A light scene offline generation method and system based on an edge lightweight neural network

Abstract

This invention provides a method and system for offline generation of lighting scenes based on a lightweight edge-end neural network, relating to the fields of intelligent lighting and IoT control technology. The method includes: acquiring initial illuminance field data of a target spatial physical model collected by a multi-dimensional illumination sensor array; calculating the spatial gradient magnitude of the illuminance field point-by-point along the three-dimensional spatial coordinate axes to obtain a local attenuation feature sequence; constructing a three-dimensional isoluminance surface based on the local attenuation feature sequence; calculating the degree of local bending deformation of the three-dimensional isoluminance surface in the spatial normal direction to obtain a spatial bending distribution matrix; and dividing the target spatial physical model into halo overlap segments and normal attenuation segments based on the deformation extreme values ​​of the spatial bending distribution matrix to obtain a spatial structure classification set. This invention improves the smoothness and real-time performance of offline generation of edge-end lighting scenes.

Description

Technical Field

[0001] This invention relates to the field of intelligent lighting and Internet of Things control technology, and in particular to an offline method and system for generating lighting scenes based on a lightweight edge neural network. Background Technology

[0002] With the widespread application of intelligent lighting technology in large indoor spaces, the automated generation of lighting scenes has gradually become a key means to enhance the spatial atmosphere and energy efficiency. Taking a large ceramic exhibition hall in a museum as an example, the exhibition hall has an interspersed display area, transition passage, and rest area, and a large number of LED physical light source nodes with adjustable color temperature and brightness are deployed. In order to achieve differentiated lighting effects under different exhibition themes, the museum has introduced a cloud-based deep learning lighting scene generation system. Illuminance sensor groups deployed in various parts of the exhibition hall collect spatial illuminance data in real time. The multi-dimensional light distribution data is uploaded to the cloud server through the public network. The convolutional neural network model running in the cloud extracts and reconstructs the spatial distribution features of the light, and generates a lighting control instruction set containing dual channels of color temperature and brightness. The instruction is then transmitted back to the local lighting control bus in the exhibition hall through the public network to drive the light source array to execute scene rendering.

[0003] The solution revealed technical flaws in actual operation: the inference latency of the cloud-based deep learning model is highly susceptible to fluctuations in the public network; when museums hold temporary special exhibitions and need to frequently switch lighting scenes, the scene switching response exhibits noticeable stuttering, affecting the smoothness of the visitor experience; simultaneously, the cloud-based solution is rigidly dependent on network connectivity; if the public network in the exhibition hall area is interrupted or base station congestion occurs, the entire lighting scene generation function is completely paralyzed; during the global illuminance field feature extraction process, the cloud model does not differentiate between the overlapping halo sections and normal attenuation sections caused by the overlap of multiple light sources in the physical space, resulting in discontinuous and overly harsh illuminance distribution in the halo overlap areas of the generated reconstructed illumination matrix, specifically manifested as visible brightness discontinuities and color temperature jumps in the transition zones between adjacent light sources, making it difficult to form a smooth and natural lighting transition effect; continuously uploading high-precision three-dimensional illuminance field data inside the exhibition hall to a third-party cloud platform poses a risk of leakage of spatial lighting environment data, making it difficult to meet the security and compliance requirements of classified venues for offline and localized operation. Summary of the Invention

[0004] This invention provides a method and system for offline generation of lighting scenes based on a lightweight neural network at the edge, which improves the smoothness and real-time performance of offline generation of lighting scenes at the edge.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a method for offline generation of lighting scenes based on a lightweight edge neural network, the method comprising: Acquire the initial illuminance field data of the target spatial physical model collected by the multi-dimensional illumination sensor array, calculate the spatial gradient magnitude of the illuminance field point by point along the three-dimensional spatial coordinate axis, and obtain the local attenuation characteristic sequence. A three-dimensional iso-illuminance surface is constructed based on the local attenuation feature sequence. The degree of local bending deformation of the three-dimensional iso-illuminance surface in the spatial normal direction is calculated to obtain the spatial bending distribution matrix. Based on the deformation extreme values ​​of the spatial bending distribution matrix, the target spatial physical model is divided into halo overlap segments and normal attenuation segments to obtain a spatial structure classification set. Based on the halo overlap section in the spatial structure classification set, the continuous settlement of the illuminance difference between adjacent physical nodes is calculated pixel by pixel to obtain halo smooth transition parameter data; The halo smooth transition parameter data and the normal attenuation segment in the spatial structure classification set are topologically stitched together to obtain a global illumination reconstruction matrix covering the entire target space. The global illumination reconstruction matrix is ​​input into a lightweight neural network pre-deployed on edge computing nodes to extract the implicit features of the spatial distribution of illumination and obtain the scene control latent vector. The scene control latent vector is executed by feature expansion and decoding to generate a multi-channel lighting control instruction set containing color temperature and brightness dimensions; the multi-channel lighting control instruction set is sent to the lighting control bus to drive the physical light source array to perform offline scene rendering, thus completing the offline generation of the lighting scene.

[0006] Secondly, an offline lighting scene generation system based on a lightweight edge neural network includes: The acquisition module is used to acquire the initial illuminance field data of the target space physical model collected by the multi-dimensional illumination sensor array, calculate the spatial gradient magnitude of the illuminance field point by point along the three-dimensional spatial coordinate axis, and obtain the local attenuation feature sequence. The partitioning module is used to construct a three-dimensional iso-illuminance surface based on the local attenuation feature sequence, calculate the degree of local bending deformation of the three-dimensional iso-illuminance surface in the spatial normal direction, and obtain the spatial bending distribution matrix; based on the deformation extreme value of the spatial bending distribution matrix, the target spatial physical model is divided into halo overlap segment and normal attenuation segment, and a spatial structure classification set is obtained. The calculation module is used to calculate the continuous settlement of the illuminance difference between adjacent physical nodes pixel by pixel based on the halo overlap section in the spatial structure classification set, so as to obtain the halo smooth transition parameter data. The stitching module is used to topologically stitch the halo smooth transition parameter data with the normal attenuation segment in the spatial structure classification set to obtain a global illumination reconstruction matrix covering the entire target space. The extraction module is used to input the global illumination reconstruction matrix into a lightweight neural network pre-deployed on edge computing nodes to extract the implicit features of the illumination spatial distribution and obtain the scene control latent vector. The driver module is used to expand and decode the scene control latent vector execution features to generate a multi-channel lighting control instruction set containing color temperature and brightness dimensions; the multi-channel lighting control instruction set is sent to the lighting control bus to drive the physical light source array to perform offline scene rendering and complete the offline generation of the lighting scene.

[0007] The above-described solution of the present invention has at least the following beneficial effects: By acquiring initial illuminance field data collected by a multi-dimensional illumination sensor array, the spatial gradient magnitude of the illuminance field is calculated point by point along the three-dimensional spatial coordinate axis, and a three-dimensional iso-illuminance surface is constructed to measure its local bending deformation in the spatial normal direction. The target spatial physical model is divided into halo overlap section and normal attenuation section, overcoming the technical defect of existing methods that do not perform differentiated modeling of the overlapping area of ​​multiple light sources, and thus realizing the identification and structural classification of halo overlap boundaries in complex illumination space.

[0008] By continuously accumulating the illuminance difference between adjacent physical nodes pixel by pixel in the halo overlap section, smooth halo transition parameter data is obtained. This data is then topologically stitched with the normal attenuation section to form a global illumination reconstruction matrix. This overcomes the problems of discontinuous and overly abrupt illuminance distribution in the halo overlap area of ​​existing solutions, thereby eliminating brightness discontinuities and color temperature jumps in the transition zone between adjacent light sources. This achieves a continuous and smooth illumination distribution representation covering the entire target space. By inputting the global illumination reconstruction matrix into a lightweight neural network pre-deployed on edge computing nodes, implicit features of the illumination spatial distribution are extracted and scene control latent vectors are generated. After feature expansion and decoding, a multi-channel lighting control instruction set is directly output. This overcomes the technical shortcomings of cloud-based solutions, such as high inference latency, reliance on public network connections, and the risk of data leakage. It enables local offline generation of lighting scenes, reduces scene switching response latency, and ensures the local security of spatial illumination environment data. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating an offline lighting scene generation method based on a lightweight edge neural network provided by an embodiment of the present invention.

[0010] Figure 2 This is a schematic diagram of an offline lighting scene generation system based on a lightweight edge neural network provided by an embodiment of the present invention.

[0011] Figure 3 This is a schematic diagram of the process for obtaining halo smooth transition parameter data provided by an embodiment of the present invention. Detailed Implementation

[0012] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0013] like Figure 1 As shown, an embodiment of the present invention proposes an offline lighting scene generation method based on a lightweight edge neural network, the method comprising the following steps: Step 1: Obtain the initial illuminance field data of the target space physical model collected by the multi-dimensional illumination sensor array, and calculate the spatial gradient magnitude of the illuminance field point by point along the three-dimensional spatial coordinate axis to obtain the local attenuation characteristic sequence. Step 2: Construct a three-dimensional iso-illuminance surface based on the local attenuation feature sequence, calculate the degree of local bending deformation of the three-dimensional iso-illuminance surface in the spatial normal direction, and obtain the spatial bending distribution matrix; based on the deformation extreme values ​​of the spatial bending distribution matrix, divide the target spatial physical model into halo overlap segments and normal attenuation segments to obtain a spatial structure classification set. Step 3: Based on the halo overlap section in the spatial structure classification set, calculate the continuous settlement amount of the illuminance difference between adjacent physical nodes pixel by pixel to obtain halo smooth transition parameter data; Step 4: Topologically stitch the halo smooth transition parameter data with the normal attenuation segment in the spatial structure classification set to obtain a global illumination reconstruction matrix covering the entire target space. Step 5: Input the global illumination reconstruction matrix into a lightweight neural network pre-deployed on edge computing nodes to extract the implicit features of the spatial distribution of illumination and obtain the scene control latent vector. Step 6: Decode the scene control latent vector execution features to generate a multi-channel lighting control instruction set containing color temperature and brightness dimensions; send the multi-channel lighting control instruction set to the lighting control bus to drive the physical light source array to perform offline scene rendering and complete the offline generation of the lighting scene.

[0014] In this embodiment of the invention, by acquiring the initial illuminance field data of the target space physical model collected by a multi-dimensional illumination sensor array, the spatial gradient magnitude of the illuminance field is calculated point by point along the three-dimensional spatial coordinate axis. A three-dimensional isoluminance surface is constructed, and its local bending deformation degree in the spatial normal direction is calculated. Based on the deformation extreme value, the target space is divided into halo overlap section and normal attenuation section. This overcomes the technical defect of existing methods that lack differentiated modeling of multi-source overlapping areas, and realizes the structural classification of halo overlap boundaries and normal attenuation areas in the illumination space. By continuously accumulating the illuminance drop of adjacent physical nodes pixel by pixel for halo overlap sections, halo smooth transition parameter data is generated, and topologically spliced ​​with the normal attenuation section to form a global illumination reconstruction matrix. This overcomes the problems of discontinuous and overly harsh illuminance distribution in halo overlap areas in traditional schemes, eliminates brightness discontinuities and color temperature jumps between adjacent light sources, and achieves a natural and smooth illumination transition effect covering the entire target space.

