Intelligent control method and system of intelligent lamp based on industrial cloud platform

Abstract

This invention discloses an intelligent control method and system for smart lighting fixtures based on an industrial cloud platform. The invention relates to the technical field of industrial cloud platforms. In this system, a multidimensional state tensor is input into a corresponding spatiotemporal graph neural network. Using smart lighting fixtures as nodes and the light spillover interference and power line carrier coupling relationships between these fixtures as edges, a spatial topology graph is further constructed, improving its accuracy. Message passing is performed on this topology graph, and a temporal convolutional network is used to capture the long-term evolution of lighting demand, thereby outputting a multidimensional decision space representing the collaborative control of the smart lighting fixture cluster. Within this multidimensional decision space, the luminous efficacy quality index of the smart lighting fixtures, the system energy efficiency ratio, and the lighting light decay index are used as triple constraints. A Pareto optimal frontier search is performed simultaneously to generate a global intelligent control strategy for the smart lighting fixture cluster, improving the accuracy of the global intelligent control strategy.

Description

Technical Field

[0008] , ,

[0007] , ,

[0006] , ,

[0009]

[0001] The present invention relates to the technical field of industrial cloud platforms, and particularly to an intelligent control method and system for intelligent lamps based on an industrial cloud platform. Background Art

[0002] With the rapid development of Internet of Things technology and smart home, intelligent lighting systems have evolved from single lamp dimming control to collaborative control of a whole-house intelligent lamp cluster. Existing intelligent lighting systems usually rely on edge gateways or cloud platforms to perform simple linkage control on multiple lamps according to ambient light sensor data.

[0003] When constructing a group control network for existing intelligent lighting systems, most of them are logically networked based on simple device lists, room divisions or pure geographical spatial distances. This topological structure that is divorced from physical electromagnetics and photometry completely ignores the "power line carrier coupling impedance" differences between different lamps in the same household power grid due to the sharing of phase lines, and also ignores the "light spill interference" existing in three-dimensional space between lamps with different light distribution curves, affecting the accuracy of the spatial topology map and resulting in a relatively low accuracy of the global intelligent control strategy. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art. The present invention provides an intelligent control method and system for intelligent lamps based on an industrial cloud platform.

[0005] An embodiment of the present invention provides an intelligent control method for intelligent lamps based on an industrial cloud platform, including:

[0006] Collect multi-source heterogeneous data of intelligent lamps, trigger a multi-scale feature extraction network along the associated content of the multi-source heterogeneous data in the spatio-temporal dimension, and map multiple state features corresponding to the multi-source heterogeneous data into a multi-dimensional state tensor;

[0007] In the industrial cloud platform, input the multi-dimensional state tensor into the corresponding spatio-temporal graph neural network, and further construct a spatial topology map with intelligent lamps as nodes and the light spill interference and power line carrier coupling relationships between each intelligent lamp as edges;

[0008] Perform message passing on the spatial topology map, and capture the evolution law of long-term lighting requirements in combination with a temporal convolutional network, so as to output a multi-dimensional decision space representing the collaborative control of an intelligent lamp cluster, and the multi-dimensional decision space covers a physical execution dimension, an environment-energy efficiency dimension and a service life dimension;

[0009] Within the multi-dimensional decision space, use the light efficiency quality index, system energy efficiency ratio and lighting light decay index of intelligent lamps as triple constraint conditions, and synchronously perform a Pareto optimal front search to generate a global intelligent control strategy for the intelligent lamp cluster.

[0010] This invention provides an intelligent control system for smart lighting fixtures based on an industrial cloud platform. This intelligent control system applies the aforementioned intelligent control method for smart lighting fixtures based on an industrial cloud platform. The intelligent control system for smart lighting fixtures based on an industrial cloud platform includes:

[0011] The multidimensional state tensor module is used to collect multi-source heterogeneous data from smart lighting fixtures, trigger a multi-scale feature extraction network along the spatiotemporal correlation of the multi-source heterogeneous data, and map multiple state features corresponding to the multi-source heterogeneous data into a multidimensional state tensor.

[0012] The industrial cloud platform module is used to input multidimensional state tensors into the corresponding spatiotemporal graph neural network in the industrial cloud platform, and further construct a spatial topology graph with smart lamps as nodes and the light spillover interference and power line carrier coupling relationship between smart lamps as edges.

[0013] The multidimensional decision module is used to pass messages on the spatial topology graph and combine it with a temporal convolutional network to capture the long-term evolution of lighting demand, thereby outputting a multidimensional decision space that represents the collaborative control of intelligent lighting clusters. The multidimensional decision space covers the physical execution dimension, the environmental-energy efficiency dimension, and the service life dimension.

[0014] The intelligent control module is used to simultaneously perform Pareto optimal frontier search within a multi-dimensional decision space, using the luminous efficacy quality index of the intelligent lighting fixtures, the system energy efficiency ratio, and the lighting light decay index as triple constraints, in order to generate a global intelligent control strategy for the intelligent lighting fixture cluster.

[0015] Compared with the prior art, the beneficial effects of the present invention are:

[0016] (1) Collect multi-source heterogeneous data of smart lamps, trigger a multi-scale feature extraction network along the spatiotemporal correlation of the multi-source heterogeneous data, and map the multiple state features corresponding to the multi-source heterogeneous data into multi-dimensional state tensors; in the industrial cloud platform, input the multi-dimensional state tensors into the corresponding spatiotemporal graph neural network, take the smart lamps as nodes and the light spillover interference and power line carrier coupling relationship between each smart lamp as edges, further construct the spatial topology graph, introduce the multi-dimensional state tensor, further control the industrial cloud platform, and improve the accuracy of the spatial topology graph.

[0017] (2) Message passing is performed on the spatial topology graph, and the evolution of long-term lighting demand is captured by combining the temporal convolutional network, thereby outputting a multi-dimensional decision space representing the collaborative control of the intelligent lighting cluster. The multi-dimensional decision space covers the physical execution dimension, the environment-energy efficiency dimension and the service life dimension. In the multi-dimensional decision space, the luminous efficacy quality index of the intelligent lighting fixtures, the system energy efficiency ratio and the lighting light decay index are used as triple constraints. Pareto optimal frontier search is performed simultaneously to generate a global intelligent control strategy for the intelligent lighting cluster. The multi-dimensional decision space is further controlled, and the triple constraints and the intelligent lighting cluster are fully considered, which improves the accuracy of the global intelligent control strategy. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the intelligent control method for smart lighting fixtures based on an industrial cloud platform according to an embodiment of the present invention.

[0019] Figure 2 This is a flowchart illustrating step S11 in the intelligent control method for intelligent lighting fixtures based on an industrial cloud platform according to an embodiment of the present invention.

[0020] Figure 3 This is a flowchart illustrating step S12 in the intelligent control method for intelligent lighting fixtures based on an industrial cloud platform according to an embodiment of the present invention.

[0021] Figure 4 This is a flowchart illustrating step S13 in the intelligent control method for intelligent lighting fixtures based on an industrial cloud platform according to an embodiment of the present invention.

[0022] Figure 5 This is a flowchart illustrating step S14 in the intelligent control method for intelligent lighting fixtures based on an industrial cloud platform according to an embodiment of the present invention.

[0023] Figure 6 This is a schematic diagram of the structural composition of an intelligent control system for smart lighting fixtures based on an industrial cloud platform, as described in an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0025] Please see Figures 1 to 6 A smart control method for smart lighting fixtures based on an industrial cloud platform is proposed and applied to industrial cloud platform scenarios. The smart control method for smart lighting fixtures based on an industrial cloud platform includes:

[0026] Step S11: Collect multi-source heterogeneous data of smart lamps, trigger a multi-scale feature extraction network along the spatiotemporal correlation of the multi-source heterogeneous data, and map the multiple state features corresponding to the multi-source heterogeneous data into multi-dimensional state tensors.

[0027] Step S12: In the industrial cloud platform, the multidimensional state tensor is input into the corresponding spatiotemporal graph neural network. The spatial topology graph is further constructed with smart lamps as nodes and the light spillover interference and power line carrier coupling relationship between smart lamps as edges.

[0028] Step S13: Message passing is performed on the spatial topology graph, and the long-term evolution of lighting demand is captured by combining a temporal convolutional network, thereby outputting a multi-dimensional decision space that represents the collaborative control of intelligent lighting clusters. The multi-dimensional decision space covers the physical execution dimension, the environment-energy efficiency dimension, and the service life dimension.

[0029] Step S14: In the multi-dimensional decision space, using the luminous efficacy quality index of the intelligent lighting fixtures, the system energy efficiency ratio, and the lighting light decay index as triple constraints, simultaneously perform Pareto optimal frontier search to generate a global intelligent control strategy for the intelligent lighting fixture cluster.

[0030] refer to Figure 2 In step S11, the specific steps are as follows:

[0031] S111: Multi-source heterogeneous data is synchronously collected through the sensing array built into the smart lamp. This multi-source heterogeneous data includes real-time illuminance, node voltage and current waveforms, and driving junction temperature. Based on the identification of multi-source heterogeneous data in the spatiotemporal dimension, the corresponding related content is determined and the related content is input into the pre-constructed multi-scale feature extraction network.

[0032] S112: By using parallel dilated convolution kernels and cross-modal attention mechanisms in the multi-scale feature extraction network, nonlinear projection and alignment of multi-source heterogeneous data are performed to filter out transient noise in the environment and extract high-order interaction features. These features are then combined with multiple state features corresponding to the multi-source heterogeneous data and reorganized according to the preset physical semantic dimension to map them into a multi-dimensional state tensor that represents the coupling relationship between the current operating mode of the lamp and the local environment.

[0033] In the embodiments of this application, multi-source heterogeneous data is synchronously collected through the sensing array built into the smart lamp. The multi-source heterogeneous data includes real-time illuminance flux, node voltage and current waveforms, and driving junction temperature. Based on the identification of multi-source heterogeneous data in the spatiotemporal dimension, the corresponding related content is determined. The related content is input into a pre-constructed multi-scale feature extraction network, which is compatible with the overall consideration of multi-source heterogeneous data in the spatiotemporal dimension and ensures the accuracy of the corresponding related content.

