Building light environment neural network control system based on BIM

Abstract

The invention discloses a building light environment neural network control system based on BIM, relates to the technical field of building intelligent control, and is used for solving the problem of poor intelligent illumination adaptive control. The method comprises the following steps: constructing a structure enhanced semantic map based on a BIM model, extracting a spatial structure attribute and a communication relationship to generate a node vector and an edge weight, and describing an illumination propagation relationship; the method comprises the following steps: constructing an original state vector by fusing sensing data and structural information, carrying out time sequence modeling by adopting a gated recurrent neural network, dividing four types of subspaces to express environment evolution characteristics, identifying a lighting task scene according to a current activity and time period, matching a control target, reasoning an adjustment direction, and screening a candidate action path; according to the method, the optimal strategy is generated and executed, user feedback is collected to construct behavior response factors, user preferences are updated, a preference closed-loop mechanism is constructed, illumination control cooperatively driven by structure perception, state evolution and user preferences is achieved, and the comfort, intelligence and energy saving performance of the building light environment are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of building intelligent control, and more specifically, to a building light environment neural network control system based on BIM. Background Art

[0002] Building Information Modeling (BIM), as a digital modeling technology that integrates spatial geometric information, structural characteristics and material properties, has been widely used in architectural design, construction and operation and maintenance. BIM models provide high-precision three-dimensional information about building component orientation, window-to-wall ratio, shading structure, interior space topology, etc., providing a rich data foundation for the simulation and prediction of the building physical environment. In particular, in light environment regulation, BIM models can serve as an important basis for establishing light propagation paths, spatial illumination distribution and shading impact analysis.

[0003] Deficiencies in existing technologies: Existing lighting control technologies lack the ability to fine-grainedly model the differences between the current state and the target state, resulting in policy responses often being over-adjusted or delayed. This makes it impossible to select the optimal adjustment path based on the real-time environment and behavioral context. In the action decision-making process, the comprehensive assessment of energy consumption impact and disturbance intensity is often ignored, leading to a lack of systematic optimization of control strategies. This leads to problems such as low energy efficiency and impaired user experience in lighting adjustment behaviors. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, there is the following solution to solve the problem of poor adaptive control of intelligent lighting in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The BIM-based building light environment neural network control system includes a building lighting space division module, a lighting scene data acquisition module, a lighting strategy generation module, and a lighting feedback control module. Each module is connected through signals.

[0007] The architectural lighting space division module divides space units according to the BIM model and extracts five types of structural attributes to form node vectors. It identifies spatial connectivity to generate a Boolean matrix, and constructs a modulation function based on spatial spacing, occlusion overlap ratio, and light-transmitting cross-section to generate edge weights and form a structure-enhanced semantic map.

[0008] The lighting scene data acquisition module is used to collect sensory information, integrate spatial structural attributes to construct the original state vector, and use a gated recurrent neural network to perform temporal modeling on the sensory information, extract evolutionary features, and divide it into four subspaces: immediate perception, behavioral context, historical trajectory, and individual preference.

[0009] The lighting strategy generation module identifies the lighting task scenario and matches the control target based on the current activity type and time period. It adjusts the direction by reasoning the difference between the state and the target, screens and evaluates candidate action paths, generates the optimal strategy, and executes the control.

[0010] The lighting feedback control module collects user feedback signals, constructs a behavior response factor to determine the preference matching degree of the current control strategy, and updates and records the preference status.

[0011] In a preferred embodiment, the BIM model is divided into spatial units and five types of structural attributes are extracted to form node vectors. The specific steps are as follows:

[0012] Divide the entire building area into a group of spatial units, each of which represents a control area with independent lighting adjustment requirements;

[0013] After the spatial unit division is completed, the structural attributes are extracted, including orientation angle, surface reflectivity, light transmittance, shading complexity level and clear height;

[0014] The structural attributes are combined into a five-dimensional node structure vector to characterize the light environment structure state of each spatial unit.

[0015] In a preferred embodiment, spatial connectivity is identified to generate a Boolean matrix, and a modulation function is constructed by combining spatial spacing, occlusion overlap ratio, and light cross section to generate edge weights to form a structure-enhanced semantic graph. The specific steps include:

[0016] Extract the topological boundaries and opening positions between spaces from the BIM model, analyze the physical contact and visual distance between topological boundary components, and establish a Boolean matrix to represent the connection status between each spatial unit;

[0017] A modulation function is constructed based on the spatial distance, the occlusion-projection overlap ratio, and the minimum light-transmitting cross-section of the connection path, and is used as the edge weight input of the structure-enhanced semantic graph. Each node in the structure-enhanced semantic graph corresponds to a spatial unit.

[0018] In a preferred embodiment, the steps for collecting sensory information and fusing spatial structure attributes to construct an original state vector are as follows:

[0019] Sensors deployed in spatial units collect environmental status and user behavior data as perception information. The perception information includes the illuminance value, color temperature value, occupant status identifier, behavior tag, system time, and outdoor natural light intensity within the current spatial unit.

[0020] The perceptual information is combined with the spatial structure attributes to form an original state vector.

