A Fast Filtering and Matching Method for Lighting Fixture Databases Based on a Six-Dimensional Tag System

By employing a rapid filtering and matching method based on a six-dimensional label system for lighting databases, and utilizing a dimension-specific semantic encoder and dynamic envelope model, semantic anchor point drift is detected and lightweight calibration is implemented. This solves the problems of insufficient real-time adaptability and response latency in high-frequency interaction scenarios in existing technologies, achieving efficient and stable multi-dimensional matching and recommendation.

CN123115373APending Publication Date: 2026-11-17CLING GUANGZHOU ELECTRONIC TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611085302.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for rapid filtering and multidimensional matching of lighting databases struggle to adapt to subtle changes in user intent in real time. This leads to deviations between intent understanding and label mapping results and the original semantics, affecting the accuracy and consistency of multidimensional matching. Furthermore, high-frequency interaction scenarios require significant computing power and have long response times, making it impossible to distinguish between normal semantic evolution and abnormal drift, resulting in model oscillations and short-term label failures.

Method used

A method based on a six-dimensional labeling system is adopted. An initial feature vector is generated through a dimension-specific semantic encoder, a dynamic envelope model is constructed, semantic anchor point drift is detected, and a directional constraint reprojection and hierarchical weighting strategy are implemented to achieve lightweight semantic calibration and incremental updates, avoiding full image reconstruction.

Benefits of technology

It enhances the system's adaptability to designers' frequent leaps and progressive language behaviors, reduces response latency, improves the system's usability and fluency in real-world design dialogue scenarios, and ensures semantic stability and robustness, making it suitable for creative design fields.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115373A_ABST
    Figure CN123115373A_ABST
Patent Text Reader

Abstract

This invention relates to a rapid filtering and matching method for lighting databases based on a six-dimensional tagging system. For natural language requirement texts, a trained professional semantic encoder and a six-dimensional vector space are used to decouple and extract multi-dimensional tags such as time, space, people, context, emotion, and style. After semantic anchor point clustering, interactive dynamic calibration, and hierarchical weighting processing, a semantic matching model with steady-state adaptive capabilities is constructed. Quantile trajectory modeling is used to detect and correct semantic anchor point drift, achieving dimensional semantic drift closed-loop correction, effectively improving the accuracy of retrieval recommendations and the long-term adaptability of the system. Finally, standardized six-dimensional comprehensive features are efficiently matched with database materials, and combined with structural feasibility verification and BOM cost estimation, an optimized design scheme that meets spatial and usage constraints is output. This invention can accurately analyze complex requirements and improve matching quality, enhancing the robustness and self-learning ability of the recommendation system.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of semantic tag space dynamic calibration and multidimensional matching technology, and in particular to a fast screening and matching method for a lighting database based on a six-dimensional tag system. Background Technology

[0002] In the current field of rapid filtering and multi-dimensional matching of lighting fixture databases, mainstream solutions mostly adopt technical approaches such as knowledge graph-based tag embedding, dynamic graph structure updates, RAG-enhanced semantic retrieval, or large language model-based prompting and fine-tuning. These approaches establish a static mapping relationship between lighting fixture attribute tags and design semantics to understand user intent and match candidate materials. In practical implementation, industry-standard methods generally rely on lighting fixture design knowledge graphs, scene ontology alignment, and multi-modal tag collaboration to improve the accuracy of multi-tag matching and scene coverage. Recent development trends tend to further optimize the intelligence level of intent adaptation by integrating deep semantic encoders, multi-dimensional intent parsing models, and database retrieval engines, and attempt to introduce adaptive weights or incremental knowledge update mechanisms to respond to changes in user habits and scenarios.

[0003] Existing dynamic calibration techniques for semantic label spaces primarily focus on the accurate mapping of labels to intents in complex design tasks, concentrating on static learning of semantic knowledge, vector embedding alignment, and full graph reconstruction. For example, in multi-label matching in the lighting industry, common techniques train multi-level semantic encoders based on historical project labels and expert annotation data. Leveraging the node structure and upstream / downstream dependencies of the knowledge graph, they periodically reconstruct or fuse semantic mapping relationships to maintain the synchronous evolution of the label system and user intent. Some recent advancements have begun to address the lack of adaptability of the label space as the scenario changes through semantic drift detection, model fine-tuning, or dynamic adjustment of label weights. However, "knowledge representation updates" remain the mainstream approach, and related technologies often face practical challenges such as high computational consumption, poor real-time performance, and model oscillations.

[0004] Multi-dimensional tag spatial mapping applications for lighting fixtures involve not only the parsing and embedding of multi-dimensional tags (such as time, space, people, context, emotion, and style), but also the accurate positioning of users' natural language intent within the multi-dimensional space. Typically, industry solutions are applicable to scenarios such as rapid retrieval of lighting fixture materials, structural design adaptation, AI-assisted lighting fixture solution recommendations, and emotional scene lighting design. Typical applications of these solutions require support for personalized expression by designers across multiple interaction rounds and ensure the stability and evolution of multi-dimensional semantic tags throughout the design cycle.

[0005] However, current technologies have several significant shortcomings: First, because most existing methods rely on static knowledge graphs or periodically reconstructing label spaces, they struggle to adapt to subtle changes in user intent in real time. This leads to "semantic drift" between intent understanding and the labels in the lighting database. Specifically, during multiple rounds of natural language interaction or scene switching, the label mapping results often gradually deviate from the original semantics, affecting the accuracy and consistency of multidimensional matching. Second, solutions that rely entirely on full model updates, knowledge graph reconstruction, or fine-tuning of prompts require high computational power and have long response latency, making them unsuitable for high-frequency, real-time, and fine-grained intent calibration interaction scenarios. Third, traditional dynamic semantic weight adjustment mechanisms often fail to distinguish between normal semantic evolution and abnormal drift, easily causing model "oscillations" or even short-term label failures, lacking the ability to differentiate and recover from semantic drift. Furthermore, in the absence of human annotation feedback or scene cognitive inertia, current technologies lack the elasticity and flexibility of dynamic semantic space labeling, making it difficult to guarantee semantic stability and the system's personalized adaptability across multiple rounds of interaction. Summary of the Invention

[0006] This application provides a fast filtering and matching method for a lighting database based on a six-dimensional tag system, which aims to solve one of the problems or issues of the existing technology mentioned in the background.

[0007] This application provides a fast filtering and matching method for lighting fixture databases based on a six-dimensional tag system, specifically including:

[0008] S1: Obtain the user's natural language requirement text, and use a preset dimension-specific semantic encoder to decouple the user's natural language requirement text to generate a six-dimensional initial dimension feature vector.

[0009] S2: Based on the six-dimensional initial feature vector, calculate the cluster center of all tag instances in each dimension, define the cluster center as the initial semantic anchor point of the six-dimensional tag system, and extract the six-dimensional sub-vectors of the current interaction round as the current activation coordinates;

[0010] S3: For the current active coordinates of each dimension, call the historical anchor point coordinate sequence within the sliding time window to construct a dynamic envelope model and establish the semantic stable region of each dimension;

[0011] S4: If there exists a dimension where the current active coordinate exceeds the boundary of the corresponding semantically stable region for three consecutive rounds, then mark the dimension as the drift dimension where semantic anchor point drift occurs, and generate a drift detection signal.

[0012] S5: In response to the drift detection signal, freeze the five dimensional coordinates that have not drifted, perform a direction-constrained reprojection operation only on the current sub-vector of the drifted dimension, apply a semantic restoring force that is non-linearly proportional to the deviation distance in the opposite direction of the deviation, and generate calibrated dimensional coordinates.

[0013] S6: Based on the calibrated dimensional coordinates and the original coordinates of the un-drifted dimensions, construct a six-dimensional comprehensive feature vector;

[0014] S7: Perform cosine similarity calculation on the six-dimensional comprehensive feature vector, and output the matching results of lighting materials sorted by similarity, as well as the corresponding structural feasibility verification data and BOM cost estimation data;

[0015] S8: After completing ten rounds of effective interaction, extract new anchor point samples that are still in the semantically stable region after calibration in this session, and perform incremental smooth updates on the dynamic envelope model of each dimension to generate the updated semantically stable region.

[0016] This application provides a fast filtering and matching method for lighting fixture databases based on a six-dimensional tag system, which has the following beneficial effects:

[0017] (1) By modeling the six-dimensional scene tags (time / space / person / context / emotion / style) as semantic anchors with physical plasticity rather than static symbol nodes, this solution effectively overcomes the problem of delayed intent response and context breakage caused by the semantic solidification of tags in traditional lighting design recommendation systems. Existing technologies mostly rely on global knowledge graph reconstruction or large model fine-tuning to cope with the evolution of user intent, which has a large computational cost and is prone to semantic oscillation. However, this solution constructs a lightweight semantic encoder specific to each dimension in the initialization stage, and generates initial cluster centers based on historical project corpora and expert annotations as semantic anchors for each dimension, so that each tag has a measurable and adjustable vector representation basis. Combined with quantile trajectory modeling within the sliding time window, the system dynamically depicts the semantically stable regions of each dimension, gets rid of the dependence on the Gaussian distribution assumption, and significantly improves the adaptability to non-stationary interaction sequences. It is especially suitable for the language behavior characteristics of designers who frequently exhibit leaps in expression and progressive focus during the creative exploration process, thereby achieving high sensitivity perception of semantic drift without introducing complex model training.

[0018] (2) In response to the detected semantic anchor point drift events, this scheme innovatively introduces a local elastic calibration mechanism and a directional constraint reprojection strategy to achieve synergistic optimization of semantic evolution and cross-round consistency. When a certain dimension sub-vector exceeds its semantic stability boundary for three consecutive rounds, the system determines that a substantial semantic shift has occurred, but does not trigger full-image reconstruction or parameter relearning. Instead, it freezes the coordinates of the other five dimensions and applies a "semantic restoring force" that is non-linearly proportional to the deviation distance to pull it back into the stable region. This prevents the uncontrolled spread of semantics while preserving reasonable extension space, simulating the cognitive inertia of human designers who accept minor adjustments of inspiration while maintaining the coherence of the core concept. Furthermore, a hierarchical weighting strategy is adopted in subsequent multi-dimensional similarity matching: the non-drifted dimension uses the original coordinates to participate in cosine calculation to ensure accuracy, while the drifted dimension uses the calibrated coordinates and is superimposed with a weight coefficient that dynamically decays with the calibration intensity, avoiding the sudden change in recommendation caused by over-correction. This mechanism significantly reduces the latency of hundreds of milliseconds caused by frequent retraining or retrieval enhancement in traditional methods, achieving millisecond-level online response. At the same time, it can complete closed-loop adaptive adjustment without manual annotation and feedback, significantly improving the system's practicality and fluency in real-world dialogue design scenarios.

[0019] (3) By periodically and incrementally updating semantically stable regions, this scheme constructs a dynamic memory system that combines robustness and evolutionary capability. After every ten rounds of effective interaction, the system only selects new anchor point samples that remain stable within the current envelope region after calibration, and uses them to smoothly update the quantile trajectory boundaries of each dimension. This effectively filters short-term noise interference while absorbing long-term trend changes, avoiding the overall model oscillation caused by a single abnormal input. This design fully utilizes the characteristics of strong scene continuity and relative orthogonality between dimensions in the field of lighting design, transforming semantic control from the traditional "knowledge representation update" paradigm to the "multi-dimensional spatial mechanics control" paradigm, completely avoiding high-cost technical paths such as tag graph reconstruction, embedding alignment, RAG retrieval, and multimodal encoding. The overall architecture not only has good interpretability—the semantic drift and calibration process can be visualized through anchor point trajectories—but also exhibits excellent scalability, making it easy to migrate to other creative design fields that require multi-dimensional scene modeling. Ultimately, the system achieved accurate tracking and robust response to the designer's intent with low computing power consumption, effectively supporting an efficient, natural, and continuous human-computer collaborative design process from concept description to lighting fixture form suggestions.