[0034] At this time, the drive control motherboard of the smart lighting fixture integrates a multi-source sensing array, which triggers a synchronous sampling interrupt under a unified system clock beat. At this time, the real-time illuminance flux analog signal after the ambient space is mixed is collected by the front-mounted micro photodiode, the high-frequency transient signal of the node AC / DC voltage and current waveform is collected by the Hall current sensor and voltage divider resistor network connected in series in the LED load circuit, and the slowly changing physical signal of the drive junction temperature is collected by the thermistor or single-bus temperature sensor attached to the aluminum substrate or the surface of the drive MOSFET. The above three types of signals are converted into digital sequences by a high-frequency analog-to-digital converter (ADC) and then forcibly aligned to the same timestamp to form an initial heterogeneous dataset.

[0035] Because the aforementioned heterogeneous data have significant differences in sampling frequency and physical response delay, such as electrical waveforms being transient at the millisecond level and junction temperatures being slow-state at the second or even minute level, it is necessary to perform correlation decoupling and identification in the spatiotemporal dimension. In the time dimension, time series slices based on sliding windows are constructed to identify the time difference constant between the phase change points of voltage and current waveforms and the step change points of illuminance flux, so as to establish the dynamic hysteresis correlation between electrical response and optical output.

[0036] In the spatial dimension, based on the heat transfer and optical topology of the lamp, a physical coupling matrix of "chip junction temperature - lamp cavity temperature - spatial illuminance flux" is established. This identifies the hidden radiation interference that the heat accumulation caused by changes in electrical power consumption may cause to the optical sensor in space. Through the cross-validation of the above spatiotemporal dimensions, asynchronous pseudo-signals caused by the physical characteristics of the sensor itself are eliminated, and a set of related content that can truly reflect the external environmental state and internal operating state of the lamp is extracted.

[0037] The established set of related content is standardized and tensor-encapsulated. For related content with high-frequency fluctuation characteristics, such as illuminance flux and electrical waveform, Z-score standardization is used to eliminate hardware gain differences between different batches of lamps. For related content with monotonically increasing characteristics, such as junction temperature, max-min normalization is used to constrain it to the sensitive range of a specific activation function. The processed multidimensional related content is spliced ​​and recombined according to a preset channel order to form a three-dimensional input tensor Xin∈RC×W with a fixed-dimensional structure, where C represents the number of physical channels of the related content, such as illuminance channel, electrical channel, and temperature channel, and W represents the length of the time sliding window. Finally, this input tensor is fed into a pre-constructed multi-scale feature extraction network for subsequent nonlinear feature decoupling.

[0038] Optionally, a multi-scale feature extraction network is triggered along the spatiotemporal correlations of multi-source heterogeneous data. Specifically, spatiotemporal correlation identification is achieved as follows: In the time dimension, the system segments the continuously sampled light flux sequence, voltage and current waveform sequence, and driving junction temperature sequence into frames with a sliding window step of 50 milliseconds and a total window length of 500 milliseconds. For the voltage and current waveform sequence within each time window, a first-order differential thresholding method is used to detect waveform abrupt changes: the amplitude difference between adjacent sampling points is calculated. When this difference exceeds a preset amplitude change rate threshold, the location is marked as the time of an electrical event. Similarly, in the light flux sequence, when the illuminance change rate of adjacent sampling points exceeds an ambient light fluctuation threshold, it is marked as the time of an optical event. The system further calculates the time difference between the electrical event and the optical event. If the absolute value of this time difference falls within a preset electro-optical response delay interval, it is determined that the electrical event and the optical event have a dynamic lag correlation; otherwise, it is marked as an asynchronous pseudo-signal and discarded.

[0039] In the spatial dimension, the system pre-defines a thermal-optical-electrical coupling matrix based on the internal physical topology of the luminaire. The rows of this matrix correspond to changes in the driving junction temperature, and the columns correspond to changes in luminous flux and electrical power consumption. The system substitutes the measured gradients of junction temperature, luminous flux, and electrical power consumption within the current window into this coupling matrix and calculates the spatial coupling coefficients between the physical quantities through matrix inversion. When the residual between the measured change of a certain sensor channel and the change predicted by the coupling matrix exceeds a preset physical reasonable threshold, it is determined that the channel has latent radiative interference caused by local environmental interference, and the associated content of that channel is marked as invalid. The system uses data segments that have passed both temporal correlation verification and spatial coupling verification as the final set of true associated content and outputs it to the multi-scale feature extraction network. All the preset thresholds and window parameters can be calibrated offline or adaptively adjusted online according to the hardware response characteristics of the luminaire and the installation scenario.

[0040] Specifically, on a summer evening, a user is reading under a smart light fixture in the living room when a dark cloud suddenly drifts by and blocks the sunset. At the same time, the user turns on a floor fan that blows directly onto the smart light fixture. The microcontroller inside the smart light fixture triggers synchronous sampling at time t0. At this moment, the photoelectric sensor collects real-time illuminance data; the voltage and current sensors collect the current waveform data driving the LED beads, maintaining a stable DC low ripple state; and the temperature sensor collects the junction temperature data of the drive, which shows an abnormally rapid downward trend due to the direct fan blowing.

[0041] The control system of the intelligent lighting fixture identifies the correlation between the collected phenomena in the spatiotemporal dimensions. In the temporal dimension, the system recognizes that the "drastic drop in illuminance" and the "fan turning on" coincide in the same time window t0, but the "drive current waveform" does not show a compensatory increase in response to insufficient illumination. Therefore, in terms of temporal logic, it is determined that the drop in illuminance is not caused by a fault in the lighting fixture itself, but by the disappearance of the external light source. In the spatial dimension, the system identifies, based on the preset "thermal-optical coupling matrix," that the current "rapid drop in junction temperature" is, in spatial physics, a drastic change in the local microenvironment of the sensor caused by external forced convection. This temperature drop is not accompanied by an increase in the actual luminous efficiency of the LED chip, and belongs to the "pseudo-temperature correlation" in the spatial dimension. Based on this, the system removes the temperature anomaly caused by direct fan blowing and the light spikes as interference items and extracts the true correlation content: the external base illuminance drops sharply, but the lighting fixture itself is in a healthy thermodynamic state and has sufficient electrical margin.

[0042] The edge computing chip of the smart lighting fixture performs tensor mapping on the filtered real-relation content; fills the "net illumination descent gradient after removing fan blade occlusion burrs" into the illumination channel C1 of the tensor; fills the "stable voltage and current RMS values ​​and ripple characteristics" into the electrical channel C2; and fills the "real junction temperature reference value after removing the forced cooling effect of the fan" into the temperature channel C3, finally forming a 3×128 (3 physical semantic channels, 128 sampling point windows) standard input tensor, which is then input into the subsequent multi-scale feature extraction network.

[0043] Furthermore, by using parallel dilated convolutional kernels and cross-modal attention mechanisms within the multi-scale feature extraction network, nonlinear projection and alignment of multi-source heterogeneous data are performed to filter out transient environmental noise and extract high-order interaction features. These features are then combined with multiple state features corresponding to the multi-source heterogeneous data and reorganized according to a preset physical semantic dimension, mapping them into a multi-dimensional state tensor that represents the coupling relationship between the current operating mode of the lighting fixture and the local environment. This introduces a multi-dimensional state tensor that represents the coupling relationship between the current operating mode of the lighting fixture and the local environment.

[0044] At this point, the standard input tensor is fed into the multi-scale feature extraction network and processed by multiple sets of parallel dilated convolution kernels. These parallel sets of dilated convolution kernels are configured with different dilation coefficients to construct differentiated multi-scale temporal content under the same computing resources. Convolution kernels with smaller dilation coefficients focus on capturing high-frequency transient change signals, while convolution kernels with larger dilation coefficients focus on extracting low-frequency slowly changing trend features.

[0045] In the multi-scale feature extraction network, a nonlinear activation function is introduced to perform cross-dimensional nonlinear projection on heterogeneous data, forcing data from different modalities to be mapped to a unified latent feature space. By setting a dynamic threshold, high-frequency violent fluctuation features captured in small receptive field branches are suppressed and smoothed out at the feature level, while retaining low-frequency effective features that reflect the evolution of the real environmental baseline. High-frequency violent fluctuation features are environmental transient noises caused by uncontrollable external factors, such as airflow cutting and rapid shadow occlusion.

[0046] The initial features, after noise filtering, are input into the cross-modal attention mechanism module for deep fusion. In this module, the illumination flux feature sequence is set as the main query vector, and the voltage and current waveform feature sequences and the driving junction temperature feature sequence are jointly set as the key-value pair vector. By calculating the dot product interaction between the query vector and the key vector, the cross-modal correlation weight matrix of different physical modes at the current time step is obtained. Based on this weight matrix, the value vector is weighted and aggregated, enabling the network to spontaneously learn and extract the high-order interaction features hidden behind a single mode. For example, when the illumination demand changes, the network can accurately assess the degree of constraint of the current electrical response capability and thermodynamic boundary on the illumination change, thereby outputting a high-order interaction feature vector that integrates multiple physical constraints of "light-electricity-heat".

[0047] The extracted high-order interactive feature vectors are flattened and, based on the underlying physical control logic of the lamp, are forcibly reorganized according to the preset physical semantic dimensions. At this point, the features are reorganized into three orthogonal physical semantic dimensions: the first dimension represents optical execution semantics, including state features such as target compensation illuminance and color temperature offset; the second dimension represents power conversion semantics, including state features such as the current efficiency inflection point and transient power margin; and the third dimension represents thermodynamic lifetime semantics, including state features such as junction temperature change rate and thermal resistance accumulation effect. The reorganized multidimensional feature matrix is ​​linearly transformed by a fully connected layer and finally mapped to a high-dimensional multidimensional state tensor. This tensor accurately represents the deep coupling relationship between the operating mode of the smart lamp at the current microscopic moment and its local environment on the mathematical manifold.

[0048] Specifically, the edge computing chip of the smart lighting fixture receives an input tensor containing high-frequency lighting anomalies caused by fan blades cutting light and abnormal temperature drops caused by forced convection from the fan. The data enters a parallel dilated convolutional layer: a small-scale convolutional kernel with an expansion coefficient of 1 keenly captures the regular dips that occur every 0.1 seconds in the lighting channel, the physical obstruction of the fan blades, and the high-frequency jitter in the temperature channel; while a large-scale convolutional kernel with an expansion coefficient of 8 bypasses these glitches and observes a continuous downward trend in the overall lighting channel, as well as the attenuation of the natural light baseline caused by clouds obstructing the sunset. Through nonlinear projection and threshold suppression, the network judges the "high-frequency obstruction of fan blades" and "forced cooling of the fan" captured by the small-scale convolutional kernel as transient noise without physical meaning and smoothly removes them, successfully extracting the trend feature that "the ambient basic illuminance is undergoing a substantial decrease in the low-to-medium frequency range".