[0021] In a preferred embodiment, a gated recurrent neural network is used to perform temporal modeling of the perception information, extract evolutionary features, and divide the information into four subspaces: immediate perception, behavioral context, historical trajectory, and personality preference. The following steps are included:

[0022] A gated recurrent neural network is used to model the temporal sequence of sensory information. The currently embedded raw state vector is fed into the GRU unit for joint processing with the hidden state at the previous moment. Gated logic judgment is then performed to determine the change in illumination state, which is then used as the behavioral response expression of the spatial unit at the current moment.

[0023] After completing the time-recursive modeling, the state space is divided into four subspaces: immediate perception, behavioral context, historical trajectory, and personality preference.

[0024] In a preferred embodiment, after completing the time-recursive modeling, the state space is divided into four subspaces: immediate perception, behavioral context, historical trajectory, and personality preference. The specific steps are as follows:

[0025] The immediate perception subspace is used to extract the components directly related to the current sensor value;

[0026] Behavior context subspace, used to extract feature components related to behavior labels and time periods;

[0027] The historical trajectory subspace is used to analyze the state change rate and trend, and to determine whether the lighting state is drifting, stable, or abnormal;

[0028] Personalized preference subspace, which is constructed by combining user interaction records or feedback information to dynamically adjust preference parameters in the control strategy;

[0029] Strategy calls are made according to different subspaces.

[0030] In a preferred embodiment, according to the current activity type and time period, the lighting task scene is identified and the control target is matched. The specific steps are as follows:

[0031] Obtain the user's activity type and time period, and determine the lighting task scenario of the current space based on the joint rule base between the behavior type and time period;

[0032] After the current lighting task scene is identified, it is mapped to a set of lighting control targets, which include the brightness level required for the space, the color characteristics of the light source, the energy consumption control standards, and the adjustment preferences previously expressed by the user.

[0033] In a preferred embodiment, the direction is adjusted by reasoning about the difference between the state and the goal, candidate action paths are screened and evaluated, the optimal strategy is generated, and control is executed. The specific steps are as follows:

[0034] After the lighting control target is determined, the strategy generation phase begins. The difference between the current state expression and the expected target is converted into the adjustment intention, including adjustment direction identification, candidate action extraction, and optimal action path selection.

[0035] After the candidate actions are determined, they are summarized as a set of candidate action paths. The expected adjustment effect, system energy consumption change and state disturbance intensity of each action combination are evaluated to select the optimal action path.

[0036] In a preferred embodiment, user feedback signals are collected, a behavior response factor is constructed to determine the preference matching degree of the current control strategy, and the preference status is updated and recorded. The specific steps are as follows:

[0037] The difference between the current state and the target lighting state is used as the expected adjustment effect;

[0038] Based on the power model and spatial layout structure of each lamp type in the BIM, combined with the switching behavior and brightness level adjustment range involved in the action path, the overall power fluctuation value corresponding to the action combination is calculated as the system energy consumption change;

[0039] The state disturbance intensity is used to measure the extent of the intervention of the action combination on the stability of the current spatial light environment. It is obtained by analyzing the mutation rate of the state variables caused by each action per unit time and the propagation effect in the subspace.

[0040] In a preferred embodiment, user feedback signals are collected, a behavior response factor is constructed to determine the preference matching degree of the current control strategy, and the preference status is updated and recorded. The specific steps are as follows:

[0041] User feedback signals include the user's subjective evaluation of the current lighting environment through the terminal device, the user's manual adjustment of the lighting system, and changes in the user's behavior duration or behavior pattern in the current lighting environment;

[0042] According to the type, intensity, timeliness and behavior results of user feedback signals, behavior response factors are constructed one by one, and a normalized response value is formed. After receiving the user behavior response result, the current user preference state is updated.

[0043] The technical effects and advantages of the BIM-based building light environment neural network control system of the present invention are as follows:

[0044] The present invention constructs a structurally enhanced spatial semantic graph through the BIM model, combines five types of lighting-related structural attributes with spatial connectivity to establish node and edge weights, and realizes the modeling of light propagation relationships; integrates environmental perception information and structural priors to construct the original state vector, and uses a gated recurrent neural network for time series modeling, extracts evolutionary features and divides them into four subspaces to achieve a multi-dimensional semantic expression of the lighting environment; matches lighting task scenarios based on activity types and time periods, infers control targets and identifies adjustment directions, screens candidate action paths and comprehensively evaluates adjustment effects, energy consumption changes, and state disturbance intensity, generates optimal strategies, and accurately executes control actions; simultaneously collects user feedback signals to construct behavioral response factors, dynamically updates individual preference vectors, and establishes a closed-loop adjustment mechanism. The technical effect is reflected in the realization of a dynamic lighting control system that deeply couples spatial structure with user behavior, possesses refined perception, semantic modeling, and adaptive optimization capabilities, and significantly improves the comfort, intelligence, and energy efficiency of lighting. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a structural schematic diagram of the BIM-based building light environment neural network control system of the present invention. DETAILED DESCRIPTION

[0046] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0047] In order to achieve the above objectives, Figure 1 The structural diagram of the BIM-based building light environment neural network control system of the present invention is given. Specifically, it includes a building lighting space division module, a lighting scene data acquisition module, a lighting strategy generation module, and a lighting feedback control module. The modules are connected by signals.