[0020] In summary, this method, by constructing a spatial control mechanism based on semantic anchors and dynamic envelopes, achieves stable parsing and continuous adaptation of complex and dynamic user intentions in lighting design scenarios without the need for external feedback and large-scale computational support. It balances real-time response, semantic flexibility, and system robustness, forming a lightweight and adaptive new paradigm for semantic understanding in the creative field. Attached Figure Description

[0021] Figure 1 This is the main flowchart of a fast filtering and matching method for lighting databases based on a six-dimensional tag system;

[0022] Figure 2 This is a sub-flowchart of a fast filtering and matching method for lighting fixture databases based on a six-dimensional tag system;

[0023] Figure 3 This is another sub-flowchart of a fast filtering and matching method for lighting databases based on a six-dimensional tag system;

[0024] Figure 4 This is an application environment diagram of a fast filtering and matching method for a lighting database based on a six-dimensional tag system, as described in one embodiment.

[0025] Figure 5 This is a schematic diagram of the structure of a computer device provided in one embodiment of this application. Detailed Implementation

[0026] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0027] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0028] like Figure 1 As shown, this application provides a fast filtering and matching method for a lighting database based on a six-dimensional tag system, specifically including:

[0029] S1: Obtain the user's natural language requirement text, and use a pre-set dimension-specific semantic encoder trained on a historical lighting project corpus and professional designer labeled data to decouple the user's natural language requirement text and generate a six-dimensional initial dimensional feature vector corresponding to the six dimensions of time, space, person, context, emotion and style.

[0030] S2: Based on the six-dimensional initial feature vector, calculate the cluster center of all tag instances within each dimension, define the cluster center as the initial semantic anchor point of the six-dimensional tag system, and extract the six-dimensional sub-vectors of the current interaction round as the current activation coordinates. The dimensions include time, space, person, context, emotion, and style mentioned above.

[0031] S3: For the current active coordinates of each dimension, call the historical anchor point coordinate sequence within the sliding time window, and use the quantile trajectory modeling method to construct a dynamic envelope model containing the upper tenths line and the lower tenths line, so as to establish the semantic stable region of each dimension.

[0032] S4: If there exists a dimension where the current active coordinate exceeds the boundary of the corresponding semantically stable region for three consecutive rounds, then mark the dimension as the drift dimension where semantic anchor point drift occurs, and generate a drift detection signal.

[0033] S5: In response to the drift detection signal, freeze the five dimensional coordinates that have not drifted, perform a directional constraint reprojection operation only on the current subvector of the drifted dimension, apply a semantic restoring force in the opposite direction of the deviation direction that is non-linearly proportional to the deviation distance, and generate calibrated dimensional coordinates.

[0034] S6: Based on the calibrated dimension coordinates and the original coordinates of the un-drifted dimension, construct a six-dimensional comprehensive feature vector under a hierarchical weighting strategy, wherein the calibrated dimension coordinates are superimposed with a decay weight coefficient that is dynamically adjusted with the calibration intensity.

[0035] S7: Perform cosine similarity calculation on the six-dimensional comprehensive feature vector, and output the matching results of lighting materials sorted by similarity, as well as the corresponding structural feasibility verification data and BOM cost estimation data.

[0036] S8: After completing ten rounds of effective interaction, extract new anchor point samples that are still in the semantically stable region after calibration in this session, perform incremental smooth updates on the dynamic envelope model of each dimension, and generate updated semantically stable regions to complete closed-loop optimization.

[0037] Step S1: Obtain the user's natural language requirement text. Using a pre-set dimension-specific semantic encoder trained on a historical lighting project corpus and professional designer-annotated data, decouple the user's natural language requirement text to generate a six-dimensional initial feature vector corresponding to six dimensions: time, space, person, context, emotion, and style. Specifically, this includes:

[0038] S1.1: Obtain the natural language requirement text input by the user, and perform word segmentation processing and stop word filtering on the natural language requirement text using a pre-established professional dictionary in the lighting field, so as to generate a standard lighting requirement word sequence containing keywords of time, space, person, scenario, emotion and style.

[0039] Receive the original natural language requirement text input by the user, which contains unstructured colloquial descriptions, mixed professional terms and potential noise characters, and serves as the initial data source for semantic parsing. Perform preprocessing and cleaning operations on the original natural language requirement text, use regular expression matching to remove special symbols, residual HTML tags and meaningless blank characters in the text, generate a standardized plain text sequence, and eliminate the interference of format noise on subsequent word segmentation accuracy. Call the pre-established professional dictionary in the lighting field, which covers the professional term libraries of time dimension (such as early morning, late night), space dimension (such as living room, corridor), person dimension (such as the elderly, children), scenario dimension (such as reading, party), emotion dimension (such as warm, calm) and style dimension (such as Nordic, industrial style), perform forward maximum matching word segmentation processing on the standardized plain text sequence, and split the continuous character stream into a set of candidate lexical units with semantic independence.

[0040] Based on the set of candidate lexical units, perform stop word filtering operation, load the general stop word list and the special stop word list for the lighting field, remove functional words and high-frequency function words that have no substantial semantic contribution such as "de", "le" and "yi ge", retain the content lexical units carrying the core intention, and generate a preliminarily screened key feature word list. For each lexical unit in the key feature word list, calculate its semantic attribution probability in each dimension of the six-dimensional label system, use the pre-trained lightweight classifier to identify the target dimension category to which the lexical unit belongs, mark ambiguous lexical units that do not hit any dimension or have a confidence lower than the preset threshold as noise and remove them, so as to ensure that all remaining lexical units have clear dimensional directivity. According to the preset dimension order of time, space, person, scenario, emotion and style, perform structured reorganization and sorting on the noise-removed lexical units, aggregate the lexical units of the same dimension into a sub-sequence, and finally generate a standard lighting requirement word sequence containing six dimensions of keywords, realizing the conversion from unstructured text to structured semantic units.

[0041] S1.2: Based on the standard lighting requirement word sequence, call the dimension-specific semantic encoder trained on the historical lighting project corpus and professional designer annotation data, and perform context embedding encoding processing on the standard lighting requirement word sequence, so as to generate a high-dimensional original semantic tensor fused with context information.

[0042] The system receives a standard lighting requirement word sequence generated in step S1.1. This sequence contains a set of keywords related to time, space, people, context, emotion, and style, filtered by a domain dictionary and stop words. This set serves as the input data source for the dimension-specific semantic encoder. A pre-trained six-dimensional dimension-specific semantic encoder model is invoked. This model is built on the Transformer architecture and integrates a multi-head self-attention mechanism and a positional encoding module to capture long-distance dependencies and contextual semantic relationships between words in the lighting requirement text. An embedding lookup operation is performed on each lexical element in the standard lighting requirement word sequence. The corresponding initial word embedding vector is retrieved from a predefined lighting domain word vector table, and the positional encoding vector is superimposed on the initial word embedding vector to generate an enhanced word representation vector containing sequence position information.

[0043] The enhanced word representation vectors are input into a multi-layered stacked attention layer of the encoder. A self-attention mechanism is used to calculate the relevance weight matrix between word units, and global contextual information is aggregated on the input sequence to generate a deep semantic hidden state sequence incorporating contextual features. For the deep semantic hidden state sequence, average pooling or max pooling strategies are applied for sequence-level feature compression to eliminate the influence of sequence length differences, extracting a global contextual feature vector that can represent the core semantics of the user's needs for the entire sentence. A non-linear activation function is used to transform the global contextual feature vector, enhancing the non-linear expressive power of the feature space. A dropout regularization layer is used to suppress overfitting, generating a highly robust intermediate semantic representation tensor.

[0044] A fully connected projection layer maps the intermediate layer semantic representation tensor to a high-dimensional latent space. The dimension of this latent space is set to be wide enough to cover the semantic distribution of the six-dimensional label system, ensuring the orthogonality and separability of semantic information across different dimensions. In the high-dimensional latent space, a dimension-aware gating mechanism is introduced. Based on pre-defined prior knowledge of the six-dimensional label categories, preliminary channel weighting is applied to the high-dimensional features, strengthening semantic feature components strongly correlated with the lighting scenario and suppressing noise interference. The output is a high-dimensional original semantic tensor that incorporates contextual information. This tensor retains the complete semantic details and implicit intentions of the user's natural language requirements, providing a context-rich underlying data foundation for subsequent six-dimensional orthogonal decoupling.

[0045] S1.3: Based on the high-dimensional original semantic tensor, a linear projection and feature separation operation is performed on the high-dimensional original semantic tensor using a preset six-dimensional orthogonal decoupling matrix to generate six independent subspace feature vectors corresponding to the time dimension, space dimension, character dimension, context dimension, emotion dimension and style dimension, respectively.

[0046] The system receives the high-dimensional original semantic tensor from the output of step S1.2, which represents the overall semantic distribution of the user's natural language needs in a unified high-dimensional space. However, it has not yet distinguished the independent semantic features of the six specific dimensions: time, space, person, situation, emotion, and style. A pre-set six-dimensional orthogonal decoupling matrix is ​​invoked. This matrix is ​​a linear transformation operator trained during the system initialization phase based on a historical lighting project corpus and professional designer-annotated data, using the principle of maximizing mutual information and minimizing redundancy. Its row vectors correspond to the orthogonal basis directions of the six dimensions, ensuring the statistical independence of each dimension's features in the vector space. The high-dimensional original semantic tensor is expanded into a one-dimensional column vector form, matrix multiplication is performed, and the original semantic tensor is linearly projected using the six-dimensional orthogonal decoupling matrix. This maps the mixed semantic information into six mutually orthogonal subspaces, achieving initial separation of semantic features.

[0047] Sub-vector segments corresponding to the time dimension are extracted from the projected result vector. These segments retain time-related semantic components from the original text, such as "night," "early morning," or "holiday," while suppressing interference noise from other dimensions such as space and emotion. Sub-vector segments corresponding to the space dimension are extracted from the projected result vector. These segments focus on spatial location features such as "living room," "corridor," or "outdoors," ensuring the purity of spatial semantic expression. Sub-vector segments corresponding to the person dimension are extracted from the projected result vector. These segments extract semantic features of target user groups such as "children," "elderly," or "visitors," providing basic data for person profile matching. Sub-vector segments corresponding to the context dimension are extracted from the projected result vector. These segments capture the semantics of behavioral scenarios such as "reading," "gathering," or "sleeping," reflecting the specific functional context of the lighting fixture's use. Sub-vector segments corresponding to the emotion dimension are extracted from the projected result vector. These segments analyze emotional connotations such as "warm," "calm," or "enthusiastic," quantifying the user's subjective psychological expectations. Extract sub-vector segments corresponding to the style dimension from the projected result vector. These segments identify aesthetic style semantics such as "modern minimalism", "new Chinese style" or "industrial style" to establish the visual tone of the lighting design.

[0048] The six sub-vector segments are subjected to dimension reshaping operations to restore them to a preset fixed-length feature vector format, forming six independent subspace feature vectors, each containing only a single dimension of semantic information.

[0049] S1.4: Based on the six independent subspace feature vectors, L2 norm normalization is used to perform modulus standardization on the six independent subspace feature vectors to generate a six-dimensional initial dimension feature vector with a unified metric scale, thus completing the transformation from unstructured text to a standardized vector space.

[0050] Step S2: Based on the initial six-dimensional feature vectors, calculate the cluster centers of all tag instances within each dimension, define the cluster centers as the initial semantic anchors of the six-dimensional tag system, and extract the six-dimensional sub-vectors of the current interaction round as the current activation coordinates. The dimensions include the aforementioned time, space, person, context, emotion, and style. Specifically, they include:

[0051] S2.1: Obtain the six initial dimensional feature vectors generated by the previous steps, and use K-means clustering to iteratively cluster the tag instance feature sets corresponding to each dimension in the lighting database corpus to calculate the initial cluster center vector representing the semantic distribution center of each dimension.

[0052] Obtain the six initial feature vectors generated in the preceding step S1.4 as the baseline input data for constructing semantic anchors. This step is located at the beginning of the main step S2 and aims to extract representative semantic distribution centers from the massive lighting database corpus, providing a static reference system for the subsequent establishment of the dynamic envelope model. Read the feature sets of tag instances corresponding to each dimension in the lighting database corpus, and establish independent feature index tables for each of the six dimensions: time, space, person, context, emotion, and style. Perform data cleaning operations on the feature sets of each dimension, removing missing values ​​and outliers to ensure the integrity and consistency of the clustering input data. Initialize the parameter configuration of the K-means clustering algorithm, setting the number of clusters K to a preset semantic subdivision granularity, usually determined based on the number of natural classifications of each dimension's tags. Randomly select the feature vectors of K tag instances as initial cluster centers, or use the K-means++ optimization strategy to select initial center points that are far apart, to accelerate convergence and avoid local optima.

