensing distance.

[0070] For details, see Figure 8 As shown, in a specific embodiment of the present invention, the multimodal sensor data is first fused to accurately determine the presence status of personnel, and the PIR sensor output is then processed. Microwave radar sensor output and the change in light intensity detected by the ambient light sensor We perform weighted fusion and calculate the overall probability of a person's presence using learnable weights: In the formula, This represents the overall probability of the presence of personnel. This is the status output of the PIR sensor. This is the status output of the radar sensor, with a value of 0 or 1. This represents the change in ambient light intensity, and the value changes significantly when people enter and cause obstruction. Learnable fusion weights for PIR sensors; Learnable fusion weights for radar sensors; Learnable fusion weights for ambient light sensors; b This is a bias term. When... The presence of people is determined at that time, thus obtaining a reliable status of their presence.

[0071] Then, the ambient temperature is extracted from the multimodal sensor data. ,humidity airflow speed and obstacle density Environmental parameters, etc., to establish the sensor's nominal sensing distance The environmental correction function is obtained by using a multiple regression model to analyze the effects of these environmental parameters. This function is obtained by fitting historical data and outputs a sensing distance correction coefficient, thus expressing the effective sensing distance as the product of the nominal distance and the correction coefficient: In the formula, Indicates the effective sensing distance after environmental correction; This indicates the sensor's nominal sensing distance under standard conditions; This represents the environment correction function, which takes various environment parameters as input and outputs correction coefficients.

[0072] Next, a scene classifier is used to identify scene categories based on current time, population density, and historical activity patterns. The scene classifier employs a softmax regression model, using state feature embeddings to represent the scene categories. Given the input, output the probability distribution for each scenario: In the formula, A probability distribution vector representing the scene category; An embedding representation of state features; This represents the weight matrix of the scene classifier; This represents the bias vector of the scene classifier. Scene categories include high-density office, standard office, low-density office, transition period, and idle period, etc. Each scene corresponds to a set of pre-set baseline sensing distance parameters, and then the corresponding baseline sensing distance is selected based on the classification results.

[0073] Finally, the real-time trigger accuracy is calculated based on the obtained personnel presence status. and false trigger rate ,in To ensure the correct number of triggers, This refers to the number of times the event was accidentally triggered. This represents the total number of trigger checks. When the trigger accuracy falls below a preset lower threshold... Increase the sensing distance when the false trigger rate exceeds the preset upper limit threshold. The sensing distance is reduced in time, and the reference sensing distance is fine-tuned online. The adjustment amount is calculated using the following formula: In the formula, Indicates the online fine-tuning amount of the sensing distance; To trigger an adjustment of the step size parameter to indicate insufficient accuracy, To address the issue of excessively high false trigger rates, adjust the step size parameter. This represents the lower threshold for trigger accuracy. This represents the upper limit threshold of the false trigger rate. This online fine-tuning mechanism continuously converges the effective sensing distance to the optimal value adapted to the current environment.

[0074] Understandably, this embodiment determines the presence of personnel by using multi-sensor weighted fusion, reducing the probability of false detection and missed detection by a single sensor, and improving the accuracy and robustness of personnel detection. By establishing a multivariate regression model of the influence of sensing distance on temperature, humidity, airflow, and obstacle density, the sensing distance is automatically corrected as the environment changes, improving the sensor's adaptability under different physical conditions. By using a scene classifier to identify scene categories based on time, personnel density, and historical patterns and setting a benchmark sensing distance, the sensing range is matched with the actual needs of different office scenarios, improving the system's intelligence level. Through an online fine-tuning mechanism for real-time trigger accuracy and false trigger rate, the sensing distance continuously converges to the optimal value, improving the stability and reliability of the system in long-term operation in dynamic environments.

[0075] Based on the above technical solutions, those skilled in the art can make corresponding equivalent improvements according to the specific characteristics of the application scenario or system requirements. For example, the fusion method of PIR sensor and microwave radar sensor can be replaced by attention fusion to dynamically adjust the contribution of each sensor; or the adjustment step size parameter in online fine-tuning can be set to adaptive change to accelerate the convergence speed; or the thresholds of trigger accuracy and false trigger rate can be dynamically set according to the historical data distribution to balance sensitivity and false alarm rate.

