An artificial intelligence-based intelligent screening method for harmful substances in consumer products

Abstract

The application discloses a kind of based on artificial intelligence's consumer goods harmful substance intelligent screening method, it is related to artificial intelligence technical field, it is difficult to effectively identify and analyze the technical problem that a large number of natural language, commercial name or commonly known as etc. non-standardized description exists in ingredient list, it is also difficult to accurately convert these fuzzy text information into the standardized terminology that can be accurately matched by regulation database;Including the following steps: the fuzzy ingredient list containing natural language or commercial name description is preprocessed by multi-task text regularization model;Then, utilize embedding semantic similarity model and the multi-evidence fusion presumption unit of improved attention mechanism, fuzzy component is intelligently analyzed and matched, and standard substance identification and confidence are output;Finally, through multi-agent risk decision framework combines real-time regulation database, dynamic compliance screening and risk assessment are carried out;The application improves the intelligent level of consumer goods safety compliance management.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to an intelligent screening method for hazardous substances in consumer products based on artificial intelligence. Background Technology

[0002] Consumer products, especially those in the chemical, toy, textile, and automotive sectors, are directly related to consumer health and the environment in terms of safety and compliance. Strict regulations restricting hazardous substances have been enacted in various regions around the world.

[0003] However, in actual supply chain management, compliance screening of products faces the following prominent technical challenges: Many ingredients in the ingredient lists provided by upstream suppliers use commercial names, common names, or vague descriptions instead of standard chemical names or CAS numbers, making it impossible for automated systems to directly and accurately match with regulatory databases. This necessitates heavy reliance on manual identification, resulting in low efficiency and a high risk of errors. Existing screening systems are mostly based on simple keyword matching or rule engines, unable to handle the contextual semantics of ingredients, product usage scenarios, and historical supply chain data. They lack the ability to reason about complex and ambiguous situations, leading to high false positive and false negative rates. Traditional risk assessments are often based on a single standard or static database, making it difficult to integrate dynamic, sometimes conflicting, evidence from multiple dimensions such as regulations, health, environment, and supply chain. They lack an intelligent, adaptive fusion decision-making mechanism, making it unable to make sensitive and accurate judgments on emerging risks or risks in specific scenarios. Existing systems lack an effective mechanism for continuously learning from user feedback and optimizing their knowledge base and models, resulting in rigid system capabilities and difficulty adapting to rapidly changing market and regulatory environments.

[0004] Therefore, an intelligent screening method for hazardous substances in consumer products based on artificial intelligence has emerged. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; to this end, this invention proposes an intelligent screening method for hazardous substances in consumer products based on artificial intelligence, to solve the following technical problem: Existing automated screening systems struggle to effectively identify and parse the numerous non-standardized descriptions in ingredient lists, such as natural language, commercial names, or common names. They also struggle to accurately and automatically convert these ambiguous and heterogeneous textual information into standard chemical identifiers that can be precisely matched to regulatory databases. Furthermore, they lack a dynamic fusion decision-making mechanism that can quantify the uncertainty of evidence, intelligently weigh conflicts, and adapt to the risk preferences of different products. This results in technical problems that make it difficult to make sensitive, accurate, and consistent risk level determinations for emerging hazardous substances or specific application scenarios.

[0006] To address the above problems, a first aspect of the present invention provides an intelligent screening method for hazardous substances in consumer products based on artificial intelligence, comprising the following steps: S1: Receive the ingredient list of the product to be screened, the ingredient list containing vague ingredient items described in natural language or commercial names; S2: Preprocess the ingredient list using a multi-task text regularization model, and then classify and archive it. S3: Analyze the fuzzy component terms based on an artificial intelligence model: S31: Utilize a deep learning industry terminology mapping library with an embedded semantic similarity calculation model for accurate matching, and output standard substance identifiers and their confidence scores; S32: If the matching fails, the improved attention mechanism multi-evidence fusion inference unit is activated to extract four types of features, including fuzzy component text features, product category prior knowledge features, component list context semantic features and supplier historical submission features. The candidate list and its inferred confidence are output through intelligent weighted fusion. S33: Submit the output results to the user for verification, and use the feedback samples to dynamically update the mapping library and model parameters; S4: Use artificial intelligence search algorithms to query real-time synchronized hazardous substance regulatory databases, use risk level assessment models to conduct compliance screening and risk assessment, and output standardized screening reports.