[0048] The architectural lighting space division module divides space units according to the BIM model and extracts five types of structural attributes to form node vectors. It identifies spatial connectivity to generate a Boolean matrix, and constructs a modulation function based on spatial spacing, occlusion overlap ratio, and light-transmitting cross-section to generate edge weights and form a structure-enhanced semantic map.

[0049] The lighting scene data acquisition module is used to collect sensory information, integrate spatial structural attributes to construct the original state vector, and use a gated recurrent neural network to perform temporal modeling on the sensory information, extract evolutionary features, and divide it into four subspaces: immediate perception, behavioral context, historical trajectory, and individual preference.

[0050] The lighting strategy generation module identifies the lighting task scenario and matches the control target based on the current activity type and time period. It adjusts the direction by reasoning the difference between the state and the target, screens and evaluates candidate action paths, generates the optimal strategy, and executes the control.

[0051] The lighting feedback control module collects user feedback signals, constructs a behavior response factor to determine the preference matching degree of the current control strategy, and updates and records the preference status.

[0052] To enable neural network control systems to perceive the structural characteristics of light propagation within buildings, a semantic state map should first be established at the architectural spatial level, encompassing structural attributes, connectivity relationships, and shading effects. This map not only reflects the geometric and functional connectivity between spaces, but also encodes key structural parameters that affect light flux transmission, such as shading factors, material properties, and spatial orientation.

[0053] Step 1: Construct a structure-enhanced spatial semantic state map. The specific steps are as follows:

[0054] At the initial stage of graph construction, the entire building area needs to be divided into a limited set of spatial units. Each spatial unit represents a control area with independent lighting adjustment requirements, such as a room, corridor, or shared area. This process is based on the extraction of spatial annotation entities (such as room and zone) in the BIM model to ensure the consistency of the physical boundaries of the division. Each spatial unit exists as a node in the graph structure in subsequent processing.

[0055] After the spatial unit division is completed, the system extracts structural attributes closely related to the lighting environment from the BIM model. Structural attributes mainly include the following five categories:

[0056] The orientation angle is the angle between the main window surface or facade component of the space and the north direction, which is used to describe the main direction of the natural light path;

[0057] Surface reflectivity refers to the average reflectance coefficient of walls, ceilings, and major surfaces, which can be extracted from the BIM material definition table, such as interior wall paint, metal reflective surfaces, etc.

[0058] Light transmittance, which measures the ability of materials such as glass and translucent components to allow natural light to penetrate;

[0059] The shading complexity level is a discrete indicator extracted from the structural layer, which is used to describe whether there are light-blocking components such as blinds, sunshades, and curtain walls in the space;

[0060] Clear height, the vertical distance from the floor to the ceiling, affects the dimension and uniformity of the light diffusion path;

[0061] The above structural attributes are combined into a five-dimensional node structure vector to characterize the light environment structure state of each spatial unit. The five-dimensional node structure vector will be used in the neural network modeling stage to provide spatial structure prior information.

[0062] When constructing spatial connections, it is important to identify direct connectivity between different spatial units within a building. These connections typically take the form of doorways, open areas, or glass walls. The topological boundaries and opening locations between spaces must be extracted from the BIM model. By analyzing the physical contact and visual distance between topological boundary components, a Boolean matrix can be constructed to represent the connectivity between spatial units. This Boolean matrix is ​​then used to construct the edge structure of the semantic graph.

[0063] After analyzing the structural connection relationship, the light shading path is modeled. The light shading path considers the following three factors: spatial distance, shading projection overlap ratio, and the minimum light cross-section of the connection path;

[0064] Spatial distance represents the geometric path length required for light to propagate from one space to another. The longer the distance, the more significant the light energy attenuation. The light shielding effect of components between the two spaces (such as walls, doors, and windows) needs to be estimated based on the orientation, shape, and projected overlapping area of ​​the components. If there is a light-permeable connection between the spaces (such as a glass partition), the minimum light-permeable cross-section in the connection is considered. The smaller the value, the weaker the light transmission ability.

[0065] The above three factors jointly determine the attenuation intensity of the degree of illumination influence between spatial units. The construction logic is: when the spatial units are far apart, the occlusion overlap area is large, and the light path cross-section is small, the light transmission efficiency should be significantly reduced; on the contrary, when the spaces are adjacent, there is no occlusion between the components, or there are abundant light-transmitting structures, the lighting linkage between the two spatial units should be enhanced. To this end, a modulation function is designed to combine these three spatial properties. This modulation function has the following properties: it is sensitive to spatial distance, sensitive to occlusion overlap area, and can be nonlinearly adjusted according to the light path cross-section;

[0066] The modulation function uses the spatial distance multiplied by the square of the projected overlapping area as the suppression factor, and expresses the degree of openness of the light channel in the form of reciprocal square root, giving it a strong nonlinear attenuation characteristic. At the same time, a minimum positive constant is introduced in the calculation process to avoid the situation where the function has a singular solution when the light-transmitting area is zero. The smaller the output value of this function, the more serious the shading relationship between the spaces and the lower the possibility of light transmission.

