Intelligent risk early warning system and method based on multi-dimensional information fusion

Abstract

The invention discloses an intelligent risk early warning system and method based on multi-dimensional information fusion. The intelligent risk early warning system comprises a name and obligee similarity calculation module, a trademark commodity service comparison module, an image feature extraction module, a multi-dimensional feature database and a comprehensive risk calculation module. According to the method, four core dimension data including trademark names, obligee information, commodity service classification and identification images are organically fused, and a collaborative strategy of text multi-dimension comparison, multi-channel representation, multi-agent rule matching and image multi-view feature extraction is adopted, so that the approximate risk level of the trademark is accurately judged; and key technical support is provided for trademark registration examination, infringement monitoring and intellectual property management.

Description

Technical Field

[0001] This invention belongs to the field of trademark risk warning technology in the field of intellectual property. Specifically, it relates to an intelligent risk warning system based on multi-dimensional information fusion. This technology organically integrates four core dimensions of data: trademark name, rights holder information, goods and services classification, and logo image. It adopts a collaborative strategy of multi-dimensional text comparison, multi-channel representation, multi-agent rule matching, and multi-view feature extraction of images to accurately determine the risk level of trademark similarity, providing key technical support for trademark registration examination, infringement monitoring, and intellectual property management. Background Technology

[0002] With the deepening of global economic integration and the awakening of brand awareness among market players, trademark registration applications, as an important carrier of core intellectual property rights and brand value, are experiencing explosive growth. In the face of massive trademark data and a complex market competition environment, trademark registration examination, infringement monitoring, and full lifecycle risk management face severe challenges. Existing trademark risk prediction technologies primarily serve examiners or agencies, with their core objective being to maintain market competition order and protect the legitimate rights and interests of rights holders. However, facing increasingly complex new risk scenarios such as cross-category related infringements, malicious registration, confusion due to identical names in different classes, and non-standard descriptions of goods and services, traditional technical means are proving inadequate and unable to meet the actual needs for high accuracy and intelligence.

[0003] Existing technologies have significant limitations in single-dimensional feature extraction and comparison. In the text dimension, traditional methods often rely on string edit distance (Levenshtein Distance) or simple word vector matching for similarity calculations, focusing only on literal morphological overlap while neglecting deep semantic connections such as polyphonic characters, similar-looking characters, and similar meanings. This makes it difficult to identify implicit infringement risks such as "same sound, different character" or "same meaning, different form." In the image dimension, existing methods based on traditional computer vision (CV) feature extraction algorithms (such as SIFT and SURF) or shallow neural networks are mostly limited to comparing the overall outline or color distribution of images. They lack the ability to independently isolate the salient parts of combined text and image trademarks, making it difficult to capture subtle visual infringement behaviors such as minor adjustments to font design styles or partial graphic copying. Furthermore, the lack of a rights holder-level association analysis mechanism is a major weakness. Existing systems often view each applicant in isolation, unable to identify defensive cross-class layouts by the same entity or malicious hoarding behavior by related companies through knowledge graph construction. This leads to misjudging normal defensive registrations as infringement or overlooking complex infringement risks from related entities.

[0004] In the core step of comparing goods and services, the bottlenecks of existing technologies are particularly prominent. Traditional methods for determining similarity in goods and services mainly rely on fixed-code matching from the "Similar Goods and Services Classification Table" or rule-based keyword matching. However, in actual business activities, descriptions of goods and services vary greatly, often including non-standard expressions, industry jargon, emerging concepts, or semantically ambiguous descriptions. Traditional Natural Language Processing (NLP) technologies often suffer from ambiguity when processing such short texts due to a lack of contextual understanding. For example, simple keyword matching cannot distinguish the contextual differences between "apple (fruit)" and "apple (electronic product)," nor can it handle matching biases caused by missing core attributes or synonym substitutions (such as "smart terminal" and "mobile phone"). Existing technologies generally lack mature Large Language Model (LLM) empowerment mechanisms, failing to leverage the massive general knowledge base and powerful logical reasoning capabilities (such as thought chain analysis) of large models to extract the core elements of goods (function, purpose, target consumer), resulting in severely insufficient robustness when handling complex semantic scenarios and persistently high false positive and false negative rates.

[0005] In the image dimension, the core shortcoming of existing technologies lies in the lack of effective image-text separation processing logic, which is also the key issue for the insufficient accuracy of visual risk identification. Traditional methods often directly compare pixels or extract features from the overall trademark image, without designing a dedicated processing link for the composite structure of "graphic + text" trademarks. This fails to isolate the interference of complex backgrounds and text areas, leading to confusion and interference between pure graphic features and glyph features, making it difficult to accurately locate and extract visual information that reflects the core recognizability of the trademark. This "hybrid" processing approach prevents key visual elements such as the graphic design creativity and font style characteristics of the trademark from being analyzed in isolation, resulting in a lack of specificity in the judgment of visual similarity risks and failing to meet the refined judgment requirements for graphic and glyph similarity in trademark examination.

[0006] More importantly, the multi-dimensional fusion logic of existing technologies is too simplistic and crude, lacking dynamic adaptability. Most mainstream risk assessment models currently employ simple linear weighted summation or direct multiplication to fuse similarity across graphic, text, product, and rights holder dimensions. This rigid mathematical model suffers from a fatal "weakest link" or "zero-vote veto" flaw. For example, in a multiplicative model, once the product / service category is determined to be dissimilar (score of 0), even if the trademark image and text are completely copied and there is suspicion of malicious registration, the overall risk score will be forcibly reduced to zero, leading to serious underreporting. Simultaneously, existing algorithms lack a priority response mechanism for strong features (such as identical trademark names) and cannot dynamically adjust the protection threshold based on the reputation of the warning trademark itself. This results in the system lacking necessary flexibility and accuracy when facing special scenarios such as cross-class protection of well-known trademarks or name confusion.

[0007] In summary, existing trademark risk prediction technologies have significant shortcomings in terms of the depth of feature extraction, the accuracy of product semantic understanding, and the logical rigor of multi-dimensional fusion. There is an urgent need for an innovative solution that can deeply integrate multimodal information, introduce large-scale model semantic reasoning capabilities at key stages, and possess dynamic weight adjustment and circuit breaker protection mechanisms to construct a robust, end-to-end trademark risk assessment system. Summary of the Invention

[0008] This invention addresses the shortcomings of existing trademark risk warning technologies, such as low accuracy, high false positive rate, and poor interpretability when handling multilingual, cross-category, complex semantic, and graphic-text combination trademarks. It combines Large Language Modeling (LLM), Multi-Agent Systems, and deep computer vision technology to provide an intelligent risk warning system and method based on multi-dimensional information fusion. This invention aims to overcome the limitations of traditional single-dimensional comparison by organically integrating the form, sound, and semantic features of trademark names, multi-channel information from rights holders, deep semantic relationships between goods and services, and multi-view visual features of logo images. This effectively solves technical challenges such as cross-language semantic gaps, non-standardized matching of product descriptions, and difficulties in extracting graphic features in complex backgrounds. Furthermore, this invention is dedicated to providing high-precision and robust technical support for trademark registration examination, corporate brand monitoring, and intellectual property management, helping examiners and rights holders to more comprehensively assess potential infringement risks, improve handling efficiency, and thus maintain a fair market competition order.