[0053] The iterative clustering process begins, calculating the Euclidean distance between the feature vector of each label instance and the current K cluster centers. Each label instance is then assigned to the cluster of the nearest cluster center, completing the first round of cluster partitioning. This process calculates the proximity of the sample to the center using the following distance formula:

[0054]

[0055] in, This represents the feature vector of the i-th label instance. Let represent the vector of the j-th cluster center, and D be the dimension of the feature vector. and These are the components of the corresponding dimension.

[0056] Based on the current cluster partitioning, the mean of the feature vectors of all label instances within each cluster is recalculated, and the cluster center positions are updated. The new cluster center vectors are composed of the arithmetic mean of all member vectors within the cluster, ensuring that the new centers minimize the sum of squared errors within the cluster. This update process is performed using the following mean formula:

[0057]

[0058] in, This represents the set of label instances belonging to the j-th cluster. This represents the number of samples in the set.

[0059] To determine if the cluster centers have converged, calculate the displacement distance between the old and new cluster center vectors. If the displacement distances of the cluster centers in all dimensions are less than the preset convergence threshold, or the maximum number of iterations is reached, terminate the iteration process. Otherwise, return to the distance calculation step and continue the next round of cluster partitioning and center updating using the updated cluster centers. Extract the K cluster center vectors from the final converged state as the initial cluster center vectors representing the semantic distribution core of that dimension. Perform the above clustering process on each of the six dimensions to generate an initial set of cluster center vectors containing six dimensions: time, space, person, context, emotion, and style.

[0060] S2.2: Based on the initial cluster center vector, perform a semantic anchor definition operation to map the initial cluster center vector of each dimension to the initial semantic anchor of the corresponding dimension in the six-dimensional label system, so as to generate an initial semantic anchor set containing six dimensions: time, space, person, context, emotion and style.

[0061] S2.3: Receive the user's natural language parsing result for the current interaction round, and use the dimensional decoupling mapping mechanism to separate six independent dimensional sub-vectors from the overall intent vector to generate a sequence of six dimensional sub-vectors corresponding to the current state of the six-dimensional label system.

[0062] The system receives a six-dimensional initial feature vector generated in the preceding steps. This vector represents the overall mapping state of the user's natural language requirements in a unified semantic space and serves as the input benchmark for dimensional decoupling processing. A pre-defined six-dimensional orthogonal projection operator matrix is ​​invoked. This matrix is ​​a linear transformation matrix obtained through singular value decomposition optimization during system initialization, based on the six-dimensional label independence constraints annotated by lighting domain experts. It is used to achieve a redundancy-free mapping from the high-dimensional semantic space to six low-dimensional subspaces. Matrix multiplication is performed between the six-dimensional initial feature vector and the six-dimensional orthogonal projection operator matrix, projecting the overall intent vector into six mutually orthogonal semantic subspaces, initially separating the original component signals of each dimension. For the projected original components of each dimension, a dimension-specific attention masking mechanism is introduced. This mechanism utilizes the attention weight distribution output by the dimension-specific semantic encoder trained in S1.2 to suppress noise interference from non-target dimensions and enhance the saliency of the current dimension's semantic features.

[0063] Element-wise multiplication is used to apply attention mask weights to the corresponding original dimensional components, eliminating feature contamination caused by cross-dimensional semantic coupling and generating six independent dimensional sub-vectors after noise suppression. Residual correction is performed on each independent dimensional sub-vector, calculating the deviation vector between the sub-vector and the initial semantic anchor point of that dimension (generated by S2.2) to capture the personalized offset of the user intent relative to the standard label center. The deviation vector is then superimposed onto the corresponding independent dimensional sub-vector to complete semantic recentering, ensuring that the generated sub-vectors retain the uniqueness of the user intent while remaining within the relative coordinate framework of the global label system.

[0064] S2.4: Based on the six-dimensional sub-vector sequence, perform coordinate extraction and formatting operations to convert each dimensional sub-vector into coordinate point data in a high-dimensional space, so as to generate six-dimensional currently active coordinates that represent the current user intent position in the six-dimensional semantic space.

[0065] The system receives a sequence of six dimensional sub-vectors generated in the preceding step S2.3. This sequence contains feature data for six independent dimensions: time, space, person, context, emotion, and style. These serve as the raw input for coordinate extraction and formatting. A dimensional traversal operation is performed on the six dimensional sub-vector sequence, sequentially extracting the feature vector components corresponding to each dimension. This ensures that the data for each dimension remains orthogonal and independent in subsequent processing, avoiding cross-dimensional semantic interference. For each extracted dimensional sub-vector, a pre-defined coordinate mapping transformation function is invoked to project the high-dimensional sparse feature representation onto specific coordinate axes of the six-dimensional semantic space, achieving dimensionality reduction from the abstract feature space to the concrete geometric coordinate space. During the projection process, a linear transformation matrix is ​​used to weight the dimensional sub-vectors. This transformation matrix is ​​derived based on the co-occurrence frequency statistics of each dimension's labels in a historical lighting design corpus to enhance the coordinate saliency of high-frequency semantic features.

[0066] The Euclidean norm of the projected coordinates in six-dimensional space is calculated. Normalization is then used to eliminate numerical deviations caused by differences in feature dimensions, ensuring that the coordinate values ​​of each dimension are within a unified [0,1] interval. The normalized six-dimensional coordinate values ​​are then structurally assembled according to a preset dimension index order to construct a six-tuple data structure, where each element corresponds to the currently active coordinate value of a dimension. An integrity check is performed on the constructed six-tuple data structure to check for null values ​​or infinite values. If outliers are detected, interpolation is performed using the historical coordinate average from adjacent rounds to correct them, ensuring the continuity of the coordinate data.

[0067] S2.5: Integrate the initial set of semantic anchor points with the six currently active coordinates to construct a six-dimensional semantic space context data packet containing static baselines and dynamic states, so as to output a standardized anchor point coordinate data structure for subsequent dynamic envelope model construction module to call.

[0068] like Figure 2 As shown, step S3: For the current active coordinates of each dimension, the historical anchor point coordinate sequence within the sliding time window is called, and a dynamic envelope model containing the upper and lower tenths digits is constructed using the quantile trajectory modeling method to establish the semantically stable region of each dimension. Specifically, this includes:

[0069] S3.1: Obtain the current active coordinates of a specific dimension in the current interaction round, and call the sliding time window buffer corresponding to that specific dimension to extract the historical anchor point coordinate sequence that has been calibrated in the past several rounds of interaction, so as to form a raw time series data set for trajectory analysis.

[0070] The system receives the currently active coordinate data of the six-dimensional label system output from the preceding step S2, locates the specific dimension index, and extracts the standardized feature vector value of that dimension in the current interaction round as the target anchor point coordinates to be analyzed. It then calls the sliding time window buffer bound to that specific dimension and reads historical interaction records of a preset length N. This buffer uses a circular queue data structure to store the calibrated historical anchor point coordinate sequence, ensuring constant memory usage and efficient data access. The system performs validity checks on the read historical anchor point coordinate sequence, removing null values ​​and illegal values ​​caused by system initialization or abnormal interruptions, and retaining valid coordinate data with complete semantic mapping relationships to form a preliminary candidate time-series dataset.

[0071] The candidate time-series datasets are reverse-sorted according to timestamp order to ensure that the calibrated coordinates generated by the latest round of interaction are at the beginning of the sequence. This reflects the recent priority principle of user intent evolution and enhances the sensitivity of the dynamic envelope model to changes in the current context. The most recent N consecutive valid anchor point coordinates are extracted from the sorted sequence to construct an original time-series dataset containing the current time t and the past N-1 times. This dataset fully records the semantic trajectory fluctuations of this dimension within a local time window. The original time-series dataset is encapsulated into a standardized array structure, with accompanying dimension identifiers and time window metadata, serving as the direct input source for the subsequent quantile trajectory modeling module.

[0072] S3.2: Perform sorting preprocessing on the original time series data set, arrange the historical anchor point coordinate sequence in ascending order according to the numerical value, and generate an ordered historical anchor point coordinate queue to provide a standardized input basis for subsequent quantile calculation.

[0073] S3.3: Based on the ordered historical anchor point coordinate queue, and based on non-parametric quantile statistical processing, locate the coordinates of the lower quantile point with a cumulative probability of 0.1 and the upper quantile point with a cumulative probability of 0.9 in the ordered historical anchor point coordinate queue, and generate the lower tenths threshold and the upper tenths threshold.

[0074] The system receives a pre-sorted, ordered queue of historical anchor point coordinates. This queue contains calibrated historical anchor point coordinates for specific dimensions from past interactions within a sliding time window, serving as the input dataset for non-parametric quantile statistics. The total sample size N of the current sliding time window is determined, and the index positions corresponding to the next decile and the previous decile are calculated according to a pre-defined quantile definition rule. The index positions are precisely calibrated using linear interpolation or the nearest neighbor rank method to ensure the location of continuous cumulative probability boundaries in the discrete data sequence. For the location of the next decile, the following non-parametric quantile statistical processing is used to calculate the lower quantile threshold:

[0075]

[0076] in, The threshold value for the next tenth place is [value]. The i-th element in the ordered queue of historical anchor coordinates Index for the integer part, The weight of the decimal part is specifically determined by... and Confirmed, cumulative probability here. The value is 0.1, and N is the total number of valid historical sample coordinates contained in the ordered historical anchor point coordinate queue within the current sliding time window.

[0077] Based on the index positions calculated above, the corresponding coordinate values ​​are extracted from the ordered historical anchor point coordinate queue. If the index is not an integer, a precise lower quantile threshold is generated using a linear weighted average of two adjacent rank coordinate values. This threshold represents the lower bound of the semantic distribution in this dimension, excluding the bottom 10% of extremely low-frequency or abnormal semantic fluctuations. For the location of the upper quantile points, the same non-parametric quantile statistical processing logic is reused, the cumulative probability parameter is adjusted to 0.9, and the upper quantile threshold is calculated:

[0078]

[0079] in, This is the threshold value for the tenth place. and These are integer indices and decimal weights calculated based on a probability of 0.9, respectively. The coordinate values ​​of the corresponding positions in the queue are obtained by extraction or interpolation.

[0080] The generated upper decile threshold represents the upper boundary of the semantic distribution of this dimension, excluding the top 10% of extremely high-frequency or outlier semantic jumps. Together with the lower decile threshold, it forms a confidence interval covering the middle 80% of historically stable semantic states. The calculated lower and upper decile thresholds are encapsulated into structured data objects, including dimension identifiers, timestamps, lower and upper bound values, and sample confidence labels, forming the dynamic boundary determination parameters for this dimension in the current interaction round. Non-parametric quantile statistical processing is used to extract the bilateral boundaries of the ordered historical anchor point coordinate queue, transforming the sorted data from the previous step into specific numerical indicators representing the upper and lower bounds of the semantically stable region. This achieves a robust quantitative definition of the fluctuation range of user intent in the lighting design context, effectively avoiding boundary estimation bias caused by the failure of the Gaussian distribution assumption.

[0081] S3.4: Using the lower tenths threshold and the upper tenths threshold as spatial boundary constraints, a closed dynamic envelope geometry is constructed in the six-dimensional feature space to generate a dynamic envelope model instance that includes the semantic fluctuation tolerance range.

[0082] The lower and upper decimal place thresholds generated in step S3.3 are received and used as the basic boundary parameters for constructing the dynamic envelope geometry, establishing the extreme value constraint range of the current dimension in the semantic feature space. Based on the vector space characteristics of the six-dimensional label system, the lower and upper decimal place thresholds are mapped to the coordinate axes of the corresponding dimensions, defining the left and right boundary points of the semantically stable region of that dimension, forming a one-dimensional interval constraint. Considering the independence and orthogonality of the semantic features of each dimension in the lighting design scenario, the above boundary mapping operation is performed on the six dimensions of time, space, person, context, emotion, and style respectively, generating six independent one-dimensional semantically stable intervals.

[0083] Using Cartesian product logic, six one-dimensional semantically stable intervals are combined to construct a hyperrectangular geometry in a six-dimensional high-dimensional feature space. This geometry serves as the dynamic envelope model instance in the current interaction state. To ensure the dynamic envelope model has reasonable tolerance for semantic fluctuations, a semantic density weighting mechanism is introduced to calculate the distribution concentration of historical anchor point coordinate sequences within the hyperrectangular geometry, generating a semantic fluctuation tolerance coefficient. This semantic fluctuation tolerance coefficient is then used to fine-tune and expand the boundary of the hyperrectangular geometry. Specifically, the upper decimal threshold is shifted upwards by a small amount, and the lower decimal threshold is shifted downwards by a small amount, with the shift amount proportional to the tolerance coefficient. This generates the final dynamic envelope model instance that encompasses the semantic fluctuation tolerance range.