[0076] Preferably, step S500 includes: Step S510: Construct a shared feature extraction network to extract shared representations from the multimodal sensor data to support the determination of target control parameters, the target control parameters including the results of the automatic partitioning and grouping, the dynamically adjusted delay time, the lighting gradation speed and the lighting gradation speed of the lamps, and the dynamically adjusted effective sensing distance; Step S520: Based on the shared representation, perform multi-task joint training on the target control parameters. During the training process, an uncertainty weighting method is used to automatically balance the loss weights of each task, and a penalty function method is introduced to encode the collaborative constraints between the control parameters. Step S530: Generate the lighting control command by combining the target control parameters output after the multi-task joint training.

[0077] For details, see Figure 9As shown, in a specific embodiment of the present invention, four functional modules—partitioning and grouping, delay optimization, gradual speed adjustment, and sensing distance adjustment—are integrated into an end-to-end joint optimization framework. Joint optimization of the four target control parameters is achieved through shared representation learning and multi-task collaborative training. First, a shared feature extraction network is constructed. This network adopts an encoder architecture, consisting of two one-dimensional convolutional layers and one bidirectional long short-term memory network, used to extract shared representations valuable for all control tasks from the original sensor time series. The original sensor time series... The input is a shared feature extraction network, which captures local temporal patterns through two one-dimensional convolutional layers, and then a bidirectional long short-term memory network captures long-term temporal dependencies, finally outputting a shared representation vector: In the formula, This represents the raw, input sensor time-series data. This represents a one-dimensional convolution operation used to extract feature patterns within a local time window; This represents a bidirectional long short-term memory network that encodes sequences in both forward and reverse directions to capture complete contextual temporal dependencies; Denotes the extracted shared representation vector, where The shared representation vector is fed in parallel into four task-specific policy head networks, responsible for outputting partitioning / grouping decisions, delay time parameters, gradual velocity parameters, and sensing distance parameters, respectively. Through the shared feature extraction network, parameter sharing and cross-task information transfer are achieved between tasks, reducing the total number of parameters and the risk of overfitting, and enabling explicit modeling of the synergistic effects between different parameters.

[0078] During training, the four control tasks employ a multi-task collaborative training framework, and the total loss function is a weighted sum of the losses from each task: In the formula, This represents the total loss function value for multi-task joint training; Indicates the partitioning and grouping loss; This indicates the loss due to delayed control. Indicates the loss of gradual velocity; This indicates the loss of sensing distance; Indicates the first i The weighting coefficients for each task; This represents the regularization term, used to prevent overfitting. To automatically balance the loss weights for each task, an uncertainty-weighted method is used, adaptively adjusting the weights based on the homoscedastic uncertainty of each task. In the formula, Indicates the firsti The uncertainty parameters of each task are learnable variables; This indicates the automatic weighting factor for the task; tasks with higher uncertainty automatically receive lower weights. This represents a regularization term to prevent the uncertainty parameter from degenerating to zero. This uncertainty-weighted method avoids the drag on overall training from difficult tasks, enabling all tasks to converge collaboratively at a balanced pace.

[0079] Simultaneously, a penalty function method is introduced in the joint optimization to explicitly encode the collaborative constraints between control parameters; the constraints include that the effective sensing distance difference of each lamp within the same control group does not exceed a preset threshold, and that the gradual change speed corresponding to different spatial function types is constrained within their respective preset ranges. For the first... k Inequality constraints in Representing the set of all learnable parameters, the penalty function transforms a constrained optimization problem into an unconstrained optimization problem. In the formula, This represents the total loss function after introducing constraints; Indicates the first k An inequality constraint function whose value is negative or zero when the parameters satisfy the constraints; Indicates the first k The penalty function coefficients for each constraint; This indicates that a positive value is taken when a constraint is violated, and zero is taken otherwise.

[0080] When control parameters violate constraints, the penalty function term generates additional loss penalties, guiding the optimization process to converge in the direction that satisfies the constraints. Finally, by combining the partitioning and grouping results, delay time, gradient speed, and sensing distance output after multi-task joint training, a unified lighting control command is generated and issued to each lamp for execution.

[0081] Understandably, this embodiment unifies the extraction of low-level features from four control tasks into a shared representation through a shared feature extraction network, thereby improving the degree of information sharing among tasks and the parameter efficiency of the model; it automatically balances the loss weights of each task through an uncertainty weighting method, avoiding the tedious and difficult task of manual parameter tuning and its negative impact on the overall training, thus improving the stability and convergence efficiency of multi-task joint training; and it explicitly encodes the physical and scenario constraints between control parameters through a penalty function method, so that the joint optimization results naturally meet the collaborative constraint requirements in actual deployment, thereby improving the engineering feasibility and control coordination of the system.