[0015] By inputting the global illumination reconstruction matrix into a lightweight neural network pre-deployed on edge computing nodes, the implicit features of the spatial distribution of illumination are extracted to obtain the scene control latent vector. After feature expansion and decoding, a multi-channel lighting control instruction set is directly generated to drive the light source array to perform offline rendering. This overcomes the technical defects of cloud solutions, such as reliance on public network connections, high inference latency, and the risk of data leakage. It realizes the localization and low-latency offline generation of lighting scenes, ensures the local security of spatial illumination environment data, and reduces the consumption of computing resources by edge deployment.

[0016] In a preferred embodiment of the present invention, step 1 above may include: Step 11: Read the raw illuminance readings of each physical node from the multi-dimensional illumination sensor array according to the preset spatial sampling grid, and establish the mapping relationship between each physical node and its three-dimensional spatial coordinate position to obtain a spatial discrete illuminance sampling set. Specifically, this includes: extracting the three-dimensional coordinate position of each physical node in the preset spatial sampling grid based on the raw illuminance readings of each physical node in the multi-dimensional illumination sensor array, and establishing the mapping relationship between the three-dimensional coordinate position and the corresponding illuminance reading. The multi-dimensional illumination sensor array is deployed in the target spatial physical model in a regular spatial grid form. The sampling space of the target spatial physical model is divided into several rows at equal intervals along the first coordinate axis, several columns at equal intervals along the second coordinate axis, and several layers at equal intervals along the third coordinate axis. For the spatial sampling grid located at the first... line, number Column, No. The physical nodes of the layer, where Integers from 1 to the total number of rows. Integers from 1 to the total number of columns. For each element, an integer from 1 to the total number of layers, the three-dimensional spatial coordinates of the physical node are denoted as a coordinate triplet, and the original illuminance reading of the physical node is denoted as the corresponding illuminance value. The spacing between physical nodes in the horizontal, vertical, and lateral directions is determined by the actual installation parameters of the sensor array. In a typical exhibition hall deployment scenario, the horizontal spacing ranges from 0.3 meters to 1.5 meters, the vertical spacing ranges from 0.3 meters to 1.5 meters, and the lateral spacing ranges from 0.5 meters to 2.0 meters. The node spacing in all three directions is set to 0.5 meters to achieve a balance between spatial resolution and sensor deployment cost. After completing the coordinate illuminance mapping of all physical nodes, a spatial discrete illuminance sampling set is obtained. The spatial discrete illuminance sampling set is a set of values ​​arranged regularly in three dimensions. Each element contains a three-dimensional spatial coordinate position and the illuminance value directly collected by the sensor array at that position.

[0017] Step 12: Based on the spatial discrete illuminance sampling set, extract the illuminance difference between adjacent physical nodes along the first coordinate axis of the three-dimensional spatial coordinate system to obtain the illuminance difference component in the first direction. Specifically, this includes: for nodes located at the first coordinate axis... line, number Column, No. The physical node of the layer is extracted horizontally along the horizontal direction to find the position of the physical node in the same row, the same layer, and the adjacent column (i.e., the first). The illuminance difference between the physical nodes of the column. The formula for calculating the illuminance difference component in the first direction is as follows: ; In the formula, This indicates that the spatial discrete illuminance sampling set is located at the first... Line 1 Liede The illuminance difference components of the physical nodes of the layer in the horizontal direction (i.e., the direction of the first coordinate axis), in lux; Indicates that it is located at the th Line 1 Liede The illuminance value of the physical nodes of the layer, in lux, is derived from the coordinate illuminance mapping relationship; Indicates that it is located at the th Line 1 Liede The illuminance values ​​of adjacent physical nodes in the layer, in lux, are calculated by performing the above difference operation on each node in the spatial sampling grid that has adjacent physical nodes in the horizontal direction. If a physical node does not have any adjacent nodes in the horizontal direction (i.e., the node is located at the horizontal boundary of the sampling grid), the first direction illuminance difference component is not calculated for that node. All the effectively calculated first direction illuminance difference components are arranged according to the spatial position of the corresponding physical nodes and summarized into a first direction illuminance difference component set. The first direction illuminance difference component set records the piecewise change of illuminance along the horizontal direction in the target spatial physical model.

[0018] Step 13: Based on the spatial discrete illuminance sampling set, extract the illuminance difference between adjacent physical nodes along the second and third coordinate axes of the three-dimensional spatial coordinate system, respectively, to obtain the illuminance difference component in the second direction and the illuminance difference component in the third direction. Summarize these components to obtain the set of illuminance difference components in the three axes. Specifically, this includes: based on the same spatial discrete illuminance sampling set, extracting the illuminance difference between adjacent physical nodes along the horizontal and vertical directions, respectively. The calculation formula for the illuminance difference component along the horizontal direction (i.e., the second coordinate axis direction) is as follows: ; In the formula, This indicates that the spatial discrete illuminance sampling set is located at the first... Line 1 Liede The illuminance difference components of the physical nodes of the layer in the horizontal longitudinal direction (i.e., the direction of the second coordinate axis), in lux; Indicates that it is located at the th Line 1 Liede Illuminance values ​​of adjacent physical nodes of a layer, in lux; Indicates that it is located at the th Line 1 Liede The illuminance values ​​of the physical nodes of the layer, in lux, are calculated using the following formula for the illuminance difference components along the vertical direction (i.e., the direction of the third coordinate axis): ; In the formula, This indicates that the spatial discrete illuminance sampling set is located at the first... Line 1 Liede The illuminance difference components of the physical nodes of the layer in the vertical direction (i.e., the direction of the third coordinate axis), in lux; Indicates that it is located at the th Line 1 Liede The illuminance values ​​of adjacent physical nodes in a layer, in lux, are calculated. The second-direction illuminance difference component is calculated for all nodes in the spatial sampling grid that are adjacent in the horizontal direction, and the third-direction illuminance difference component is calculated for all nodes that are adjacent in the vertical direction. Nodes located at the boundaries of each axis of the sampling grid participate in the calculation only in the direction where they have adjacent nodes.

[0019] The first directional illuminance difference component set output in step 12, the second directional illuminance difference component set calculated in this step, and the third directional illuminance difference component set are summarized to obtain the three-axis directional illuminance difference component set. The three-axis directional illuminance difference component set records the illuminance difference in the three coordinate axes for each physical node that has adjacent nodes in all three directions, forming complete original data of illuminance spatial variation.

[0020] Step 14: Based on the three-axis illuminance difference component set, the illuminance difference components of each physical node in the first, second, and third directions are compared with the distance between physical nodes in the corresponding coordinate axis directions to obtain the local attenuation feature sequence. Specifically, this includes: based on the three-axis illuminance difference component set, dividing the illuminance difference components of each physical node in the three directions by the spatial distance between adjacent physical nodes in the corresponding coordinate axis directions to obtain the directional gradient components of the physical node in the three directions. The formula for calculating the horizontal directional gradient components is as follows: ; In the formula, Indicates that it is located at the th Line 1 Liede The gradient component of the physical nodes of the layer in the horizontal direction, in lux per meter, reflects the amount of illuminance attenuated per unit distance in the horizontal direction. This represents the spatial distance between adjacent physical nodes in the horizontal direction, in meters. This value is determined by the actual physical installation spacing of the multi-dimensional illumination sensor array. The formula for calculating the directional gradient component in the horizontal direction is as follows: ; In the formula, Indicates that it is located at the th Line 1 Liede The gradient components of the physical nodes of the layer in the horizontal and vertical directions, in lux per meter; This represents the spatial distance between adjacent physical nodes in the horizontal direction, in meters. The formula for calculating the vertical direction gradient component is as follows: ; In the formula, Indicates that it is located at the th Line 1 Liede The directional gradient component of the physical nodes of the layer in the vertical direction, in lux per meter; This represents the spatial distance between adjacent physical nodes in the vertical direction, in meters. Based on the components in the three directions mentioned above, the formula for calculating the comprehensive spatial gradient feature value at each physical node is as follows: ; In the formula, Indicates that it is located at the th Line 1 Liede The comprehensive spatial gradient feature value at the physical node of the layer, in lux per meter, reflects the rate of light intensity decay in three-dimensional space along each direction at the node location. The three-dimensional grid coordinates of all physical nodes and their corresponding comprehensive spatial gradient feature values ​​are organized into a three-dimensional numerical array according to the spatial arrangement relationship to obtain the local decay feature sequence. Each entry in the local decay feature sequence contains the spatial grid index and comprehensive spatial gradient feature value of the physical node.

[0021] In this embodiment of the invention, the original illuminance readings of each physical node in the multidimensional illumination sensor array are read according to a preset spatial sampling grid and a mapping relationship with the three-dimensional spatial coordinates is established. The illuminance difference between adjacent physical nodes is extracted along the three coordinate axes, and the ratio of the illuminance difference component in each direction to the distance between physical nodes in the corresponding coordinate axis direction is calculated. This overcomes the technical problem that discrete sensor data is difficult to accurately characterize the rate of change of continuous spatial illumination gradient, and obtains a local attenuation feature sequence that can accurately reflect the instantaneous attenuation law of the illuminance field along each direction in three-dimensional space.

[0022] In a preferred embodiment of the present invention, step 2 above may include: Step 21: Based on the illuminance values ​​of each physical node in the local attenuation feature sequence, extract adjacent physical nodes with equal illuminance values. Connect the extracted adjacent physical nodes using a spatial mesh in a three-dimensional spatial coordinate system to construct a three-dimensional isoluminance surface covering the target spatial physical model. Specifically, this includes: extracting the illuminance values ​​of each physical node in the local attenuation feature sequence; connecting adjacent physical nodes with equal illuminance values ​​using a spatial mesh in a three-dimensional spatial coordinate system to construct a three-dimensional isoluminance surface covering the target spatial physical model; and determining a set of preset isoluminance level values. The isoluminance level values ​​are generated by uniformly sampling with a fixed step size between the global minimum and maximum illuminance values ​​of the target spatial physical model. The minimum illuminance value in the target spatial physical model is set as the minimum illuminance value, and the maximum illuminance value is set as the maximum illuminance value. The preset number of isoluminance levels is the preset number of levels. The formula for calculating the illuminance interval between two adjacent isoluminance levels is as follows: ; In the formula, Indicates the illuminance interval between two adjacent equal illuminance levels, in lux; This represents the maximum illuminance value of all physical nodes in the target space physical model, in lux. This value is obtained by searching for the global maximum value of the spatial discrete illuminance sampling set. This represents the minimum illuminance value of all physical nodes in the target space physical model, in lux. This value is obtained by performing a global minimum value search on the spatial discrete illuminance sampling set. This represents the preset number of levels, dimensionless, with a value range of 8 to 32. Its value is determined based on: when... When the value is 8, the spacing between isoilluminance surfaces is relatively large, resulting in lower spatial resolution. This is suitable for scenarios with gradual changes in spatial illuminance and minimal overlap of multiple light sources, leading to lower computational overhead. When the value is 32, the spacing between iso-illuminance surfaces is relatively close, which can more finely depict the spatial morphological changes of iso-illuminance surfaces. This is suitable for scenes with high light source density and large gradients in illuminance. Choosing 16, this value strikes a balance between spatial resolution and computational cost under typical indoor exhibition space illuminance variation gradients. The formulas for calculating the illuminance values ​​of each equal illuminance level are as follows: ; In the formula, Indicates the first Illuminance values ​​for each isolux level, in lux; This indicates the sequence number of the iso-illuminance level, dimensionless, and ranging from 1 to... For each iso-illuminance level, within the spatial sampling grid covered by the local attenuation feature sequence, a grid edge traversal operation is performed: Each grid edge connecting two adjacent physical nodes is examined sequentially. When the illuminance values ​​at the two endpoints of this grid edge are exactly located on either side of the illuminance value of the current iso-illuminance level, an iso-illuminance point is determined to exist on this grid edge. The illuminance value at this iso-illuminance point is equal to the illuminance value of the current iso-illuminance level. For each grid edge determined to contain an iso-illuminance point, a linear interpolation method is used along this grid edge to determine the precise three-dimensional spatial coordinates of the iso-illuminance point. The interpolation calculation formula is as follows: ; In the formula, This represents the three-dimensional spatial coordinates of the iso-illuminance point obtained through interpolation, in meters; This represents the three-dimensional spatial coordinates of the grid endpoint A, in meters; This represents the three-dimensional spatial coordinates of the grid edge endpoint B, in meters; This represents the illuminance value at endpoint A, in lux. This represents the illuminance value at endpoint B, in lux. Indicates the number of processes currently being processed. The illuminance values ​​of each isoluminance level are calculated in lux. After traversing and interpolating all mesh edges, all isoluminance points belonging to the same isoluminance level are connected by triangular meshes in the three-dimensional coordinate system according to their spatial proximity. The rule for connecting the triangular meshes is: isoluminance points that are spatially adjacent and located in the same isoluminance level are connected into triangular patches according to the principle of shortest side length priority. The set of all triangular patches constitutes a continuous surface covering the isoluminance level. The mesh edge traversal, interpolation positioning and triangular mesh connection operations are repeated for each isoluminance level, and finally a set of three-dimensional isoluminance surfaces consisting of multiple isoluminance level surfaces is obtained, covering the entire illuminance range within the target spatial physical model.