[0009] The technical solution adopted by this invention to solve its technical problem includes the following steps:

[0010] In a first aspect, embodiments of this application provide an intelligent risk warning system based on multi-dimensional information fusion, including a name and rights holder similarity calculation module, a trademark goods and services comparison module, an image feature extraction module, a multi-dimensional feature database, and a comprehensive risk calculation module.

[0011] The name and rights holder similarity calculation module is used to perform a deep comparison between the name of the trademark to be inspected and the name of the rights holder and the base data in the multidimensional feature database, and calculate the corresponding similarity score.

[0012] The trademark-product-service comparison module constructs a two-layer processing architecture of "vector coarse screening - intelligent agent fine ranking" to make a final similarity judgment on the product and service descriptions.

[0013] The image feature extraction module performs refined processing on the original trademark image and calculates the trademark image similarity.

[0014] The multidimensional feature database adopts a hybrid storage architecture of "MySQL relational database + vector database" to uniformly store the basic metadata and multidimensional feature vectors of the base database and the trademarks to be inspected, supporting millisecond-level retrieval and multimodal association analysis. The relational database stores standardized basic metadata, and the vector database stores feature vectors extracted from multiple dimensions.

[0015] The comprehensive risk calculation module calculates the final risk score by combining the similarity of mixed text and images, the dynamic commodity coefficient, and the difference between rights holders.

[0016] In one possible implementation, the name-to-rights-owner similarity calculation module operates as follows:

[0017] For trademark names, a dual-track parallel processing mechanism of "original text + translation" is constructed. A translation model is used to generate translation fields; this model is a neural network translation model capable of supporting multilingual translation while retaining the original text fields. Based on this, a weighted fusion of form similarity based on edit distance and visual hash, sound similarity based on pinyin or Romanized transliteration, and semantic similarity based on multilingual sentence vectors is used to obtain the comprehensive similarity of the trademark names. For the rights holder's name, multi-channel feature extraction is implemented. The similarity calculation of the rights holder's name is divided into three channels: full name, abbreviation, and pinyin of the abbreviation. The string edit distance and visual hash similarity are calculated for each channel, and semantic similarity is calculated by combining sentence vector encoding. Finally, the comprehensive similarity of the rights holder's name is obtained through multi-channel weighted fusion.

[0018] In one possible implementation, the trademark goods and services comparison module operates as follows:

[0019] A two-stage retrieval mechanism of "coarse screening-fine ranking" is adopted. First, the semantic vector model is used to vectorize and encode the descriptions of goods and services and calculate cosine similarity to complete the candidate set recall. Then, a multi-agent system is constructed, and each agent node works together to extract the rule basis of the "Similar Goods and Services Distinction Table", perform vector retrieval and semantic fine ranking, and make a final similarity judgment by integrating rule information.

[0020] In one possible implementation, the image feature extraction module operates as follows:

[0021] A deep learning-based image-text separation processing chain is constructed. First, the GrabCut algorithm based on energy minimization is used to segment the foreground of the original trademark image to be inspected. Second, the text region in the foreground is located by combining a deep learning-based optical character recognition module and extracted as a character view. Then, the text region is filled with an image inpainting algorithm to obtain a pure graphic view. Finally, high-dimensional feature vectors of the original image view, character view and pure graphic view are extracted by a deep visual neural network and are collectively referred to as multi-view feature vectors. The obtained multi-view feature vectors are then compared with the multi-view feature vectors of the base database in the multi-dimensional feature database to calculate the similarity of the trademark image.

[0022] In one possible implementation, the comprehensive risk calculation module operates as follows:

[0023] First, a non-linear fusion of image and text similarity is performed. For the similarity between trademark images and trademark names, a three-element fusion strategy of "maximum value as the primary factor, arithmetic mean as the auxiliary factor, and harmonic mean as the fallback factor" is adopted to calculate the image-text mixed similarity. Second, a dual-path optimization calculation of dynamic commodity coefficient is performed, with the "soft gating coefficient" and "circuit breaker upgrade coefficient" calculated in parallel, and the maximum value of the two is used as the dynamic commodity coefficient. Finally, the difference of the rights holder is introduced as an exclusivity factor, and the final risk score is synthesized through multiplicative logic.

[0024] Secondly, embodiments of this application provide a method for an intelligent risk early warning system based on multi-dimensional information fusion, comprising the following steps:

[0025] Step 1: Constructing a multidimensional feature database:

[0026] The multidimensional feature database adopts a hybrid storage architecture of "MySQL relational database + vector database" to uniformly store the basic metadata and multidimensional feature vectors of the base database and the trademarks to be inspected, supporting millisecond-level retrieval and multimodal association analysis. The relational database stores standardized basic metadata, and the vector database stores feature vectors extracted from multiple dimensions.

[0027] Step 2: Perform a deep comparison between the name of the trademark to be inspected and the name of the right holder and the base data in the multidimensional feature database, and calculate the corresponding similarity score.

[0028] For trademark names, a dual-track parallel processing mechanism of "original text + translation" is constructed. A translation model is used to generate translation fields; this model is a neural network translation model capable of supporting multilingual translation while retaining the original text fields. Based on this, a weighted fusion of form similarity based on edit distance and visual hash, sound similarity based on pinyin or Romanized transliteration, and semantic similarity based on multilingual sentence vectors is used to obtain the comprehensive similarity of the trademark names. For the rights holder's name, multi-channel feature extraction is implemented. The similarity calculation of the rights holder's name is divided into three channels: full name, abbreviation, and pinyin of the abbreviation. The string edit distance and visual hash similarity are calculated for each channel, and semantic similarity is calculated by combining sentence vector encoding. Finally, the comprehensive similarity of the rights holder's name is obtained through multi-channel weighted fusion.

[0029] Step 3: Construct a two-layer processing architecture of "vector coarse screening - intelligent agent fine ranking" to make the final similarity judgment on the product and service descriptions.

[0030] A two-stage retrieval mechanism of "coarse screening-fine ranking" is adopted. First, the semantic vector model is used to vectorize and encode the descriptions of goods and services and calculate cosine similarity to complete the candidate set recall. Then, a multi-agent system is constructed, and each agent node works together to extract the rule basis of the "Similar Goods and Services Distinction Table", perform vector retrieval and semantic fine ranking, and make a final similarity judgment by integrating rule information.

[0031] Step 4: Perform fine-grained processing on the original trademark image and calculate the trademark image similarity.

[0032] A deep learning-based image-text separation processing chain is constructed. First, the GrabCut algorithm based on energy minimization is used to segment the foreground of the original trademark image to be inspected. Second, the text region in the foreground is located by combining a deep learning-based optical character recognition module and extracted as a character view. Then, an image inpainting algorithm is used to fill the text region to obtain a pure graphic view. Finally, high-dimensional feature vectors of the original image view, character view, and pure graphic view are extracted by a deep visual neural network and collectively referred to as multi-view feature vectors. The obtained multi-view feature vectors are then compared with the multi-view feature vectors of the base database in the multi-dimensional feature database to calculate the similarity of the trademark image.