[0067] Finally, a spatial shielding influence matrix is ​​constructed. This spatial shielding influence matrix serves as the edge weight input of the graph structure and is used to represent the structural strength of the mutual light flux between spaces. Each node in the graph corresponds to a spatial unit, and its eigenvector is the five-dimensional structural attribute vector extracted previously. The weight of each edge is determined by the calculation result of the modulation function, and the existence of the edge is determined by the connection relationship between the spaces.

[0068] Constructing a Boolean matrix is ​​to determine whether there is structural connectivity between each pair of spatial units, such as door openings, glass partitions, open areas, etc. As the basic basis for generating graph edges in the semantic state graph, the Boolean matrix can be regarded as a symmetrical two-dimensional structure with true or false values, which is used to indicate whether space i in the structure is directly connected to space j.

[0069] The Boolean matrix construction process is:

[0070] Extract all enclosed space units (such as rooms, corridors, shared areas, etc.) in the building from the BIM model, assign a unique identifier to each space, and form a space set. The total number of spaces is set to N;

[0071] Traverse all components defined as "walls," "doors," "windows," "glass curtain walls," and other types in the BIM model. For each spatial unit, extract the set of components that constitute its boundary and record the relationship between the components and the space.

[0072] For any two spaces i and j, determine whether they share at least one boundary component. If there is a set of components belonging to the common boundary of space i and space j (such as sharing a wall), proceed to the next step; otherwise, the space pair is directly marked as disconnected;

[0073] For space pairs with a common boundary, determine whether their boundaries contain components with connectivity, such as doorways or open door components, transparent glass walls or low-shading partitions, or wall components with physical openings. If at least one of the above conditions is met, the space pair is considered to have connectivity.

[0074] The connectivity of each pair of spatial cells is entered as a Boolean value (true or false) in a Boolean matrix. The matrix is ​​a symmetric matrix, and the main diagonal elements represent the connectivity of the spatial cell itself to itself, which is fixed to true.

[0075] The modulation function is used to quantify the attenuation intensity of the lighting effect between any two spatial units in a building. This function simulates the combined effects of geometric attenuation, shading interference, and path transparency restrictions that occur when natural or artificial light propagates from one space to another under real building structure conditions. This function is used to calculate the weight of the edges in the semantic graph.

[0076] The modulation function is calculated as:

[0077] S1. Calculate spatial distance. This involves extracting the centroid coordinates of each spatial unit based on the geometric description data of each spatial unit in the BIM model. In BIM structural data, spatial geometry is usually composed of multiple facets. The geometric centroid of the entire space can be obtained by weighted averaging the centroids of the facets. For any two spaces A and B, calculate the Euclidean distance between their centroid coordinates to measure the length of the transmission path from the light source to the other space. The longer the distance, the more significant the attenuation of light during propagation.

[0078] S2. Evaluate the occlusion-projection overlap ratio, which assesses the distribution of occluding components in the main direction between the two spaces. Project the two-dimensional image outline of the components from space A facing space B, and analyze the projection overlap with the receivable area corresponding to the entrance direction of space B. The occlusion ratio is defined as the ratio of the component projection overlap area to the target receiving projection area. The larger the ratio, the more severe the blockage of direct light received by space B, reflecting the degree of occlusion interference. Here, a nonlinear influence mechanism (such as squaring) is used to enhance the weight of high occlusion areas.

[0079] S3. Extract the minimum light-transmitting cross-section of the connection path. Given that there is a structural connection between the two spaces, extract all possible light-transmitting paths (such as doorways, glass partitions, ventilation windows, etc.) between the two spaces based on the BIM model and identify the location with the weakest light transmission capacity. The minimum light-transmitting cross-section refers to the optical channel area with the smallest cross-sectional area among all the connection channels. Its physical significance is that even if the spaces are relatively close, if the connection point is extremely narrow or highly shielded, the light flux transmission capacity will also be limited;

[0080] S4. After obtaining the spatial distance, the occlusion-projection overlap ratio, and the minimum light-passing cross section, a nonlinear modulation function that meets the following conditions is constructed;

[0081] When the spatial distance increases, the function value needs to decrease exponentially; when the occlusion-projection overlap ratio increases, the function value needs to be significantly weakened; when the minimum light cross section decreases, the function value should show a rapid decline characteristic; all input features should be positive;

[0082] Using spatial distance as the main modulator, its value increases linearly. A nonlinear enhancement term is constructed in the form of the square of the occlusion-projection overlap ratio, which significantly reduces the effect when the occlusion-projection overlap ratio is high. Using the minimum light-transmitting cross-section as the denominator influencing term, a fractional attenuation logic is constructed to ensure that small cross-sections lead to high impedance. At the same time, a small positive bias is introduced in the denominator to avoid function singularity caused by a zero light-transmitting area.

[0083] After generating the modulation values ​​between all spatial pairs, they are uniformly normalized to fall into a unified edge weight range (such as 0 to 1) and constructed as a set of graph edge attributes as the input of the spatial semantic graph.

[0084] In summary, a structure-enhanced semantic graph containing spatial nodes, connecting edges, node attributes and edge attributes is constructed. This graph can be used as a structural prior condition for neural network input, enabling the system to perceive the lighting impact path between spaces in lighting state modeling and control decision-making.