[0082] Based on the above technical solutions, those skilled in the art can make corresponding equivalent improvements according to the specific characteristics of the application scenario or system requirements. For example, the uncertainty weighting method can be replaced with the gradient normalization method to balance the learning speed of each task from the perspective of gradient magnitude; or the initial value of the task weight coefficient can be preset according to the historical convergence speed of each task to accelerate the convergence in the early stage of training.

[0083] See Figure 10 The second aspect of this invention proposes an office lighting fixture configuration system based on deep reinforcement learning and graph neural networks, comprising: The lighting fixture partitioning and grouping module is used to acquire multimodal sensor data deployed in the office environment, and construct a lighting fixture map based on the spatial coordinates of the lighting fixtures and personnel flow information extracted from the multimodal sensor data; the module learns the embedding representation of each node in the lighting fixture map through a graph neural network to automatically partition and group the lighting fixtures to form multiple lighting fixture control groups; The delay adaptive optimization module is used to input the data from the lighting control group and the multimodal sensor, and continuously learn and dynamically adjust the triggerless lighting delay time of each lighting control group through reinforcement learning algorithm; The gradual speed adjustment module is used to predict the user's lighting needs based on the multimodal sensor data and the user's historical behavior using a time series model, and to adaptively adjust the lighting speed and the turning-off speed of the lamps according to the prediction results. The sensing distance adjustment module is used to integrate environmental parameters from the multimodal sensor data and dynamically adjust the effective sensing distance of each lamp or sensor in the control group. The joint optimization and control module is used to integrate the results of the automatic partitioning and grouping, the delay time of the non-trigger lighting, the lighting gradation speed and the lighting gradation speed of the lamps, and the effective sensing distance to generate lighting control commands to control the switching on and off and brightness adjustment of the corresponding lamps.

[0084] A third aspect of the present invention also provides a storage medium storing an office lighting fixture configuration processing program based on deep reinforcement learning and graph neural networks. When the office lighting fixture configuration program based on deep reinforcement learning and graph neural networks is executed by a processor, it implements the steps of the office lighting fixture configuration method based on deep reinforcement learning and graph neural networks as described in any of the above embodiments.

[0085] See Figure 11 As shown, a fourth aspect of the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the office lighting fixture configuration method based on deep reinforcement learning and graph neural networks as described in any embodiment of the first aspect.

[0086] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. This processor is the control center of the office lighting fixture configuration system based on deep reinforcement learning and graph neural networks, connecting various parts of the operational device through various interfaces and lines.

[0087] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the office lighting configuration system based on deep reinforcement learning and graph neural networks by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0088] Compared with the prior art, the beneficial effects of the present invention include at least the following: This invention provides an office lighting fixture configuration method and system based on deep reinforcement learning and graph neural networks. It automatically learns the spatial and functional relationships between lighting fixtures using graph neural networks and combines this with spectral clustering to achieve manual partitioning and grouping, improving deployment efficiency and adapting to dynamic changes in office space layout. Reinforcement learning transforms the latency from a fixed configuration to a context-aware dynamic strategy, reducing ineffective lighting energy consumption while ensuring user comfort. Time series models predict user lighting needs and adaptively adjust the gradation speed accordingly, improving visual comfort and user experience of lighting changes. Multi-sensor fusion and environmental impact modeling dynamically adjust the sensing distance, improving the accuracy and robustness of personnel detection and reducing false and missed triggers. Shared representation learning and multi-task joint optimization achieve end-to-end coordination of four parameters, enabling the system to automatically converge to the globally optimal configuration without manual intervention.