[0033] Step 5: Calculate the final risk score by combining the similarity of mixed text and images, the dynamic commodity coefficient, and the difference between rights holders.

[0034] First, a non-linear fusion of image and text similarity is performed. For the similarity between trademark images and trademark names, a three-element fusion strategy of "maximum value as the primary factor, arithmetic mean as the auxiliary factor, and harmonic mean as the fallback factor" is adopted to calculate the image-text mixed similarity. Second, a dual-path optimization calculation of dynamic commodity coefficient is performed, with the "soft gating coefficient" and "circuit breaker upgrade coefficient" calculated in parallel, and the maximum value of the two is used as the dynamic commodity coefficient. Finally, the difference of the rights holder is introduced as an exclusivity factor, and the final risk score is synthesized through multiplicative logic.

[0035] This invention makes the following contributions to research and innovation compared to existing technologies:

[0036] 1. Unlike existing technologies that rely solely on single textual features for trademark name or rights holder comparison, this invention proposes a multi-channel, multi-feature fusion-based trademark subject comparison technology. By constructing a trademark name comparison model encompassing "form, sound, and meaning" across all dimensions, and a rights holder comparison model using "full name, abbreviation, and pinyin" across multiple channels, this invention effectively overcomes challenges such as cross-language homomorphic spoofing, transliteration infringement, and concealed registration by related entities. Particularly in product and service comparison, it innovatively introduces a multi-agent collaborative semantic reasoning mechanism and semantic vectorization technology, achieving a leap from "keyword matching" to "semantic rule collaborative reasoning," significantly reducing missed and incorrect judgments caused by non-standard product descriptions or differences in industry slang.

[0037] 2. This invention introduces a rule-driven semantic reasoning mechanism in the product and service comparison process. Unlike traditional "black box" AI judgment, this invention uses a large model to automatically extract the rule basis from the "Similar Goods and Services Differentiation Table" in the product and service comparison stage, and provides clear rule references and risk attribution in the output results. This "data-driven + rule-guided" model not only outputs a high-confidence risk level, but also provides examiners and legal personnel with detailed judgment basis, effectively shortening the manual review time and improving the efficiency of business decision-making.

[0038] 3. To address the problem in existing technologies that process combined graphic and text trademarks as a whole and cannot distinguish between text and graphic areas, this invention provides a multi-view trademark image processing method that separates text and graphics. The proposed image-text separation link based on GrabCut and Inpainting can accurately separate pure graphic and font design features. This multi-view (original image, pure image, font) feature extraction strategy avoids the impact of background noise and text interference on graphic similarity calculations, enabling the model to more sensitively detect subtle infringement behaviors such as partial graphic plagiarism or font style imitation, greatly enhancing the robustness of visual dimension comparison.

[0039] 4. This invention adopts a modular system architecture design, setting trademark name processing, product and service comparison, image processing and risk fusion calculation as relatively independent functional modules, so that each module can be replaced or expanded without affecting the overall process, which is suitable for trademark risk assessment needs of different scales and application scenarios. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0041] Figure 1 This is a flowchart of the workflow of the module for calculating the similarity between the name of this invention and the rights holder.

[0042] Figure 2 This is a flowchart of the trademark, goods and services comparison module of this invention.

[0043] Figure 3 This is a flowchart of the image feature extraction module of the present invention.

[0044] Figure 4 This is the overall architecture diagram of the comprehensive risk calculation module of this invention. Detailed Implementation

[0045] The present invention will be further described below with reference to the accompanying drawings.

[0046] like Figure 1-4 As shown, this invention proposes an intelligent risk early warning system and method based on multi-dimensional information fusion. This method achieves high-precision prediction of complex trademark infringement risks by integrating large language model semantic reasoning, deep computer vision processing and multi-agent collaborative mechanism.

[0047] This application provides a method for an intelligent risk early warning system based on multi-dimensional information fusion, including the following steps:

[0048] Step 1: Constructing a multidimensional feature database:

[0049] The multidimensional feature database adopts a hybrid storage architecture of "MySQL relational database + vector database" to uniformly store the basic metadata and multidimensional feature vectors of the base database and the trademarks to be inspected, supporting millisecond-level retrieval and multimodal association analysis. The relational database stores standardized basic metadata, and the vector database stores feature vectors extracted from multiple dimensions.

[0050] Step 2: Perform a deep comparison between the name of the trademark to be inspected and the name of the right holder and the base data in the multidimensional feature database, and calculate the corresponding similarity score.

[0051] For trademark names, a dual-track parallel processing mechanism of "original text + translation" is constructed. A translation model is used to generate translation fields to enhance cross-language semantic comparability. This translation model is a neural network translation model capable of supporting multilingual translation, such as a large language translation model based on the Transformer structure. Simultaneously, the original text field is retained for minority language detection and visual rendering. Based on this, a weighted fusion of form similarity based on edit distance and visual hash, sound similarity based on pinyin or Romanized transliteration, and semantic similarity based on multilingual sentence vectors is used to obtain the comprehensive similarity of the trademark names. For the rights holder's name, multi-channel feature extraction is implemented. The similarity calculation of the rights holder's name is divided into three channels: full name, abbreviation, and pinyin of the abbreviation. The string edit distance and visual hash similarity are calculated for each channel, and semantic similarity is calculated in conjunction with sentence vector encoding. Finally, the comprehensive similarity of the rights holder's name is obtained through multi-channel weighting.

[0052] Step 3: Construct a two-layer processing architecture of "vector coarse screening - intelligent agent fine ranking" to make the final similarity judgment on the product and service descriptions.

[0053] A two-stage retrieval mechanism of "coarse screening-fine ranking" is adopted. First, the semantic vector model is used to vectorize and encode the descriptions of goods and services and calculate cosine similarity to complete the candidate set recall. Then, a multi-agent system is constructed, and each agent node works collaboratively: extracting the rule basis of the "Similar Goods and Services Distinction Table", performing vector retrieval and semantic fine ranking, and making a final similarity judgment by integrating rule information. This step organically integrates the generalization ability of neural networks with the logic of symbolic rules, breaking through the limitations of traditional keyword matching in handling non-standard product descriptions.

[0054] Step 4: Perform fine-grained processing on the original trademark image and calculate the trademark image similarity.

[0055] A deep learning-based image-text separation processing chain is constructed to address the problem of complex background interference. First, the GrabCut algorithm, based on energy minimization, is used to segment the foreground of the original trademark image to be inspected. Second, a deep learning-based optical character recognition module is used to locate the text regions in the foreground and extract them separately as character views. An image inpainting algorithm is then used to fill in the text regions, resulting in a pure graphic view. Finally, a deep visual neural network is used to extract high-dimensional feature vectors from the original image view, character view, and pure graphic view, collectively referred to as multi-view feature vectors. These multi-view feature vectors are then compared with the multi-view feature vectors in the base database of a multi-dimensional feature database to calculate the trademark image similarity, providing a clean feature representation and trademark image similarity score for subsequent multi-dimensional visual comparison.