[0049] The decreasing illumination trend feature serves as the main query vector, which interacts with voltage and current features and junction temperature features in the cross-modal attention module for computation. When calculating the attention weights, the network discovers that while illumination demand is increasing rapidly, the attention mechanism also extracts important information from the key-value pairs—due to direct fan blowing, the current junction temperature is abnormally low, for example, only 35°C, far below the safe threshold of 65°C, and the voltage and current waveform features show that the power supply is in a half-load state with extremely low ripple. Based on this strong correlation, the cross-modal attention mechanism outputs a set of high-order interactive features, which accurately express the following physical logic: "There is currently a huge demand for illumination compensation, and due to abnormal external forced cooling, the thermodynamic and electrical bottlenecks of the lamps have been completely eliminated, enabling full-power output."

[0050] The edge chip slices and reassembles the above-mentioned high-order interaction features according to physical semantics: "The PWM duty cycle needs to be increased significantly immediately to compensate for the illuminance loss of about 150 lx" is written in the "optical execution semantics" channel; "Currently in the high conversion efficiency range, it can directly jump to the rated power" is written in the "electrical energy conversion semantics" channel; "Due to fan disturbance, the thermal resistance model has shifted, and the traditional slow-start dimming and anti-light decay strategy can be exempted" is written in the "thermodynamic lifetime semantics" channel. The features of these three dimensions are fused and mapped into a high-dimensional multidimensional state tensor. This tensor completely gets rid of the appearance of fan interference and shows the subsequent spatiotemporal graph neural network in mathematical language: "The smart lamp is currently in a special mode of sudden drop in ambient light but with perfect heat dissipation conditions. The local environmental coupling relationship has been solved and an aggressive zero-delay brightness compensation strategy can be executed at any time."

[0051] refer to Figure 3 In step S12, the specific steps are as follows:

[0052] S121: Real-time monitoring of industrial cloud platform, marking smart home scene of smart lighting, combined with multi-dimensional state tensor input to the corresponding spatiotemporal graph neural network along the communication channel, taking each smart lighting in the smart home scene as a graph node, synchronously obtaining the light spillage interference coefficient between each smart lighting, and using the corresponding power line carrier coupling impedance as edge content, further constructing a spatial topology graph that represents the strong correlation characteristics of physical space.

[0053] S122: For light overflow interference, retrieve the IES light distribution files of each smart lamp in the industrial cloud platform, combine the coordinates and installation posture of the smart lamp to determine the light radiation content, and track and count the distribution of photons of the target smart lamp on the illuminated surface of adjacent smart lamps, thereby determining the unexpected light contribution and normalizing it to obtain the light overflow interference content.

[0054] S123: For power line carrier coupling impedance, a broadband probe carrier signal is orthogonally injected into the operating current of the smart lamp, the signal attenuation and phase shift characteristics between nodes under the same phase line are extracted, the channel frequency response including distributed capacitance and inductance is fitted, and the power line carrier coupling impedance is dynamically determined.

[0055] In the embodiments of this application, the industrial cloud platform is monitored in real time, and the smart home scene of smart lamps is marked. The multi-dimensional state tensor is input to the corresponding spatiotemporal graph neural network along the communication channel. Each smart lamp in the smart home scene is used as a graph node. The light spillover interference coefficient between each smart lamp is obtained synchronously and used as the corresponding power line carrier coupling impedance as the edge content. The spatial topology graph representing the strong correlation characteristics of physical space is further constructed. The overall consideration of using the corresponding power line carrier coupling impedance as the edge content is compatible, which ensures the accuracy of the spatial topology graph representing the strong correlation characteristics of physical space.

[0056] At this point, the data bus of the industrial cloud platform is monitored in real time, and the device identification codes and metadata reported by the smart lighting cluster are parsed. Based on this, the lighting cluster in the current control cycle is marked as a specific smart home scenario. The multidimensional state tensors generated by all smart lights in this scenario at the local edge are aggregated to the industrial cloud platform along the encrypted communication channel. After receiving the tensor stream in the cloud, the tensors are indexed and concatenated according to the device identification codes to construct a global latent variable matrix containing the state information of all nodes in the smart home scenario. This matrix is ​​then used as the input source for the initial node features of the subsequent graph neural network.

[0057] When constructing the edge relationships of the graph, the simple measurement of pure spatial distance is abandoned, and dual coefficients representing the deep physical interference between nodes are obtained simultaneously. First, the light spillover interference coefficient is obtained by retrieving the spatial three-dimensional coordinates and light distribution IES files of each smart lamp pre-placed in the cloud, and using the inverse square law of illuminance and the superposition integral algorithm of luminous flux, calculating the cross contribution ratio of illuminance generated by any two lamps on the target working surface, and quantifying it as the light spillover interference coefficient. Second, the power line carrier coupling impedance is obtained by analyzing the topology wiring structure of the home smart grid, identifying lamp groups in the same phase line or the same circuit, extracting the line distributed capacitance, inductance parameters of the circuit, and the signal-to-noise ratio and attenuation function on the current channel, and converting them into normalized power line carrier coupling impedance. The above two are combined to form a composite edge feature vector with clear physical meaning.

[0058] Using each smart light fixture in a smart home scenario as a vertex of a graph network, the global latent variable matrix is ​​assigned to the corresponding vertex as the initial node feature. The substantial physical interference between the various smart lights is used as the connection condition. When the light spillover interference coefficient or power line carrier coupling impedance between two lights is greater than the preset perception threshold, an undirected or weighted edge is established between the two vertices, and the composite edge feature vector is used as the attribute weight of the edge. By traversing all light fixture node pairs in the scenario, a spatial topology graph jointly characterized by the node feature matrix and the composite adjacency matrix is ​​finally generated in the cloud memory. This graph highly restores the strong correlation characteristics of "light signal interweaving" and "electrical signal coupling" between lights in physical space in terms of mathematical structure.

[0059] Specifically, there is a TV background light B in the living room, and a dining room chandelier C connected to the same main circuit in the living room as the smart light fixture. The industrial cloud platform detects that the MAC addresses of the devices all belong to "User Zhang San - Living Room Area" and immediately marks the current calculation slice as "Smart Home - Living Room Scene". The cloud platform receives the high-dimensional multidimensional state tensor of "needs to immediately and significantly increase PWM duty cycle" generated by the smart light fixture in step S112 through the Wi-Fi gateway, and at the same time receives the idle state tensors of TV background light B (currently in low brightness ambient mode) and dining room chandelier C (currently in off state). The cloud concatenates and aligns these tensors in the order of smart light fixture A, TV background light B, and dining room chandelier C, and prepares to inject them into the graph neural network.

[0060] The cloud performed precise physical deduction when calculating the "edge content" between smart light fixture A and surrounding lights:

[0061] Light spillover interference coefficient calculation: The IES light distribution curves of smart light fixture A (main lighting, high power, wide beam angle) and TV background light B (TV background light, low power, narrow beam angle) were retrieved from the cloud. The calculation showed that if the brightness of smart light fixture A is significantly increased according to the aforementioned decision, about 18% of its scattered light will spill over onto the TV background wall, seriously destroying the low-light atmosphere created by TV background light B. Therefore, it was determined that the light spillover interference coefficient of smart light fixture A on TV background light B is extremely high. However, there is a physical partition between smart light fixture A and the dining room chandelier C, and the light spillover coefficient is close to 0.

[0062] Power line carrier coupling impedance calculation: Cloud analysis of the home electrical box topology revealed that smart light fixture A and dining room chandelier C are connected to the same live wire branch under the same microcircuit breaker. Since high-power appliances such as air conditioners and refrigerators may cause harmonic interference in the evening during summer, the cloud estimated the channel attenuation of this circuit and calculated that the power line carrier coupling impedance between smart light fixture A and dining room chandelier C is at a high level. This means that if smart light fixture A and dining room chandelier C simultaneously issue large-amplitude PWM dimming commands, severe carrier signal conflict and congestion will occur on the power line. In contrast, the TV background light B is controlled by a separate smart switch, and its carrier coupling impedance with smart light fixture A is extremely low.

[0063] In the virtual computing space of the industrial cloud platform, a miniature spatial topology graph containing 3 nodes was generated: Node layer: The first node was given the tensor feature of "aggressive compensation", the second node was given the feature of "maintaining tranquility", and the third node was given the feature of "standby"; Edge layer: The first node and the second node are tightly connected by a "high-weight light overflow edge"; The first node and the third node are tightly connected by a "high-weight carrier impedance edge"; while there is almost no edge connection between the second node and the third node.

[0064] Furthermore, regarding the light spillover interference, the IES light distribution files of each smart lamp in the industrial cloud platform are retrieved. The light radiation content is determined by combining the coordinates and installation posture of the smart lamps, and the distribution of photons from the target smart lamp on the illuminated surfaces of adjacent smart lamps is tracked and statistically analyzed. This determines the unexpected light contribution and performs normalization processing, thereby obtaining the light spillover interference content. This approach incorporates the overall consideration of tracking and statistically analyzing the distribution of photons from the target smart lamp on the illuminated surfaces of adjacent smart lamps, ensuring the accuracy of the unexpected light contribution and normalization processing.

[0065] At this point, within the industrial cloud platform, for the target smart luminaire and its adjacent luminaires, the pre-stored IES light distribution file is retrieved to obtain the core parameters of the luminous intensity distribution of the light source at various spatial solid angles. Simultaneously, combined with the three-dimensional spatial coordinates and installation attitude of each smart luminaire in the digital twin model, such as pitch angle, yaw angle, and roll angle, a local light radiation field matrix based on a specific coordinate system is constructed. By performing a spatial affine transformation between the polar coordinate light intensity distribution data in the IES file and the attitude matrix, the abstract light distribution curve is transformed into a directional three-dimensional luminous flux radiation vector field under an absolute spatial reference system, thereby establishing the initial direction and energy reference of the photons emitted by the target luminaire into the surrounding three-dimensional space.

[0066] After establishing the light radiation vector field, the luminous center of the target smart lamp is used as the origin of ray emission. Monte Carlo ray tracing or gridded ray projection is performed based on the luminous flux weight in the radiation field. For the massive amount of projected light rays, the physical geometric shell parameters of adjacent smart lamps and the plane equation of their illuminated surfaces in three-dimensional space are combined to perform rigorous ray-plane intersection calculations. Rays that do not hit the illuminated surfaces are filtered out. The number of light rays that successfully hit the illuminated surfaces of adjacent smart lamps and their corresponding incident angles and landing point coordinates are accurately tracked and counted. This generates a two-dimensional landing point matrix that characterizes the spatial distribution density of photons on the illuminated surfaces of adjacent lamps. Illuminated surfaces include the light-transmitting panel, reflector cup mouth, or diffuser surface of the lamp.