[0023] Step 22: Based on the normal direction of each spatial location point on the three-dimensional iso-illuminance surface, calculate the curvature change amplitude of the iso-illuminance surface along the normal direction at the normal direction location point to obtain a spatial curvature distribution matrix reflecting the undulation of the spatial form. Specifically, this includes: based on the set of three-dimensional iso-illuminance surfaces, for each spatial location point on the iso-illuminance surface, calculate the curvature change amplitude of the iso-illuminance surface along the normal direction at the normal direction location point to obtain a spatial curvature distribution matrix reflecting the undulation of the spatial form. The implicit surface properties are analyzed, and the first and second partial derivatives at the grid nodes are calculated using a three-dimensional central difference scheme. The partial derivative values ​​at the target points of the isolating surface are obtained through trilinear interpolation. These values ​​are then substituted into the implicit surface mean curvature formula to obtain the degree of bending deformation at that point. A three-dimensional central difference operation is performed on all grid nodes of the spatial discrete illuminance sampling set to solve for the nine partial derivatives of the illuminance field at each grid node, including the first partial derivative. , , and second-order partial derivatives , , , , , The first-order partial derivatives in the horizontal direction For example, located at the Line 1 Liede The formula for calculating the first-order partial derivative at the grid nodes of the layer is as follows: ; In the formula, This represents the first-order partial derivative of the illuminance at the grid node along the horizontal direction, in lux per meter; , These represent the illuminance values ​​of the left and right adjacent nodes of the given node, respectively, in lux; The horizontal grid spacing is in meters, and the vertical first-order partial derivative is... First-order partial derivatives in the vertical direction The same central difference formula is used to calculate the grid spacing for each direction. , The second-order partial derivative is calculated based on the difference result of the first-order partial derivative, using the horizontal second-order partial derivative. For example, the calculation formula is as follows: ; In the formula, This represents the second-order partial derivative of the illuminance at the grid node along the horizontal direction, in lux per square meter. (Mixed second-order partial derivative) , , The calculation employs a cross-directional center difference scheme, with the grid spacing in the corresponding direction participating in the computation. After calculating the nine partial derivatives for all grid nodes, for any point on the 3D iso-illuminance surface that does not coincide with any grid node, a trilinear interpolation method is used. Based on the partial derivative values ​​of the eight neighboring grid nodes around that point, the nine partial derivative values ​​at that target point are interpolated. These nine interpolated partial derivatives are then substituted into the implicit surface mean curvature formula to calculate the mean curvature of the iso-illuminance surface at that point, as shown in the following formula: ; In the formula, It represents the average curvature of the iso-illuminance surface at that location point, in units of meters, reflecting the overall degree of bending deformation of the iso-illuminance surface at that point; , , These represent the first-order partial derivatives of the illuminance field at that point along the horizontal, vertical, and lateral directions, respectively, in lux per meter. , , These represent the second-order pure partial derivatives of the illuminance field at that point along the three coordinate axes, in lux per square meter; , , Let represent the second-order mixed partial derivatives of the illuminance field at that point, in lux per square meter. When the value is less than a preset threshold, the isoilluminance surface at that point is determined to be close to a plane, and the average curvature of that point is directly assigned to zero to avoid the risk of division by zero. Dimensionally verified, the molecular weight dimension is lux. 3 per meter 4 The denominator has dimensions of lux. 3 per meter 3 The result of the calculation is in units of meters, which is consistent with the physical dimensions of curvature.

[0024] The larger the absolute value of the average curvature, the more severe the bending deformation of the iso-illuminance surface near that point, and the higher the probability of overlapping illumination from multiple light sources in the corresponding region. When the average curvature is close to zero, the iso-illuminance surface is close to a plane, and the illumination attenuation in the corresponding region follows a monotonic variation law. The calculated absolute values ​​of the average curvature at each spatial location are organized and arranged according to a three-dimensional grid coordinate system to obtain a spatial curvature distribution matrix, the mathematical representation of which is: In the formula, Indicates the location of the grid index in the spatial curvature distribution matrix. The element value at that location, in units of per meter; Indicates the grid position The absolute value of the average curvature calculated at that location. For grid points that do not fall directly on the iso-illuminance surface, the curvature value of the nearest iso-illuminance surface to that location is used for assignment.

[0025] Step 23: Based on the curvature change amplitude of each spatial location point in the spatial curvature distribution matrix, extract locations where the curvature change amplitude exceeds a preset curvature limit and classify them into the halo overlap segment; extract locations where the curvature change amplitude does not exceed the preset curvature limit and classify them into the normal attenuation segment, thus obtaining a spatial structure classification set. Specifically, this includes: based on the spatial curvature distribution matrix, for each spatial location point in the matrix, classify the location point into the corresponding segment according to whether its curvature change amplitude exceeds the preset curvature limit. The formula for calculating the preset curvature limit is as follows: ; In the formula, This indicates the preset bending limit, in meters, which is the boundary threshold for distinguishing between halo overlap sections and normal attenuation sections; This represents the bending limit adjustment coefficient, which is dimensionless and ranges from 1.5 to 3.0. This represents the standard deviation of all element values ​​in the spatial curvature distribution matrix, expressed in meters, reflecting the overall dispersion of the curvature deformation of the iso-illuminated surface across the entire target space. (Bending limit adjustment coefficient) The calibration was performed using the following method: a calibration scene with similar light source density and spatial scale to the target space was selected, and the boundary of the manually marked halo overlap section was used as the calibration benchmark. Candidate halo segments with an interval of 0.1 within the range of 1.5 to 3.0 were selected sequentially. The intersection-union ratio (IoU) of the segment recognition result corresponding to each candidate value and the manually labeled baseline is calculated. The candidate value with the highest IoU is selected as the final value. Value selection. In a typical exhibition hall scenario, the value is selected as follows: When the value is 2.0, the segment identification results are highly consistent with the manually labeled benchmark, achieving a good compromise between the identification sensitivity of halo overlapping segments and the reasonableness of the number of segments. The calculation formula is as follows: ; In the formula, This represents the arithmetic mean of all element values ​​in the spatial curvature distribution matrix, expressed in meters. This represents the total number of elements in the spatial curvature distribution matrix, and is dimensionless. This indicates that for all matrix elements Summation, then checking each grid position in the spatial curvature distribution matrix one by one. element value The classification rules are as follows: when the following conditions are met... When the condition is met, the grid position is determined to belong to the halo overlap section; when the condition is met... When the grid location is determined to be in the normal attenuation zone, after classifying all grid locations, the results are summarized into a spatial structure classification set. Each record in the set contains the three-dimensional spatial coordinates of the location point and the segment type identifier to which it belongs.

[0026] In this embodiment of the invention, by extracting equivalent adjacent nodes based on the illuminance values ​​of each physical node in the local attenuation feature sequence and connecting them in a spatial grid to construct a three-dimensional iso-illuminance surface, the curvature change amplitude of each position point on the iso-illuminance surface along the spatial normal direction is calculated to obtain a spatial curvature distribution matrix, and the segments are divided according to whether the curvature change amplitude exceeds the preset curvature limit. This overcomes the technical problem that existing methods cannot automatically identify and distinguish between overlapping areas of multiple light sources and normal attenuation areas, and realizes the objective and quantitative spatial structure classification of halo overlap segments and normal attenuation segments.

[0027] In a preferred embodiment of the present invention, step 3 above may include: like Figure 3 As shown: Step 31, extract all physical nodes within the halo overlap section from the spatial structure classification set, and determine the spatial arrangement order of all physical nodes along the halo expansion direction. Specifically, this includes: extracting all location points marked as halo overlap sections from the spatial structure classification set, and sorting these location points according to the spatial arrangement order along the halo expansion direction. The halo expansion direction is defined as the spatial direction vector from the edge of the halo overlap section, i.e., the spatial boundary between the halo overlap section and the normal attenuation section, to the center of the halo overlap section, i.e., the location where the overlap of multiple light sources is most dense and the illuminance superposition effect is most significant. For the halo overlap section, first determine the position of the halo center within the section. The halo center position is determined by calculating the weighted center of gravity of the spatial curvature distribution matrix elements at each location point within the halo overlap section. The formula for calculating the horizontal coordinate of the halo center is as follows: ; In the formula, The horizontal coordinate of the center of the halo, in meters; Indicates grid position The horizontal coordinate of the location, in meters; This represents the element value of the location point in the spatial curvature distribution matrix; This represents the summation of all grid points within the halo overlap area. The horizontal and vertical coordinates of the halo center are calculated using the same weighted average method as the formula above. After determining the halo center, each point within the halo overlap area is projected and sorted along the radial direction from the halo center to that point. The sorting index is the spatial Euclidean distance between that point and the halo center. The formula for calculating the spatial Euclidean distance is as follows: ; In the formula, This represents the spatial Euclidean distance between the location and the center of the halo, in meters; These represent the coordinates of the location point in the horizontal, vertical, and longitudinal directions, respectively, in meters. All physical nodes within the overlapping halo section are arranged in ascending order of spatial Euclidean distance. The node with the smallest distance, i.e., the node at the center of the halo, is placed at the beginning of the sequence, and the node with the largest distance, i.e., the node at the edge of the halo, is placed at the end of the sequence, thus obtaining the spatial arrangement order from the center of the halo to the edge of the halo along the direction of halo expansion.