[0056] Step 5: Calculate the final risk score by combining the similarity of mixed text and images, the dynamic commodity coefficient, and the difference between rights holders.

[0057] First, a non-linear fusion of image and text similarity is performed. For the similarity between trademark images and trademark names, the linear average algorithm is abandoned, and a three-element fusion strategy of "maximum value as the main factor, arithmetic average as an auxiliary factor, and harmonic average as a fallback" is adopted to calculate the image-text mixed similarity, so as to keenly capture the extreme value risk of a single dimension. Second, a dual-path optimization calculation of dynamic commodity coefficient is performed, and the "soft gate coefficient" and "circuit breaker upgrade coefficient" are calculated in parallel. The maximum value of the two is calculated as the dynamic commodity coefficient to solve the problem of missed detection of cross-class malicious registration and confusion due to the same name. Finally, the difference of the rights holder is introduced as an exclusivity factor, and the final risk score is synthesized through multiplicative logic.

[0058] This application also provides an intelligent risk warning system based on multi-dimensional information fusion, including a name and rights holder similarity calculation module, a trademark goods and services comparison module, an image feature extraction module, a multi-dimensional feature database, and a comprehensive risk calculation module.

[0059] The name and rights holder similarity calculation module is used to perform a deep comparison between the name of the trademark to be inspected and the name of the rights holder and the base data in the multidimensional feature database, and calculate the corresponding similarity score.

[0060] The trademark-product-service comparison module constructs a two-layer processing architecture of "vector coarse screening - intelligent agent fine ranking" to make a final similarity judgment on the product and service descriptions.

[0061] The image feature extraction module performs refined processing on the original trademark image and calculates the trademark image similarity.

[0062] The multidimensional feature database adopts a hybrid storage architecture of "MySQL relational database + vector database" to uniformly store the basic metadata and multidimensional feature vectors of the base database and the trademarks to be inspected, supporting millisecond-level retrieval and multimodal association analysis. The relational database stores standardized basic metadata, and the vector database stores feature vectors extracted from multiple dimensions.

[0063] The comprehensive risk calculation module calculates the final risk score by combining the similarity of mixed text and images, the dynamic commodity coefficient, and the difference between rights holders.

[0064] In one possible implementation, the input data for the name and rights holder similarity calculation module consists of the trademark name and rights holder name of the trademark to be examined (i.e., the original text fields). First, basic normalization processing is performed on the original text fields to obtain standard fields. This basic normalization processing includes null value filling, removal of leading and trailing whitespace, whitespace compression, and case normalization. This processing does not change the semantics and preserves the original text fields, denoted as TM_orig (trademark) and Owner_orig (right holder), respectively.

[0065] Then, the translation big model (the Hunyuan-MT-7B translation big model is used in this embodiment) is called to generate the corresponding translation fields TM_trans and Owner_trans for the standard fields. During translation, a structure protection and consistency control mechanism is adopted, specifically: (1) proper names, numbers and parentheses are replaced with placeholders and backfilled after translation is completed; (2) texts of the same language or length are batch-processed, and the random probability of the output is reduced by adjusting the decoding parameters to ensure consistency; (3) the placeholder integrity rate and number retention rate of the translated text are checked. If they fail, they are retranslated. If they still fail the check after retranslation, they are set to empty. The translation fields are only newly added to the storage and do not overwrite the original text fields, ensuring the subsequent detection of minor languages ​​and rendering of the original text.

[0066] The base database improves online throughput through offline pre-computation and caching. Specifically: for TM_orig, normalized text is used to pre-store edit distance, pre-store approximate pronunciation strings, use Unicode judgment tags (non-Latin letters and CJK Chinese characters are tags) for minority language recognition, and use visual hash fingerprints to calculate visual shape similarity; for TM_trans, semantic vector matrices are batch encoded using a multilingual sentence vector model (distiluse-base-multilingual-cased-v2). For Owner_orig, the full name is preserved and an abbreviation and its pinyin are generated to form a three-channel text, generating normalized text and visual hash fingerprints for each channel; for Owner_trans, semantic vectors are batch encoded, and the abbreviation and pinyin channels are encoded to support subsequent semantic fusion.

[0067] During the online search phase, the similarity in form, sound, and meaning between the trademark name of the trademark to be searched and each trademark name in the database is calculated and fused. The form similarity is calculated solely based on the standard fields of the trademark names in the database and the trademark name to be searched.

[0068] First, calculate the normalized similarity of the string edit distance:

[0069]

[0070] in, The normalized similarity between the trademark to be tested (a) and the base database data (b); The string edit distance between the trademark to be inspected (a) and the base database data (b); The maximum value is obtained from the modulus of the trademark to be inspected (a) and the base database data (b).

[0071] The standard fields of the trademark names in the trademark database and the trademarks to be inspected are rendered into grayscale images using rasterized text technology. A fixed-length fingerprint is obtained by calculating the difference hash, and the Hamming distance is calculated after performing an XOR operation on the fingerprint. The visual hash similarity is obtained as follows:

[0072]

[0073] in The hash bit count is preferably 64. The two are then factored together. The fusion yields the shape similarity:

[0074]

[0075] Phonetic similarity is also calculated solely based on the standard fields of the trademark name: the standard fields of the trademark name to be examined and the trademark name in the database are romanized to form a pronunciation string, and the normalized similarity of the edit distance is calculated on the pronunciation string to obtain the phonetic similarity. .

[0076] Semantic similarity calculation: The translation fields of the trademark to be examined and the base database data are encoded into vectors respectively. and Calculate the cosine similarity:

[0077]

[0078] Map the cosine similarity to Obtain semantic similarity (Semantic similarity when the translation field is empty) (The value is 0). Simultaneously, semantic similarity is directly calculated based on the standard fields of the trademark name. A two-path fusion strategy is used to determine the final semantic similarity. :

[0079]

[0080] The final weighted average similarity of the trademark names is obtained by combining the results:

[0081]

[0082] Overall similarity of trademark names Sort the data in descending order and output the results. The output items include the similarity scores of each component and the overall similarity score to support interpretation and verification.

[0083] The similarity scores of the trademark holder's name and the semantic similarity scores of each trademark holder's name in the database are calculated and fused. The calculation of the similarity scores of the trademark holder's name incorporates multiple channels for abbreviations and their pinyin spellings. First, the abbreviations and their pinyin spellings are generated based on the standard fields of the trademark holder's name. Abbreviation extraction includes removing the region segment within parentheses, removing the prefix region, and removing the organizational form suffix (Limited Company, Joint-Stock Company, etc.). If industry keywords exist, the abbreviation is truncated before the industry keywords to obtain the trade name. For example, "XX Province A Sanitary Ware Co., Ltd." is extracted as "A", and "B (XX Province) Joint-Stock Company" is extracted as "B". The abbreviations are then transliterated into pinyin and separated by spaces to generate the pinyin spellings. Subsequently, the similarity scores are calculated for the full name, abbreviation, and pinyin spellings. For each channel c, the channel similarity score is obtained using the following formula. :

[0084]

[0085] in, Normalized similarity for the edit distance of the string in channel c; Let be the visual hash similarity of channel c.