[0067] For Monte Carlo ray tracing, the system constructs a normalized probability density function based on the spatial distribution of light intensity described in the IES light distribution file of the target smart luminaire. This function guides the sampling of ray emission directions: the greater the light intensity of the solid angle region, the higher the probability that it will be selected as the ray emission direction. The system uses the luminous center of the target smart luminaire as the ray emission origin and generates the initial direction vector of each ray using an inverse transform sampling method based on this probability density function. Simultaneously, it records the relative luminous flux weight carried by each ray, which is proportional to the light intensity value in that direction. The system presets a total of 100,000 ray emissions for a single tracing iteration and uses the convergence criterion of stopping emission when the sum of the relative luminous flux weights carried by all rays converges to 95% of the total luminous flux.

[0068] The system tracks rays in parallel, one thousand rays per batch. For each ray, the system uses the Möller-Trumbore ray-triangle mesh intersection algorithm, based on its current direction vector and position coordinates, to calculate the intersection with each triangular facet in the 3D geometric shell model of adjacent smart lamps. The system pre-creates a triangular mesh model for each adjacent smart lamp, consisting of its illuminated surfaces, including the outer surface of the lamp's light diffuser, the reflector opening, and the visible surface of the heat sink. When the intersection algorithm determines that a ray intersects with a triangular facet and the intersection point is inside that facet, the system records the 3D coordinates of the intersection point, the incident angle, and the relative luminous flux weight carried by the ray, marking the ray as a hit. For rays that do not hit any illuminated surface, the system continues to propagate along a straight line until it exceeds the preset tracking radius or the cumulative number of reflections exceeds a set limit, at which point it is discarded to save computational resources.

[0069] After the system completes the tracing of 10,000 rays, it calculates the relative rate of change between the total relative luminous flux weight of the currently hit rays and the total relative luminous flux weight of the hit rays before the tracing of the previous 10,000 rays. When the rate of change is less than one percent for three consecutive iterations, the system determines that the ray tracing results have converged and stops emitting the remaining rays. The system classifies and aggregates all hit intersections according to their adjacent smart lamps and the triangular facet numbers they hit. For all intersections within the same illuminated surface, a two-dimensional histogram matrix is ​​generated based on the two-dimensional projection coordinates of the intersections on the illuminated surface. Each cell of this matrix records the sum of the relative luminous flux weights carried by the rays falling into that cell, thus forming a two-dimensional landing point matrix that characterizes the spatial distribution density of photons on the illuminated surfaces of adjacent smart lamps.

[0070] Based on the landing point matrix and combined with the apparent spectral reflectance or transmittance parameters of the materials of the illuminated surfaces of adjacent smart lamps, the sum of the actual luminous flux overflowing from the target smart lamp and acting on the surfaces of adjacent smart lamps is calculated, and this is defined as the unexpected illumination contribution of the target lamp to the adjacent lamps. The unexpected illumination contribution is then compared with the rated outgoing luminous flux of the adjacent smart lamps in the current control cycle, and normalized using the Sigmoid function or extreme value scaling algorithm to compress the absolute physical value of luminous flux into a continuous bounded interval of [0,1]. The final output value is used as the edge weight of the "light overflow interference content" connecting the two nodes in the spatial topology graph. The closer the value is to 1, the more severe the light coupling interference.

[0071] Specifically, the industrial cloud platform determines that the smart lighting fixtures need to significantly increase their brightness to compensate for the light loss caused by cloud cover. The cloud platform immediately retrieves the IES light distribution file of the smart lighting fixtures. Assuming that smart lighting fixture A is a living room main light with a deep anti-glare honeycomb mesh, its beam angle is mainly concentrated in the main light-emitting area at a downward 60-degree angle, but there is a slight edge light leakage between 75 and 85 degrees. Combining the parameters that smart lighting fixture A is installed in the center of the ceiling and its installation posture is vertically downward (0 degrees pitch angle, 0 degrees yaw angle), the cloud platform generates a three-dimensional light radiation vector field with the center of the ceiling as the origin, the main energy downward, but with a slight outward scattering at the edges through spatial affine transformation.

[0072] The cloud platform initiates virtual ray tracing for smart lamp A. Among tens of thousands of projected rays, the vast majority fall vertically downwards onto the user's desk area. However, the cloud platform also rigorously tracks the spillover rays that fall within the 75-85 degree angle. These rays travel 2 meters outwards and downwards before intersecting with the plane equations of the physical shell of TV background lamp B and its upper diffuser. The calculation results show that approximately 1,500 virtual photons successfully hit the surface of the upper diffuser of TV background lamp B, forming an asymmetric Gaussian distribution of impact points biased towards the left side of the center on the diffuser of TV background lamp B.

[0073] The cloud platform reads the material parameters of TV background light B, assuming that the reflection / scattering coefficient of its white acrylic diffuser is 0.6. The cloud platform calculates that the luminous flux carried by these 1500 photons, after being scattered by the diffuser of TV background light B, results in approximately 20 lumens of unintended illumination contribution being forcibly injected into the optical cavity or surface of TV background light B. At this time, in order to create an atmosphere, TV background light B's rated output luminous flux is only 80 lumens. The cloud platform performs a ratio calculation: 20 lumens (unintended interference from smart light A) / 80 lumens (TV background light B's own output) = 0.25.

[0074] Through the above steps, the industrial cloud platform transmits a very clear physical warning to the subsequent decision network on the graph topology edge constructed between smart light fixture A and TV background light B: if the large dimming command of smart light fixture A is blindly executed, the light leaked from smart light fixture A will seriously pollute the light field of TV background light B, causing the atmosphere of the TV background wall to be completely destroyed. This forces the subsequent algorithm to consider coordinating the adjustment of smart light fixture A and TV background light B. For example, smart light fixture A can fine-tune its brightness, while instructing TV background light B to reduce its color temperature or change its light emission angle to resist the light intrusion of smart light fixture A.

[0075] Therefore, for power line carrier coupling impedance, a broadband probe carrier signal is orthogonally injected into the operating current of the smart lighting fixture, the signal attenuation and phase shift characteristics between nodes under the same phase line are extracted, and the channel frequency response content including distributed capacitance and inductance is fitted to dynamically determine the power line carrier coupling impedance. This approach takes into account the overall consideration of the channel frequency response content of capacitance and inductance, ensuring the accuracy of the power line carrier coupling impedance. At the same time, a multi-dimensional state tensor is introduced to further control the industrial cloud platform and improve the accuracy of the spatial topology map.

[0076] At this time, in the power line communication front-end module of the smart lighting fixture, a set of broadband probe carrier signals that are completely orthogonal to the normal operating current of the lighting fixture in the spectrum are generated. For example, a spread spectrum signal with a specific pseudo-random code sequence is used, whose frequency band falls in the PLC communication frequency band of 1MHz to 30MHz. The broadband probe signal is converted into an analog signal by a digital-to-analog converter (DAC) and injected into the AC or DC power supply bus of the smart lighting fixture in a non-intrusive superposition manner through a coupling transformer or capacitive coupling circuit. This injection process ensures that the energy of the broadband probe signal is extremely low and will not cause substantial electromagnetic interference to the constant current and constant voltage control loop of the LED driver circuit. At the same time, it can establish a full-band excitation response in the power line channel.

[0077] At the same moment the broadband probe carrier signal is injected, the industrial cloud platform wakes up the PLC receiving front-end of adjacent smart lighting nodes on the same phase line by issuing a collaborative sampling command. The adjacent nodes synchronously capture the weak probe signal transmitted through the power line within a preset precise time window. By windowing and performing Fast Fourier Transform (FFT) on the captured time-domain signal, it is converted to the frequency domain. The frequency domain amplitude spectrum and phase spectrum of the received signal are aligned and compared with the reference amplitude and reference phase of the original injected probe signal stored locally, and the signal attenuation and phase shift characteristics of the communication link at each frequency point are accurately extracted to form the discrete frequency domain response feature set of the node pair.

[0078] The extracted discrete frequency domain response feature set is input into a pre-constructed lumped parameter or distributed parameter circuit fitting model. This model treats the power lines inside the home as equivalent to a two-port network consisting of a series of infinitesimal resistors, distributed inductance (characterizing the energy stored in the magnetic field of the conductors), and distributed capacitance (characterizing the parasitic capacitance between conductors and to ground). The least squares method or gradient descent algorithm is used to adjust the distributed capacitance and inductance values ​​inside the model to minimize the residual between the theoretical frequency response curve output by the model and the discrete frequency domain response feature set. When the fitting converges, based on the optimal distributed capacitance and inductance parameters obtained from the fitting, combined with the real-time operating frequency of the current power line, the complex value of the power line carrier coupling impedance in a specific frequency band between the two nodes at the current moment is dynamically calculated, and this dynamic impedance value is used as the edge feature in the spatial topology graph.

[0079] Specifically, the industrial cloud platform decides to conduct a real-time assessment of the "electrical coupling relationship" between the smart lighting fixture and the restaurant chandelier C. The cloud platform issues a detection command to the driver chip of the smart lighting fixture. While the smart lighting fixture outputs PWM current to light up the LED beads, its internal PLC module generates a weak pseudo-random code detection signal with a bandwidth of 2MHz-20MHz. This signal does not interfere with the driving current of the lighting fixture in the spectrum. Through the live wire terminal of the smart lighting fixture A base, it is connected to the live wire network in the living room and propagates to the distribution box and branch lines.

[0080] Meanwhile, the cloud platform notified the restaurant chandelier C to enter "channel monitoring" mode. The chandelier C was not lit (in standby), but its internal circuitry remained connected to the live wire. The PLC module of chandelier C captured the detection signal from the smart light A on the live wire. However, due to the peak household electricity consumption in the summer evening, the refrigerator compressor was starting, and there were numerous parallel branches in the wiring between the living room and dining room walls, such as sockets and other stationary appliances. These factors caused severe distortion in the signal received by chandelier C. The DSP chip inside chandelier C performed FFT analysis on the transmitted and received signals, calculating that at 2MHz, there was a -15dB attenuation and a 30-degree phase shift; at 10MHz, due to parasitic capacitance, the attenuation reached -40dB and the phase shift was 120 degrees. Chandelier C packaged these "attenuation-phase shift" data and sent them back to the industrial cloud platform.