[0089] Furthermore, this invention also improves the alignment between partitioning and grouping and actual usage patterns by constructing a lighting graph through the fusion of spatial proximity and personnel flow edge weights; enhances the representation ability of complex spatial layouts by automatically learning neighbor node weights through a graph attention network; improves the flexibility and adaptability of system deployment by automatically determining the optimal number of groups through contour coefficients; improves the hierarchical representation ability of the embedding representation of the topology by aggregating multi-hop neighbor information layer by layer through a three-layer graph attention network; improves the discriminative power and grouping accuracy of the embedding representation by reconstructing the pre-training task through self-supervised graphs; improves the adaptability of the delay policy to dynamic environments by modeling delay control as a Markov decision process and designing a composite reward function; improves the training stability and representation ability of the policy network by enhancing state representation through an Actor-Critic architecture and self-attention mechanism; improves the accuracy and update stability of policy gradient estimation by updating parameters through generalized advantage estimation and PPO truncation mechanism; improves the cold start efficiency of new scenario deployment by obtaining meta-initial parameters through model-independent meta-learning; and improves the efficiency of cold start deployment in new scenarios by learning through courses and... Prioritize experience replay training to improve the generalization ability of the strategy and the learning efficiency of difficult samples; improve the smoothness and safety of the control strategy by adjusting the delay time through incremental updates and truncation functions; improve the predictability and accuracy of gradual speed adjustment by predicting user lighting needs through long short-term memory networks; improve the matching degree between lighting changes and user personalized preferences by optimizing the gradual strategy through multi-source comfort feedback closed loop; improve the accuracy and robustness of personnel detection by determining the presence status of personnel through multi-sensor weighted fusion; improve the adaptability of sensors in different physical environments and scenarios by dynamically adjusting the sensing distance through environmental multivariate regression models and scene classifiers; improve the long-term stability of sensing distance in dynamic environments by fine-tuning the real-time trigger accuracy and false trigger rate online; improve information sharing and parameter efficiency among control tasks through shared feature extraction and multi-task joint training; improve the stability and convergence efficiency of multi-task training by automatically balancing task losses through uncertainty weighting; and improve the engineering feasibility and control coordination of joint optimization results by encoding parameter collaborative constraints through penalty function method.

[0090] In summary, the present invention proposes an office lighting fixture configuration method and system based on deep reinforcement learning and graph neural networks, which solves the technical problems of complex manual configuration and difficulty in adapting static configuration to dynamic environmental changes in existing office lighting systems.

[0091] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention. Equivalent structural transformations made using the description and drawings of the present invention, or direct / indirect applications in other related technical fields, are all included within the scope of patent protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

[0092] It should be noted that the embodiments implemented in the office lighting fixture configuration system based on deep reinforcement learning and graph neural networks in this invention can be referenced in conjunction with the embodiments implemented in the office lighting fixture configuration method based on deep reinforcement learning and graph neural networks. These embodiments will not be described in detail here.

[0093] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for configuring office lighting fixtures based on deep reinforcement learning and graph neural networks, characterized in that, include: Step S100: Acquire multimodal sensor data deployed in the office environment, and construct a lighting fixture map based on the spatial coordinates of the lighting fixtures and the personnel flow information extracted from the multimodal sensor data; The embedding representation of each node in the luminaire graph is learned by a graph neural network to automatically partition and group the luminaires into multiple luminaire control groups; Step S200: Using the data from the lighting control group and the multimodal sensor as input, continuously learn and dynamically adjust the triggerless lighting delay time of each lighting control group through reinforcement learning algorithm; Step S300: Based on the multimodal sensor data and user historical behavior, use a time series model to predict user lighting needs, and adaptively adjust the lighting speed and the extinguishing speed of the lamps according to the prediction results. Step S400: Integrate the environmental parameters from the multimodal sensor data to dynamically adjust the effective sensing distance of each lamp or sensor in the control group; Step S500: At least by combining the results of the automatic zoning and grouping, the non-trigger lighting delay time, the lighting gradation speed and the lighting gradation speed of the lamps, and the effective sensing distance, a lighting control command is generated to control the switching on and off and brightness adjustment of the corresponding lamps.

2. The office lighting fixture configuration method based on deep reinforcement learning and graph neural networks as described in claim 1, characterized in that, Step S100 includes: Step S110: Calculate the Euclidean distance between the lamps based on their spatial coordinates, establish spatial proximity edges between two lamps whose distance is less than a preset spatial threshold, and determine the edge weights of the spatial proximity edges based on the Gaussian kernel function of the distance. Step S120: Based on the frequency of personnel movement between different lighting areas extracted from the multimodal sensor data, determine the functional association edge weights used to characterize the degree of personnel flow association between lighting fixtures; Step S130: The edge weights of the spatially adjacent edges and the functionally associated edge weights are weighted and fused to construct the adjacency matrix of the lighting graph; Step S140: Use a graph attention network to learn node representations for the lighting graph, and aggregate neighbor node information through a multi-head attention mechanism to obtain low-dimensional embedding representations of each lighting node; Step S150: Construct a similarity matrix based on the low-dimensional embedding representation, use spectral clustering algorithm to cluster the lamp nodes, and calculate the contour coefficient under different group numbers. Take the group number corresponding to the maximum contour coefficient as the optimal group number to generate the multiple lamp control groups.