[0086] The final shape similarity is obtained by fusing the three channels element by element based on their maximum values.

[0087]

[0088] in, The similarity is calculated using three channels: the full name (full), the abbreviation (short), and the abbreviation (pinyin) (py).

[0089] The method for calculating the semantic similarity of each channel of the rights holder's name is the same as the method for calculating the semantic similarity of the trademark name; the semantic similarity of the three channels is fused according to the maximum value of each element to obtain the final semantic similarity.

[0090] The final form similarity and final meaning similarity of the right holder's name are combined on an equal basis to obtain the comprehensive similarity of the right holder's name.

[0091] In one possible implementation, the trademark-goods-services comparison module is specifically implemented as follows:

[0092] First, vectorized recall is performed. A pre-trained semantic vector encoding model (Qwen3-Embedding-0.6B model is used in this embodiment) is used to transform the product / service descriptions of the trademarks to be inspected into dense vector representations. These dense vector representations characterize the semantic features of the products / services. Based on these dense vector representations, cosine similarity is calculated to quickly select the Top-K semantic candidates (i.e., product / service descriptions) from the base database, completing an efficient coarse screening.

[0093] Based on this, a multi-processing module mechanism (i.e., a multi-agent system) is used to conduct similarity checks and risk assessments on the product and service descriptions of the trademark under inspection through multi-agent collaborative work, and output the corresponding risk level. Specifically: First, the semantic fine-ranking agent inputs the product and service descriptions of the trademark under inspection and the candidate set into a large language model (the embodiment uses the Claude large language model) for multiple sorting and fine-tuning of similarity scores, thereby obtaining a more reliable candidate pool and providing higher quality input for the rule comparison stage. Subsequently, the rule extraction agent uses the large language model (the embodiment uses the Claude large language model) to find the most relevant rule paragraphs and annotations in the "Similar Goods and Services Classification Table" for the product and service descriptions of the trademark under inspection and the Top-K semantic candidates, and summarizes the corresponding similarity group information and explanations accordingly. Finally, the rule comparison and adjudication agent uses the large language model (the embodiment also uses the Claude large language model) to comprehensively analyze the similarity group relationship between the text content of the "Similar Goods and Services Classification Table", the similarity group information of the trademark under inspection and the candidates, and their respective explanations, and outputs the final risk level. The risk levels are categorized as follows: High risk indicates that the item to be tested and the candidate item fall into the same similar group, which is reflected in the consistency at the sub-category level, and there is a possibility of direct conflict. Medium risk indicates that the item to be tested and the candidate item are not in the same similar group, but belong to the same major category or have clear cross-group association rules to support them, usually reflected in similar uses or functions or easy association by consumers, and there is a possibility of indirect conflict. Low risk indicates that there is no direct relationship between the similar group to which the item to be tested and the candidate item belong, nor is there any clear cross-group association rule to support them, and the possibility of conflict is low.

[0094] In one possible implementation, the image feature extraction module specifically achieves refined processing of the original trademark image through three consecutive sub-processes: foreground segmentation, image-text separation, and feature vectorization.

[0095] First, foreground segmentation is performed. The original trademark image z of the trademark to be inspected is received and a rectangular box covering the main body of the image is added. Based on the GrabCut algorithm with energy minimization (which quantifies the pixel classification cost by defining an energy function and integrates pixel color consistency and spatial continuity), an energy optimization model for distinguishing foreground and background is constructed. The energy function is minimized through multiple rounds of iteration, and finally a binary foreground mask is output to achieve accurate separation of trademark foreground and background.

[0096] Next, image-text separation is performed. The optical character recognition module based on deep learning is used to locate the coordinate set B of the text region in the foreground image. On the one hand, the character shape view is cropped according to the coordinates for analyzing the font design style; on the other hand, the coordinate set B of the text region guides the image inpainting algorithm, which expands inward with the text region as the boundary and uses the image inpainting algorithm to fill the text region with content to obtain a pure graphic view.

[0097] Finally, high-dimensional feature vectors are extracted from the original image view, glyph view, and pure graphic view using a deep visual neural network, providing clean feature representations for subsequent multi-dimensional visual comparison. Feature vectorization and storage are then performed. The original image view, pure graphic view, and glyph view are input into a pre-trained deep visual neural network (in this embodiment, the multimodal visual-semantic feature extraction model CLIP) to extract corresponding high-dimensional feature vectors. These vectors are assigned unique identifiers according to the rule "Original Image ID + Content Type (i.e., original image view, pure graphic view, and glyph view)," and stored in the vector database of the multi-dimensional feature database. The distance between the high-dimensional feature vector q of the trademark to be inspected and the high-dimensional feature vector d of the base database is calculated based on the cosine similarity formula and used as the trademark image similarity.

[0098]

[0099] In one possible implementation, the construction of the multidimensional feature database is as follows:

[0100] With non-root privileges, a user-level isolated environment named "db_test" is created using Conda (running only in the current user's namespace and completely decoupled from the system-level environment). A MySQL relational database and a vector database are deployed to form a hierarchical storage architecture of "basic metadata + feature vectors" to ensure the security and isolation of data storage.

[0101] During the initialization of the base database data, the base database data in the local XLSX source file is first parsed. Basic metadata, including the standard fields and corresponding translation fields of the trademark name and rights holder name, the product / service description, and the original image view, glyph view, and pure graphic view of the original trademark image, is imported into a preset empty table structure in a MySQL relational database to complete the structured storage of basic information. Then, multi-dimensional feature extraction is performed synchronously on the base database data. The trademark name undergoes dual-track processing ("original text + translation") and multi-dimensional parsing (form, sound, and meaning) to generate a multilingual semantic vector. The rights holder name undergoes multi-channel parsing ("full name, abbreviation, and pinyin of abbreviation") and form and meaning feature extraction to generate a multi-channel feature vector. The trademark image undergoes foreground segmentation and image-text separation processing to extract multi-view feature vectors from the three types of views. These three types of feature vectors are then uniformly stored in a vector database, establishing a multi-level index structure based on multi-modal features to ensure retrieval efficiency. After the import is completed, the number of basic metadata entries in MySQL and the amount of data in the source file database are checked, and the number of feature vector entries in the vector database and the amount of data in the source file database are checked, to ensure that there are no omissions or redundancies in the two types of data, and to achieve complete initialization of the database data.

[0102] When updating early warning data, the storage architecture and deployment environment of "MySQL relational database + vector database" are fully reused from the base database, with the trademark data to be inspected and the base database coexisting in the same database instance. The basic metadata of the trademark data to be inspected (in the same format as the base database's basic metadata) is parsed and imported into the corresponding MySQL data table. At the same time, various feature vectors of the trademark data to be inspected are extracted and stored in the vector database. By adding a "data type" field for status identification, the trademark data to be inspected is clearly marked as "early warning data". This part of the data maintains strict consistency with the base database in terms of data structure, and unified storage, classification management and fast retrieval of the trademark data to be inspected are achieved only through logical partitioning.