[0084] Through the above processing method, the quantile threshold of the previous step is transformed into six-dimensional dynamic envelope geometry data with physical boundary meaning and tolerance elasticity, realizing the accurate definition and quantitative description of the semantically stable region of user intent, and providing a reliable spatial benchmark for subsequent drift detection.

[0085] S3.5: Map the dynamic envelope model instance to the semantic coordinate system of the current dimension, define the continuous spatial region enclosed by the lower tenths threshold and the upper tenths threshold, and output the final judgment boundary defined as the semantically stable region of this dimension.

[0086] The system receives the lower and upper decimal thresholds generated in the previous steps and uses them as boundary constraint parameters in the current dimension's semantic space, preparing for a spatial region mapping operation. The lower and upper decimal thresholds are mapped to the normalized semantic coordinate system of the current dimension, determining the lower and upper bound coordinate points of this dimension in the feature space. Based on these lower and upper bound coordinate points, a one-dimensional closed interval geometric structure is constructed, which is represented in the multi-dimensional vector space as a superrectangular slice extending along the current dimension axis. The set of all continuous coordinate points within the closed interval is defined as a semantically stable region, where any coordinate point is considered a legitimate semantic state conforming to the historical distribution pattern of the current user intent.

[0087] The boundary coordinate data of the semantically stable region is encapsulated into a standardized decision boundary object, which includes a dimension index, a lower bound value, an upper bound value, and a confidence level identifier. This decision boundary object is output to the drift detection module as a spatial reference for subsequently determining whether semantic anchor point drift has occurred at the currently active coordinate.

[0088] By using the aforementioned spatial mapping and region delimitation methods, the statistical quantile threshold is transformed into a semantically stable region determination boundary with clear geometric meaning. This enables precise quantitative differentiation between normal fluctuation ranges and abnormal drift states in the dynamic semantic space, providing robust spatial constraints for subsequent drift detection.

[0089] For example, regarding the "emotion" dimension in the six-dimensional tagging system, assuming the sliding time window size N is set to 20, the calibrated anchor point coordinate sequence of the emotion dimension is extracted from the past 20 rounds of interaction. After sorting this sequence in ascending order, the coordinate value of the lower quantile with a cumulative positioning probability of 0.1 is -0.45, and the coordinate value of the upper quantile with a cumulative positioning probability of 0.9 is 0.62. -0.45 is mapped to the lower boundary point in the semantic coordinate system of the emotion dimension, and 0.62 is mapped to the upper boundary point. A closed interval consisting of [-0.45, 0.62] is constructed as the semantically stable region of the emotion dimension. If the activated coordinate of the emotion dimension obtained in the current round is 0.15, since 0.15 is within the interval [-0.45, 0.62], the dimension is determined to be in a semantically stable state; if the current activated coordinate is 0.75, since 0.75 is greater than the upper bound of 0.62, the coordinate is determined to be outside the boundary of the semantically stable region. The system encapsulates the range [-0.45, 0.62] into a decision boundary object containing the dimension index "emotion", a lower bound of "-0.45", an upper bound of "0.62", and a confidence level of "high", and outputs it to downstream modules. This process ensures that minor fluctuations in emotional semantics (such as from calm to slight pleasure) are contained within a stable region, while drastic semantic jumps (such as from calm to extreme anger) are accurately identified as potential drifts, significantly improving the stability and accuracy of intent understanding.

[0090] like Figure 3 As shown, step S4: If there exists a dimension where the currently active coordinate exceeds the boundary of the corresponding semantically stable region for three consecutive rounds, then this dimension is marked as the drift dimension where semantic anchor point drift has occurred, and a drift detection signal is generated. Specifically, this includes:

[0091] S4.1: Obtain the six-dimensional sub-vectors generated in the current interaction round as the activation coordinates to be detected, and call the historical anchor point coordinate sequence stored in the sliding time window to calculate the upper tenths threshold and lower tenths threshold corresponding to each dimension, and generate dynamic envelope model data containing upper and lower boundaries.

[0092] The six-dimensional initial feature vector generated in step S1 for the current interaction round is obtained and used as the activation coordinates to be detected, inputting them into the drift monitoring module. Simultaneously, the historical anchor point coordinate sequence of past interaction rounds stored in the sliding time window cache is read from system memory. This historical anchor point coordinate sequence contains time-series data in six dimensions: time, space, person, context, emotion, and style. Independent data cleaning and integrity checks are performed on each of the six dimensions of the historical anchor point coordinate sequence, removing null values ​​or outliers caused by network latency or parsing errors, ensuring that the time-series data for each dimension is of consistent length and valid values, forming a standardized raw time-series data set.

[0093] For the original time-series data sets of each dimension, quicksort is used to sort them in ascending order according to their numerical values, generating an ordered queue of historical anchor coordinates. This queue eliminates timestamp dependence and only retains the distribution characteristics of the values, providing a basic data structure for subsequent non-parametric statistical calculations. Based on the ordered queue of historical anchor coordinates, linear interpolation is applied to calculate the coordinates of the lower quantile with a cumulative probability of 0.1 and the upper quantile with a cumulative probability of 0.9. The specific calculation formula is as follows:

[0094]

[0095] in, The threshold value for the next tenth place is [value]. The total number of samples within the sliding window. Let i be the value of the i-th element in the sorted queue. This indicates the floor function.

[0096] Similarly, the threshold for the tenths place is calculated using the following formula:

[0097]

[0098] in, This is the threshold value for the tenth percentile; the meanings of the other parameters are the same as above.

[0099] The calculated lower tenths threshold is combined with the upper tenths threshold to construct a dynamic envelope model data containing upper and lower boundaries. This data defines the stable region range of the current dimension in the semantic space, which is used to determine whether the current activated coordinate has exceeded the limit.

[0100] By using the quantile trajectory modeling method described above, the historical anchor point coordinate sequence of the previous step is transformed into statistically significant dynamic boundary threshold data, which realizes the quantitative definition of the semantic fluctuation tolerance range of each dimension of the six-dimensional label system, and provides a robust judgment benchmark for subsequent drift detection.

[0101] S4.2: Based on the upper tenths threshold and lower tenths threshold in the dynamic envelope model data, perform boundary exceedance judgment processing on the values ​​of each dimension in the active coordinate to be detected. If the value is greater than the upper tenths threshold or less than the lower tenths threshold, generate a single exceedance status flag for that dimension.

[0102] The system receives the dynamic envelope model data generated in the preceding step S4.1. This data includes the upper and lower decimal thresholds for each dimension in the six-dimensional label system. Simultaneously, it obtains the six-dimensional sub-vectors after decoupling processing for the current interaction round as the activation coordinates to be detected. It iterates through each dimension index in the six-dimensional label system, extracting the scalar value or feature vector magnitude of the current dimension from the activation coordinates to be detected, using this as the current semantic state observation. It then calls the boundary parameters of the dynamic envelope model corresponding to that specific dimension, reads the pre-calculated and cached upper decimal threshold as the upper bound constraint of the semantically stable region, and reads the lower decimal threshold as the lower bound constraint of the semantically stable region.

[0103] The boundary violation judgment logic compares the current semantic state observation value with the upper bound constraint. If the observation value is strictly greater than the upper tenths threshold, the current dimension is determined to have a positive semantic drift trend. The boundary violation judgment logic compares the current semantic state observation value with the lower bound constraint. If the observation value is strictly less than the lower tenths threshold, the current dimension is determined to have a negative semantic drift trend. If the current semantic state observation value is between the lower and upper tenths thresholds (including boundary values), the current dimension is determined to be within the semantically stable region and has not experienced significant drift. Based on the above judgment results, a Boolean-type single-violation status flag is generated for the current dimension. When a positive or negative semantic drift trend is determined, the flag is set to the true value 1; otherwise, it is set to the false value 0. Using bitmasks or array mapping, the single-violation status flags of the six dimensions are integrated into a six-dimensional violation status vector to ensure that the judgment results for each dimension are independent and traceable.

[0104] By using boundary comparison and logical mapping, continuous semantic coordinate values ​​are transformed into discrete single-time out-of-bounds status flags, enabling a rapid binary representation of whether user intent deviates from the historical stable distribution within a single interaction round, and providing standardized input data for subsequent temporal continuity verification.

[0105] For example, taking the emotion dimension as an example, assume that the system's preset emotion dimension feature space is a one-dimensional scalar space with a value range of [-1, 1]. In the current T-th round of interaction, the user inputs "I feel somewhat depressed." After the encoder parses the data, the current activation coordinate value of the emotion dimension is -0.85. The system calls the historical anchor point coordinate sequence of the past N=20 rounds of interaction within the sliding time window. Through step S3, the lower decimal threshold of this dimension is calculated to be -0.60, and the upper decimal threshold is 0.75. The system compares the current activation coordinate -0.85 with the lower bound -0.60. Since -0.85 < -0.60, the judgment condition is met, and a single-time limit violation status flag of the emotion dimension is generated as 1. At the same time, the current activation coordinate of the time dimension is 0.2, and its corresponding lower decimal threshold is -0.1, and the upper decimal threshold is 0.9. Since -0.1 ≤ 0.2 ≤ 0.9, it is determined that the time dimension has not violated the limit, and a single-time limit violation status flag is generated as 0. Similarly, the same judgment is performed on the remaining four dimensions: space, character, context, and style. The final generated six-dimensional out-of-bounds state vector is [0,0,0,0,1,0] (assuming only the emotion dimension is out of bounds). This process is completed in milliseconds, effectively isolating single-round noise interference, marking only dimensions that significantly deviate from the historical distribution, avoiding misjudgments caused by slight semantic fluctuations, and ensuring the accuracy of subsequent drift detection.

[0106] S4.3: Read the sequence of single out-of-limit status flags generated during three consecutive rounds of interaction, perform a sequential continuous logic AND operation on the single out-of-limit status flags of the same dimension, and if any dimension shows an out-of-limit state in three consecutive rounds, generate a continuous drift confirmation signal for that dimension.

[0107] The system reads the single-out-of-limit status flag sequences generated in the current interaction round and the previous two rounds of historical interactions, constructs a temporal logic judgment window for each dimension of the six-dimensional tag system, and obtains boundary out-of-limit status data for three consecutive rounds. The single-out-of-limit status flags within the temporal logic judgment window are subjected to dimensional isolation processing, extracting the flag sequences for the six dimensions of time, space, person, context, emotion, and style into independent Boolean vector sequences to ensure the orthogonality and independence of semantic drift judgments for each dimension. For any specific dimension's Boolean vector sequence, a temporal continuity logic AND operation is performed. Through bitwise operations or logic gate circuit simulation, it is determined whether that dimension is in an out-of-limit state in the current round t, the previous round t-1, and the two previous rounds t-2.

[0108] If the single out-of-limit status flag for a certain dimension is true in all three consecutive rounds of interaction, the semantic coordinate deviation of that dimension is determined to be persistent and trend-based, rather than instantaneous noise interference, thus generating a continuous drift confirmation signal for that dimension. If the single out-of-limit status flag for any dimension is false in any of the three consecutive rounds, the semantic fluctuation of that dimension is determined to be within the range of random disturbance or has returned to a stable region, and no continuous drift confirmation signal is generated, maintaining the normal mapping state of that dimension. All dimensions of the six-dimensional label system are traversed, and the indices of all dimensions that generate continuous drift confirmation signals are summarized to form a set of drifting dimensions. Dimensions that do not generate continuous drift confirmation signals are marked as stable dimensions.

[0109] Through the above-mentioned sequential continuity logic and operation processing, the discrete single-time out-of-limit state flag is transformed into a continuous drift confirmation signal that represents the semantic evolution trend. This achieves robust detection of semantic anchor point drift, effectively suppresses misjudgment caused by accidental deviations in user input, and improves the accuracy and stability of dynamic spatial calibration.

[0110] S4.4: Based on the continuous drift confirmation signal, mark the label instance of the corresponding dimension as the drift dimension in which the semantic anchor point drift has occurred, freeze the coordinate state of the other five dimensions that have not received the continuous drift confirmation signal, and generate a drift detection signal containing the drift dimension index.