[0028] Step 32: Extract the illuminance values ​​of adjacent physical nodes in the spatial arrangement order, and calculate the illuminance value difference between adjacent physical nodes one by one to obtain a segmented illuminance difference sequence. Specifically, this includes: extracting the illuminance values ​​of each pair of adjacent physical nodes in the spatial arrangement order, calculating the illuminance value difference between adjacent nodes one by one, and for the sequence of physical nodes in the sorted halo overlap segment, let the illuminance difference be the first pair of adjacent physical nodes in the sequence. The illuminance value of the nth physical node is the current illuminance value, and the illuminance value of the nth physical node is the current illuminance value. The illuminance value of each physical node, i.e., the next node adjacent to the current node along the direction of halo expansion, is the next illuminance value. The formula for calculating the segment-by-segment illuminance difference between the two is as follows: ; In the formula, This indicates the spatial arrangement order within the overlapping halo region. The segment, that is, from the first The node to the first The illuminance difference between segments between nodes, in lux. A positive value indicates that the illuminance increases along the direction of halo expansion, and a negative value indicates that the illuminance decreases along the direction of halo expansion. Indicates the order of spatial arrangement. The illuminance value of a physical node, in lux; Indicates the order of spatial arrangement. The illuminance value of a physical node, in lux; The segment number in the spatial arrangement is dimensionless and ranges from 1 to the total length of the sequence minus 1. The above illuminance difference calculation is performed on all adjacent physical nodes in the spatial arrangement within the halo overlap segment. The calculation results are arranged in sequence according to the spatial arrangement direction to obtain the segment illuminance difference sequence.

[0029] Step 33: The illuminance difference values ​​in the segmented illuminance difference sequence are continuously accumulated and calculated along the spatial arrangement direction to obtain the continuous settlement amount from the halo edge to the halo center; the continuous settlement amounts of all physical nodes within the halo overlap section are integrated to obtain halo smooth transition parameter data reflecting the gradual change trend of the halo edge, including: Step 33.1: Based on the starting physical node in the spatial arrangement sequence, set the illuminance value of the starting physical node as the cumulatively calculated reference illuminance value to obtain the halo edge reference illuminance. Specifically, this includes: identifying the physical node located at the halo edge in the spatial arrangement sequence based on the segmented illuminance drop sequence and the spatial arrangement sequence. The physical node at the halo edge is defined as: the physical node farthest from the halo center in the spatial arrangement sequence, located at the spatial boundary between the halo overlap section and the normal attenuation section. Let the physical node located at the halo edge in the spatial arrangement sequence be node zero. Its illuminance value in the spatial discrete illuminance sampling set is the halo edge reference illuminance. The formula for determining the halo edge reference illuminance is as follows: ; In the formula, This represents the reference illuminance at the edge of the halo, in lux. This value is the starting reference value for the cumulative calculation. Its physical meaning is the illuminance level at the outermost edge of the halo overlap section before it is distorted by the multi-source overlapping effect. It is the unified reference origin for the calculation of continuous settlement at all locations. The selection of this value directly determines the overall offset of the halo smooth transition parameter data. This represents the illuminance value of the physical node ranked zero in the spatial arrangement, in lux. This value is directly read from the spatial discrete illuminance sampling set according to the node coordinate index. The halo edge reference illuminance corresponds to the measured illuminance value of the outermost physical node at the spatial boundary between the halo overlap section and the normal attenuation section. This position is at the transition point from the absence of halo superposition effect to its presence.

[0030] Step 33.2: Based on the segmented illuminance difference sequence and the halo edge reference illuminance, the illuminance difference value is progressively decreased and accumulated from the reference illuminance along the spatial arrangement direction to obtain the continuous settlement amount of each physical node relative to the edge within the halo overlap section. Specifically, this includes: based on the halo edge reference illuminance and the segmented illuminance difference sequence, along the spatial arrangement direction from the halo edge to the halo center, the signed illuminance difference values ​​in the segmented illuminance difference sequence are accumulated. Combined with the halo edge reference illuminance, the path cumulative illuminance change of each node is obtained. The continuous settlement amount is defined as the difference between the reference illuminance and the path cumulative illuminance, reflecting the cumulative net illuminance change from the edge to the node, while retaining the reference illuminance's offset control effect on the overall value. Segmented Illuminance Difference Preserving the original symbol, when the illuminance decreases along the direction of halo expansion. When the value is negative, the illuminance increases along the direction of halo expansion. A positive value indicates that the position in the spatial arrangement is [position number]. The formula for calculating the continuous settlement at a given physical node is as follows: ; In the formula, Indicates the order of spatial arrangement. The continuous settlement at the physical node of the halo, measured in lux, reflects the cumulative net change in illuminance value relative to the edge reference from the edge of the halo to the node location. It represents the reference illuminance at the edge of the halo, in lux, and serves as the starting reference value for cumulative calculations, directly determining the overall offset level of the settlement value; Indicates the order of spatial arrangement. The original measured illuminance value of the physical node, in lux; Represents the th segment in the sequence of illuminance difference. The illuminance difference of the segment, in lux, retains the original sign to reflect the trend of illuminance increase or decrease; Represents the node position in the spatial arrangement, dimensionless, and its value ranges from 1 to the total number of physical nodes in the halo overlap section; This represents a cumulative summation index, which is dimensionless. This represents a summation operation, with the summation range from the first paragraph to the second paragraph. The formula involves segment-by-segment signed accumulation calculations to obtain a continuous settlement sequence for all physical nodes within the halo overlap segment. For each halo overlap segment, a reference illuminance determination and segment-by-segment accumulation operation are performed to obtain an independent continuous settlement sequence corresponding to each halo overlap segment. The reference illuminance at the halo edge in this formula... It directly participates in numerical calculations, and its value changes will synchronously change the overall offset of the continuous settlement of all nodes.

[0031] Step 33.3: Based on the continuous settlement of all physical nodes within the halo overlap section, the continuous settlement is mapped to the corresponding three-dimensional spatial coordinates according to the spatial arrangement order to construct a complete spatial transition compensation surface, obtaining halo smooth transition parameter data reflecting the gradual change trend of the halo edge, including: Step 33.31: Based on the continuous settlement amount and spatial arrangement order of all physical nodes within the halo overlap section, extract the position coordinates of each physical node in the three-dimensional spatial coordinate system to obtain the node spatial coordinate set. Specifically, this includes: based on the continuous settlement amount sequence and spatial arrangement order of all physical nodes within the halo overlap section, extracting the position coordinates of each physical node in the three-dimensional spatial coordinate system in the sequence, for each node in the spatial arrangement order... qThe extraction method for the three-dimensional spatial coordinates of a physical node is as follows: using the coordinate record of the physical node in the spatial coordinate-illuminance mapping relationship as the retrieval entry point, the coordinate values ​​of the node in the horizontal, vertical, and lateral directions are read to obtain the three-dimensional spatial coordinates of the node. After the extraction operation is completed, a set of node spatial coordinates is obtained. Each entry in the set of node spatial coordinates contains the ordinal position of the physical node in the spatial arrangement. q And its corresponding three-dimensional spatial coordinates. The data structure of this set is a list of length equal to the total number of physical nodes within the halo overlap segment, where the first node in the list is the first physical node. q The three coordinate components of an element are denoted as the horizontal coordinate, the vertical coordinate, and the transverse coordinate of the node, respectively, all in meters.

[0032] Step 33.32: Based on the set of node spatial coordinates, spatially bind the continuous settlement data with the location coordinates to obtain spatial node data with settlement attributes. Specifically, this includes: based on the set of node spatial coordinates and the continuous settlement sequence, binding nodes with the same position in the spatial arrangement order... The corresponding three-dimensional spatial coordinates are spatially bound to the continuous settlement values ​​at that position to obtain spatial node data with settlement attributes. For the node in the spatial arrangement... For a physical node, the spatial binding operation integrates the node's four attributes—horizontal horizontal coordinate, horizontal vertical coordinate, vertical coordinate, and continuous settlement—into a single spatial node record. This record is represented as follows: ; In the formula, Indicates the first Each spatial node record contains both the spatial location information and settlement attribute information of that node. They represent the first The horizontal, vertical, and coordinate axes of each physical node are given in meters. Indicates the first The continuous settlement at each physical node, in lux, is measured. The above spatial binding operation is performed on all physical nodes from the first to the last position within the halo overlap segment to obtain a spatial node dataset consisting of all spatial node data with settlement attributes.

[0033] Step 33.33: Based on the spatial node data with settlement attributes, connect adjacent spatial nodes along the spatial arrangement direction to obtain a continuous spatial transition compensation surface. Specifically, this includes: based on the spatial node data with settlement attributes, linearly connecting adjacent spatial nodes in the spatial arrangement order pair by pair along the halo expansion direction; simultaneously, laterally connecting adjacent expansion paths along the spatial adjacency direction perpendicular to the halo expansion direction, forming a continuous spatial transition compensation surface covering the entire halo overlap area. The construction process of the spatial transition compensation surface is divided into two stages: the longitudinal connection stage and the lateral interpolation stage. Along the halo expansion direction (i.e., the spatial arrangement direction from the halo edge to the halo center), connect two adjacent spatial nodes in the spatial arrangement order with straight line segments. For the sequence number... nodes and order position nodes The linear interpolation formula for calculating the settlement value at any point on the line connecting the two is as follows: ; In the formula, Indicates at node and Position parameters on the connection line Interpolated settlement at the location, in lux; This represents the interpolation position parameter, which is dimensionless and ranges from 0 to 1. When the value is 0, the interpolation point is located at node 0. Place; when When the value is 1, the interpolation point is located at node 1. Place; when When the value is 0.5, the interpolation point is located at the midpoint of the spatial line connecting the two nodes; Represents a node The continuous settlement at the location, in lux; Represents a node The continuous settlement at the point is expressed in lux. The spatial coordinates of the interpolation point on the line connecting the points are also determined using the same linear interpolation formula as the settlement interpolation. That is, the coordinates of the interpolation point are equal to the weighted average of the coordinates of the two endpoints, and the weights are the same as those for the settlement interpolation.

[0034] In the target space physical model, the halo overlap section typically contains a single spatial path from the halo edge to the halo center, including multiple physical nodes distributed in a planar pattern on a spatial cross-section perpendicular to the halo expansion direction. For the same set of nodes along the same halo expansion path, their settlement variation trends are similar; however, different halo expansion paths may exhibit different settlement variation curves due to differences in local light source distribution. To construct a continuous surface covering the entire halo overlap section in three-dimensional space, lateral interpolation needs to be performed between different halo expansion paths. The lateral interpolation uses an inverse distance weighted average method. For any target interpolation point on the spatial transition compensation surface, the inverse distance weighted interpolation formula for its settlement is as follows: ; In the formula, The coordinates on the spatial transition compensation surface are: The settlement compensation value at the target interpolation point, in lux; This represents the number of neighboring spatial nodes participating in the interpolation. It is dimensionless and is set to 8 in this embodiment. The basis for this value is that in a three-dimensional spatial grid, the 8 nearest grid vertices around a target point are sufficient to cover the information of all spatial directions around the target point, and the computational cost of 8 points does not impose an excessive burden on the edge computing nodes. Indicates the first The interpolation weights of the spatial nodes participating in the interpolation to the target interpolation point are dimensionless and are calculated using the following formula: ; In the formula, Indicates the first The spatial Euclidean distance between each spatial node and the target interpolation point is expressed in meters. This represents the distance decay exponent, which is dimensionless and ranges from 1.0 to 3.0. Its value is determined based on: when... When the weight is set to 1.0, the weight decreases linearly with distance. Distant spatial nodes still have a significant impact on the target interpolation point, resulting in a smoother interpolation surface but potentially losing local details. When the weight is set to 3.0, the weight decreases with the cube of the distance. The influence of distant spatial nodes on the target interpolation point weakens rapidly, and the interpolation surface is closer to the settlement value of the nearest neighbor node. Local details are well preserved, but the surface may have local undulations. The value of 2.0 is optimal in the theoretical interpolation error analysis. After analyzing the spatial node distribution in the actual halo overlap section, the value of 2.0 can achieve a balance between surface smoothness and local detail fidelity. Indicates the first The continuous settlement at each spatial node involved in the interpolation, measured in lux, is introduced into the distance calculation using a numerical stability constant. The modified interpolation weight calculation formula is as follows: ; In the formula, This represents the numerical stability constant, expressed in meters, and has a value of [value missing]. The value of meters is selected based on the following criteria: The value should be much smaller than the spacing between spatial sampling grid nodes to avoid any noticeable impact on the weight calculation between spatial nodes whose normal distance is much greater than zero; at the same time The value should be greater than the rounding error level of double-precision floating-point numbers in computer floating-point arithmetic, effectively protecting against zero division in actual engineering calculations. This value is about six orders of magnitude smaller than the minimum node spacing and about nine orders of magnitude smaller than the floating-point rounding error, meeting the requirements. After completing the longitudinal connection and transverse interpolation stages, a continuous spatial transition compensation surface covering the entire three-dimensional space of the halo overlap section is obtained. The domain of this compensation surface is the three-dimensional spatial region of the halo overlap section, and the value range is the continuous settlement at each point within this region. Each point on the surface... Corresponding to a unique settlement compensation value .