[0081] After receiving discrete frequency domain data from smart light fixture A to restaurant chandelier C, the industrial cloud platform initiates a fitting algorithm. The platform discovers that compared to the baseline data measured during the day, the current high-frequency band exhibits abnormally severe attenuation and a significant phase shift. By adjusting the "distributed capacitance" and "distributed inductance" parameters for curve fitting, the algorithm finds that the "distributed capacitance" parameter in the model must be significantly increased to make the theoretical curve coincide with the current poor frequency response data. The platform infers that the high humidity in the summer evening alters the capacitance characteristics of the insulation layer of the cables inside the wall, or that a household appliance in standby mode is effectively connected to a large capacitive load on that phase line. Based on the deteriorated distributed capacitance and inductance parameters, the platform calculates that the power line carrier coupling impedance from smart light fixture A to restaurant chandelier C in the current control command frequency band is extremely low, for example, an impedance modulus of only 5 ohms.

[0082] The industrial cloud platform assigns "extremely low coupling impedance (high channel attenuation)" edge characteristics to the graph topology edges constructed between smart lamp A and restaurant chandelier C. This technical detail conveys a communication warning to the subsequent spatiotemporal graph neural network: smart lamp A and restaurant chandelier C are currently in an extremely poor PLC communication channel. If the cloud platform simultaneously sends large dimming commands to both smart lamp A and restaurant chandelier C, the extremely low coupling impedance and channel congestion will cause severe packet loss or even command crosstalk in the control messages of the two lamps, resulting in lamp flickering or system crashes. This forces the subsequent decision-making algorithm to adopt a "time-division multiplexing" strategy, controlling smart lamp A first and then controlling restaurant chandelier C after the channel is restored, or finding an alternative communication path.

[0083] refer to Figure 4 In step S13, the specific steps are as follows:

[0084] S131: Perform multi-hop neighbor message passing based on gating mechanism on the spatial topology graph to obtain cross-node optical-electric coupling state evolution information, aggregate the cross-node optical-electric coupling state evolution information, and combine it with the extended causal temporal convolutional network embedded in the spatiotemporal graph neural network to perform parallel sliding window convolution on the historical multidimensional state tensor sequence along the time axis. Through gradient normalization and nonlinear activation, capture the evolution law of user lighting demand over a long time span.

[0085] S132: Acquire aging hysteresis events of intelligent lighting clusters and perform multi-level decoupling in combination with the evolution law of illumination demand, thereby decoupling and outputting a multi-dimensional decision space characterizing the collaborative control of intelligent lighting clusters. The multi-dimensional decision space covers the physical execution dimension including dimming depth and color temperature step frequency, the environmental-energy efficiency dimension including target illuminance compliance rate and reactive power loss, and the lighting life dimension including cumulative heat generation and the growth rate of equivalent series resistance of electrolytic capacitors.

[0086] In the embodiments of this application, multi-hop neighbor message passing based on a gating mechanism is performed on the spatial topology graph to obtain cross-node optical-electric coupling state evolution information. The cross-node optical-electric coupling state evolution information is aggregated and processed. Combined with the extended causal temporal convolutional network embedded in the spatiotemporal graph neural network, parallel sliding window convolution is performed on the historical multidimensional state tensor sequence along the time axis. The evolution law of user lighting demand over a long time span is captured by gradient normalization and nonlinear activation.

[0087] At this point, on the constructed spatial topology graph, a multi-hop neighborhood search is performed along the edge connection with the target smart lighting node as the anchor point. During the node feature interaction process of each hop, a gated recurrent unit or a gated linear attention mechanism is introduced. At this point, the multidimensional state tensor of the current node is used as the reference input for the update gate and the reset gate. The feature vectors passed from the first-order and multi-order neighbor nodes along the edges of "light spillover interference" and "power line carrier coupling impedance" are used as candidate hidden states. Through adaptive learning of the gated scalar, unreliable neighbor features caused by high coupling impedance are dynamically suppressed, and strong physical correlation features caused by high light spillover interference are amplified. After a preset number of hops, all the gated neighbor features that converge to the anchor point are weighted and summed or max pooled to generate a local spatial aggregation tensor containing cross-node optical-electrical depth interference and restraint information.

[0088] The generated local spatial aggregation tensor is concatenated with the multidimensional state tensor sequence cached by the anchor node itself within the historical time slice in the channel dimension to form a spatiotemporal joint tensor. This joint tensor is then input into the dilated causal temporal convolutional network (TCN) embedded in the later stage of the graph neural network. Multilayer one-dimensional causal convolutional kernels with exponentially increasing dilation coefficients (e.g., 1, 2, 4, 8…) are set to ensure that the convolution operation strictly follows the chronological order and prevents the leakage of future information. Through the gradually increasing dilation coefficient, the receptive field of the convolutional kernel expands exponentially, enabling the network to cover and process historical tensor sequences spanning extremely long time spans in one go in a parallel sliding window manner without increasing the network depth and the number of parameters.

[0089] After each layer of computation in the dilated causal temporal convolution, layer normalization is introduced layer by layer to eliminate gradient vanishing or gradient explosion problems caused by excessively long historical tensor sequences, thus stabilizing the joint training process of the deep spatiotemporal network. Nonlinear mapping capability is introduced through nonlinear activation functions to remove periodic physical noise (such as fixed light variations during day and night) from historical data, extracting user behavior pattern features with nonlinear and time-varying characteristics hidden over long time spans. Finally, a highly condensed global spatiotemporal hidden vector is output, which accurately represents the evolution of users' real lighting needs over long periods under the current complex cross-node optical-electric coupling constraints.

[0090] Specifically, the spatiotemporal graph neural network on the industrial cloud platform begins processing smart lamp node A; smart lamp A initiates message transmission to its one-hop neighbors, TV background light B and dining room chandelier C; the message from TV background light B: TV background light B transmits the state of "maintaining a low brightness atmosphere, the light-receiving surface is severely interfered with"; the network's light overflow gating mechanism determines that this interference is real and fatal, assigning an extremely high gating weight, such as 0.9, allowing this feature to flood into smart lamp A; the message from dining room chandelier C: dining room chandelier C transmits the state of "my channel impedance is extremely low, the communication environment is poor"; the network's carrier gating mechanism determines that this belongs to the unstable factors at the communication layer, and in order to prevent noise pollution of spatial decision-making, assigns an extremely low gating weight, such as 0.1, blocking the feature of dining room chandelier C from the aggregation gate.

[0091] To determine how smart light fixture A should compensate for the dim light caused by the clouds, the network doesn't just look at the present moment (t0), but initiates an expanded causal temporal convolution to trace back the multidimensional state tensor sequence of the past hour (t−60 to t0). The convolutional layer with an expansion coefficient of 1 captures the features of the most recent few minutes and discovers that "the clouds caused a sudden drop in light intensity within seconds." The convolutional layer with an expansion coefficient of 4 expands the receptive field to more than ten minutes ago and captures the "abnormal temperature drop trend caused by the fan being turned on." The convolutional layer with an expansion coefficient of 8 expands the receptive field to half an hour ago or even longer. The network scans in parallel in the historical tensor sequence and captures a key pattern: every summer evening in the past, when natural light disappeared, the user would manually adjust smart light fixture A to a reading mode with a color temperature of 4000K and 1200 lumens, and the user would turn on the fan in advance before each such need occurred.

[0092] By backtracking through an hour of data, the network flattens tensor features of different dimensions, such as "summer natural light decay," "temperature drift noise from fan operation," and "the state of TV backlight B," into the same gradient space through layer normalization, thus avoiding gradient vanishing. After GELU nonlinear activation, the network filters out irrelevant spikes, elevating the aforementioned linear physical changes to a high-dimensional semantic understanding. The global spatiotemporal hidden vector output by the network accurately depicts the following long-term demand evolution patterns:

[0093] "The current environment has suddenly lost natural light. Although the fan's activation has caused a false thermal state and electrical interference, according to long-term behavioral statistics, the user has entered the deep reading preparation stage. However, due to the insurmountable light overflow red line of the TV background light B, the smart light fixture A absolutely cannot perform a global large-scale brightening. It is necessary to find another non-linear control path that can compensate for the desktop illuminance without polluting the light field of the TV background light B. For example, reducing the anti-glare angle of the smart light fixture A / reducing the power density to increase the center illuminance, or coordinating with the TV background light B to synchronously fine-tune the color temperature to offset the overflow effect." This vector is used as the most core prior knowledge and is sent to the final multi-dimensional decision space generation stage.

[0094] Furthermore, the aging hysteresis events of the intelligent lighting cluster are obtained, and multi-level decoupling is performed in combination with the evolution law of illumination demand. This decouples and outputs a multi-dimensional decision space that characterizes the collaborative control of the intelligent lighting cluster. The multi-dimensional decision space includes the physical execution dimension, which includes dimming depth and color temperature step frequency; the environmental-energy efficiency dimension, which includes target illuminance achievement rate and reactive power loss; and the lighting lifespan dimension, which includes the growth rate of accumulated heat generation and the equivalent series resistance of electrolytic capacitors. The lighting lifespan dimension, which includes the growth rate of accumulated heat generation and the equivalent series resistance of electrolytic capacitors, is introduced.

[0095] At this point, the digital twin module of the industrial cloud platform continuously collects microscopic operational data of the smart lighting cluster throughout its entire lifecycle to extract aging hysteresis events. For each lighting fixture, the actual luminous flux attenuation rate curve under the same set power, the equivalent series resistance drift trajectory of the electrolytic capacitor in the drive circuit under frequent thermal stress impacts, and the junction temperature-luminous efficacy nonlinear shift data of the LED PN junction caused by the accumulation of hot carrier injection are extracted. The above physical parameters reflecting the irreversible degradation and thermodynamic hysteresis effects of the material are transformed into aging hysteresis feature vectors with time indexes, and these are used as static or quasi-static bias terms and superimposed on the real-time state tensor of the corresponding smart lighting fixture to form an augmented state tensor containing historical degradation imprints.

[0096] The global spatiotemporal hidden vector, which represents the evolution of users' long-term demand, output from the previous steps, is decoupled from the augmented state tensor through multi-level orthogonal projection in the latent space. The first level of decoupling is "environmental demand and physical boundary decoupling," which extracts the environmental demand components that cannot be met due to device aging by calculating the direction cosine of the demand vector and the aging hysteresis feature in the augmented tensor. The second level of decoupling is "optical-electric coupling decoupling between nodes," which orthogonally separates the entangled cross-node optical effect superposition demand and carrier channel congestion risk along the edge feature direction of the spatial topology graph. After multi-level decoupling, the originally highly nonlinear coupled cluster state is projected into three mutually independent physically orthogonal subspaces.