[0111] The system receives a persistent drift confirmation signal generated in step S4.3. This signal contains a set of specific dimension indices identifying the semantic anchor point drift and the current active coordinate state of the corresponding dimension. The system parses the dimension index information in the persistent drift confirmation signal, traverses the preset dimension list of the six-dimensional label system, marks dimensions with matching indices as drift dimensions to be calibrated, and marks the remaining five dimensions with unmatched indices as state-maintained dimensions. For the five dimensions marked as state-maintained, the system extracts the original dimension sub-vector coordinates corresponding to them in the current interaction round, performs a read-only locking operation to prevent subsequent calibration modules from making any numerical modifications or weight adjustments to this part of the coordinate data, and generates a frozen coordinate set. For the target dimension marked as a drift dimension, the system extracts the deviation direction vector of its current active coordinates relative to the boundary of the semantically stable region, and constructs a structured drift detection data package containing the drift dimension index, deviation direction identifier, and current timestamp. The frozen coordinate set and the structured drift detection data package are encapsulated and integrated to form a standard drift detection signal object, which clearly distinguishes between the reference dimensions that need to be kept static and the drift dimensions that need to be dynamically corrected.

[0112] By using the aforementioned dimensional state separation and signal encapsulation processing methods, the drift determination result of the previous step is transformed into drift detection signal data with clear operation instructions. This achieves precise freezing and protection of the semantic state of the non-drifted dimension and accurate positioning of the calibration trigger of the drifted dimension, ensuring that subsequent reprojection operations only apply to the abnormal dimension, avoiding unnecessary disturbances in the global semantic space, and significantly improving the locality and stability of the system response.

[0113] Step S5: In response to the drift detection signal, freeze the coordinates of the five dimensions that have not drifted, perform a direction-constrained reprojection operation only on the current sub-vector of the drifted dimension, apply a semantic restoring force that is non-linearly proportional to the deviation distance along the opposite direction of the deviation, and generate calibrated dimension coordinates. Specifically, this includes:

[0114] S5.1: Obtain the drift detection signal and the six-dimensional initial dimension feature vector of the current round. Based on the drift dimension index in the drift detection signal, perform dimension masking on the six-dimensional initial dimension feature vector to generate a frozen coordinate set containing the original coordinates of five non-drift dimensions and a drift dimension sub-vector to be calibrated.

[0115] S5.2: Extract the coordinates of the drift dimension sub-vector to be calibrated and the semantically stable region boundary of its corresponding dimension, and calculate the deviation distance vector of the drift dimension sub-vector to be calibrated relative to the semantically stable region boundary based on Euclidean distance, so as to generate a deviation distance scalar representing the semantic drift amplitude and a unit direction vector representing the drift direction.

[0116] The system receives the drift dimension index specified in the drift detection signal, extracts the corresponding drift dimension sub-vector to be calibrated from the frozen coordinate set, and simultaneously reads the current upper and lower decimal thresholds for that dimension from the dynamic envelope model data to establish the boundary constraints of the semantically stable region. It determines the deviation direction of the value of the drift dimension sub-vector to be calibrated relative to the boundary of the semantically stable region. If the sub-vector value is greater than the upper decimal threshold, it is determined that the upper boundary has been exceeded, and the upper decimal threshold is determined as the nearest boundary coordinate; if the sub-vector value is less than the lower decimal threshold, it is determined that the lower boundary has been exceeded, and the lower decimal threshold is determined as the nearest boundary coordinate. Based on the determined nearest boundary coordinates and the drift dimension sub-vector to be calibrated, a one-dimensional Euclidean distance calculation model is constructed, and the absolute difference between the two is calculated to generate a deviation distance scalar representing the semantic drift amplitude. This scalar quantifies the degree to which the current user intent deviates from the historical stable distribution.

[0117] A direction vector pointing into the semantically stable region is constructed. If the upper bound is exceeded, the unit direction vector is defined as negative, indicating that regression towards the direction of decreasing value is required; if the lower bound is exceeded, the unit direction vector is defined as positive, indicating that regression towards the direction of increasing value is required, ensuring that the correction operation is always performed along the convergence path. The generated unit direction vector is normalized to ensure that its magnitude is 1, eliminating the influence of dimensions and forming a standardized drift direction indicator for subsequent vector synthesis of semantic restorative power.

[0118] Through the above processing method, the drift detection results of the previous step are transformed into a deviation distance scalar representing the drift amplitude and a unit direction vector representing the correction path, thereby realizing the accurate quantification of the geometric features of semantic drift and providing basic physical parameters for the subsequent application of nonlinear semantic restoring force.

[0119] For example, in the semantic mapping scenario of lighting design style dimension, the system detected a drift in the style dimension sub-vector corresponding to the "minimalism" label. In the current interaction round, the value of the style dimension sub-vector is 0.85, while the dynamic envelope model built from previous rounds of interaction shows that the upper decimal threshold for this dimension is 0.75, and the lower decimal threshold is 0.25. Since 0.85 is greater than 0.75, the system determines that the upper bound has been exceeded and selects 0.75 as the nearest boundary coordinate. Using Euclidean distance, the deviation distance scalar is calculated, i.e., |0.85-0.75| = 0.10. This value represents the extent to which the current design style preference exceeds the conventional minimalist range. At the same time, because it belongs to the upper bound exceeding the limit, the system constructs a unit direction vector pointing in the direction of decreasing value, denoted as -1.0, indicating that the correction operation needs to pull the semantic coordinates back to the more conservative minimalist style region. This process transforms the abstract semantic offset into a specific 0.10 offset distance and -1.0 direction vector, ensuring that the subsequent applied semantic restoring force can accurately guide the design intent back into a stable semantic envelope, avoiding the matching results from deviating from the user's core needs due to excessive pursuit of extreme styles.

[0120] S5.3: Receive the deviation distance scalar, and perform nonlinear transformation processing on the deviation distance scalar using a preset nonlinear hyperbolic tangent mapping function to generate semantic restoring force that is adaptively enhanced with the degree of deviation, ensuring that it provides weak correction when drifting at a small magnitude and strong pullback when drifting at a large magnitude.

[0121] The semantic restoring force refers to the correction coefficient used to pull the drifted dimensional subvector back into the semantically stable region, and its magnitude adapts non-linearly with the deviation distance. It is obtained as follows:

[0122] The system receives the deviation distance scalar, representing the semantic drift amplitude, output from step S5.2. This scalar is used as the input variable of the nonlinear mapping function to initiate the dynamic calculation process of semantic restoring force. A hyperbolic tangent activation function model, pre-configured in system memory, is invoked. This model is configured as a nonlinear transformer with a specific saturation threshold to map linear spatial distances to a nonlinear range of mechanical coefficients. Normalization preprocessing is performed on the input deviation distance scalar, dividing it by a preset maximum allowable drift threshold to generate a relative drift ratio between 0 and 1, thus eliminating the influence of dimensional differences between different dimensions on the restoring force calculation. The relative drift ratio is substituted into the hyperbolic tangent function expression, and the basic restoring force coefficient is calculated using the following formula:

[0123]

[0124] Where k represents the generated semantic restorative power, and tanh is the hyperbolic tangent function. As a sensitivity adjustment factor, This is the relative drift ratio after normalization.

[0125] A sensitivity adjustment factor α is introduced, which is statically set according to the semantic rigidity requirements of the lighting design scenario. This factor controls the steepness of the increase in restoring force with drift distance, ensuring that the curve is smooth during small drifts and rapidly approaches the saturation value of 1 during large drifts. The calculated basic restoring force coefficient k is truncated, forcibly limiting its value to the closed interval [0,1] to prevent coefficient overflow or negative values ​​caused by extreme abnormal inputs, thus ensuring the determinism of the semantic correction direction. The truncated semantic restoring force is multiplied by the preset maximum correction step size parameter to generate the final semantic restoring force acting on the drift dimension sub-vector to be calibrated. This force value directly determines the displacement amplitude of the subsequent reprojection operation.

[0126] By using the above-mentioned nonlinear hyperbolic tangent mapping processing method, the linear deviation distance scalar of the previous step is transformed into a semantic restoring force with adaptive characteristics. This achieves the technical effect of providing weak correction to preserve subtle changes in user intent during small semantic drifts and providing strong pullback to suppress semantic divergence during large semantic drifts.

[0127] S5.4: Perform a weighted product operation between the reverse vector of the unit direction vector and the semantic restoring force to generate a semantic correction increment, and superimpose the semantic correction increment onto the drift dimension sub-vector to be calibrated to generate calibrated dimension coordinates that fall within the semantically stable region.

[0128] The system receives the drift dimension unit direction vector, semantic restoring force, and drift dimension sub-vector to be calibrated generated in the previous steps as input data for constructing the orientation-constrained reprojection operator. Based on the unit direction vector, a vector inversion operation is performed to generate a reverse correction direction vector pointing into the semantically stable region. This vector establishes the geometric path to pull the drift coordinates back to the stable region. The scalar and vector multiplication operator is invoked to perform element-wise multiplication of the semantic restoring force and the reverse correction direction vector to generate a semantic correction increment vector representing the correction magnitude and direction. The magnitude of the semantic correction increment vector is determined by the semantic restoring force, and its direction is strictly limited by the reverse correction direction vector, ensuring that the correction operation is performed only in the reverse direction of the deviation trajectory, avoiding the introduction of noise interference in orthogonal directions. Vector addition is used to superimpose the semantic correction increment vector onto the drift dimension sub-vector to be calibrated, and a spatial coordinate translation operation is performed to generate preliminary calibrated dimension coordinate candidate values.

[0129] Boundary constraint checks are performed on the initially calibrated candidate dimensional coordinates to determine whether they fall entirely within the semantically stable region defined by the upper and lower decimal thresholds. If the initially calibrated coordinates are still outside the stable region, the semantic restorative power gain is iteratively increased, the semantic correction increment is recalculated and superimposed, until the coordinates fall within the stable region boundary or the maximum iteration limit is reached. If the initially calibrated coordinates have fallen within the stable region, these coordinates are directly determined as the final calibrated dimensional coordinates, preserving the reasonable semantic extension in the user's intent and avoiding intent distortion caused by over-correction.

[0130] S5.5: Integrate the original coordinates of the five undrifted dimensions in the frozen coordinate set with the calibrated dimension coordinates, and perform vector recombination operation according to the predefined dimension order of the six-dimensional label system to generate a six-dimensional comprehensive feature vector for subsequent multi-dimensional spatial similarity matching calculation.

[0131] The predefined dimension order refers to a statically defined, fixed arrangement order for the six semantic dimensions during system initialization or construction of the six-dimensional tag system. This predefined dimension order is determined during the system design phase based on business logic or technical specifications and is persistently stored as a system configuration parameter. This ensures that the positions of each dimension sub-vector within the long vector remain strictly consistent during the generation, storage, matching, and recombination of the six-dimensional comprehensive feature vector, thereby guaranteeing the correctness and comparability of vector operations such as cosine similarity calculation.

[0132] Step S6: Based on the calibrated dimensional coordinates and the original coordinates of the un-drifted dimension, construct a six-dimensional comprehensive feature vector under a hierarchical weighting strategy, wherein a decay weight coefficient dynamically adjusted with the calibration intensity is superimposed on the calibrated dimensional coordinates. Specifically, this includes:

[0133] S6.1: Obtain the current sub-vector coordinates of the drift dimension and the calibrated dimension coordinates generated by the direction-constrained reprojection operation, and calculate the spatial deviation distance between the two using Euclidean distance to generate vector displacement magnitude data representing the semantic correction magnitude.

[0134] S6.2: Based on the vector displacement modulus data, a preset nonlinear mapping function is called to perform attenuation coefficient derivation processing, converting the spatial deviation distance into a dynamic attenuation weight coefficient in the range of zero to one, so as to generate a weight scalar parameter characterizing the degree of confidence loss.