[0085] Step 2: Establish a context-driven, multi-layer state space modeling mechanism that can integrate spatial structure information, environmental physical state, behavioral context, and temporal rhythms in real time to form a multi-level, evolvable state representation, thereby providing a reliable semantic foundation for lighting control strategies. The specific steps are as follows:

[0086] The state perception vector is constructed by deploying multi-source sensors in the spatial unit to collect environmental status and user behavior data as perception information. This includes the following six types of information:

[0087] The illuminance value in the current space unit represents the actual light intensity under the combined effect of natural light and artificial lighting;

[0088] The current color temperature value reflects the warm or cold properties of the light source in visual perception;

[0089] Personnel status identification, provided by human body sensing equipment, determines whether there are people in the space;

[0090] Behavior tags, generated through image recognition models or user-device interactions, identify the user's current activity (e.g., working, resting);

[0091] The current system time reflects the lighting control period (such as morning rush hour, night time);

[0092] Outdoor natural light intensity, providing a lighting background reference.

[0093] This perception information is combined with the spatial structural attributes extracted in the previous stage (including spatial orientation, reflectivity, transmittance, shading complexity, and clear height) to form a raw state vector. This vector integrates the current physical state of the environment and the prior structural characteristics of the space, forming the input basis for lighting control decisions.

[0094] Since the various dimensional features in the collected original state input vectors are heterogeneous in terms of physical meaning, numerical scale, distribution form, etc., it is necessary to project the original feature vectors into a unified semantic space to generate an intermediate state representation with high expressive power.

[0095] Temporal modeling of perceptual information:

[0096] The light environment state in the building space is not only affected by the current perception information, but also closely related to the environmental evolution at several moments in the past. A gated recurrent neural network (GRU) is used for temporal state modeling.

[0097] The core idea of ​​the GRU structure is to introduce two types of gating mechanisms: reset gate and update gate, which are used to control the fusion relationship between the current input state and the historical state. During the state recursion process, the GRU automatically determines whether the current input is a key change point and adjusts the state memory path accordingly.

[0098] The specific implementation process is as follows: the current embedded original state vector is sent to the GRU unit and jointly processed with the hidden state at the previous moment. The system internally performs a set of gated logic judgments: if the current state changes significantly (such as a sudden increase in illumination or a behavior switch), the update gate will increase the input ratio; if the state is continuous and stable, the GRU will strengthen the maintenance of historical memory. Through this structure, the model can capture the dynamic laws of lighting state changes while ignoring short-term meaningless disturbances.

[0099] The hidden state vector generated by recursion, as the behavioral response expression of the spatial unit at the current moment, not only contains the current state characteristics, but also integrates the historical behavior trajectory and change trend, providing time context support for the control strategy.

[0100] After completing the time-recursive modeling, the unified hidden state representation is functionally split to meet the needs of different control tasks. The state space is divided into four subspaces with clear semantic boundaries, as follows:

[0101] The immediate perception subspace is used to extract components directly related to the current sensor value, supporting immediate control responses based on mutation judgment, such as turning off the lighting immediately after a person leaves.

[0102] The behavior context subspace is used to extract feature components related to the behavior label and time period, which is used to infer the current lighting task type (such as whether strong lighting or soft lighting is required);

[0103] The historical trajectory subspace is used to analyze the state change rate and trend, and to determine whether the lighting state is drifting, stable, or abnormal;

[0104] The personalized preference subspace is constructed by combining user interaction records or feedback information to dynamically adjust the preference parameters in the control strategy, such as whether the light is warmer or cooler.

[0105] For example, consider a conference room in a smart office building. The system has identified the space's structural information and usage tags based on the BIM model. The current time is 9:00 AM, a meeting has begun, and several people are present. The user has previously adjusted the lighting preference to "cool white light + high brightness" in the same conference room multiple times.

[0106] After modeling the time-recursive state, the system obtains a complete spatial state. This spatial state is divided into the following four subspaces, each of which provides a strategy call as follows:

[0107] Instant Perception Subspace: The current illuminance sensor value is 220 lux, which is lower than the recommended value for conferences. The human presence sensor triggers a occupancy signal. The noise level is normal, and there is no abnormal activity. The control response triggers a brightness increase strategy to maintain visual clarity.

[0108] Behavior context subspace: The current behavior label is meeting, the time period is working hours, and the environmental demand is predicted to be "high brightness + cool color temperature". The control response switches to the conference mode lighting configuration, that is, the proportion of cool white light and direct light source is increased;

[0109] Historical trajectory subspace: Over the past 30 minutes, the lighting adjustment changes slowly, the illumination remains stable, and there is no frequent adjustment behavior. The state trend is predicted to be in a stable period. The control response is that no further frequent adjustment is required and stable output is maintained.

[0110] Personal preference subspace: The user has manually adjusted the light to cool colors (color temperature > 5000K) in the past five meetings, with an average illumination setting of 350 lux. There is a deviation between the current user setting and the system default. The control response is to incorporate the user's historical preferences and automatically increase the color temperature to the preferred range to maintain a consistent experience.