[0035] Step 33.34: Based on the continuous spatial transition compensation surface, extract the settlement compensation values ​​at each location on the surface to obtain halo smooth transition parameter data reflecting the gradual change trend of the halo edge. Specifically, this includes: based on the continuous spatial transition compensation surface, substituting the three-dimensional spatial coordinates of all discrete physical nodes in the halo overlap section of the target spatial physical model into the surface, extracting the settlement compensation values ​​at each physical node location, and for the halo overlap section... uEach physical node's 3D spatial coordinates are input as query coordinates into a continuous spatial transition compensation surface. An inverse distance weighted interpolation formula is used to calculate the corresponding settlement compensation value for that node on the surface. For spatial points within the halo overlap section located between physical nodes but not directly coinciding with any physical sensor location, the target spatial coordinates are substituted into the compensation surface, and a uniform interpolation method is used to extract the settlement compensation value for that point, ensuring that all spatial locations within the halo overlap section can obtain a consistent settlement compensation reference. The 3D spatial coordinates of all locations within the halo overlap section and their corresponding settlement compensation values ​​are summarized and organized to obtain halo smooth transition parameter data. Each record in this data contains three fields: the horizontal coordinates, the vertical coordinates, and the coordinates of the location point, as well as the settlement compensation value extracted from the spatial transition compensation surface at that location point. The halo smooth transition parameter data provides a consistent spatial continuous settlement compensation reference for the entire halo overlap section for topology splicing. In conjunction with this data, there is a spatial boundary at the edge of each halo overlap section that intersects with the normal attenuation section, and the settlement compensation value on this boundary approaches zero.

[0036] In this embodiment of the invention, by extracting all physical nodes in the halo overlap section and determining their spatial arrangement order along the halo expansion direction, the illuminance difference between adjacent physical nodes is calculated one by one to obtain a segmented illuminance drop sequence. The illuminance drop is continuously accumulated and calculated segment by segment along the spatial arrangement direction to obtain the continuous settlement from the edge of the halo to the center. This overcomes the technical problem that the gradual change trend of illuminance in the halo overlap area is difficult to quantify and describe, and obtains halo smooth transition parameter data that can characterize the smooth transition law of the halo boundary.

[0037] In a preferred embodiment of the present invention, step 4 above may include: Step 41: Extract the cumulative settlement amount at each location point in the halo smooth transition parameter data, and determine the illuminance correction amount for each location point within the halo overlap section relative to the normal attenuation trend. Specifically, this includes: extracting the continuous settlement amount at each location point in the halo smooth transition parameter data, and constructing a reference illuminance distribution with a normal attenuation trend. Based on the building information model data of the target space physical model, the IES light distribution files corresponding to each physical light source are imported. A ray tracing algorithm is used to simulate the spatial illuminance distribution under independent operating conditions of each light source. The simulation results of all light sources operating independently are linearly superimposed to obtain the normal attenuation reference illuminance distribution for the entire target space. The ray tracing simulation process takes into account environmental factors such as reflections from interior walls and furniture obstruction, and the deviation between the simulation results and the measured illuminance data in the normal attenuation section is controlled within a preset range. Each grid node in the reference illuminance distribution... A theoretical illuminance value corresponds to the normal illuminance level at that location under conditions of no multi-source overlapping distortion. For any point within the halo overlap zone, the continuous settlement amount is used as a control variable for smooth transition. The corrected illuminance is directly generated based on the reference illuminance distribution, and the correction formula is as follows: ; In the formula, This represents the illuminance value at that location after correction and compensation, in lux. This represents the theoretical illuminance value at that location under normal attenuation trends, expressed in lux. This represents the continuous settlement at that location point, measured in lux. It is derived from the halo smooth transition parameter data. The continuous settlement at the edge of the halo overlap section approaches zero, and the corrected illuminance is seamlessly connected with the reference illuminance of the normal attenuation section. Towards the center of the halo, the continuous settlement gradually changes, leading to a smooth transition of the corrected illuminance and avoiding illuminance jumps.

[0038] Step 42: Apply the illuminance correction amount to the original illuminance values ​​at each location point within the halo overlap section for point-by-point correction and compensation to obtain the corrected illuminance distribution within the halo overlap section. Specifically, this includes: applying the illuminance correction amount to the original illuminance values ​​at each location point within the halo overlap section, and performing point-by-point correction and compensation for each location point. The calculation formula for point-by-point correction and compensation is as follows: ; In the formula, This represents the illuminance value at that location after correction and compensation, in lux. This represents the original measured illuminance value at this location before correction, in lux; This represents the illuminance correction amount at that location point, in lux. The above correction and compensation calculation is performed on each location point within the halo overlap section. When the corrected illuminance value of a certain location point is negative, the corrected illuminance value of that location point is truncated to zero. After completing the point-by-point correction and compensation for the entire area, the corrected illuminance values ​​of all location points and their corresponding three-dimensional spatial coordinates are summarized to obtain the corrected illuminance distribution of the halo overlap section. The corrected illuminance distribution of the halo overlap section smoothly connects with the normal attenuation trend at the edge of the halo, and retains the actual illuminance level of multiple light sources superimposed at the center of the halo.

[0039] Step 43: Extract the position coordinate correspondence between the corrected illuminance distribution of the halo overlap section and the original illuminance distribution of the normal attenuation section at the spatial boundary. Specifically, this includes: extracting the position coordinate correspondence between the corrected illuminance distribution of the halo overlap section and the normal attenuation section in the spatial structure classification set at the spatial boundary; traversing all grid edges in the spatial structure classification set to identify the following types of crossing edges: one endpoint of the grid edge is located within the halo overlap section, and the other endpoint is located within the normal attenuation section. These crossing edges constitute the spatial boundary between the halo overlap section and the normal attenuation section. For each identified crossing edge, record the following information: the spatial coordinates of one endpoint of the halo overlap section and its corrected illuminance value in the corrected illuminance distribution of the halo overlap section; the spatial coordinates of one endpoint of the normal attenuation section and its illuminance value in the original spatial discrete illuminance sampling set; pairing and storing the coordinates and corresponding illuminance values ​​of the two endpoints of the crossing edge to establish a position coordinate correspondence table between the corrected illuminance distribution of the halo overlap section and the original illuminance distribution of the normal attenuation section at the spatial boundary. Each record in the location coordinate correspondence table contains a pair of matching spatial location coordinates, one of which comes from the corrected illuminance distribution of the halo overlap section and the other from the original illuminance distribution of the normal attenuation section. The two are spatially adjacent and located on both sides of the section boundary.

[0040] Step 44: Topologically stitch the corrected illuminance distribution of the halo overlap section and the original illuminance distribution of the normal attenuation section in the three-dimensional spatial coordinate system according to the position coordinate correspondence to obtain a global illumination reconstruction matrix covering the entire target spatial physical model. Specifically, this includes: performing topological stitching in the three-dimensional spatial coordinate system based on the corrected illuminance distribution of the halo overlap section, the original illuminance distribution of the normal attenuation section in the spatial structure classification set, and the position coordinate correspondence. The specific process of the topological stitching operation is as follows: guided by the established position coordinate correspondence, the corrected illuminance values ​​in the corrected illuminance distribution of the halo overlap section are filled into the illuminance data slots of the corresponding halo overlap section positions in the global grid, and the original illuminance values ​​of each position point in the normal attenuation section are filled into the illuminance data slots of the corresponding normal attenuation section positions in the global grid. At the stitching boundary, to avoid numerical jumps between the corrected illuminance values ​​and the original illuminance values ​​at that point, a linear blending transition is used between the grid points on both sides of the boundary crossing edge. For a transition point located on the boundary of a spatial intersection and at a distance from the boundary at a specific ratio, the formula for calculating its spliced ​​illuminance value is as follows: ; In the formula, This represents the final illuminance value at this location within the transition area of ​​the splicing boundary, in lux; This represents the distance weighting coefficient, which is dimensionless and ranges from 0 to 1. Its value is determined by the following rule: when... A value of 0 indicates that the location is on one side of the halo overlap area. A value of 1 indicates that the location is on one side of the normal attenuation zone, within the transition region. Values ​​are taken according to linear changes; This represents the corrected illuminance value corresponding to this location point in the corrected illuminance distribution within the halo overlap area, in lux; This represents the original illuminance value at that location point within the normal attenuation zone, in lux. After filling in the illuminance data for all zones and all boundary transition areas, a three-dimensional numerical matrix is ​​obtained, with the number of rows equal to the number of horizontal vertical grid layers, the number of columns equal to the number of horizontal horizontal grid columns, and the depth equal to the number of vertical grid layers. The value of each element in this three-dimensional numerical matrix represents the reconstructed illuminance value at the corresponding grid location in the target space physical model.

[0041] In this embodiment of the invention, the illuminance correction amount is determined by the cumulative settlement amount of each position point in the halo smooth transition parameter data, and the halo overlap section is corrected and compensated point by point. The position coordinate correspondence between the corrected illuminance distribution and the original illuminance distribution of the normal attenuation section at the boundary is extracted, and topological stitching is performed in the three-dimensional spatial coordinate system accordingly. This overcomes the technical problem that the corrected halo overlap section and the normal attenuation section have data separation and are difficult to seamlessly integrate at the spatial boundary, and obtains a global illumination reconstruction matrix that covers the entire target spatial physical model and has a continuous transition.

[0042] In a preferred embodiment of the present invention, step 5 above may include: Step 51: Arrange and reorganize the global illumination reconstruction matrix according to the input channel format of the lightweight neural network to obtain channel-adapted input data. Specifically, this includes: arranging and reorganizing the matrix according to the input channel format of the lightweight neural network pre-deployed on edge computing nodes, based on the global illumination reconstruction matrix. The lightweight neural network has determined the standard shape of its input tensor during the training phase. Let this standard shape be a three-dimensional array, with its three dimensions corresponding to the height dimension, width dimension, and channel dimension, respectively. The height dimension is equal to the number of grid layers in the target space in the horizontal direction, the width dimension is equal to the number of grid columns in the target space in the horizontal direction, and the channel dimension is equal to the number of grid layers in the target space in the vertical direction. The specific operation of data arrangement and reorganization is: mapping each layer of two-dimensional illuminance slice data of the global illumination reconstruction matrix along the vertical direction to an independent channel of the input tensor. For the first... k Layer depth slices, the two-dimensional illuminance distribution data on these slices are filled up to the input tensor in the original mesh order. kThe corresponding spatial positions of each channel, after reconstruction, the channel-adapted input data obtained is consistent with the global illumination reconstruction matrix in terms of numerical content, and meets the requirements of lightweight neural network input layer in terms of data arrangement format.