[0135] The vector displacement magnitude data output from step S6.1 is received. This data characterizes the spatial position correction magnitude of the drift dimension during calibration and serves as the basic input for confidence assessment. A pre-defined nonlinear mapping function library is invoked, selecting a variant of the sigmoid activation function or hyperbolic tangent function with smooth transition characteristics and a value range strictly limited to the zero-to-one interval as the core operator for deriving the attenuation coefficient, ensuring the numerical stability and physical interpretability of the output weights. The vector displacement magnitude data is substituted into the independent variable position of the nonlinear mapping function, and a combination of exponential and division operations is performed. The steepness of the attenuation curve is controlled by adjusting the slope parameter of the function, so that small displacements generate high weights close to 1, while large displacements generate low weights close to 0. Specifically, the following mathematical model is used to calculate the dynamic attenuation weight coefficients:

[0136]

[0137] in, For the generated dynamic decay weighting coefficients, For the input vector displacement magnitude data, To control the gain coefficient of attenuation sensitivity, As a reference point for triggering a significant attenuation of displacement, It is a natural constant.

[0138] The calculated initial weight values ​​are truncated. If the calculation result slightly exceeds the [0,1] interval due to floating-point precision error, it is forcibly clamped to the endpoint of the interval to ensure the validity of the weight coefficients and prevent numerical overflow or sign reversal anomalies in subsequent weighting operations. The finally determined dynamic decay weight coefficients are encapsulated as scalar parameters. These parameters directly reflect the degree of confidence loss of the current calibrated dimension coordinates relative to the original user intent. The larger the displacement, the lower the weight, indicating that the semantic matching reliability of that dimension is weaker.

[0139] S6.3: Receive the calibrated dimension coordinates and the weight scalar parameters, perform scalar multiplication to perform weighted compression processing on the calibrated dimension coordinates, so as to generate a drift dimension weighted feature vector after confidence correction.

[0140] The system receives the calibrated dimensional coordinate vector generated by the direction-constrained reprojection operation. This vector represents the new spatial position of the drift dimension after the semantic restorative force is applied. Simultaneously, it acquires the dynamic decay weight coefficient scalar calculated from the previous steps, which quantifies the degree of confidence loss caused by semantic drift correction. A weighted compression operator for the drift dimension is constructed, performing element-wise multiplication of each component element in the calibrated dimensional coordinate vector with the dynamic decay weight coefficient scalar. Through a linear transformation mechanism of the vector space by the scalar field, the overall contraction of the feature magnitude of this dimension is achieved. During the operation, if the dynamic decay weight coefficient scalar is less than 1, the magnitude of the calibrated dimensional coordinate vector is correspondingly reduced. This reduction is not a simple numerical decrease, but rather symbolizes a decrease in the certainty of this dimension in the current user intent expression, thereby reducing its contribution weight in subsequent multidimensional spatial similarity matching.

[0141] Specifically, the confidence-corrected drift-dimensional weighted feature vector is calculated using the following formula:

[0142]

[0143] in, This is a feature vector weighted by the drift dimension after confidence correction. The dynamic decay weighting coefficient is between 0 and 1. For the calibrated dimensional coordinate vector, the sign is... This represents the multiplication operation between a scalar and a vector.

[0144] This operation ensures that when the semantic drift is large, resulting in a high calibration intensity, the weight coefficient is significantly reduced, thereby greatly compressing the influence of the feature vector in that dimension and preventing noise introduced by over-calibration from interfering with the overall matching result. Conversely, when the drift is small, the weight coefficient is close to 1, preserving most of the original semantic information. A drift-dimensional weighted feature vector with confidence correction is generated. While maintaining the original semantic direction, the energy distribution of this vector has been adaptively adjusted according to semantic stability, realizing the technical transformation from geometric space calibration to semantic confidence mapping.

[0145] S6.4: Extract the original dimension sub-vectors corresponding to the five dimensions that have not drifted as the original feature vectors of the undrifted dimensions, and perform dimension concatenation processing on the weighted feature vectors of the drifted dimensions that have been corrected by confidence and the original feature vectors of the undrifted dimensions to generate a preliminary fused six-dimensional heterogeneous feature sequence.

[0146] The system receives the confidence-corrected drift dimension weighted feature vector from the output of previous step S6.3. This vector represents the semantic state of a specific dimension after semantic restorative calibration and its confidence decay information. From the dimension index mapping table of the six-dimensional label system, the system retrieves the identifiers of the remaining five dimensions not marked as semantic anchor drift in the current interaction round, constructing a set of non-drift dimension indexes. Based on this set of non-drift dimension indexes, the system extracts the corresponding five original non-drift dimension sub-vectors from the frozen coordinate set generated in step S5, keeping their original high-dimensional space coordinate values ​​unchanged, forming a set of original non-drift dimension feature vectors. A dimension consistency check is performed on the set of original non-drift dimension feature vectors to ensure that the dimension length of each sub-vector is consistent with the dimension length of the drift dimension weighted feature vector. If a dimension mismatch exists, zero-padding or truncation preprocessing is triggered to align the feature space.

[0147] Based on the predefined physical order of the six-dimensional tagging system—time, space, person, context, emotion, and style—a standard splicing template for six-dimensional feature vectors is established, determining the fixed position index of each dimension in the final composite vector. The confidence-corrected drift dimension weighted feature vector is filled into the corresponding drift dimension position index in the standard splicing template. The five sub-vectors from the original feature vector group of the non-drift dimensions are respectively filled into the corresponding non-drift dimension position index in the standard splicing template. Vector splicing is performed, concatenating the six-dimensional sub-vectors end-to-end according to the order of the standard splicing template to construct a high-dimensional preliminary fused six-dimensional heterogeneous feature sequence. The preliminary fused six-dimensional heterogeneous feature sequence is encapsulated into a data structure, adding session ID, timestamp, and drift dimension marker metadata to generate a standardized intermediate data object for subsequent normalization processing.

[0148] S6.5: Perform normalization and standardization processing on the initially fused six-dimensional heterogeneous feature sequence to eliminate the differences in the dimensions of each dimension and unify the feature distribution range, so as to generate the final six-dimensional comprehensive feature vector under the hierarchical weighting strategy for subsequent similarity calculation.

[0149] Step S7: Perform cosine similarity calculation on the six-dimensional comprehensive feature vector, and output the matching results of lighting fixture materials sorted by similarity, along with the corresponding structural feasibility verification data and BOM cost estimation data. As a preferred embodiment, the system outputs the top 20 lighting fixture material matching results with the highest similarity. Those skilled in the art will understand that the aforementioned number 20 can be configured to other values ​​according to actual needs. Specifically, this includes:

[0150] S7.1: Obtain the calibrated dimension coordinates generated by the directional constraint reprojection operation and the original coordinates of the dimension that has not drifted. Apply a decay weight coefficient that is dynamically adjusted with the calibration intensity to the calibrated dimension coordinates based on a hierarchical weighting strategy to construct a six-dimensional comprehensive feature vector with semantic stability to eliminate the interference of semantic anchor drift on the subsequent retrieval accuracy.

[0151] The system receives a six-dimensional comprehensive feature vector from the output of step S6, which is based on a hierarchical weighting strategy. This vector contains the original coordinate components of five non-drift dimensions and a drift dimension component that has been calibrated by directional constraint reprojection and subjected to dynamic attenuation weights. Semantic consistency verification is performed on the six-dimensional comprehensive feature vector. The weight flags of each dimension component in the vector are read to confirm whether the attenuation weight coefficients corresponding to the drift dimension have been correctly embedded into the feature amplitude, ensuring that the semantic contribution of this dimension is suppressed by confidence loss during subsequent retrieval. The verified six-dimensional comprehensive feature vector is encapsulated into a standardized retrieval request object, where each dimension component carries its corresponding semantic label category index and normalized numerical coordinates, forming a query anchor with complete semantic context information.

[0152] The feature indexing engine of the lighting database is invoked to load a pre-stored high-dimensional feature matrix of lighting materials. This matrix consists of six-dimensional label feature vectors representing time, space, person, context, emotion, and style for all lighting instances in the database. Each dimension has been normalized using the same L2 norm as the query vector to eliminate dimensional differences. Based on the query anchor and the lighting material feature matrix, a hierarchical weighted cosine similarity calculation process is initiated. For non-drifted dimensions, their original coordinates are directly extracted for dot product operations. For drifted dimensions, coordinates compressed with attenuated weights are extracted for dot product operations, thus mathematically implementing a penalty mechanism for unstable semantic dimensions.

[0153] S7.2: Read the six-dimensional comprehensive feature vector and the pre-stored feature matrix of lighting materials in the lighting material database containing six-dimensional tag attributes of time, space, people, context, emotion and style. Use the improved cosine similarity calculation to perform high-dimensional space vector angle operation to generate a multi-dimensional space similarity score sequence that represents the degree of semantic association between user intent and each lighting material.

[0154] The system reads a six-dimensional comprehensive feature vector constructed using a hierarchical weighting strategy, which serves as the input query vector for the user's current intent in the standardized semantic space. It then calls upon a pre-stored feature matrix from the lighting database, containing standardized label feature vectors for all lighting instances in the database across six dimensions: time, space, person, context, emotion, and style. A dimension-wise dot product operation is performed between the query vector and each lighting material feature vector in the database to calculate the local semantic similarity components in their respective independent subspaces. For dimensions where semantic anchor point drift has occurred and been calibrated, a dynamic decay weight coefficient is introduced to weight and correct the local similarity components, suppressing the uncertainty in confidence caused by semantic reprojection. The weighted local similarity components across the six dimensions are summed to generate an original similarity score representing the degree of global semantic association between the user's intent and a specific lighting material. An improved cosine similarity processing method is used to normalize the vector magnitude of the original similarity score, eliminating the interference of differences in the length of feature vectors from different lighting materials on the matching results. The multidimensional similarity score between the i-th lighting material and the user's query vector is calculated using the following formula:

[0155]

[0156] Where S is the final multidimensional spatial similarity score; k is the dimension index, ranging from 1 to 6, corresponding to the time, space, person, context, emotion, and style dimensions respectively; w k For the k-th dimension, the dynamic weight coefficients are given by the unshifted dimension w. k Set to 1, for the drift dimension w k The value is the dynamic decay weight coefficient generated in the previous step; q k d represents the feature component of the user query vector in the k-th dimension; i,k Let ||Q||2 be the feature component of the i-th lighting fixture material in the k-th dimension; ||Q||2 be the L2 norm of the user query vector; ||D i ||2 represents the L2 norm of the feature vector of the i-th lighting fixture material. The similarity calculation process is repeated across all material instances in the lighting fixture database, generating a multidimensional spatial similarity score sequence containing all candidate lighting fixture material identifiers and their corresponding similarity scores. Through an improved cosine similarity calculation and dynamic weight correction method, the hierarchical weighted six-dimensional comprehensive feature vector constructed in the previous step is transformed into a multidimensional spatial similarity score sequence representing the semantic association between user intent and each lighting fixture material, achieving the expected high-efficiency matching effect that balances semantic stability and retrieval accuracy.

[0157] S7.3: Receive the multidimensional spatial similarity score sequence, perform fast sorting processing according to the score values ​​from high to low, extract the top 20 lighting material identifiers and their corresponding six-dimensional tag attribute data from the sorting results, and generate an initial Top 20 lighting material matching result set to meet the business needs of efficient retrieval and accurate recommendation.

[0158] S7.4: Call the geometric structure parameters and installation environment constraints of each lamp material in the initial Top20 lamp material matching result set, start the structural feasibility verification engine to perform collision detection and mechanical support analysis, and generate structural feasibility verification data containing installation compatibility judgment conclusions and potential structural risk warning information to ensure the physical feasibility of the design scheme.

[0159] The system receives an initial set of Top 20 lighting fixture material matching results, parses the 3D geometric model data, installation interface type parameters, and preset physical constraints of each lighting fixture material, and constructs a set of lighting fixture physical attribute objects to be verified. For each lighting fixture material in the set of lighting fixture physical attribute objects, its bounding box size data, centroid coordinate vector, and mounting point normal direction vector are extracted as the basic geometric input parameters for structural feasibility verification. The system obtains the installation environment spatial model implicitly or explicitly specified in the current user interaction context. This spatial model includes ceiling / wall material properties, load-bearing limit thresholds, reserved pipeline location coordinates, and obstacle geometric distribution information.

[0160] The mounting point coordinates of the luminaire material are mapped to the local coordinate system of the installation environment space model. Rigid body transformation matrix operations are performed to calculate the absolute spatial pose of the luminaire in its installation posture, generating a luminaire-environment coupled spatial state matrix. Based on this luminaire-environment coupled spatial state matrix, collision detection based on the Separating Axis Theorem (SAT) is initiated, traversing all potential contact surfaces between the luminaire geometry and environmental obstacles, and calculating the minimum penetration depth vector.