[0103] In one possible implementation, the risk assessment of the integrated risk calculation module includes three core phases:

[0104] Part 1: Calculation of Basic Similarity between Nonlinearly Fused Image and Text Dimensions

[0105] First, based on the image feature extraction module and the name and rights holder similarity calculation module, the similarity scores of the trademark image dimension and the trademark name dimension are obtained. and (The highest similarity score in the trademark image dimension and the highest similarity score in the trademark name dimension are obtained after comparing the trademark to be inspected with the database data one by one.) To simulate the mechanism for capturing significant features in manual examination, the system does not use a linear arithmetic mean algorithm, but instead executes a non-linear fusion logic based on the maximum value. Simultaneously, weight coefficients are defined during initialization. , and ( , , The similarity between the image and text is calculated using the following formula: ):

[0106]

[0107] In this formula, The function is used to extract the maximum value from the similarity results between the trademark image dimension and the trademark name dimension to respond to the most significant infringement features; The function is used to calculate the average similarity results between the trademark image dimension and the trademark name dimension, reflecting the overall average level; This function calculates the harmonic mean of the similarity results between the trademark image dimension and the trademark name dimension, providing a conservative estimate when their differences are significant. This calculation formula ensures the accuracy of the output. It can keenly respond to extreme features in any single dimension, enabling it to derive relatively accurate image-text mixing similarity regardless of whether the confusion is due to visual graphics or auditory pronunciation.

[0108] Second: Dual-path optimization calculation of dynamic commodity coefficients. In this stage, the system abandons the static calculation method that relies solely on the commodity classification table, and instead adopts a dynamic mechanism that uses parallel calculation of "soft-gated path" and "strong feature circuit breaker path" and then selects the best output to determine the final dynamic commodity coefficients. ).

[0109] In the soft-gating path, the base value corresponding to the brand awareness of the trademark to be inspected is first predefined. (Determined based on trademark reputation; the higher the reputation, the larger the base value), the base value is used to assess the similarity of goods and services. (Similarity scores are determined by risk level; in this example, high, medium, and low risks are set to 0.9, 0.6, and 0.3, respectively.) Linear interpolation is performed to obtain soft gating coefficients, ensuring bottom-line risk in cross-class scenarios; in the circuit breaker path, the comprehensive similarity of the trademark name dimension ( ) and the predefined upgrade threshold coefficient ( Multiplying these results yields the circuit breaker upgrade coefficient, establishing a forced mapping between trademark name similarity and product relevance. Finally, the dynamic product coefficient is determined using the following optimal formula:

[0110]

[0111] This formula ensures that the system can adaptively handle different scenarios: in cross-class preemptive registration scenarios ( ), maintaining bottom-line risk through soft gating mechanisms; in the case of name confusion ( ), by forcibly raising the risk level through the circuit breaker mechanism, so that It remains at a high level.

[0112] Third: Introducing the rights holder's difference to synthesize the final score. After obtaining the image-text mixed similarity and dynamic product coefficient, the rights holder's difference is introduced (…). (Calculated based on the comprehensive similarity of the rights holder's name, the formula is as follows:) =1 - Comprehensive similarity of rights holder names) is used as the final exclusivity adjustment factor. Multiplicative logic is used to synthesize the indicators from three dimensions: image-text hybrid similarity, dynamic product coefficient, and rights holder difference. The final trademark similarity risk score is calculated using the following formula ( ):

[0113]

[0114] The physical meaning of this final comprehensive risk score calculation formula lies in: the degree of difference among rights holders ( It acts as a "logic switch," meaning that when the rights holders are the same, Approaching zero, causing Zeroing out automatically filters defensive trademarks; when the rights holders are different, the risk score is mainly based on ontology similarity ( ) and dynamic commodity coefficient ( The formula is jointly determined by both the risk assessment of cross-class risks and the priority recall of trademarks with strong similar characteristics, thus achieving accurate measurement and quantitative classification of complex trademark infringement risks.

[0115] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0116] The various embodiments in this specification are described in a related manner. Each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other.

[0117] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

Claims

1. An intelligent risk early warning system based on multi-dimensional information fusion, characterized in that, It includes a name and rights holder similarity calculation module, a trademark goods and services comparison module, an image feature extraction module, a multi-dimensional feature database, and a comprehensive risk calculation module; The name and right holder similarity calculation module is used to perform a deep comparison between the name of the trademark to be inspected and the name of the right holder with the base data in the multidimensional feature database, and calculate the corresponding similarity score. The trademark-product-service comparison module constructs a two-layer processing architecture of "vector coarse screening - intelligent agent fine ranking" to make a final similarity judgment on the product and service descriptions; The image feature extraction module performs refined processing on the original trademark image and calculates the trademark image similarity. The multidimensional feature database adopts a hybrid storage architecture of "MySQL relational database + vector database" to uniformly store the basic metadata and multidimensional feature vectors of the base database and the trademarks to be inspected, supporting millisecond-level retrieval and multimodal association analysis. The relational database stores standardized basic metadata, and the vector database stores feature vectors extracted from multiple dimensions. The comprehensive risk calculation module calculates the final risk score by combining the similarity of mixed text and images, the dynamic commodity coefficient, and the difference between rights holders.

2. The intelligent risk early warning system based on multi-dimensional information fusion according to claim 1, characterized in that, The operation of the name and rights holder similarity calculation module is as follows: For trademark names, a dual-track parallel processing mechanism of "original text + translation" is constructed. A translation model is used to generate translation fields. The translation model is a neural network translation model that can support multilingual translation while retaining the original text fields. On this basis, a weighted fusion of form similarity based on edit distance and visual hash, sound similarity based on pinyin or Romanized transliteration, and semantic similarity based on multilingual sentence vectors is used to obtain the comprehensive similarity of the trademark name. For the rights holder's name, multi-channel feature extraction is implemented. The similarity calculation of the rights holder's name is divided into three channels: full name, abbreviation, and pinyin of the abbreviation. The string edit distance and visual hash similarity of each channel are calculated separately, and the semantic similarity is calculated in combination with sentence vector encoding. Finally, the comprehensive similarity of the rights holder's name is obtained through multi-channel weighting.

3. The intelligent risk early warning system based on multi-dimensional information fusion according to claim 1, characterized in that, The operation of the trademark goods and services comparison module is as follows: A two-stage retrieval mechanism of "coarse screening-fine ranking" is adopted. First, the semantic vector model is used to vectorize and encode the descriptions of goods and services and calculate cosine similarity to complete the candidate set recall. Then, a multi-agent system is constructed, and each agent node works together to extract the rule basis of the "Similar Goods and Services Distinction Table", perform vector retrieval and semantic fine ranking, and make a final similarity judgment by integrating rule information.