[0043] Step 52: Input the channel adaptation input data into the depthwise separable convolutional layer of the lightweight neural network to extract the local correlation patterns of the spatial distribution of illumination step by step, and obtain an intermediate feature map set; input the intermediate feature map set into the pooling layer of the lightweight neural network for spatial downsampling processing to obtain the dimensionality-reduced downsampled feature map. Specifically, this includes: according to the channel adaptation input data, inputting the data into the depthwise separable convolutional layer and pooling layer of the lightweight neural network for step-by-step processing; inputting the channel adaptation input data into the first-level depthwise separable convolutional layer of the lightweight neural network; the training method of the lightweight neural network is as follows: construct a training dataset, each sample contains a set of global illumination reconstruction matrices obtained by physical simulation or on-site measurement as input samples, and a set of corresponding ideal lighting scene control instructions compiled by artificial lighting design experts; the scene control latent vector obtained by encoding the control instruction set is used as the sample label; during the training process, mean squared error is used as the loss function, and the input of the loss function is the predicted latent vector output by the neural network and the sample label latent vector, which is used to quantify the numerical deviation between the two. The Adam optimizer is used to perform iterative parameter updates. The initial learning rate is set to 0.001, and the batch size is set to 32. During training, the learning rate decays at fixed intervals. Iterative training continues until the loss function converges to below a preset threshold, completing the training of the lightweight neural network. After training, the network parameters are quantized and compressed, then deployed to edge computing nodes for feature extraction in the offline inference stage. The depthwise separable convolutional layer consists of a cascaded convolutional sub-layer and a pointwise convolutional sub-layer. In the convolutional sub-layer, each input channel uses a set of independent two-dimensional convolutional kernels for spatial domain filtering. The formula for calculating the intermediate output after channel-by-channel convolution of each input channel is as follows: ; In the formula, The intermediate feature map representing the output of the channel-wise convolutional sub-layer is shown below. individual channels, spatial locations The element value at that position; Indicates the first The two-dimensional convolution kernel with each channel has an offset. The weight values ​​at each location are determined through learning during the training phase of the lightweight neural network; Indicates the first input data of the channel adapter individual channels, spatial locations Illuminance value at that location; This represents the half-window radius of the convolution kernel, dimensionless, typically taking values ​​of 1 or 2, corresponding to... or kernel size; , These represent the row and column indices on the output feature map, respectively, in grid cells. Following the channel-wise convolutional sub-layer, the pointwise convolutional sub-layer uses a size of... The convolution kernel performs cross-channel information mixing across all channels. The formula for calculating pointwise convolution is as follows: ; In the formula, The first feature map representing the output of the pointwise convolutional sub-layer individual channels, spatial locations The element value at that position; The weight matrix representing pointwise convolution connects the first... The input channel and the first The element values ​​of each output channel are determined through parameter learning during the training phase; The input channel is represented by the number of channels that adapt to the input data. After being processed through multiple layers of depthwise separable convolutional layers, an intermediate feature map set containing multi-level spatial local correlation pattern information is obtained. This intermediate feature map set is then input into the pooling layer of a lightweight neural network for spatial downsampling. The pooling layer spatially aggregates the elements within each non-overlapping pooling window for each channel of the input feature map. A typical value for the pooling window size is... With a step size of 2, the spatial size after downsampling is reduced to half of the original size in each dimension. After processing by the pooling layer, a downsampled feature map with reduced spatial resolution but retained effective feature information is obtained.

[0044] Step 53: Input the downsampled feature map into the fully connected layer of the lightweight neural network. This compresses the spatial distribution information in the downsampled feature map into a fixed-dimensional numerical vector, obtaining the scene control latent vector. Specifically, this includes: inputting the downsampled feature map into the fully connected layer of the lightweight neural network to compress the spatial distribution information into a fixed-dimensional numerical vector; expanding the element values ​​of all spatial positions in all channels of the downsampled feature map into a one-dimensional long vector in a fixed order. The total length of the expanded one-dimensional long vector is equal to the product of the number of channels in the downsampled feature map, the downsampled height, and the downsampled width. Then, input the expanded one-dimensional long vector into the fully connected layer. Through matrix multiplication and bias addition, the high-dimensional spatial distribution information is compressed into a fixed-dimensional numerical vector. The calculation formula for the fully connected layer is as follows: ; In the formula, The scene control latent vector is a fixed-dimensional numerical vector. Each element of the vector represents a certain dimension of the implicit lighting spatial distribution feature extracted from the global lighting reconstruction matrix after compression and encoding by a neural network. This represents the weight matrix of the fully connected layer. The number of rows in this matrix is ​​equal to the preset dimension of the scene control latent vector, and the number of columns is equal to the length of the one-dimensional long vector after the downsampled feature map is expanded. The value of each element in the matrix is ​​learned and determined by the backpropagation algorithm during the training phase of the lightweight neural network. This represents a one-dimensional long vector after the downsampled feature map is expanded; This represents the bias vector of the fully connected layer, with a length equal to the preset dimension of the scene control latent vector. The value of each element in the vector is learned and determined during the training phase. The preset dimension of the scene control latent vector ranges from 32 to 256. The reason for this choice is as follows: when the dimension is 32, the compression is relatively high, and the latent vector has limited ability to express the characteristics of the illumination spatial distribution, making it suitable for spaces with relatively simple illumination scene patterns. When the dimension is 256, the latent vector can retain more detailed information about the illumination spatial distribution, making it suitable for spaces with complex illumination scene patterns, but the computational cost increases accordingly. The preset dimension is 128. Experimental verification shows that this value can achieve a reasonable balance between the sufficiency of feature expression and the computational resource consumption of edge computing nodes under the illumination scene pattern complexity of typical exhibition hall spaces.

[0045] In this embodiment of the invention, the global illumination reconstruction matrix is ​​rearranged and reorganized according to the input channel format of a lightweight neural network. The local correlation patterns of the illumination spatial distribution are extracted step by step through depthwise separable convolutional layers and spatial downsampling is performed through pooling layers. The spatial distribution information is compressed into a fixed-dimensional scene control latent vector by a fully connected layer. This overcomes the technical problem that it is difficult to efficiently extract global latent features from a large-scale illumination matrix under the condition of limited computing power at the edge. While reducing the computational complexity and the number of parameters, a compact scene control latent vector with rich expressive power is obtained.

[0046] In a preferred embodiment of the present invention, step 6 above may include: Step 61: The decoder restores the compressed information of each dimension of the scene control latent vector into color temperature control components and brightness control components to obtain a multi-channel lighting control instruction set. Specifically, this includes: inputting the latent vector into the decoder based on the scene control latent vector, performing feature expansion decoding to restore the compressed information of each dimension of the latent vector into color temperature control components and brightness control components. The decoder is a fully connected network module in a lightweight neural network that is symmetrical to the encoder structure. The decoder receives the scene control latent vector as input and expands it step by step from the preset dimension to the dimension of the output space through multi-layer fully connected expansion operations. The first fully connected layer of the decoder expands the scene control latent vector from the preset dimension to the first intermediate dimension. The value of the first intermediate dimension is twice the preset dimension of the scene control latent vector. The calculation formula of the first fully connected layer is as follows: ; In the formula, This represents the output vector of the first fully connected layer of the decoder, with a length equal to the first intermediate dimension. Each element in the vector is a real number and is dimensionless. This represents the weight matrix of the first fully connected layer of the decoder. The number of rows in this matrix is ​​equal to the first intermediate dimension, and the number of columns is equal to the preset dimension of the scene control latent vector. The value of each element in the matrix is ​​learned and determined by the backpropagation algorithm during the training phase of the lightweight neural network. This represents the scene control implicit vector, with a preset dimension of 128, and each element in the vector is a real number. This represents the bias vector of the first fully connected layer of the decoder. Its length is equal to the first intermediate dimension. The value of each element in the vector is learned and determined during the training phase. The initial value is set to an all-zero vector. The activation function is represented by a leaky linear rectified function. This function directly outputs the original value for positive inputs and multiplies the original value by a small leaky slope coefficient for negative inputs. The leaky slope coefficient ranges from 0.01 to 0.02, and its selection is based on the following criteria: When the leaky slope coefficient is 0.01, the slope of the negative half-axis is extremely small, and the activation function behaves similarly to the standard linear rectified function, exhibiting good sparsity during training but potentially leading to permanent neuron inactivation; when the leaky slope coefficient is 0.20, more information is retained in the negative half-axis, making neurons less prone to inactivation but reducing sparsity; and when the leaky slope coefficient is 0.10, this value is a commonly used empirical value in deep learning, and extensive experiments have verified that it achieves a good balance between preventing neuron inactivation and maintaining feature sparsity. The mathematical expression of the leaky linear rectified function is as follows: ; In the formula, The input value for the activation function is a real number. This represents the leakage slope coefficient, which is dimensionless. In this embodiment, it is taken as zero. The second fully connected layer of the decoder receives the output vector of the first fully connected layer. This is further expanded to the dimension of the output space. The dimension of the output space is equal to twice the number of light source channels in the physical light source array. Let the total number of light sources in the physical light source array be denoted as the number of light source channels. The calculation formula for the second fully connected layer is as follows: ; In the formula, This represents the original output vector of the decoder, with a total length equal to twice the number of light source channels. Each element in the vector is a real number, and its value needs to be converted into the actual physical control parameters through a subsequent linear mapping step. This represents the weight matrix of the second fully connected layer of the decoder. The number of rows in this matrix is ​​equal to twice the number of light source channels, and the number of columns is equal to the first intermediate dimension. The value of each element in the matrix is ​​learned and determined during the training phase of the lightweight neural network. This represents the output vector of the first fully connected layer of the decoder; This represents the bias vector of the second fully connected layer of the decoder. Its length is equal to twice the number of light source channels. The value of each element in the vector is learned and determined during the training phase, and the initial value is set to an all-zero vector.

[0047] From the original output vector The rule for parsing the control components of each channel is as follows: the elements of the front light source channel of this vector correspond sequentially to the original color temperature control value of each light source channel, and the elements of the back light source channel correspond sequentially to the original brightness control value of each light source channel. For the ... One physical light source channel, The value range is from 1 to an integer representing the number of light source channels. The location for extracting the original color temperature control value is... The index in the vector is The element, the extraction position of the original value of brightness control is The index in the vector is the number of light source channels plus... The elements, due to The values ​​of each element in the vector are unbounded, while the actual physical light source control parameters have a defined physical range. Therefore, the original output values ​​need to be transformed to the target control range through a linear mapping. The linear mapping formula for the color temperature control components is as follows: ; In the formula, Indicates the first The final color temperature control component of each physical light source channel, in Kelvin, is the color temperature instruction value actually written into the physical light source driver circuit. express The first vector The value of the nth element, i.e., the nth... The original color temperature control values ​​corresponding to each light source channel are real numbers and dimensionless. express The minimum value among several elements of the front light source channel in the vector is a real number, dimensionless, and is obtained by... The extreme value is obtained by segmenting the vector into corresponding segments; express The maximum value among several elements of the front light source channel in the vector is a real number, dimensionless, and the retrieval method is the same as above; This represents the lower limit of the color temperature control range. In this embodiment, it is set to 2700 Kelvin, which corresponds to the warm white light color temperature. This represents the upper limit of the color temperature control range. In this embodiment, it is taken as 6500 Kelvin, corresponding to the cool white light color temperature. The linear mapping formula for the brightness control component is as follows: ; In the formula, Indicates the first The final brightness control component of each physical light source channel is a percentage of the maximum brightness of that light source, dimensionless, and this value is the brightness command value actually written into the physical light source driver circuit. This represents the number of light source channels, i.e., the total number of light sources in the physical light source array, and is dimensionless. express The first vector The value of the nth element, i.e., the nth... The original brightness control values ​​corresponding to each light source channel are real numbers and dimensionless. express The minimum value among several elements of the rear light source channel in the vector is a real number and dimensionless. express The maximum value among several elements of the rear light source channel in the vector is a real number and dimensionless. The lower limit of the brightness control range is 5% in this embodiment. The basis for this value is: when the brightness control value is lower than 5%, the output luminous flux of the LED physical light source is extremely low, and the human eye can hardly perceive its lighting effect. At the same time, the pulse width modulation duty cycle of the LED driving circuit is too small in this low brightness range, resulting in unstable output. Therefore, the lower limit is cut off at 5%. This represents the upper limit of the brightness control range, which is 100% in this embodiment, corresponding to the maximum rated output brightness of the light source. After extracting the original values ​​and performing linear mapping transformation for all light source channels, the color temperature control components and brightness control components of each channel are organized according to the light source number. The organization method is as follows: a control record is generated for each physical light source channel. This record contains three fields: physical light source number, color temperature control component, and brightness control component. The set of all control records constitutes the multi-channel lighting control instruction set.