[0097] The decoupled feature components are injected into three pre-defined orthogonal subspaces for dimensional expansion, generating a multi-dimensional decision space: Physical execution dimension: mapping to an execution constraint tensor containing dimming depth and color temperature step frequency; Environment-energy efficiency dimension: mapping to an energy efficiency evaluation tensor containing target illuminance compliance rate and reactive power loss; Lamp lifespan dimension: mapping to a lifespan prediction tensor containing the growth rate of accumulated heat generation and the equivalent series resistance of electrolytic capacitors. These three tensors are orthogonal on the mathematical manifold, jointly spanning a multi-dimensional decision space representing the upper and lower limits of the collaborative control of intelligent lighting clusters. Dimming depth is represented by the lower limit of the PWM duty cycle and the dynamic adjustment step size; color temperature step frequency is represented by the time gradient of warm and cool color temperature switching and smoothing filter parameters; target illuminance compliance rate is represented by the probability distribution of the illuminance of the work surface falling within the standard range; reactive power loss is represented by the harmonic distortion rate and reactive power ratio of the drive power supply under non-rated operating conditions. Accumulated heat generation is such as the integral area of ​​the chip junction temperature exceeding the safety threshold; the growth rate of the equivalent series resistance of electrolytic capacitors is such as the slope of the accelerated ESR degradation caused by the thermal shock induced by the current dimming strategy.

[0098] Specifically, when the clouds obscure the light fixture and smart lamp A prepares to respond with compensation, the industrial cloud platform does not blindly issue instructions. Instead, it first retrieves the "aging hysteresis event" file of smart lamp A. The cloud platform discovers that smart lamp A has frequently dimmed at high temperatures over the past three years, and the equivalent series resistance (ESR) of its internal driving electrolytic capacitor has increased by 25% compared to when it left the factory. This means that if smart lamp A is required to instantly jump from 30% power to 100% power, the surge current will cause the ESR to generate extremely high Joule heat, which can easily cause the capacitor to burst or permanently fail. The cloud platform marks this "high ESR hysteresis risk characteristic" as a red label and affixes it to the augmented state tensor of smart lamp A.

[0099] The cloud platform orthogonally decouples the user's need for high illuminance without interfering with the TV background light B, derived from S131, with the "high ESR hysteresis risk" of smart light fixture A: First-level decoupling: The algorithm finds a serious conflict between the user's required "high illuminance (1200 lumens)" and the smart light fixture A's current "physical limit (limited by aging capacitors, not allowing instantaneous high current loading)." The algorithm forcibly decouples the "absolute illuminance requirement" into an "achievable equivalent visual illuminance requirement." Second-level decoupling: The algorithm further decouples the smart light fixture A's "light overflow interference" to the TV background light B from the "low carrier impedance" of the dining room chandelier C, concluding that high-frequency PWM dimming cannot be used (to avoid generating a large amount of high-frequency harmonic interference on the poor channel of the dining room chandelier C), and a low-frequency or DC analog dimming strategy must be adopted.

[0100] Based on the above calculations, the cloud platform defines an extremely convergent multidimensional decision space on the mathematical manifold for smart light fixture A, TV background light B, and dining room chandelier C. The specific dimensions are as follows: Physical Execution Dimension: The decision space stipulates that the dimming depth of smart light fixture A is controlled to "prohibit full-load transitions," only increasing slowly by 2% duty cycle per second; the color temperature step frequency is limited to "zero," meaning that switching between warm and cool color temperatures is prohibited at this time, as color temperature switching would trigger additional transient power consumption in the drive circuit, potentially damaging aging capacitors. Environmental-Energy Efficiency Dimension: The decision space shows that because the smart light fixture ALED itself has a three-year light decay and cannot operate at full power, the target illuminance compliance rate is predicted to be only 85%. However, to ensure system stability, reactive power loss is strictly limited to within 5% to avoid pollution to the household power grid. Light fixture lifespan dimension: The decision space draws a hard red line: Due to the false temperature perception caused by the direct blowing of the fan, the true junction temperature needs to be re-estimated through the electrical compensation algorithm. The cumulative heat generation caused by this dimming strategy must be clamped to a zero growth state; the ESR growth rate of the electrolytic capacitor must be forcibly constrained to a negative number. By using the slow climbing process, the forced cold air of the fan can actually remove the historical residual heat on the surface of the capacitor, thus achieving "thermal repair".

[0101] refer to Figure 5 In step S14, the specific steps are as follows:

[0102] S141: Real-time monitoring of the multi-dimensional decision space to obtain the luminous efficacy quality index, system energy efficiency ratio and lighting light decay index of the smart lamps. Based on the fusion of the luminous efficacy quality index, system energy efficiency ratio and lighting light decay index, the triple constraint conditions are determined, and the Pareto optimal frontier search based on non-dominated sorting is launched simultaneously.

[0103] S142: In the process of Pareto optimal frontier search, an adaptive allocation mechanism is introduced, and the interaction between light quality, energy consumption level and device life is dynamically balanced by combining the current working data of smart lamps. Multiple smart control items are determined by tracing along this interaction relationship. Furthermore, a global smart control strategy for smart lamp clusters is generated by combining the lighting modes of each smart lamp, so that the smart lamp cluster can achieve the optimal configuration of global performance in complex and ever-changing home environments.

[0104] In the embodiments of this application, the multidimensional decision space is monitored in real time to obtain the luminous efficacy quality index, system energy efficiency ratio, and lighting light decay index of the smart lamp. Based on the fusion of the luminous efficacy quality index, system energy efficiency ratio, and lighting light decay index, the triple constraint conditions are determined. Simultaneously, a Pareto optimal frontier search based on non-dominated sorting is initiated, which is compatible with the overall consideration of the luminous efficacy quality index, system energy efficiency ratio, and lighting light decay index, and ensures the accuracy of the triple constraint conditions.

[0105] At this point, a dynamic patrol and monitoring thread targeting the multi-dimensional decision space is constructed within the industrial cloud platform. Under this monitoring mechanism, the boundary tensors and state gradients of the three orthogonal subspaces in the multi-dimensional decision space output by step S132 are read in real time. Meanwhile, luminous efficacy quality indicators characterizing the current optical output quality are extracted from the physical execution dimension, including but not limited to the desktop illuminance uniformity gradient, the offset of the color rendering index (CRI), and the dynamic threshold of the anti-glare index (UGR). From the environment-energy efficiency dimension, the system energy efficiency ratio characterizing energy conversion economy is extracted, including the overall lumen per watt efficiency of the luminaire and the ratio of input active power to apparent power of the drive power supply under a specific PWM duty cycle. From the luminaire lifespan dimension, lighting decay indicators characterizing the thermodynamic degradation rate of materials are extracted, including the slope of the LED luminous flux maintenance rate decrease calculated based on the junction temperature integral, and the accelerated degradation rate of the equivalent series resistance (ESR) of the electrolytic capacitor due to thermal stress.

[0106] The three categories of indicators obtained are mapped with a pre-defined industrial / household standard benchmark library using difference mapping, transforming them into mathematical inequality constraints with absolute veto power. The specific construction logic is as follows: a lower limit threshold for luminous efficacy quality indicators is set as the first hard constraint to ensure basic visual health, such as the illuminance of the work surface must not be lower than 300 lx and the UGR must not be greater than 19; a minimum conversion efficiency threshold for the system energy efficiency ratio is set as the second hard constraint to ensure grid harmonic compliance and low-carbon operation, such as the overall system efficiency must not fall below a specific lumen per watt threshold to prevent inefficiency and overheating caused by deep dimming; a single action tolerance upper limit for lighting light decay indicators is set as the third hard constraint to prevent irreversible thermal runaway of devices, such as the instantaneous junction temperature rise rate caused by a single dimming command must not exceed a specific slope, and the ESR growth gradient must be negative or zero; these three constraints are then logically ANDed to form an infeasible solution elimination barrier surrounding the multi-dimensional decision space.

[0107] Within the feasible region formed by satisfying the above triple constraints, the population matrix of the multi-objective evolutionary algorithm (such as NSGA-II or NSGA-III) is initialized synchronously; the physical execution dimension parameters in the multi-dimensional decision space are encoded as the chromosomal genotypes of the individuals in the population; in the first generation of population evolution, for each individual's decoded control strategy, the corresponding objective function values ​​of luminous efficacy, system energy efficiency ratio, and illumination decay penalty function are calculated in parallel; based on the distribution of the above three objective function values ​​in the three-dimensional objective space, the non-dominated ranking logic is executed: if the first individual, under the premise of satisfying the triple constraints, has better luminous efficacy and energy efficiency than the second individual, and its light decay is worse than the second individual, then the first individual and the second individual are determined to be in a non-dominated relationship, and are assigned the same non-dominated level (ParetoRank1), thereby initially outlining a mutually competitive and uncompromising Pareto optimal frontier on the boundary of the feasible region, and formally starting the iterative search process.

[0108] Specifically, the monitoring thread of the industrial cloud platform locked the current multi-dimensional decision space of smart lamp A: Luminous efficacy quality index acquisition: If smart lamp A is brightened alone, due to three years of light decay, the center illuminance on the desktop can only barely reach 320lx, and the edge illuminance is extremely uneven. The UGR (glare) index may soar to 22 due to concentrated light; System energy efficiency ratio acquisition: Because low-frequency dimming must be used to avoid channel congestion of the restaurant chandelier C, the switching loss of the smart lamp A's driver chip increases. If the power is forcibly increased at this time, the system energy efficiency ratio will plummet to 60lm / W, while it should normally be above 85lm / W; Light decay index acquisition: Combining the aging capacitor of smart lamp A, a fatal indicator is deduced—if a low-frequency high-current jump occurs at this moment, the instantaneous heat growth rate of ESR will exceed the safety red line, causing accelerated light decay.

[0109] The cloud platform immediately established three red lines based on the above indicators: The first constraint (luminous efficiency baseline): Constraint 1: The average desktop illuminance is ≥300lx and UGR≤19, rejecting the passive strategy of "not adjusting the brightness at all in order to protect the capacitor"; The second constraint (energy efficiency baseline): Constraint 2: The system energy efficiency ratio is ≥75lm / W, rejecting the crude strategy of "ignoring channel congestion and forcibly driving at high frequency and full load", because that would cause the energy efficiency ratio to fall below the baseline and interfere with the restaurant chandelier C; The third constraint (lifespan baseline): Constraint 3: The single action ESR growth rate is ≤0 and the instantaneous junction temperature slope is ≤5℃ / s, directly judging the conventional compensation strategy of "instantly increasing the PWM duty cycle from 30% to 90%", because the instantaneous surge would directly damage the aging smart lamp A capacitor.