4. The intelligent risk early warning system based on multi-dimensional information fusion according to claim 1, characterized in that, The image feature extraction module operates as follows: A deep learning-based image-text separation processing chain is constructed. First, the GrabCut algorithm based on energy minimization is used to segment the foreground of the original trademark image to be inspected. Second, the text region in the foreground is located by combining a deep learning-based optical character recognition module and extracted as a character view. The text region is then filled with an image inpainting algorithm to obtain a pure graphic view. Finally, high-dimensional feature vectors of the original image view, character view, and pure graphic view are extracted by a deep visual neural network and are collectively referred to as multi-view feature vectors. The obtained multi-view feature vectors are then compared with the multi-view feature vectors of the base database in the multi-dimensional feature database to calculate the similarity of the trademark image.

5. The intelligent risk early warning system based on multi-dimensional information fusion according to claim 1, characterized in that, The operation of the comprehensive risk calculation module is as follows: First, a non-linear fusion of image and text similarity is performed. For the similarity between trademark images and trademark names, a three-element fusion strategy of "maximum value as the main factor, arithmetic mean as the auxiliary factor, and harmonic mean as the backup factor" is adopted to calculate the image-text mixed similarity. Second, a dual-path optimization calculation of dynamic commodity coefficient is performed, and the "soft gating coefficient" and "circuit breaker upgrade coefficient" are calculated in parallel, and the maximum value of the two is calculated as the dynamic commodity coefficient. Finally, the difference of the rights holder is introduced as an exclusivity factor, and the final risk score is synthesized through multiplicative logic.

6. The intelligent risk early warning system based on multi-dimensional information fusion according to claim 2, characterized in that, The input data for the similarity calculation module is the trademark name and the name of the right holder of the trademark to be tested, i.e., the original text field. First, the original text field is subjected to basic normalization processing to obtain the standard field. The basic normalization processing includes null value filling, removal of leading and trailing whitespace, whitespace compression, and case normalization. The processing does not change the semantics and retains the original text field. Then, the translation big model is called to generate the corresponding translation fields TM_trans and Owner_trans for the standard fields; during translation, a structure protection and consistency control mechanism is adopted, specifically: (1) proper names, numbers and parentheses are replaced with placeholders and backfilled after translation is completed; (2) texts of the same language or length are batch-processed, and the random probability of the output is reduced by adjusting the decoding parameters to ensure consistency; (3) the placeholder integrity rate and number retention rate of the translation are checked. If they fail, they are retranslated. If they still fail the check after retranslation, they are set to empty. The similarity of the trademark name to be inspected with the form, sound, and meaning of each trademark name in the database is calculated and then fused. The form similarity is calculated only based on the standard fields of the trademark names in the database and the trademark name to be inspected. First, the normalized similarity of the string edit distance is calculated. Then, the standard fields of the trademark name in the target trademark and the base database are rendered into grayscale images using font rasterization text technology. A fixed-length fingerprint is obtained by calculating the difference hash. After performing an XOR operation on the fingerprint, the Hamming distance is calculated. Visual hash similarity is obtained; the two are then fused by coefficient to obtain shape similarity. Phonetic similarity is also calculated solely based on the standard fields of the trademark name: the standard fields of the trademark name to be examined and the trademark name in the base database are romanized and transliterated to form a pronunciation string. The normalized similarity of the edit distance is calculated on the pronunciation string to obtain the phonetic similarity. Semantic similarity calculation: The translation fields of the trademark to be examined and the base database are encoded into vectors respectively, and the cosine similarity is calculated; then the cosine similarity is mapped to... Obtain semantic similarity Simultaneously, semantic similarity is directly calculated based on standard fields of trademark names. A two-path fusion strategy is used to determine the final semantic similarity. : Finally, the similarity of form, sound, and meaning is combined according to weights to obtain the overall similarity of trademark names; the overall similarity of trademark names is sorted in descending order and output, and the output includes the similarity of each component and the overall similarity; The similarity in form and meaning between the name of the trademark owner to be inspected and each name of the owner in the database is calculated and fused. In the calculation of the similarity in form of the name of the owner, the abbreviation and the pinyin of the abbreviation are introduced into a multi-channel approach. First, the abbreviation and the pinyin of the abbreviation are generated based on the standard fields of the name of the trademark owner to be inspected. The abbreviation extraction includes removing the area segment in parentheses, removing the prefix area, stripping the organizational form suffix, and truncating before the industry keyword when there is an industry keyword to obtain the trade name. The abbreviation is transliterated into pinyin and separated by spaces to generate the pinyin abbreviation. Then, the shape similarity is calculated for the full name, abbreviation, and pinyin abbreviation channels. For each channel c, the channel shape similarity is obtained using the following formula. : in, Normalized similarity for the edit distance of the string in channel c; Let be the visual hash similarity of channel c; The final shape similarity is obtained by fusing the three channels element by element based on their maximum values. in, The similarity scores are calculated using three channels: the full name (full), the abbreviation (short), and the abbreviation (pinyin) (py). The method for calculating the semantic similarity of each channel of the rights holder's name is the same as the method for calculating the semantic similarity of the trademark name; the semantic similarity of the three channels is fused according to the maximum value of each element to obtain the final semantic similarity; The final form similarity and final meaning similarity of the right holder's name are combined on an equal basis to obtain the comprehensive similarity of the right holder's name.

7. The intelligent risk early warning system based on multi-dimensional information fusion according to claim 3, characterized in that, The trademark-goods-services comparison module is implemented as follows: First, vectorized recall is performed; a pre-trained semantic vector encoding model is used to transform the product and service descriptions of the trademarks to be inspected into dense vector representations to characterize the semantic features of the products and services; cosine similarity is calculated based on the dense vector representations to select the Top-K semantic candidates from the base database. Based on this, a multi-processing module mechanism with multi-agent collaboration is used to conduct similarity checks and risk assessments on the product and service descriptions of the trademark under inspection, and output the corresponding risk levels. Specifically: First, the semantic fine-ranking agent inputs the product and service descriptions of the trademark under inspection and the candidate set into a large language model for multiple sorting and fine-tuning of similarity scores, thereby obtaining a more reliable candidate pool and providing higher-quality input for the rule comparison stage. Subsequently, the rule extraction agent uses the large language model to find the most relevant rule paragraphs and annotations in the "Similar Goods and Services Classification Table" for the product and service descriptions of the trademark under inspection and the Top-K semantic candidates, and summarizes the corresponding similarity group information and explanations accordingly. Finally, the rule comparison and adjudication agent uses the large language model to comprehensively analyze the similarity group relationship between the two, based on the text content of the "Similar Goods and Services Classification Table", the similarity group information of the trademark under inspection and the candidates, and their respective explanations, and outputs the final risk level. The risk levels are classified as follows: high risk indicates that the item to be tested and the candidate items fall into the same similar group; medium risk indicates that the item to be tested and the candidate items are not in the same similar group, but are in the same major category or have clear cross-group association rules to support them. Low risk indicates that there is no direct correlation between the item to be tested and the candidate item to which they belong, nor is there any clear cross-group association rule to support it.