3. The office lighting fixture configuration method based on deep reinforcement learning and graph neural networks as described in claim 2, characterized in that, Step S140 includes: Step S141: Input the original features of each lamp node into the first graph attention layer, and map the original features to the hidden space through a learnable weight matrix to obtain the first hidden representation; Step S142: Input the first hidden representation into the second graph attention layer, calculate the attention coefficient between each lamp node and its neighbor nodes through a multi-head attention mechanism, and aggregate the information of multi-hop neighbor nodes based on the attention coefficients to obtain the second hidden representation; Step S143: Input the second hidden representation into the third graph attention layer, and project the aggregated node features into the low-dimensional embedding space through a single-head attention mechanism to obtain the low-dimensional embedding representation of each lamp node; Step S144: During the training of the graph attention network, the reconstructing of the adjacency matrix using the low-dimensional embedding representation is used as a self-supervised pre-training task. This makes the light fixture nodes with close spatial proximity or personnel flow associations closer in the low-dimensional embedding space. The reconstruction loss adopts the form of cross-entropy. In the formula, This represents the reconstruction loss; This represents an edge in the lighting fixture diagram; This represents the edge weight between node i and node j in the adjacency matrix; Represents the low-dimensional embedding representation vector of the lamp node j; Represents the low-dimensional embedding representation vector of lamp node i; This represents the inner product of the embedding vectors of lamp node i and lamp node j; This represents the sigmoid activation function, used to map the inner product of the embedding vectors to the predicted probability that there is a connecting edge between two nodes; Let represent the set of edges in the luminaire graph.

4. The office lighting fixture configuration method based on deep reinforcement learning and graph neural networks as described in claim 1, characterized in that, Step S200 includes: Step S210: For each lighting control group, model the control problem of the delay time of the lighting without trigger as a Markov decision process, define a state vector that includes the environmental state, user state, lighting state and scene context, and use the adjustment amount of the delay time as the action vector; Step S220: Design a reward function to give a positive reward for the situation that there is no human activity in the area corresponding to the control group after the delay ends, and to impose a negative penalty for the behavior of manually turning on the lights due to the delay being too short. Step S230: Construct a policy network with the state vector as input and the action vector as output, calculate the advantage function using the reward function, and train the policy network based on the advantage function using a proximal policy optimization algorithm; during the training process, an experience replay mechanism and a course learning strategy are adopted, and the delay time of each control group is dynamically adjusted through incremental updates.

5. The office lighting fixture configuration method based on deep reinforcement learning and graph neural networks as described in claim 4, characterized in that, In step S230, the steps of constructing a policy network with the state vector as input and the action vector as output, calculating the advantage function using the reward function, and training the policy network based on the advantage function using a proximal policy optimization algorithm include: Step S231: Construct the policy network of the Actor-Critic architecture, with the state vector as input, wherein the Actor network outputs the mean and logarithmic variance of the action vector, and the Critic network outputs the state value estimate; Step S232: A self-attention mechanism is introduced after the input layer of the policy network. The state vector is mapped to a query matrix, a key matrix, and a value matrix through a learnable projection matrix. After scaling dot product attention calculation, the state vector is residually connected to the original state vector to obtain an enhanced state representation. Step S233: Input the enhanced state representation into the Actor network and the Critic network respectively, the Actor network outputs the mean and logarithmic variance of the action vector, and the Critic network outputs the state value estimate; Step S234: Based on the state value estimate output by the Critic network and the reward function, calculate the advantage function using generalized advantage estimation; use the advantage function to update the parameters of the policy network through the truncation mechanism of the near-end policy optimization algorithm.

6. The office lighting fixture configuration method based on deep reinforcement learning and graph neural networks as described in claim 4, characterized in that, In step S230, the step of employing an experience replay mechanism and a course learning strategy during the training process, and dynamically adjusting the delay time of each control group through incremental updates, includes: Step S235: Before the formal training begins, a model-independent meta-learning framework is used to perform meta-training on multiple office scenario tasks to obtain the meta-initial parameters of the policy network, and the meta-initial parameters are used as the initial parameters of the policy network. Step S236: During the formal training process, the training of the policy network is iteratively executed in order from a single standardized scenario to a variety of scenarios with increasing complexity. Step S237: When updating the parameters of the policy network in each iteration, a priority experience replay mechanism is adopted to assign sampling priority to the experience samples according to the temporal difference error of each experience sample, and training samples are extracted from the experience replay pool according to the sampling priority. Step S238: After each update of the parameters of the policy network, the action vector is incrementally added to the current delay time, and the updated delay time is limited to a preset effective range by a truncation function.