[0111] Based on the status results of the above four subspaces, the system generates the following control strategies: switch the lighting mode to the conference scene preset, increase the illumination to 300 lux (to meet perception and preference), and increase the color temperature to 5200K (to meet the preference subspace), while maintaining the stability of status monitoring and not triggering redundant updates.

[0112] Step 3: Based on the task orientation and user behavior, a strategy is generated and adjusted. According to the current state and task scenario, a lighting control strategy with a clear goal orientation is inferred and converted into a specific executable lighting adjustment action. The specific steps are as follows:

[0113] In lighting control scenarios, the user's activity type and time period determine the lighting function that should be implemented in the unit space. First, the activity type within the current unit space is determined, such as working, reading, resting, or temporarily leaving. At the same time, the time period is determined based on the system clock, such as early morning, noon, evening, or late night. The system presets a joint rule base between activity type and time period to determine the lighting task scenario for the current space.

[0114] After the system identifies the current lighting task scenario, it maps it to a set of predefined lighting control targets, which usually include the required brightness level of the space, the color characteristics of the light source, the energy consumption control standards, and the user's previous adjustment preferences.

[0115] For example, when the system determines that the current behavior is "static reading at night" through the task scene recognition mechanism, it maps it to a set of multi-dimensional lighting control targets. This target set covers aspects such as brightness level, light source color characteristics, energy consumption control standards and user adjustment preferences. For example, in a shared reading area, if the system recognizes that the current time is a weekday evening, there is no natural light interference in the space, and the sensor captures that the person is in a static reading state, then according to the task and parameter mapping relationship, the illumination target is set to about 420 lux, and the color temperature range is controlled between 2900K and 3100K; at the same time, according to the energy consumption strategy during non-peak hours, the main light source is activated and the edge light strips are turned off, limiting the overall power to within 80% of the rated upper limit. Further combined with user preference records, the system can also automatically enable soft light diffusion mode to optimize the light output angle of the lamp to avoid glare interfering with the reading experience.

[0116] After the lighting control objectives are determined, the system enters the strategy generation phase. The core task of this phase is to convert the discrepancy between the current state expression and the desired goal into a control intention and select an appropriate combination of control actions to narrow this discrepancy. Strategy generation does not rely on empirical rules or simple classification logic, but instead uses structured reasoning to bidirectionally align state semantics with control objectives. The specific process includes the following steps:

[0117] The system identifies the adjustment direction by comparing the deviation trend between the current state and the expected state to determine whether the current lighting environment deviates from the target, such as insufficient light, inconsistent color temperature, or excessive energy consumption. This identification process not only considers instantaneous differences but also further integrates historical evolution trends and preferred adjustment records to ensure the stability and adaptability of the adjustment direction.

[0118] Extract candidate actions. That is, based on the adjustment direction, the system selects a set of potentially feasible action paths from the control action library. These paths include lighting brightness adjustment, color temperature change, sunshade control, and area lighting switching. The system will filter and sort the actions based on task type and preference, and after determining the candidate actions, summarize them as a set of candidate action paths;

[0119] Select the optimal action path. From the set of candidate action paths, the system evaluates the expected regulation effect, system energy consumption change, and state disturbance intensity of each action combination, and selects the action sequence with the least impact, fastest response, and most in line with individual preferences as the execution plan for the current control strategy.

[0120] The expected adjustment effect is inferred based on the difference between the current state and the target lighting state. The system compares the degree of fit between the current illuminance, color temperature and spatial distribution and the target state in the characteristic dimension, and combines the control model's ability to predict the state change trend after each set of actions is executed to deduce the adjustment adaptability of each action combination; the system energy consumption change is based on the power model and spatial layout structure of each lamp type in the BIM, combined with the switching behavior and brightness level adjustment range involved in the action path, to calculate the overall power fluctuation value corresponding to the action combination, and consider the time period electricity price model to deduce its cost impact; and the state disturbance intensity is used to measure the intervention extent of the action combination on the stability of the current spatial light environment. The system models the potential disturbance generated by each set of combined actions by analyzing the mutation rate of the state variables caused by each action per unit time and its propagation effect between subspaces, thereby identifying and eliminating high-intervention and high-cost solutions.

[0121] After the control strategy is generated, the system sends control instructions to the lamp driver, color temperature adjustment module, shading control system or other related control units through the communication protocol to implement specific lighting adjustment operations. During the adjustment process, the system continuously collects new sensor data and monitors the light environment status to evaluate whether the adjustment has achieved the expected effect;

[0122] If the system finds no significant improvement in lighting conditions after adjustments, if the user triggers manual intervention, or if the state evolution trend drifts in the opposite direction, it activates a policy self-correction mechanism. This mechanism first determines whether there is a target setting error and then reassesses the rationality of the control target. If the target is reasonable but the preference response is poor, the adjustment weights in the user preferences are updated. If the path selection deviates from the stable policy, the policy structure is replaced to restore system convergence.