[0048] Step 62: Extract the physical light source number corresponding to each channel in the multi-channel lighting control instruction set. Encapsulate the color temperature control component and brightness control component of each channel into a data frame according to the communication protocol of the lighting control bus to obtain a bus driver data packet. Specifically, this includes: extracting the physical light source number and its color temperature control component and brightness control component corresponding to each channel in the multi-channel lighting control instruction set, and encapsulating the data frame according to the communication protocol format of the lighting control bus. The communication protocol of the lighting control bus defines the structure of a standard data frame. A complete data frame includes four components: a frame header identifier field, a destination address field, a control data field, and a check field. The frame header identifier field is a fixed byte sequence used to identify the start boundary and frame type of the data frame. The receiving end identifies the start position of the frame by detecting this fixed byte sequence. The destination address field is written with the unique number of the physical light source node to which the frame is targeted. This number corresponds to the channel number of the multi-channel lighting control instruction set in step 61. The control data field contains two numerical fields: the first numerical field is written with the binary value of the color temperature control component after quantization encoding, and the second numerical field is written with the binary value of the brightness control component after quantization encoding. The check field contains the cyclic redundancy check value calculated based on all bytes of the frame header identifier field, destination address field, and control data field. The receiving end recalculates and compares the check value to determine whether a bit error occurred during data frame transmission. The structure expression of the complete data frame encapsulated from each physical light source node is as follows: ; In the formula, Indicates that for the first The complete data frame obtained by encapsulating each physical light source node is stored in the transmission buffer in the form of a byte sequence; The frame header identifier field is a fixed-length byte sequence. In this embodiment, a fixed two-byte pattern is used. The probability of the two appearing consecutively in a random data stream is extremely low, which can effectively reduce frame synchronization errors caused by coincidental matching of data segment content with the frame header. Indicates the first The target address field for each physical light source node occupies one byte and can represent an address range of 0 to 255. The value of this byte is equal to the unique number of the physical light source node in the lighting control bus. ; Indicates the first The data segment of each channel's color temperature control component, after quantization and encoding, occupies two bytes. The quantization and encoding rule for the color temperature control component is as follows: the color temperature control component obtained by linear mapping in step 61... Convert to a two-byte unsigned integer value using the following formula: ; In the formula, The quantization encoding result of the color temperature control component is an unsigned 16-bit integer value, ranging from 0 to 65535, and is written to the data frame in two-byte big-endian format. This indicates the rounding operation; Indicates the first The color temperature control component of each physical light source channel, in Kelvin; This indicates the lower limit of the color temperature control range, with a value of 2700 Kelvin. This indicates the upper limit of the color temperature control range, with a value of 6500 Kelvin. This represents the maximum value of a 16-bit unsigned integer, which is 65535. Indicates the first The data segment of each channel's brightness control component, after quantization and encoding, also occupies two bytes. The quantization and encoding formula for the brightness control component is as follows: ; In the formula, The quantization encoding result of the brightness control component is an unsigned sixteen-bit integer value, ranging from 0 to 65535, and is written into the data frame in two-byte big-endian format. Indicates the first The brightness control components of each physical light source channel are percentage values ​​and are dimensionless. This indicates the lower limit of the brightness control range, with a value of 5%. This indicates the upper limit of the brightness control range, with a value of 100%. Indicates the first The cyclic redundancy check (CRC) field of each data frame occupies two bytes. This check value is identified by the frame header field. Target address field Color temperature data segment and brightness data segment All bytes are used as input data, and the standard Cyclic Redundancy Check (CRC) algorithm is used to calculate the CRC. The calculation process is as follows: the seven bytes to be checked are concatenated into a 56-bit binary sequence in the order of transmission. This binary sequence can be regarded as a coefficient taken from... The 55th degree polynomial. This polynomial is denoted as the data polynomial to be verified, and its mathematical expression is as follows: ; In the formula, This represents the polynomial of the data to be verified, and its coefficients... The value is 0 or 1, and each coefficient corresponds to a bit in the data to be verified. Indicates the first in the data to be verified The bit value of a bit can be either 0 or 1; The indefinite variables of the polynomial are formal variables, dimensionless, and the data to be verified are polynomials. Multiply The result is the 16th power of the remainder polynomial, divided by the preset generator polynomial. The sixteen coefficients of the remainder polynomial are the cyclic redundancy check values. The mathematical expression for this calculation process is as follows: ; In the formula, The calculated cyclic redundancy check value is a 16-bit unsigned integer value, written to the check field of the data frame in big-endian format (two bytes). This represents a polynomial operation that shifts left by sixteen bits, corresponding to reserving space for a sixteen-bit check value; This represents the modular operation of a polynomial, that is, taking the remainder polynomial obtained by polynomial division; The generator polynomial is represented in this embodiment, which uses the standard CRC16CCITT generator polynomial. Its coefficient form is expressed as follows: ; In the formula , , The constant term 1 corresponds to the terms with a coefficient of 1 in the generator polynomial, while the coefficients of the remaining terms are 0. The Hamming distance characteristics of the CRC16CCITT generator polynomial, verified through theoretical analysis and engineering practice, can detect all single-bit errors, all double-bit errors, all odd-numbered bit errors, all burst errors with a length not exceeding sixteen bits, and over 99.99% of burst errors with a length of seventeen bits. This is applicable to multi-channel lighting control command sets from... to For each light source channel, the above data frame encapsulation operation is performed one by one, and the data is aggregated to obtain the bus driver data packet. The bus driver data packet is a packet consisting of... A sequence of bytes consisting of data frames arranged in ascending order of light source number is prepared for transmission via the lighting control bus.

[0049] Step 63: Transmit the bus drive data packet to the corresponding physical light source node in the physical light source array via the light control bus; drive the physical light source array to perform offline scene rendering operation according to the color temperature control component and brightness control component received by the physical light source node, and complete the offline generation of the lighting scene. Specifically, this includes: according to the bus drive data packet, transmitting each data frame in the data packet to the corresponding physical light source node in the physical light source array via the light control bus, driving each physical light source node to perform offline scene rendering according to the received control parameters, and the light control bus performs addressing and routing according to the target address field in each data frame during the transmission process, so as to accurately deliver each data frame to the physical light source node that matches the target address. After receiving a complete data frame, each physical light source node verifies the data in the frame: it recalculates the cyclic redundancy check value of the received data and compares it with the check field carried in the frame. For frames that pass the verification, the physical light source node parses the quantized values ​​of the color temperature control component and the brightness control component. For frames that fail the verification, the physical light source node sends a retransmission request to the bus, triggering the retransmission of the data frame. After parsing the color temperature and brightness quantized values, the physical light source node reverse maps them into the corresponding color temperature control quantity and brightness control quantity, and writes them into the driver circuit register inside the physical light source node.

[0050] The driving circuit adjusts the driving current ratio of different color temperature light-emitting units in the physical light source according to the color temperature control value to achieve the target color temperature output. The driving circuit also adjusts the total driving power of the light-emitting units through pulse width modulation according to the brightness control value to achieve the target brightness output. Since the inference and command issuance processes of the edge computing nodes are completed within the local area network and do not rely on a public network connection, completely offline operation is achieved. All physical light source nodes synchronously execute output adjustments according to their respective received color temperature and brightness control values, completing the overall rendering of the lighting scene and the entire offline generation process of the lighting scene based on a lightweight edge neural network.

[0051] In this embodiment of the invention, the compressed information of each dimension in the scene control latent vector is restored into color temperature control components and brightness control components by the decoder, and the control components corresponding to each physical light source are encapsulated into data frames according to the communication protocol of the light control bus to generate bus driver data packets, which are transmitted to each physical light source node via the light control bus to perform offline scene rendering. This overcomes the technical problem that abstract implicit features are difficult to directly map into executable hardware control instructions, and realizes offline automatic generation of light scenes from light perception, feature extraction, instruction generation to light source driving in a closed loop.

[0052] like Figure 2 As shown, embodiments of the present invention also provide an offline lighting scene generation system based on a lightweight edge neural network, comprising: The acquisition module is used to acquire the initial illuminance field data of the target space physical model collected by the multi-dimensional illumination sensor array, calculate the spatial gradient magnitude of the illuminance field point by point along the three-dimensional spatial coordinate axis, and obtain the local attenuation feature sequence. The partitioning module is used to construct a three-dimensional iso-illuminance surface based on the local attenuation feature sequence, calculate the degree of local bending deformation of the three-dimensional iso-illuminance surface in the spatial normal direction, and obtain the spatial bending distribution matrix; based on the deformation extreme value of the spatial bending distribution matrix, the target spatial physical model is divided into halo overlap segment and normal attenuation segment, and a spatial structure classification set is obtained. The calculation module is used to calculate the continuous settlement of the illuminance difference between adjacent physical nodes pixel by pixel based on the halo overlap section in the spatial structure classification set, so as to obtain the halo smooth transition parameter data. The stitching module is used to topologically stitch the halo smooth transition parameter data with the normal attenuation segment in the spatial structure classification set to obtain a global illumination reconstruction matrix covering the entire target space. The extraction module is used to input the global illumination reconstruction matrix into a lightweight neural network pre-deployed on edge computing nodes to extract the implicit features of the illumination spatial distribution and obtain the scene control latent vector. The driver module is used to expand and decode the scene control latent vector execution features to generate a multi-channel lighting control instruction set containing color temperature and brightness dimensions; the multi-channel lighting control instruction set is sent to the lighting control bus to drive the physical light source array to perform offline scene rendering and complete the offline generation of the lighting scene.