[0110] Within the confines of the three red lines, the cloud platform generated hundreds of tentative control strategies: Strategy 1: Smart lamp A slowly increases its brightness to 80%, while TV backlight B remains unchanged; Evaluation result: Lifetime constraint is met, but desktop illuminance is only 280 lx, violating the first constraint, and it is directly eliminated; Strategy 2: Smart lamp A increases its brightness to 85%, while simultaneously instructing TV backlight B to reduce light pollution and thus benefit smart lamp A; Evaluation result: Desktop illuminance reaches 310 lx, energy efficiency ratio is 78, ESR shows no increase, this individual currently does not violate any constraints, becoming the first-generation non-dominated solution; Strategy 3: Smart lamp A only increases its brightness to 75%, but significantly reduces the anti-glare angle of smart lamp A. The illuminance was precisely 320 lx, the UGR dropped to 16 (better luminous efficacy), the energy efficiency ratio was 82 (higher energy efficiency), and the ESR showed no increase (equally safe lifespan). When comparing the second and third strategies, it was found that the third strategy completely outperformed the second strategy in terms of luminous efficacy and energy efficiency, and the lifespan index was on par. Therefore, the third strategy absolutely dominates the second strategy. The algorithm marked the third strategy and similar balancing strategies as the highest Pareto Rank 1 level, and laid out a Pareto optimal frontier in the three-dimensional target space that "makes reading bright, consumes little electricity, and preserves the lifespan of old lamps", officially starting a deeper generation of optimization.

[0111] Furthermore, during the Pareto optimal frontier search, an adaptive allocation mechanism is introduced, and the interaction between light quality, energy consumption level, and device lifespan is dynamically balanced based on the current working data of the smart lamps. Multiple smart control items are determined by tracing this interaction relationship. In addition, a global intelligent control strategy for the smart lamp cluster is generated by combining the lighting modes of each smart lamp, so that the smart lamp cluster can achieve the optimal configuration of global performance in the complex and ever-changing home environment. This further controls the multi-dimensional decision space, fully considers the triple constraints and the smart lamp cluster, and improves the accuracy of the global intelligent control strategy.

[0112] At this point, during the iterative search process of the Pareto optimal frontier, the static, equally weighted multi-objective evaluation system is abandoned, and an adaptive allocation mechanism is introduced. Real-time acquisition of the current operating data of the smart lighting fixtures, including the real-time ambient light gap ratio, the instantaneous thermal resistance state of the driver, and the real-time phase angle of the power grid, serves as a feedback source for dynamically adjusting the weights of the three objective functions: luminous efficacy quality, energy consumption level, and device lifespan. An adaptive factor based on fuzzy logic or a dynamic penalty function is constructed: when the operating data reflects a sharp increase in the ambient light gap exceeding a safety threshold, the adaptive factor exponentially increases the optimization weight of the luminous efficacy quality objective while dynamically decreasing the tolerance of the weights of the energy consumption and lifespan objectives; conversely, when the operating data reflects that the device junction temperature is approaching its physical limit, the adaptive factor forcibly locks in an absolutely high weight for the lifespan objective. Through this closed-loop feedback, the selection pressure of the non-dominated ranking is dynamically changed in the three-dimensional objective space, forcing the population evolution direction to seek the optimal balance point that conforms to the current physical conditions in real-time within the interactive game relationship between luminous efficacy quality, energy consumption level, and device lifespan.

[0113] As the adaptive allocation mechanism guides the Pareto front to gradually converge toward the optimal equilibrium point, the chromosomes of elite individuals on the convergence boundary are extracted for gene decoding. Following the interaction relationships of the decoded multi-dimensional parameters in the spatiotemporal topology graph, a reverse tracing is performed: adjusting the step size based on the decoded PWM duty cycle, tracing back to the edge in the spatial topology graph where the power line carrier coupling impedance is extremely low, establishing the "carrier time slot peak-shifting scheduling" intelligent control project; tracing back to the edge in the spatial topology graph where the light spillover interference coefficient is high based on the decoded luminous flux compensation allocation ratio, establishing the "cross-node luminous flux restraint and spatial light field reshaping" intelligent control project; and tracing back to the node with potential thermal stress accumulation based on the decoded color temperature shift constraint, establishing the "thermal derating smooth transition" intelligent control project. Thus, the abstract mathematical optimal solution is reduced in dimension and decoupled into a set of discrete intelligent control projects with clear physical action directions.

[0114] The established set of discrete intelligent control projects is semantically matched and fused with the pre-built intelligent lighting mode library in the industrial cloud platform. Based on the user's habitual characteristics in the scenario, the control projects are encapsulated into the corresponding lighting mode's instruction frame structure. Differentiated parameter allocation is applied to each intelligent lighting fixture within the cluster: for main lighting fixtures, instructions containing precise and limited dimming curves and anti-glare tightening are issued; for background lighting fixtures experiencing interference, instructions for color temperature shift and brightness base adjustment are issued; for lighting fixtures in congested areas, instructions for communication delay compensation are issued. Finally, a global intelligent control strategy message for the intelligent lighting cluster, including timestamp verification, device addressing, action sequences, and anomaly fallback mechanisms, is generated and executed through the edge gateway to achieve optimal global performance configuration for the cluster in complex and changing environments. Lighting modes include "Focused Reading Mode" and "Ambient Audio-Visual Mode," etc.

[0115] Specifically, smart light fixture A is aging; TV background light B is interfered with by light overflow from smart light fixture A; and dining room chandelier C shares a poor PLC channel with smart light fixture A. When the industrial cloud platform reaches the 50th generation in its Pareto front search, its adaptive allocation mechanism is working intensely.

[0116] Weight Adaptive: The sensor transmits current working data showing that the desktop illuminance has dropped to 150 lx (the light gap is extremely large, affecting the user's eyesight); the adaptive factor immediately increases the weight of "light efficiency quality" to 0.8, reduces the weight of "energy consumption level" to 0.1, and maintains the weight of "lifespan" at the bottom line of 0.1.

[0117] Dynamic equilibrium result: The algorithm found an "aggressive but bottom-line" solution in the population - let smart lamp A undertake the main supplementary lighting, but resolutely avoid large current jumps, and maintain the lifespan bottom line of 0.1 to prevent capacitor explosion. This dynamic equilibrium breaks the deadlock of the conventional "slow supplementary lighting to maintain lifespan" and adapts to the sudden changes in the transient environment.

[0118] The cloud platform decodes the chromosome of this radical solution and performs physical tracing along the graph topology edges; tracing the optical-electric coupling edge (smart lamp A and restaurant chandelier C): the decoding shows that smart lamp A needs to climb with a specific low-frequency PWM sequence; the cloud platform traces to the "extremely low carrier impedance edge" between AC and determines that if it is sent directly, it will inevitably lead to a message collision; therefore, control item one is established: "carrier time slot staggered scheduling item for smart lamp A and restaurant chandelier C", which stipulates that smart lamp A must send dimming commands during the millisecond interval when restaurant chandelier C is in sleep mode.

[0119] Tracing the light overflow edge (smart lamp A and TV background light B): Decoding shows that smart lamp A will seriously exceed the limit after the brightness is increased; the cloud platform traces to the "high light overflow edge" between A and B, and establishes control project two: "cross-node luminous flux restraint and light field reshaping project", which forces smart lamp A to reduce the angle of its internal anti-glare honeycomb lens while increasing the brightness, and absolutely does not allow photons to overflow to the TV wall where the TV background light B is located.

[0120] Tracing thermal stress (intelligent lamp A itself): In response to the pseudo-low temperature caused by direct fan blowing and the aging capacitor, control project three was established: "thermal derating smooth transition project". Taking advantage of the pseudo-cool state brought by the fan, the surge current is converted into a small preheating of the aging capacitor by increasing in a very small step (1% per second).

[0121] The cloud platform encapsulates the above three core control items into a policy message for "Smart Home - Deep Reading Mode": Policy investment for smart light fixture A: issue the command [Enter Reading Mode: Activate anti-glare angle contraction, climb to 85% duty cycle at an extremely low frequency of 1% per second using PWM, avoid sending in the time slot of dining room chandelier C, and utilize the cool air bonus for thermal smoothing]; Policy investment for TV background light B: issue the command [Enter Reading Collaboration Mode: In response to the light field reshaping of smart light fixture A, actively adjust the background wall color temperature from 2700K to 2600K, and reduce the brightness by 5% to absorb residual overflow light]; Policy investment for dining room chandelier C: issue the command [Enter Channel Avoidance Mode: Release bus control and enter deep listening state].

[0122] With the implementation of this global intelligent control strategy, users no longer experience "the light suddenly brightens and glares" or "the light slowly turns on and affects reading" in the living room. Instead, the light on the book gradually fills the space in an extremely smooth, flicker-free, and highly focused manner, while the atmosphere of the TV background wall remains intact. The old smart light fixture A does not emit a burnt smell due to the sudden increase in power. Through the perfect closed loop of graph neural network and Pareto optimization, the system truly achieves the optimal global performance configuration of "light, electricity, heat, and lifespan" in the complex and harsh micro-environment of the home.

[0123] Please see Figure 6 The intelligent control system for smart lighting fixtures based on an industrial cloud platform is applied to the aforementioned intelligent control method for smart lighting fixtures based on an industrial cloud platform; the intelligent control system for smart lighting fixtures based on an industrial cloud platform includes:

[0124] The multidimensional state tensor module 21 is used to collect multi-source heterogeneous data of smart lamps, trigger a multi-scale feature extraction network along the spatiotemporal correlation of the multi-source heterogeneous data, and map multiple state features corresponding to the multi-source heterogeneous data into a multidimensional state tensor.

[0125] The industrial cloud platform module 22 is used to input multidimensional state tensors into the corresponding spatiotemporal graph neural network in the industrial cloud platform, and further construct a spatial topology graph with smart lamps as nodes and the light spillage interference and power line carrier coupling relationship between smart lamps as edges.

[0126] The multidimensional decision module 23 is used to pass messages on the spatial topology graph and combine a temporal convolutional network to capture the long-term evolution of lighting demand, thereby outputting a multidimensional decision space that represents the collaborative control of intelligent lighting clusters. The multidimensional decision space covers the physical execution dimension, the environment-energy efficiency dimension, and the service life dimension.