[0161] If the minimum penetration depth vector is detected to be less than zero, geometric interference is determined, the volume percentage of the interference region and the interference depth value are recorded, and a geometric conflict marker is generated. If the minimum penetration depth in all directions is greater than or equal to zero, geometric spatial compatibility is determined. For the luminaire material that passes the geometric compatibility determination, a set of static equilibrium equations is constructed, incorporating the luminaire's self-weight vector, the reaction force vector at the mounting point, and the support force vector provided by environmental constraints into the force analysis model.

[0162] A simplified finite element method was used to calculate the stress distribution at the mounting point under the gravitational moment of the lamp. Based on the yield strength criterion of material mechanics, the structural safety of the installation interface was assessed. A structural safety factor was defined, and the structural risk index of the current installation scheme was calculated using the following formula:

[0163]

[0164] Where R is the structural risk index, F load Let L be the equivalent load weight of the lamp, L be the lever arm length, and K be the weight of the lamp. d σ is the dynamic impact coefficient. yield The installation substrate yield strength is given by A, the effective cross-sectional area under stress is given by S, and the preset safety redundancy factor is given by S.

[0165] When the calculated structural risk index R is greater than 1, the installation scheme is deemed to have a structural failure risk, and a high-risk warning signal is generated; when R is less than or equal to 1 but greater than 0.8, a low-risk warning signal is generated; when R is less than or equal to 0.8, the structure is deemed safe. Geometric conflict markers, interference depth values, structural risk indices, and risk level signals are integrated to form a structured verification data package containing installation compatibility assessment conclusions and potential structural risk warning information.

[0166] S7.5: Integrate the basic information of the bill of materials and the real-time market price database from the initial Top 20 lighting fixture material matching result set, run the BOM cost estimation model to perform multi-level material cost accumulation and labor cost conversion, generate BOM cost estimation data containing total cost estimate and itemized cost composition, and integrate the structural feasibility verification data to output the final design scheme matching result package.

[0167] The initial Top 20 lighting fixture material matching result set is obtained, and the basic information of the bill of materials corresponding to each lighting fixture material is extracted. The basic information of the bill of materials includes the standardized component codes and quantity configurations of the light source module, driver power supply, lamp body structural components, optical lenses, and connecting accessories. Based on the standardized component codes, a multi-source data retrieval is performed by calling the real-time market price database to obtain the latest purchase price, inventory status, and supplier delivery cycle of each component at the current timestamp, and to construct a dynamic cost parameter set including price fluctuation factors. For the retrieved price data, outlier filtering and missing value imputation are performed, and linear interpolation is used to fill in the blank fields caused by supplier quotation delays, ensuring the integrity and timeliness of the cost calculation benchmark data.

[0168] A multi-level material cost accumulation model is established, decomposing lighting fixture materials into three levels: direct material cost, indirect manufacturing overhead, and assembly labor cost. For direct material cost, a weighted average method is used to calculate the comprehensive unit price of the same component from different suppliers. Combined with the component quantity configuration, the total direct material cost of a single lighting fixture material is calculated using the following formula:

[0169]

[0170] Among them, C materialLet P be the total direct material cost, N be the total number of component types, and P be the total cost of materials. i Let Q be the comprehensive weighted unit price of the i-th component. i Let represent the number of configurations for the i-th component. This calculation process iterates through each lighting scheme in the Top 20 result set, generating its respective direct material cost vector.

[0171] A time cost conversion mechanism is introduced to calculate indirect manufacturing costs based on the structural complexity coefficient of the luminaire and the preset standard assembly time quota. The structural complexity coefficient is determined by the total number of luminaire parts, the proportion of irregularly shaped parts, and the number of electrical connection nodes, and is normalized to a range of 0.5 to 1.5. The assembly time cost of a single luminaire is calculated using the following formula:

[0172]

[0173] Among them, C labor For assembly time costs, T base K is the standard baseline assembly time. complex R is the structural complexity coefficient. hour The unit labor hour rate is used. The total direct material cost is added to the assembly labor cost, and a preset management allocation rate is added to obtain the preliminary BOM cost estimate for each lighting fixture material.

[0174] Integrating the structural feasibility verification data generated in previous steps, the initial BOM cost estimate is adjusted for risk. If the structural feasibility verification data shows installation compatibility warnings or potential structural risks, a cost penalty factor mechanism is triggered, adding additional reinforcement material costs or special processing fees based on the risk level. A logic judgment module identifies high-risk lighting fixture solutions, applying a non-linear premium coefficient to their BOM cost estimates to generate final BOM cost estimates that include a risk reserve. This process ensures that the cost estimate reflects not only the static material value but also potential hidden costs during implementation.

[0175] By integrating the final BOM cost estimate data, structural feasibility verification conclusions, and attribute tags of the initial Top 20 lighting fixture materials, a structured design scheme matching result package is constructed. This result package is then serialized and encapsulated, sorted in descending order of similarity score, and each scheme is labeled with its estimated total cost, cost breakdown ratio, and structural implementation risk level. Through this multi-level cost accumulation and risk correction process, the preliminary matching results from the previous step are transformed into comprehensive design scheme data with engineering feasibility reference value. This achieves a closed-loop process from semantic matching to engineering economic feasibility assessment, significantly improving the feasibility and commercial conversion rate of the recommended schemes.

[0176] Step S8: After completing ten rounds of effective interaction, extract new anchor point samples that remain within the semantically stable region after calibration in this session, and incrementally and smoothly update the dynamic envelope model of each dimension to generate the updated semantically stable region to complete the closed-loop optimization. Specifically, this includes:

[0177] S8.1: Obtain the historical anchor point coordinate sequence and corresponding drift detection signal record of the ten rounds of interaction completed in this session, use the session state filter to filter the historical anchor point coordinate sequence, remove abnormal coordinate data marked as semantic anchor point drift, and generate a candidate anchor point set containing only the original coordinates of the non-drifted dimension and the calibrated coordinates of the calibrated dimension.

[0178] S8.2: Based on the candidate anchor point set and the upper and lower quantile boundary data of the current dynamic envelope model in each dimension, spatial inclusion determination processing is used to perform region belonging verification on the coordinates of each candidate anchor point in the set, identify coordinate instances that fall completely inside the current semantically stable region, and generate a new anchor point sample set with high confidence to eliminate edge oscillation noise interference.

[0179] S8.3: For the high-confidence new anchor point sample set, the exponential weighted moving average fusion strategy is invoked to perform weighted fusion calculation between the new anchor point coordinates in the sample set and the center trajectory of the previous version of the dynamic envelope model, generating a mixed anchor point distribution sequence containing historical memory weights and the latest interaction features, so as to achieve a smooth transition between old and new semantic knowledge.

[0180] Obtain a high-confidence new anchor point sample set. This set contains coordinate instances of each dimension that fall completely within the current semantically stable region after being filtered by spatial inclusion determination. This set serves as the input data source for the exponentially weighted moving average fusion strategy. Read the center trajectory vector of the previous version of the dynamic envelope model. This vector represents the semantic mean state of the six-dimensional label system in each dimension within the historical interaction period. This vector serves as the benchmark reference point for this incremental update.

[0181] For each coordinate instance in the high-confidence new anchor point sample set, the Euclidean distance between it and the historical center trajectory vector of the corresponding dimension is calculated to quantify the degree of deviation of the new sample from the historical semantic centroid. Based on the degree of deviation, a preset time decay factor and confidence adjustment function are invoked to dynamically assign a fusion weight coefficient to each new anchor point sample, ensuring that recent and stable interaction data dominates the update, while the influence of distant or marginal data is smoothly suppressed.

[0182] An exponentially weighted moving average is used to linearly combine the weighted average of the new anchor point sample set with the center trajectory vector of the previous version, generating a hybrid anchor point distribution sequence that includes historical memory weights and the latest interaction features. Specifically, for the d-th dimension of the label space, its updated center trajectory vector is calculated using the following formula:

[0183]

[0184] in, The previous version 3D center trajectory vector; For the high-confidence new anchor point samples in the current session, in the 1st... The arithmetic mean over the dimension; This is a smoothing coefficient, with a value between 0 and 1, used to control the proportion of historical information retained; This is the updated center trajectory vector.

[0185] The smoothing coefficient is not a fixed constant, but is dynamically adjusted according to the interaction round t of the current session. The calculation formula is as follows:

[0186]

[0187] This formula ensures that in the early stages of a conversation (when t is small), the system is highly sensitive to new data and can quickly adapt to the initial intent. As the number of interaction rounds increases (t increases), the smoothing coefficient gradually approaches 0, and the system tends to maintain the established semantic stability, avoiding model oscillations caused by a single abnormal input. The above weighted fusion calculation is performed on the six dimensions of time, space, person, context, emotion, and style in the six-dimensional label system to generate updated central trajectory vectors for each dimension. The updated central trajectory vectors for the six dimensions are then reassembled into a complete hybrid anchor point distribution sequence. This sequence retains the semantic skeleton formed by the designer's long-term cognition while incorporating fine-tuning features specific to the current scenario.

[0188] By using an exponentially weighted moving average fusion processing method, high-confidence new anchor point samples and historical center trajectories are transformed into a hybrid anchor point distribution sequence with temporal smoothness characteristics, achieving a seamless transition and gradual evolution of new and old semantic knowledge. This effectively suppresses model jitter caused by semantic drift and improves the long-term robustness of the label mapping system.

[0189] S8.4: Based on the hybrid anchor point distribution sequence, re-execute the quantile trajectory modeling operation, calculate the upper tenths and lower tenths values ​​of the updated sequence respectively, construct a new boundary constraint envelope, and generate updated semantically stable region definition data to reflect the gradual evolution trend of the designer's cognition.

[0190] The system receives a hybrid anchor point distribution sequence generated after processing with an exponentially weighted moving average fusion strategy. This sequence contains weighted coordinate data of historical memory weights and the latest interaction features, serving as the input basis for quantile trajectory modeling calculations. Independent data cleaning and outlier removal operations are performed on each dimensional component of the hybrid anchor point distribution sequence to remove non-finite values ​​(NaN or Inf) caused by numerical overflow or computational errors, ensuring the numerical stability and robustness of subsequent statistical calculations.

[0191] For the cleaned mixed anchor point distribution sequence, an empirical cumulative distribution function (ECDF) is constructed using linear interpolation to map discrete anchor point coordinate samples into continuous probability distribution curves, accurately capturing subtle trends in the evolution of designers' cognition. Based on the constructed ECDF, an inverse cumulative distribution function is applied to locate the coordinate thresholds corresponding to cumulative probability values ​​of 0.1 and 0.9, where the 10th percentile represents the lower boundary of the semantically stable region, and the 0.9 quantile represents the upper boundary of the semantically stable region.

[0192] Specifically, for the d-th dimension label space, let the mixed anchor point distribution sequence be... Where M is the total number of samples, the tenths value is calculated using the following formula. and the value in the tenths place :

[0193]

[0194]

[0195] in, The empirical cumulative distribution function is defined as follows: , This is an indicator function that takes the value 1 when the condition is true and 0 otherwise.

[0196] Using the calculated upper and lower decimal places, a closed dynamic envelope geometry is constructed in a six-dimensional feature space. This geometry is bounded by six upper and lower boundary hyperplanes, forming new semantically stable region definition data. The newly generated semantically stable region definition data is compared with the parameters of the old dynamic envelope model, and the magnitude of boundary contraction or expansion is recorded to assess the directionality and intensity of the designer's cognitive evolution.

[0197] By performing quantile trajectory modeling operations, the mixed anchor point distribution sequence is transformed into new semantically stable region definition data containing upper and lower boundary constraints. This enables adaptive updating of the semantic drift boundary of labels in lighting design scenarios, ensuring the system's accurate capture of the gradual evolution trend of designers' cognition and guaranteeing long-term stability.