[0123] Step 4: Establish a human-computer interaction feedback mechanism for light environment adjustment, so that lighting control not only has adaptive capabilities but also forms a closed-loop adjustment path. The user's actual response to the lighting status is used to correct subsequent control behaviors. The specific steps are as follows:

[0124] Conduct behavioral signal collection and responsiveness generation, and synchronously collect user feedback signals from multiple sources. User feedback signals include but are not limited to the following situations: First, the user's subjective evaluation of the current lighting environment through the terminal device, such as scoring and sliding adjustment; second, the user manually adjusts the lighting system, such as opening sunshades and dimming light sources; third, the user's behavior duration or behavior pattern changes in the current lighting environment, such as reading without turning on local lighting for a long time;

[0125] Based on the type, intensity, timeliness and behavioral results of these user feedback signals, behavioral response factors are constructed one by one. Each type of behavioral response will have a specific set of weight coefficients to measure the credibility and intervention strength of the behavior in the expression of preference. The system then merges various behavioral factors to form a normalized response value, which reflects the user's overall satisfaction or preference for the control strategy execution results at the current moment. The weight coefficient is set according to actual needs.

[0126] After receiving the user's behavior response, the system updates the current user preference state. The preference state, as an embedded vector, represents the user's personalized preferences in different control target dimensions (such as illumination, color temperature, and adjustment frequency). To ensure that the preference update is continuous and resistant to interference, a nonlinear recursive state correction mechanism is adopted.

[0127] The basic logic of the update mechanism is: compare the current preference state with the previous state, and determine the direction and magnitude of the preference change based on the latest behavioral response value. In order to prevent the behavioral response value from being too small or changing frequently, which may cause system jitter, an activation interval is set within the update mechanism. Only when the behavioral feedback intensity exceeds the lower limit of this interval will the system trigger the preference state update. The size of the update amplitude has a nonlinear growth relationship with the feedback intensity. The stronger the feedback, the larger the step size of the preference correction in the target direction. When the feedback intensity is in the boundary interval, the system will adopt a slow and gradual approach to fine-tune the preference state to ensure a consistent user experience. After multiple feedback behaviors accumulate, the preference state will gradually converge to the user's long-term usage habits and be reflected in the generation of control targets, dynamically adjusting the system's lighting strategy output.

[0128] Each time a control strategy is executed, the core input information and reasoning path underlying that strategy decision are automatically recorded. This record includes the current environmental perception state, user behavior state, target lighting parameters, control preference vector, strategy candidate action set, the final selected adjustment action, and the status feedback after the action is executed.

[0129] All of this information is organized into a structured data block, forming a control strategy map. Each element node in the map is labeled with its source and impact direction, and the system can reconstruct the control decision path based on the map. For example, it can clearly identify a complete path such as "In a certain dimming situation, the system chooses maximum power dimming because of its historical preference for high brightness and its behavioral response is active fill light."

[0130] The control strategy map not only serves as a basis for backtracking within the system, but can also be used for auditing, version comparison, and playback simulation of later control strategies. It can also support the construction of multi-round strategy map sequences based on time, showing the logical links of strategy evolution and the gradual shift trend of user preferences.

[0131] To enhance user trust in the control system, a policy explanation engine can also be built. This engine abstracts and compresses the key paths in the policy traceability structure and outputs explanations in user-understandable language and graphics. The explanation includes why the current control strategy was triggered, the judgment logic for selecting the strategy, other options and the reasons for not adopting them, the deviation between the current state and the target, and the expected effect of the system after executing the strategy.

[0132] The system presents these interpretations on the user's device, such as a mobile app or dashboard, and allows the user to make responsive adjustments based on the interpretations. The user can choose to accept the current strategy, make minor adjustments to the control results, or provide feedback that the system's current control cycle did not meet expectations. All feedback information is fed into the aforementioned behavior collection mechanism for further adjustment of preference states.

[0133] The above formulas are all dimensionless and numerical calculations. The formula is a formula that is closest to the actual situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0134] The above embodiments may be implemented in whole or in part through software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0135] Those skilled in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0136] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0137] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0138] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. The BIM-based building light environment neural network control system is characterized by: It includes a building lighting space division module, a lighting scene data acquisition module, a lighting strategy generation module, and a lighting feedback control module, and each module is connected through signals; The architectural lighting space division module divides space units according to the BIM model and extracts five types of structural attributes to form node vectors. It identifies spatial connectivity to generate a Boolean matrix, and constructs a modulation function based on spatial spacing, occlusion overlap ratio, and light-transmitting cross-section to generate edge weights and form a structure-enhanced semantic map. The lighting scene data acquisition module is used to collect sensory information, integrate spatial structural attributes to construct the original state vector, and use a gated recurrent neural network to perform temporal modeling on the sensory information, extract evolutionary features, and divide it into four subspaces: immediate perception, behavioral context, historical trajectory, and individual preference. The lighting strategy generation module identifies the lighting task scenario and matches the control target based on the current activity type and time period. It adjusts the direction by reasoning the difference between the state and the target, screens and evaluates candidate action paths, generates the optimal strategy, and executes the control. The lighting feedback control module collects user feedback signals, constructs a behavior response factor to determine the preference matching degree of the current control strategy, and updates and records the preference status.

2. The BIM-based building light environment neural network control system according to claim 1, characterized in that: Divide the space units according to the BIM model and extract five types of structural attributes to form node vectors. The specific steps are as follows: Divide the entire building area into a group of spatial units, each of which represents a control area with independent lighting adjustment requirements; After the spatial unit division is completed, the structural attributes are extracted, including orientation angle, surface reflectivity, light transmittance, shading complexity level and clear height; The structural attributes are combined into a five-dimensional node structure vector to characterize the light environment structure state of each spatial unit.