[0127] The intelligent control module 24 is used to simultaneously perform Pareto optimal frontier search in a multi-dimensional decision space, using the luminous efficacy quality index of the intelligent lamps, the system energy efficiency ratio, and the lighting light decay index as triple constraints, in order to generate a global intelligent control strategy for the intelligent lamp cluster.

[0128] It should be noted that although multiple modules are mentioned in the detailed description above, this division is not mandatory; in fact, according to the embodiments of this disclosure, the features and functions of two or more modules or described above can be embodied in one module; conversely, the features and functions of one module described above can be further divided into multiple modules to be embodied.

[0129] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein; this application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein; the specification and embodiments are to be considered exemplary only.

[0130] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent control of smart lighting fixtures based on an industrial cloud platform, characterized in that, include: Collect multi-source heterogeneous data from smart lighting fixtures, trigger a multi-scale feature extraction network along the spatiotemporal correlation of the multi-source heterogeneous data, and map the multiple state features corresponding to the multi-source heterogeneous data into multi-dimensional state tensors. In the industrial cloud platform, the multidimensional state tensor is input into the corresponding spatiotemporal graph neural network. The smart lamps are used as nodes, and the light spillover interference and power line carrier coupling relationship between the smart lamps are used as edges to further construct a spatial topology graph. Message passing is performed on the spatial topology graph, and the evolution of long-term lighting demand is captured by combining a temporal convolutional network, thereby outputting a multi-dimensional decision space that represents the collaborative control of intelligent lighting clusters. The multi-dimensional decision space covers the physical execution dimension, the environmental-energy efficiency dimension, and the service life dimension. Within a multidimensional decision space, using the luminous efficacy quality index of intelligent lighting fixtures, the system energy efficiency ratio, and the lighting light decay index as triple constraints, a Pareto optimal frontier search is performed simultaneously to generate a global intelligent control strategy for intelligent lighting fixture clusters.

2. The intelligent control method for intelligent lighting fixtures based on an industrial cloud platform according to claim 1, characterized in that, The process involves collecting multi-source heterogeneous data from smart lighting fixtures, triggering a multi-scale feature extraction network along the spatiotemporal correlations of this data, and mapping multiple state features corresponding to the multi-source heterogeneous data into multi-dimensional state tensors, including: The smart lamp synchronously collects multi-source heterogeneous data through the built-in sensing array. This multi-source heterogeneous data includes real-time illuminance flux, node voltage and current waveforms, and driving junction temperature. Based on the identification of multi-source heterogeneous data in the spatiotemporal dimension, the corresponding related content is determined and input into a pre-constructed multi-scale feature extraction network.

3. The intelligent control method for intelligent lighting fixtures based on an industrial cloud platform according to claim 2, characterized in that, The process of collecting multi-source heterogeneous data from smart lighting fixtures, triggering a multi-scale feature extraction network along the spatiotemporal correlations of the multi-source heterogeneous data, and mapping multiple state features corresponding to the multi-source heterogeneous data into multi-dimensional state tensors, also includes: By using parallel dilated convolutional kernels and cross-modal attention mechanisms within a multi-scale feature extraction network, nonlinear projection and alignment of multi-source heterogeneous data are performed to filter out transient environmental noise and extract high-order interaction features. These features are then combined with multiple state features corresponding to the multi-source heterogeneous data and reorganized according to a preset physical semantic dimension, mapping them into a multi-dimensional state tensor that represents the coupling relationship between the current operating mode of the lighting fixture and the local environment.

4. The intelligent control method for intelligent lighting fixtures based on an industrial cloud platform according to claim 1, characterized in that, In the industrial cloud platform, a multidimensional state tensor is input into the corresponding spatiotemporal graph neural network. Using smart lighting fixtures as nodes and the light spillover interference and power line carrier coupling relationships between these fixtures as edges, a spatial topology graph is further constructed, including: The industrial cloud platform is monitored in real time, and smart home scenarios of smart lighting fixtures are marked. The multi-dimensional state tensor is input into the corresponding spatiotemporal graph neural network along the communication channel. Each smart lighting fixture in the smart home scenario is used as a graph node. The light spillover interference coefficient between each smart lighting fixture is obtained synchronously and used as the edge content along with the corresponding power line carrier coupling impedance. This further constructs a spatial topology graph that represents the strong correlation characteristics of physical space.

5. The intelligent control method for intelligent lighting fixtures based on an industrial cloud platform according to claim 4, characterized in that, In the industrial cloud platform, the multidimensional state tensor is input into the corresponding spatiotemporal graph neural network. Using smart lighting fixtures as nodes and the light spillover interference and power line carrier coupling relationships between smart lighting fixtures as edges, a spatial topology graph is further constructed. This also includes: To address light spillover interference, the IES light distribution files of each smart luminaire in the industrial cloud platform are retrieved. Combined with the coordinates and installation orientation of the smart luminaires, the light radiation content is determined. Furthermore, the distribution of photons from the target smart luminaire on the illuminated surfaces of adjacent smart luminaires is tracked and statistically analyzed to determine the unexpected light contribution, which is then normalized to obtain the light spillover interference content. To address the power line carrier coupling impedance, a broadband probe carrier signal is orthogonally injected into the operating current of the smart lighting fixture. The signal attenuation and phase shift characteristics between nodes under the same phase line are extracted, and the channel frequency response, including distributed capacitance and inductance, is fitted to dynamically determine the power line carrier coupling impedance.

6. The intelligent control method for intelligent lighting fixtures based on an industrial cloud platform according to claim 1, characterized in that, The process involves message passing on a spatial topology graph and combining it with a temporal convolutional network to capture the long-term evolution of lighting demand, thereby outputting a multi-dimensional decision space characterizing the collaborative control of intelligent lighting clusters. This multi-dimensional decision space encompasses the physical execution dimension, the environment-energy efficiency dimension, and the lifespan dimension, including: Multi-hop neighbor message passing based on a gating mechanism is performed on the spatial topology graph to obtain cross-node optical-electric coupling state evolution information. The cross-node optical-electric coupling state evolution information is aggregated and processed. Combined with an extended causal temporal convolutional network embedded in the spatiotemporal graph neural network, parallel sliding window convolution is performed on the historical multidimensional state tensor sequence along the time axis. Gradient normalization and nonlinear activation are used to capture the evolution of user lighting demand over a long time span.

7. The intelligent control method for intelligent lighting fixtures based on an industrial cloud platform according to claim 6, characterized in that, The process involves message passing on a spatial topology graph and combining it with a temporal convolutional network to capture the long-term evolution of lighting demand, thereby outputting a multi-dimensional decision space characterizing the collaborative control of intelligent lighting clusters. This multi-dimensional decision space encompasses the physical execution dimension, the environmental-energy efficiency dimension, and the lifespan dimension, and also includes: The aging hysteresis events of the intelligent lighting cluster are acquired and decoupled at multiple levels in combination with the evolution law of illumination demand. This decouples and outputs a multi-dimensional decision space that characterizes the collaborative control of the intelligent lighting cluster. The multi-dimensional decision space includes the physical execution dimension, which includes dimming depth and color temperature step frequency; the environmental-energy efficiency dimension, which includes the target illuminance achievement rate and reactive power loss; and the lighting lifespan dimension, which includes the growth rate of accumulated heat generation and the equivalent series resistance of electrolytic capacitors.

8. The intelligent control method for intelligent lighting fixtures based on an industrial cloud platform according to claim 1, characterized in that, The process involves simultaneously performing a Pareto optimal frontier search within a multidimensional decision space, using the luminous efficacy quality index of intelligent lighting fixtures, the system energy efficiency ratio, and the lighting light decay index as triple constraints, to generate a global intelligent control strategy for intelligent lighting fixture clusters. This strategy includes: Real-time monitoring of the multi-dimensional decision space is used to obtain the luminous efficacy quality index, system energy efficiency ratio, and lighting light decay index of the smart lamps. Based on the fusion of the luminous efficacy quality index, system energy efficiency ratio, and lighting light decay index, triple constraints are determined, and Pareto optimal frontier search based on non-dominated sorting is initiated simultaneously.

9. The intelligent control method for intelligent lighting fixtures based on an industrial cloud platform according to claim 8, characterized in that, The method of simultaneously performing Pareto optimal frontier search within a multi-dimensional decision space, using the luminous efficacy quality index of intelligent lighting fixtures, the system energy efficiency ratio, and the lighting light decay index as triple constraints, to generate a global intelligent control strategy for intelligent lighting fixture clusters, also includes: In the Pareto optimal frontier search process, an adaptive allocation mechanism is introduced, and the interaction between light quality, energy consumption level and device lifespan is dynamically balanced by combining the current working data of smart lamps. By tracing the interaction, multiple smart control items are determined. Furthermore, a global smart control strategy for the smart lamp cluster is generated by combining the lighting modes of each smart lamp, so that the smart lamp cluster can achieve the optimal configuration of global performance in the complex and ever-changing home environment.

10. An intelligent control system for intelligent lighting fixtures based on an industrial cloud platform, characterized in that, The intelligent control system for intelligent lighting fixtures based on an industrial cloud platform is applied to the intelligent control method for intelligent lighting fixtures based on an industrial cloud platform as described in any one of claims 1-9. The intelligent control system for the smart lighting fixtures based on the industrial cloud platform includes: The multidimensional state tensor module is used to collect multi-source heterogeneous data from smart lighting fixtures, trigger a multi-scale feature extraction network along the spatiotemporal correlation of the multi-source heterogeneous data, and map multiple state features corresponding to the multi-source heterogeneous data into a multidimensional state tensor. The industrial cloud platform module is used to input multidimensional state tensors into the corresponding spatiotemporal graph neural network in the industrial cloud platform, and further construct a spatial topology graph with smart lamps as nodes and the light spillover interference and power line carrier coupling relationship between smart lamps as edges. The multidimensional decision module is used to pass messages on the spatial topology graph and combine it with a temporal convolutional network to capture the long-term evolution of lighting demand, thereby outputting a multidimensional decision space that represents the collaborative control of intelligent lighting clusters. The multidimensional decision space covers the physical execution dimension, the environmental-energy efficiency dimension, and the service life dimension. The intelligent control module is used to simultaneously perform Pareto optimal frontier search within a multi-dimensional decision space, using the luminous efficacy quality index of the intelligent lighting fixtures, the system energy efficiency ratio, and the lighting light decay index as triple constraints, in order to generate a global intelligent control strategy for the intelligent lighting fixture cluster.

Images