[0198] For example, in the design session for "Modern Minimalist Living Room Main Light," after processing in step S8.3, a mixed anchor point distribution sequence containing 50 sample points is obtained for the "style" dimension. Data cleaning is performed on this sequence, removing two NaN values ​​caused by network fluctuations, leaving 48 valid samples. An empirical cumulative distribution function is constructed, calculating the lower quantile coordinates of a cumulative probability of 0.1 to 0.35 and the upper quantile coordinates of a cumulative probability of 0.9 to 0.82. Based on this, a new semantically stable region for the "style" dimension is constructed, with a boundary range of [0.35, 0.82]. Compared to the previous version's boundary range [0.30, 0.85], the new region increases by 0.05 at the lower boundary and shrinks by 0.03 at the upper boundary, reflecting a trend of user intent converging towards a purer minimalist style. The system writes this new boundary data [0.35, 0.82] into the database, replacing the old parameters. In the subsequent 11th round of interaction, if the user inputs "slightly retro elements", the parsed style dimension coordinate is 0.28. This value is lower than the new lower boundary of 0.35. The system will immediately trigger the drift detection mechanism instead of misjudging it as normal fluctuation. This significantly improves the sensitivity of intent recognition and the accuracy of matching results, avoiding recommendation bias caused by semantic boundary lag.

[0199] S8.5: Using the updated semantically stable region definition data, replace the old dynamic envelope model parameters stored in the system to complete the online hot update operation of the semantically stable regions of each dimension of the six-dimensional tag system, and generate a closed-loop optimized tag mapping environment with adaptive capabilities to support the accurate parsing of user needs in the next round.

[0200] In one embodiment, a rapid filtering and matching system for a lighting fixture database based on a six-dimensional tagging system is provided, including: a six-dimensional initial dimension feature vector acquisition module, a current active coordinate acquisition module, a semantically stable region acquisition module, a drift detection signal acquisition module, a calibrated dimension coordinate acquisition module, a six-dimensional comprehensive feature vector acquisition module, and a lighting fixture material matching and comprehensive evaluation module. Specifically,

[0201] The six-dimensional initial dimension feature vector acquisition module is used to acquire the user's natural language requirement text, and use a preset dimension-specific semantic encoder to decouple the user's natural language requirement text to generate a six-dimensional initial dimension feature vector.

[0202] The current activation coordinate acquisition module is used to calculate the cluster center of all tag instances in each dimension based on the six-dimensional initial dimension feature vector, define the cluster center as the initial semantic anchor point of the six-dimensional tag system, and extract the six-dimensional sub-vectors of the current interaction round as the current activation coordinates.

[0203] Semantic stable region acquisition module: For the current active coordinates of each dimension, it calls the historical anchor point coordinate sequence within the sliding time window, constructs a dynamic envelope model, and establishes the semantic stable regions of each dimension.

[0204] Drift detection signal acquisition module: If there is a dimension that exceeds the boundary of the corresponding semantically stable region for three consecutive rounds, then mark the dimension as the drift dimension where semantic anchor point drift occurs, and generate a drift detection signal.

[0205] The calibrated dimension coordinate acquisition module is used to freeze the five dimension coordinates that have not drifted in response to the drift detection signal, perform a direction-constrained reprojection operation only on the current sub-vector of the drift dimension, apply a semantic restoring force that is non-linearly proportional to the deviation distance in the opposite direction of the deviation, and generate calibrated dimension coordinates.

[0206] Six-dimensional comprehensive feature vector acquisition module: used to construct a six-dimensional comprehensive feature vector based on the calibrated dimension coordinates and the original coordinates of the un-drifted dimension;

[0207] Lighting fixture material matching and comprehensive evaluation module: This module performs cosine similarity calculation on the six-dimensional comprehensive feature vector and outputs lighting fixture material matching results sorted by similarity, along with corresponding structural feasibility verification data and BOM cost estimation data.

[0208] This application provides a fast filtering and matching method for lighting fixture databases based on a six-dimensional tag system, which can be applied to applications such as... Figure 4 In the application environment shown, terminal 102 communicates with server 104 via a network. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices, and server 104 can be a standalone server or a server cluster consisting of multiple servers.

[0209] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores data generated during the implementation of a rapid screening and matching method for a lighting fixture database based on a six-dimensional tag system. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a rapid screening and matching method for a lighting fixture database based on a six-dimensional tag system.

[0210] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements a fast filtering and matching method for a lighting database based on a six-dimensional tag system as described in any of the above embodiments.

[0211] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0212] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0213] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for rapid filtering and matching of lighting fixture databases based on a six-dimensional tagging system, characterized in that, include: S1: Obtain the user's natural language requirement text, and use a preset dimension-specific semantic encoder to decouple the user's natural language requirement text to generate a six-dimensional initial dimension feature vector. S2: Based on the six-dimensional initial feature vector, calculate the cluster center of all tag instances in each dimension, define the cluster center as the initial semantic anchor point of the six-dimensional tag system, and extract the six-dimensional sub-vectors of the current interaction round as the current activation coordinates; S3: For the current active coordinates of each dimension, call the historical anchor point coordinate sequence within the sliding time window to construct a dynamic envelope model and establish the semantic stable region of each dimension; S4: If there exists a dimension where the current active coordinate exceeds the boundary of the corresponding semantically stable region for three consecutive rounds, then mark the dimension as the drift dimension where semantic anchor point drift occurs, and generate a drift detection signal. S5: In response to the drift detection signal, freeze the five dimensional coordinates that have not drifted, perform a direction-constrained reprojection operation only on the current sub-vector of the drifted dimension, apply a semantic restoring force that is non-linearly proportional to the deviation distance in the opposite direction of the deviation, and generate calibrated dimensional coordinates. S6: Based on the calibrated dimensional coordinates and the original coordinates of the un-drifted dimensions, construct a six-dimensional comprehensive feature vector; S7: Perform cosine similarity calculation on the six-dimensional comprehensive feature vector, and output the matching results of lighting materials sorted by similarity, as well as the corresponding structural feasibility verification data and BOM cost estimation data.

2. The method for rapid filtering and matching of lighting fixture databases based on a six-dimensional tag system according to claim 1, characterized in that, After step S7, the following is included: S8: After completing ten rounds of effective interaction, extract new anchor point samples that are still in the semantically stable region after calibration in this session, and perform incremental smooth updates on the dynamic envelope model of each dimension to generate the updated semantically stable region.

3. The method for rapid filtering and matching of lighting fixture databases based on a six-dimensional tag system according to claim 1, characterized in that, The dynamic envelope model includes an upper tenths line and a lower tenths line.

4. The method for rapid filtering and matching of lighting fixture databases based on a six-dimensional tag system according to claim 3, characterized in that, S3 specifically includes: Obtain the current active coordinates of a specific dimension in the current interaction round, and call the sliding time window buffer corresponding to the specific dimension to extract the historical anchor point coordinate sequence that has been calibrated in the past several rounds of interaction, forming the original time series data set; The original time-series data set is sorted and preprocessed to arrange the historical anchor point coordinate sequence in ascending order according to the numerical value, thereby generating an ordered historical anchor point coordinate queue. Based on the ordered historical anchor point coordinate queue, and based on non-parametric quantile statistical processing, the coordinates of the lower quantile point with a cumulative probability of 0.1 and the coordinates of the upper quantile point with a cumulative probability of 0.9 in the ordered historical anchor point coordinate queue are located respectively, and the lower tenths threshold and the upper tenths threshold are generated. Using the lower tenths threshold and the upper tenths threshold as spatial boundary constraints, a closed dynamic envelope geometry is constructed in the six-dimensional feature space to generate a dynamic envelope model instance that includes semantic fluctuation tolerance range. The dynamic envelope model instance is mapped to the semantic coordinate system of the current dimension, defining a continuous spatial region enclosed by the lower tenths threshold and the upper tenths threshold, and outputting the final judgment boundary defined as the semantically stable region of this dimension.

5. The method for rapid filtering and matching of lighting fixture databases based on a six-dimensional tag system according to claim 4, characterized in that, S4 specifically includes: The six-dimensional sub-vectors generated in the current interaction round are obtained as the activation coordinates to be detected, and the historical anchor point coordinate sequence stored in the sliding time window is called to calculate the upper tenths threshold and lower tenths threshold corresponding to each dimension, and generate dynamic envelope model data. Based on the upper tenths threshold and lower tenths threshold in the dynamic envelope model data, the values ​​of each dimension in the active coordinate to be detected are subjected to boundary exceedance judgment processing. If the value is greater than the upper tenths threshold or less than the lower tenths threshold, a single exceedance status flag for that dimension is generated. Read the sequence of single out-of-limit status flags generated during three consecutive rounds of interaction, perform a sequential continuous logic AND operation on the single out-of-limit status flags of the same dimension, and if any dimension shows an out-of-limit state in three consecutive rounds, generate a continuous drift confirmation signal for that dimension. Based on the continuous drift confirmation signal, the label instance of the corresponding dimension is marked as the drift dimension in which semantic anchor point drift has occurred, and the coordinate states of the other five dimensions that have not received the continuous drift confirmation signal are frozen, thereby generating the drift detection signal containing the drift dimension index.

6. The method for rapid filtering and matching of lighting fixture databases based on a six-dimensional tag system according to claim 5, characterized in that, S5 specifically includes: Obtain the drift detection signal and the six-dimensional initial dimension feature vector of the current round. Based on the drift dimension index in the drift detection signal, perform dimension masking on the six-dimensional initial dimension feature vector to generate a frozen coordinate set containing the original coordinates of five non-drift dimensions and a drift dimension sub-vector to be calibrated. Extract the coordinates of the drift dimension sub-vector to be calibrated and the semantically stable region boundary of its corresponding dimension; calculate the deviation distance vector of the drift dimension sub-vector to be calibrated relative to the semantically stable region boundary based on Euclidean distance; and generate the deviation distance scalar and the unit direction vector. The deviation distance scalar is received, and a nonlinear transformation is performed on the deviation distance scalar using a preset nonlinear hyperbolic tangent mapping function to generate semantic restoring force. The semantic correction increment is generated by performing a weighted product operation between the reverse vector of the unit direction vector and the semantic restoring force, and the semantic correction increment is superimposed on the drift dimension sub-vector to be calibrated to generate calibrated dimension coordinates that fall within the semantically stable region. The original coordinates of the five undrifted dimensions in the frozen coordinate set are integrated with the calibrated dimension coordinates, and a vector recombination operation is performed according to the predefined dimension order of the six-dimensional label system to generate the six-dimensional comprehensive feature vector.

7. The method for rapid filtering and matching of lighting fixture databases based on a six-dimensional tag system according to claim 1, characterized in that, The dimensions include: time, space, people, context, emotion, and style.

8. A rapid filtering and matching system for a lighting fixture database based on a six-dimensional tagging system, comprising: The six-dimensional initial dimension feature vector acquisition module is used to acquire the user's natural language requirement text, and use a preset dimension-specific semantic encoder to decouple the user's natural language requirement text to generate a six-dimensional initial dimension feature vector. The current activation coordinate acquisition module is used to calculate the cluster center of all tag instances in each dimension based on the six-dimensional initial dimension feature vector, define the cluster center as the initial semantic anchor point of the six-dimensional tag system, and extract the six-dimensional sub-vectors of the current interaction round as the current activation coordinates. Semantic stable region acquisition module: For the current active coordinates of each dimension, it calls the historical anchor point coordinate sequence within the sliding time window, constructs a dynamic envelope model, and establishes the semantic stable regions of each dimension. Drift detection signal acquisition module: If there is a dimension that exceeds the boundary of the corresponding semantically stable region for three consecutive rounds, then mark the dimension as the drift dimension where semantic anchor point drift occurs, and generate a drift detection signal. The calibrated dimension coordinate acquisition module is used to freeze the five dimension coordinates that have not drifted in response to the drift detection signal, perform a direction-constrained reprojection operation only on the current sub-vector of the drift dimension, apply a semantic restoring force that is non-linearly proportional to the deviation distance in the opposite direction of the deviation, and generate calibrated dimension coordinates. Six-dimensional comprehensive feature vector acquisition module: used to construct a six-dimensional comprehensive feature vector based on the calibrated dimension coordinates and the original coordinates of the un-drifted dimension; Lighting fixture material matching and comprehensive evaluation module: This module performs cosine similarity calculation on the six-dimensional comprehensive feature vector and outputs lighting fixture material matching results sorted by similarity, along with corresponding structural feasibility verification data and BOM cost estimation data.

9. A computer device, characterized in that, include: A processor, a memory, and a computer program stored in the memory and executable on the processor; when executed by the processor, the computer program implements the steps of a fast filtering and matching method for a lighting database based on a six-dimensional tag system as described in any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, implements the steps of a fast filtering and matching method for a lighting database based on a six-dimensional tag system as described in any one of claims 1 to 7.