7. The office lighting fixture configuration method based on deep reinforcement learning and graph neural networks as described in claim 1, characterized in that, Step S300 includes: Step S310: Based on the multimodal sensor data and the recorded user historical behavior, construct time series features, which include time encoding, personnel presence status, ambient light intensity change rate, and user manual adjustment records; Step S320: Input the time series features into a long short-term memory network to obtain hidden states, and predict the probability of user demand for lighting adjustment and the expected gradation speed in the future time period based on the hidden states; Step S330: When the demand probability exceeds a preset threshold, the lighting transition speed and the lighting extinguishing transition speed are calculated by combining the expected transition speed, the difference between ambient light and target brightness, and the current time period information. Step S340: Obtain the user's implicit feedback and explicit satisfaction rating on the lighting changes, generate a comfort reward signal based on this, and feed the comfort reward signal back to the reinforcement learning algorithm to continuously optimize the gradual speed adjustment strategy.

8. The office lighting fixture configuration method based on deep reinforcement learning and graph neural networks as described in claim 1, characterized in that, Step S400 includes: Step S410: Perform weighted fusion on the PIR sensor data, microwave radar sensor data and ambient light sensor data in the multimodal sensor data, calculate the comprehensive probability of the presence of personnel through learnable weights, and determine the presence of personnel when the comprehensive probability is higher than a preset threshold, so as to obtain the personnel presence status. Step S420: Obtain environmental parameters from the multimodal sensor data, establish a multivariate regression influence model between the sensor's nominal sensing distance and the environmental parameters, and obtain environmental correction coefficients; the environmental parameters include ambient temperature, humidity, airflow, and obstacle density; Step S430: Use a scene classifier to identify scene categories based on current time, personnel density, and historical activity patterns, and determine a baseline sensing distance based on the scene category; Step S440: Calculate the real-time trigger accuracy and false trigger rate based on the personnel presence status, and fine-tune the baseline sensing distance online to obtain the effective sensing distance.

9. The office lighting fixture configuration method based on deep reinforcement learning and graph neural networks as described in claim 1, characterized in that, Step S500 includes: Step S510: Construct a shared feature extraction network to extract shared representations from the multimodal sensor data to support the determination of target control parameters, the target control parameters including the results of the automatic partitioning and grouping, the dynamically adjusted delay time, the lighting gradation speed and the lighting gradation speed of the lamps, and the dynamically adjusted effective sensing distance; Step S520: Based on the shared representation, perform multi-task joint training on the target control parameters. During the training process, an uncertainty weighting method is used to automatically balance the loss weights of each task, and a penalty function method is introduced to encode the collaborative constraints between the control parameters. Step S530: Generate the lighting control command by combining the target control parameters output after the multi-task joint training.

10. An office lighting fixture configuration system based on deep reinforcement learning and graph neural networks, characterized in that, include: The lighting fixture partitioning and grouping module is used to acquire multimodal sensor data deployed in the office environment, and construct a lighting fixture map based on the spatial coordinates of the lighting fixtures and personnel flow information extracted from the multimodal sensor data; the module learns the embedding representation of each node in the lighting fixture map through a graph neural network to automatically partition and group the lighting fixtures to form multiple lighting fixture control groups; The delay adaptive optimization module is used to input the data from the lighting control group and the multimodal sensor, and continuously learn and dynamically adjust the triggerless lighting delay time of each lighting control group through reinforcement learning algorithm; The gradual speed adjustment module is used to predict the user's lighting needs based on the multimodal sensor data and the user's historical behavior using a time series model, and to adaptively adjust the lighting speed and the turning-off speed of the lamps according to the prediction results. The sensing distance adjustment module is used to integrate environmental parameters from the multimodal sensor data and dynamically adjust the effective sensing distance of each lamp or sensor in the control group. The joint optimization and control module is used to integrate the results of the automatic partitioning and grouping, the delay time of the non-trigger lighting, the lighting gradation speed and the lighting gradation speed of the lamps, and the effective sensing distance to generate lighting control commands to control the switching on and off and brightness adjustment of the corresponding lamps.

Images