3. The BIM-based building light environment neural network control system according to claim 2, characterized in that: Identify spatial connectivity to generate a Boolean matrix, and then build a modulation function based on spatial spacing, occlusion overlap ratio, and light cross section to generate edge weights and form a structure-enhanced semantic graph. The specific steps include: Extract the topological boundaries and opening positions between spaces from the BIM model, analyze the physical contact and visual distance between topological boundary components, and establish a Boolean matrix to represent the connection status between each spatial unit; A modulation function is constructed based on the spatial distance, the occlusion-projection overlap ratio, and the minimum light-transmitting cross-section of the connection path, and is used as the edge weight input of the structure-enhanced semantic graph. Each node in the structure-enhanced semantic graph corresponds to a spatial unit.

4. The BIM-based building light environment neural network control system according to claim 3 is characterized by: It is used to collect perception information through sensors and fuse spatial structure attributes to construct the original state vector. The specific steps are as follows: Sensors deployed in spatial units collect environmental status and user behavior data as perception information. The perception information includes the illuminance value, color temperature value, occupant status identifier, behavior tag, system time, and outdoor natural light intensity within the current spatial unit. The perceptual information is combined with the spatial structure attributes to form an original state vector.

5. The BIM-based building light environment neural network control system according to claim 4 is characterized in that: A gated recurrent neural network is used to perform temporal modeling of the perception information, extract evolutionary features, and divide it into four subspaces: immediate perception, behavioral context, historical trajectory, and personality preference. The following steps are included: A gated recurrent neural network is used to model the temporal sequence of sensory information. The currently embedded raw state vector is fed into the GRU unit for joint processing with the hidden state at the previous moment. Gated logic judgment is then performed to determine the change in illumination state, which is then used as the behavioral response expression of the spatial unit at the current moment. After completing the time-recursive modeling, the state space is divided into four subspaces: immediate perception, behavioral context, historical trajectory, and personality preference.

6. The BIM-based building light environment neural network control system according to claim 5, characterized in that: After completing the time-recursive modeling, the state space is divided into four subspaces: immediate perception, behavioral context, historical trajectory, and personality preference. The specific steps are as follows: The immediate perception subspace is used to extract the components directly related to the current sensor value; Behavior context subspace, used to extract feature components related to behavior labels and time periods; The historical trajectory subspace is used to analyze the state change rate and trend, and to determine whether the lighting state is drifting, stable, or abnormal; Personalized preference subspace, which is constructed by combining user interaction records or feedback information to dynamically adjust preference parameters in the control strategy; Strategy calls are made according to different subspaces.

7. The BIM-based building light environment neural network control system according to claim 6, characterized in that: According to the current activity type and time period, identify the lighting task scene and match the control target. The specific steps are as follows: Obtain the user's activity type and time period, and determine the lighting task scenario of the current space based on the joint rule base between the behavior type and time period; After the current lighting task scene is identified, it is mapped to a set of lighting control targets, which include the brightness level required for the space, the color characteristics of the light source, the energy consumption control standards, and the adjustment preferences previously expressed by the user.

8. The BIM-based building light environment neural network control system according to claim 7, characterized in that: Adjust the direction by reasoning about the difference between the state and the goal, screen the candidate action paths and evaluate them, generate the optimal strategy and execute the control. The specific steps are as follows: After the lighting control target is determined, the strategy generation phase begins. The difference between the current state expression and the expected target is converted into the adjustment intention, including adjustment direction identification, candidate action extraction, and optimal action path selection. After the candidate actions are determined, they are summarized as a set of candidate action paths. The expected adjustment effect, system energy consumption change and state disturbance intensity of each action combination are evaluated to select the optimal action path.

9. The BIM-based building light environment neural network control system according to claim 8, characterized in that: Collect user feedback signals, construct a behavior response factor to determine the preference matching degree of the current control strategy, and update and record the preference status. The specific steps are as follows: The difference between the current state and the target lighting state is used as the expected adjustment effect; Based on the power model and spatial layout structure of each lamp type in the BIM, combined with the switching behavior and brightness level adjustment range involved in the action path, the overall power fluctuation value corresponding to the action combination is calculated as the system energy consumption change; The state disturbance intensity is used to measure the extent of the intervention of the action combination on the stability of the current spatial light environment. It is obtained by analyzing the mutation rate of the state variables caused by each action per unit time and the propagation effect in the subspace.

10. The BIM-based building light environment neural network control system according to claim 9, characterized in that: Collect user feedback signals, construct a behavior response factor to determine the preference matching degree of the current control strategy, and update and record the preference status. The specific steps are as follows: User feedback signals include the user's subjective evaluation of the current lighting environment through the terminal device, the user's manual adjustment of the lighting system, and changes in the user's behavior duration or behavior pattern in the current lighting environment; According to the type, intensity, timeliness and behavior results of user feedback signals, behavior response factors are constructed one by one, and a normalized response value is formed. After receiving the user behavior response result, the current user preference state is updated.

Images