8. The intelligent risk early warning system based on multi-dimensional information fusion according to claim 1, characterized in that, The construction of the multidimensional feature database is as follows: Under non-root privileges, a user-level isolated environment named "db_test" is created using Conda, and a MySQL relational database and a vector database are deployed to form a hierarchical storage architecture of "basic metadata + feature vectors" to ensure the security and isolation of data storage. During the initialization of the base database data, the base database data in the local XLSX source file is first parsed, and the corresponding basic metadata is imported into the preset empty table structure of the MySQL relational database to complete the structured storage of basic information. The basic metadata includes standard fields and corresponding translation fields of trademark name and right holder name, product and service description and original image view, font view and pure graphic view of the original trademark image. Then, multi-dimensional feature extraction is performed synchronously on the base database data—the trademark name is processed by "original text + translation" dual-track processing and multi-dimensional analysis of form, sound, and meaning to generate multilingual semantic vectors; the rights holder's name is processed by "full name, abbreviation, and pinyin of abbreviation" multi-channel analysis and extraction of form and meaning features to generate multi-channel feature vectors; the trademark image is processed by foreground segmentation and image-text separation to extract multi-view feature vectors from three types of views; the above three types of feature vectors are uniformly stored in the vector database to establish a multi-level index structure based on multi-modal features; When updating the early warning data, the storage architecture and deployment environment of "MySQL relational database + vector database" are fully reused from the base database data. The trademark data to be inspected and the base database data coexist in the same database instance. The basic metadata of the trademark data to be inspected is parsed and imported into the corresponding data table in MySQL. At the same time, various feature vectors of the trademark data to be inspected are extracted and stored in the vector database. By adding a "data type" field for status identification, the trademark data to be inspected is clearly marked as "early warning data". This part of the data maintains strict consistency with the base database data in terms of data structure. The unified storage, classification management and fast retrieval of the trademark data to be inspected are achieved only through logical division.

9. The intelligent risk early warning system based on multi-dimensional information fusion according to claim 5, characterized in that, The risk assessment in the comprehensive risk calculation module includes three core phases: Part 1: Calculation of Basic Similarity between Nonlinearly Fused Image and Text Dimensions First, based on the image feature extraction module and the name-rights similarity calculation module, similarity scores for the trademark image dimension and trademark name dimension are obtained. and This refers to the highest similarity score in the trademark image dimension and the highest similarity score in the trademark name dimension after comparing the trademark to be inspected with the base database one by one; at the same time, the weight coefficients are initialized and defined. , and The similarity between the image and text is calculated using the following formula. : In this formula, The function is used to extract the maximum value of the similarity results between the trademark image dimension and the trademark name dimension; The function is used to calculate the average of the similarity results between the trademark image dimension and the trademark name dimension; The function is used to calculate the harmonic mean of the similarity results between the trademark image dimension and the trademark name dimension; Second: The dynamic commodity coefficient is calculated by dual-path optimization. A dynamic mechanism is adopted to select the best output after parallel calculation of "soft gated path" and "strong feature circuit breaker path" to determine the final dynamic commodity coefficient. In the soft-gating path, the base value corresponding to the brand awareness of the trademark to be inspected is first predefined. The base value is used to determine the similarity of goods and services based on risk level. Linear interpolation is performed to obtain the soft gating coefficients; in the circuit breaker path, the comprehensive similarity of the trademark name dimension is calculated. Compared with the predefined upgrade threshold coefficient Multiplying these factors yields the circuit breaker upgrade coefficient; finally, the dynamic commodity coefficient is determined using the following optimal formula. : Third: Introduce the difference in rights holders to synthesize the final score; After obtaining the image-text hybrid similarity and dynamic commodity coefficient, the rights holder difference is introduced. As the ultimate exclusive regulator, =1 - Overall similarity of rights holder names; The final trademark similarity risk score is calculated by combining three dimensions—text-image similarity, dynamic product coefficient, and rights holder difference—using multiplicative logic. : 。 10. A method for an intelligent risk early warning system based on multi-dimensional information fusion, characterized in that, The steps include the following: Step 1: Constructing a multidimensional feature database: The multidimensional feature database adopts a hybrid storage architecture of "MySQL relational database + vector database" to uniformly store the basic metadata and multidimensional feature vectors of the base database and the trademarks to be inspected, supporting millisecond-level retrieval and multimodal association analysis. The relational database stores standardized basic metadata, and the vector database stores feature vectors extracted from multiple dimensions. Step 2: Perform a deep comparison between the name of the trademark to be inspected and the name of the right holder and the base data in the multidimensional feature database, and calculate the corresponding similarity score; For trademark names, a dual-track parallel processing mechanism of "original text + translation" is constructed. Translation fields are generated using a translation model, which is a neural network translation model capable of supporting multilingual translation, while retaining the original text fields. Based on this, a weighted fusion of form similarity based on edit distance and visual hash, sound similarity based on pinyin or Romanized transliteration, and semantic similarity based on multilingual sentence vectors is used to obtain the comprehensive similarity of the trademark name. For the rights holder's name, multi-channel feature extraction is implemented, and the similarity calculation of the rights holder's name is divided into three channels: full name, abbreviation, and pinyin of the abbreviation. The string edit distance and visual hash similarity of each channel are calculated separately, and semantic similarity is calculated in combination with sentence vector encoding. Finally, the comprehensive similarity of the rights holder's name is obtained through multi-channel weighting. Step 3: Construct a two-layer processing architecture of "vector coarse screening - intelligent agent fine ranking" to make the final similarity judgment on the product and service descriptions; A two-stage retrieval mechanism of "coarse screening-fine ranking" is adopted. First, the semantic vector model is used to vectorize and encode the descriptions of goods and services and calculate cosine similarity to complete the candidate set recall. Then, a multi-agent system is constructed, and each agent node works together to extract the rule basis of the "Similar Goods and Services Distinction Table", perform vector retrieval and semantic fine ranking, and make a final similarity judgment by integrating rule information. Step 4: Perform fine-grained processing on the original trademark images and calculate the trademark image similarity; A deep learning-based image-text separation processing chain is constructed. First, the GrabCut algorithm based on energy minimization is used to segment the foreground of the original trademark image to be inspected. Second, the text region in the foreground is located by combining a deep learning-based optical character recognition module and extracted as a character view. The text region is then filled with an image inpainting algorithm to obtain a pure graphic view. Finally, high-dimensional feature vectors of the original image view, character view, and pure graphic view are extracted by a deep visual neural network and collectively referred to as multi-view feature vectors. The obtained multi-view feature vectors are then compared with the multi-view feature vectors of the base database in the multi-dimensional feature database to calculate the similarity of the trademark image. Step 5: Calculate the final risk score by combining the similarity of mixed text and images, the dynamic commodity coefficient, and the difference in rights holders; First, a non-linear fusion of image and text similarity is performed. For the similarity between trademark images and trademark names, a three-element fusion strategy of "maximum value as the main factor, arithmetic mean as the auxiliary factor, and harmonic mean as the backup factor" is adopted to calculate the image-text mixed similarity. Second, a dual-path optimization calculation of dynamic commodity coefficient is performed, and the "soft gating coefficient" and "circuit breaker upgrade coefficient" are calculated in parallel, and the maximum value of the two is calculated as the dynamic commodity coefficient. Finally, the difference of the rights holder is introduced as an exclusivity factor, and the final risk score is synthesized through multiplicative logic.

Images