[0053] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0054] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for offline generation of lighting scenes based on a lightweight edge neural network, characterized in that, The method includes: Acquire the initial illuminance field data of the target spatial physical model collected by the multi-dimensional illumination sensor array, calculate the spatial gradient magnitude of the illuminance field point by point along the three-dimensional spatial coordinate axis, and obtain the local attenuation characteristic sequence. A three-dimensional iso-illuminance surface is constructed based on the local attenuation feature sequence. The degree of local bending deformation of the three-dimensional iso-illuminance surface in the spatial normal direction is calculated to obtain the spatial bending distribution matrix. Based on the deformation extreme values ​​of the spatial bending distribution matrix, the target spatial physical model is divided into halo overlap segments and normal attenuation segments to obtain a spatial structure classification set. Based on the halo overlap section in the spatial structure classification set, the continuous settlement of the illuminance difference between adjacent physical nodes is calculated pixel by pixel to obtain halo smooth transition parameter data; The halo smooth transition parameter data and the normal attenuation segment in the spatial structure classification set are topologically stitched together to obtain a global illumination reconstruction matrix covering the entire target space. The global illumination reconstruction matrix is ​​input into a lightweight neural network pre-deployed on edge computing nodes to extract the implicit features of the spatial distribution of illumination and obtain the scene control latent vector. The scene control latent vector is executed by feature expansion and decoding to generate a multi-channel lighting control instruction set containing color temperature and brightness dimensions; the multi-channel lighting control instruction set is sent to the lighting control bus to drive the physical light source array to perform offline scene rendering, thus completing the offline generation of the lighting scene.

2. The offline generation method for lighting scenes based on a lightweight edge neural network according to claim 1, characterized in that, Acquire the initial illuminance field data of the target space physical model collected by a multi-dimensional illumination sensor array, calculate the spatial gradient magnitude of the illuminance field point by point along the three-dimensional spatial coordinate axes, and obtain the local attenuation characteristic sequence, including: The original illuminance readings of each physical node in the multidimensional illumination sensor array are read one by one according to the preset spatial sampling grid. The mapping relationship between each physical node and the three-dimensional spatial coordinate position is established to obtain the spatial discrete illuminance sampling set. Based on the spatial discrete illuminance sampling set, the illuminance difference between adjacent physical nodes is extracted along the first coordinate axis of the three-dimensional spatial coordinate system to obtain the first direction illuminance difference component. Based on the spatial discrete illuminance sampling set, the illuminance difference between adjacent physical nodes is extracted along the second and third coordinate axes of the three-dimensional spatial coordinate system, respectively, to obtain the second-direction illuminance difference component and the third-direction illuminance difference component, and then the three-axis illuminance difference component set is obtained. Based on the set of illuminance difference components in the three axes, the illuminance difference components of each physical node in the first, second, and third axes are compared with the distance between the physical nodes in the corresponding coordinate axes to obtain the local attenuation feature sequence.

3. The offline generation method for lighting scenes based on a lightweight edge neural network according to claim 2, characterized in that, A three-dimensional iso-illuminance surface is constructed based on the local attenuation characteristic sequence. The degree of local bending deformation of the three-dimensional iso-illuminance surface in the spatial normal direction is calculated to obtain the spatial bending distribution matrix. Based on the deformation extrema of the spatial bending distribution matrix, the target spatial physical model is divided into halo overlap segments and normal attenuation segments, resulting in a spatial structure classification set, including: Based on the illuminance values ​​of each physical node in the local attenuation feature sequence, adjacent physical nodes with equal illuminance values ​​are extracted. In the three-dimensional spatial coordinate system, the extracted adjacent physical nodes are connected by a spatial mesh to construct a three-dimensional isoluminance surface covering the target spatial physical model. Based on the normal direction of each spatial point on the three-dimensional iso-illuminance surface, the curvature change of the iso-illuminance surface along the normal direction at the point of the normal direction is calculated, and a spatial curvature distribution matrix reflecting the undulation of the spatial morphology is obtained. Based on the curvature change amplitude of each spatial location point in the spatial curvature distribution matrix, the locations where the curvature change amplitude exceeds the preset curvature limit are extracted and classified into the halo overlap segment, and the locations where the curvature change amplitude does not exceed the preset curvature limit are extracted and classified into the normal attenuation segment, thus obtaining the spatial structure classification set.

4. The offline generation method for lighting scenes based on a lightweight edge neural network according to claim 3, characterized in that, Step 3: Based on the halo overlap segments in the spatial structure classification set, calculate the continuous settlement of the illuminance difference between adjacent physical nodes pixel by pixel to obtain halo smooth transition parameter data, including: Extract all physical nodes within the halo overlap section of the spatial structure classification set, and determine the spatial arrangement order of all physical nodes along the halo expansion direction; Extract the illuminance values ​​of adjacent physical nodes in the spatial arrangement order, calculate the difference in illuminance values ​​between adjacent physical nodes one by one, and obtain the illuminance drop sequence segment by segment. The illuminance difference values ​​in the segmented illuminance difference sequence are continuously accumulated and calculated along the spatial arrangement direction to obtain the continuous settlement amount from the halo edge to the halo center; the continuous settlement amount of all physical nodes in the halo overlap section is integrated to obtain halo smooth transition parameter data reflecting the gradual change trend of the halo edge.

5. The offline generation method for lighting scenes based on a lightweight edge neural network according to claim 3, characterized in that, The illuminance difference values ​​in the segmented illuminance difference sequence are continuously accumulated and calculated along the spatial arrangement direction to obtain the continuous settlement amount from the halo edge to the halo center; the continuous settlement amounts of all physical nodes within the halo overlap section are integrated to obtain halo smooth transition parameter data reflecting the gradual trend of the halo edge, including: Based on the starting physical node in the spatial arrangement order, the illuminance value of the starting physical node is set as the cumulative calculation reference illuminance value to obtain the reference illuminance at the edge of the halo. Based on the segmented illuminance drop sequence and the reference illuminance at the halo edge, the illuminance drop value is gradually decreased and accumulated from the reference illuminance along the spatial arrangement direction to obtain the continuous settlement amount of each physical node relative to the edge within the halo overlap section. Based on the continuous settlement of all physical nodes within the halo overlap section, the continuous settlement is mapped to the corresponding three-dimensional spatial coordinates according to the spatial arrangement order to construct a complete spatial transition compensation surface, thereby obtaining halo smooth transition parameter data that reflects the gradual change trend of the halo edge.

6. The offline generation method for lighting scenes based on a lightweight edge neural network according to claim 5, characterized in that, Based on the continuous settlement of all physical nodes within the halo overlap section, the continuous settlement is mapped to the corresponding three-dimensional spatial coordinates according to the spatial arrangement order to construct a complete spatial transition compensation surface. This yields halo smooth transition parameter data reflecting the gradual change trend of the halo edge, including: Based on the continuous settlement and spatial arrangement of all physical nodes within the halo overlap section, the position coordinates of each physical node in the three-dimensional spatial coordinate system are extracted to obtain the set of node spatial coordinates. Based on the set of node spatial coordinates, the continuous settlement is spatially bound to the location coordinates to obtain spatial node data with settlement attributes. Based on spatial node data with settlement attributes, adjacent spatial nodes are connected by mesh along the spatial arrangement direction to obtain a continuous spatial transition compensation surface. Based on the continuous spatial transition compensation surface, the settlement compensation values ​​at each position on the surface are extracted to obtain halo smooth transition parameter data that reflects the gradual change trend of the halo edge.

7. The offline generation method for lighting scenes based on a lightweight edge neural network according to claim 4, characterized in that, By topologically stitching the halo smooth transition parameter data with the normal attenuation segments in the spatial structure classification set, a global illumination reconstruction matrix covering the entire target space is obtained, including: Extract the cumulative settlement amount at each location point in the halo smooth transition parameter data, and determine the illuminance correction amount at each location point in the halo overlap section relative to the normal attenuation trend; The illuminance correction is applied to the original illuminance values ​​at each location point within the halo overlap section for point-by-point correction and compensation, resulting in the corrected illuminance distribution of the halo overlap section. Extract the position coordinate correspondence between the corrected illuminance distribution in the halo overlap section and the original illuminance distribution in the normal attenuation section at the spatial boundary; The corrected illuminance distribution in the halo overlap section and the original illuminance distribution in the normal attenuation section are topologically stitched together in a three-dimensional spatial coordinate system according to the position coordinate correspondence, to obtain a global illumination reconstruction matrix covering the entire target spatial physical model.

8. The offline generation method for lighting scenes based on a lightweight edge neural network according to claim 5, characterized in that, The global illumination reconstruction matrix is ​​input into a lightweight neural network pre-deployed on edge computing nodes to extract implicit features of the illumination spatial distribution, resulting in scene control latent vectors, including: The global illumination reconstruction matrix is ​​rearranged and reorganized according to the input channel format of a lightweight neural network to obtain channel-adapted input data; The channel-adapted input data is fed into the depthwise separable convolutional layer of the lightweight neural network to extract the local correlation patterns of the spatial distribution of illumination step by step, resulting in an intermediate feature map set. The intermediate feature map set is then fed into the pooling layer of the lightweight neural network for spatial downsampling to obtain the dimensionality-reduced downsampled feature map. The downsampled feature map is input into the fully connected layer of the lightweight neural network, and the spatial distribution information in the downsampled feature map is compressed into a fixed-dimensional numerical vector to obtain the scene control latent vector.

9. The offline generation method for lighting scenes based on a lightweight edge neural network according to claim 6, characterized in that, The scene control latent vectors are subjected to feature expansion and decoding to generate a multi-channel lighting control instruction set containing color temperature and brightness dimensions. This multi-channel lighting control instruction set is then sent to the lighting control bus to drive the physical light source array to perform offline scene rendering, completing the offline generation of the lighting scene, including: The decoder restores the compressed information of each dimension in the scene control latent vector into color temperature control components and brightness control components, thus obtaining a multi-channel lighting control instruction set. Extract the physical light source number corresponding to each channel in the multi-channel lighting control instruction set, and encapsulate the color temperature control component and brightness control component of each channel into data frames according to the communication protocol of the lighting control bus to obtain the bus driver data packet; The bus drive data packet is transmitted to the corresponding physical light source node in the physical light source array via the light control bus; the physical light source array is driven to perform offline scene rendering operations according to the color temperature control component and brightness control component received by the physical light source node, thus completing the offline generation of the light scene.

10. An offline lighting scene generation system based on a lightweight edge neural network, the system implementing the method as described in any one of claims 1 to 9, characterized in that, include: The acquisition module is used to acquire the initial illuminance field data of the target space physical model collected by the multi-dimensional illumination sensor array, calculate the spatial gradient magnitude of the illuminance field point by point along the three-dimensional spatial coordinate axis, and obtain the local attenuation feature sequence. The partitioning module is used to construct a three-dimensional iso-illuminance surface based on the local attenuation feature sequence, calculate the degree of local bending deformation of the three-dimensional iso-illuminance surface in the spatial normal direction, and obtain the spatial bending distribution matrix; based on the deformation extreme value of the spatial bending distribution matrix, the target spatial physical model is divided into halo overlap segment and normal attenuation segment, and a spatial structure classification set is obtained. The calculation module is used to calculate the continuous settlement of the illuminance difference between adjacent physical nodes pixel by pixel based on the halo overlap section in the spatial structure classification set, so as to obtain the halo smooth transition parameter data. The stitching module is used to topologically stitch the halo smooth transition parameter data with the normal attenuation segment in the spatial structure classification set to obtain a global illumination reconstruction matrix covering the entire target space. The extraction module is used to input the global illumination reconstruction matrix into a lightweight neural network pre-deployed on edge computing nodes to extract the implicit features of the illumination spatial distribution and obtain the scene control latent vector. The driver module is used to expand and decode the scene control latent vector execution features to generate a multi-channel lighting control instruction set containing color temperature and brightness dimensions; the multi-channel lighting control instruction set is sent to the lighting control bus to drive the physical light source array to perform offline scene rendering and complete the offline generation of the lighting scene.

Images