[0007] Preferably, in step S2, the construction and training of the multi-task text regularization model includes the following steps: The model comprises a shared character-level feature extraction layer and three cascaded task sub-networks, the task sub-networks including: Character correction and confidence generation subnet: Taking the character feature sequence output by the shared layer as input, based on a one-dimensional convolutional neural network and a conditional random field, the output is: the corrected character sequence and the error position confidence vector of the same length as the input sequence, where the confidence value of each position represents the probability that the character at that position has a spelling error; Vocabulary and Unit Parsing Subnetwork: Taking the corrected character sequence as input, the subnetwork performs word segmentation and part-of-speech tagging tasks based on a bidirectional long short-term memory network model. When calculating the hidden state at each time step, the subnetwork introduces a feature selection gating mechanism controlled by the error position confidence vector. The final output is: the segmented word sequence and the part-of-speech tag vector corresponding to each word. Semantic Structure Parsing Subnet: Taking the fusion result of word sequence and part-of-speech tag vector corresponding to each word as input, based on Transformer encoder and pointer network, it identifies and extracts three fields of structured substance name, concentration content and functional description; The model is trained end-to-end using a multi-task learning framework, and the overall loss function is obtained by weighting the loss of the three sub-network tasks and the auxiliary loss term.

[0008] Preferably, the introduction of the feature selection gating mechanism controlled by the error location confidence vector includes the following steps: For the first in the input sequence Character feature vectors at each position The corresponding error location confidence level is The feature vectors that are actually input to the bidirectional long short-term memory network unit are then... It is obtained through the gating calculation formula, specifically: ;in, The feature vectors actually input into the network after being adjusted by the gating mechanism. The index of the character's position in the sequence. For the first Error location confidence for each position For the first The original character feature vector at each position, and The weight parameter matrix, For the first Character feature vectors at each position, For the first Character feature vectors at each position.

[0009] Preferably, step S31 includes the following steps: The preprocessed fuzzy component text is input into the embedded semantic similarity calculation model, and the model outputs a high-dimensional semantic vector, denoted as the query vector. ; In parallel, the semantic vectors of the standard names and aliases of all standard substances are retrieved from the deep learning industry terminology mapping library to form a candidate vector set. ; The query vector is calculated using a multi-scale similarity fusion algorithm. With each candidate vector Match confidence between ; Select matching confidence The top K candidate standard substances with the highest confidence level were identified, and the highest confidence level was determined. Does it exceed the preset first threshold? The matching decision is executed as follows: like If the match is successful, the corresponding standard substance identifier will be output. As its confidence level; if Then, the top K candidate standard substances and their confidence levels are output as the fuzzy matching results, where, For the preset second threshold; if If the match fails, step S32 is triggered directly.

[0010] Preferably, the query vector is calculated using a multi-scale similarity fusion algorithm. With each candidate vector Match confidence between This includes the following steps: By calculating the query vector With each candidate vector Cosine similarity between them is used as global semantic similarity. ,in, Let L2 be the norm of the vector; Computational domain enhanced similarity Constructing the domain subspace projection matrix The column vectors of the projection matrix are formed by orthogonalizing the semantic vectors of K pre-selected high-frequency core terms in the field of consumer product chemistry; the query vector With candidate vectors Projecting each vector onto the respective neighborhood subspace yields a projection vector. and ,in, Projection matrix Transpose of the vector; calculate the normalized similarity between the two projected vectors. ,in, Projection vector and The Euclidean distance between them These are the preset scaling parameters; Dynamically determine the fusion weight based on the characteristics of the query text. The matching confidence level is obtained. Among them, fusion weight The length and term density features of the query text are extracted and then calculated using a preset nonlinear mapping function. Based on the calculated matching confidence Sets, perform matching decisions and output results.

[0011] Preferably, in step S32, four types of features are extracted, including the following steps: The extraction of the fuzzy component text features involves inputting the fuzzy component into a pre-trained language model encoder to obtain its context-dependent semantic representation, which serves as the fuzzy component text feature vector. ; The extraction of prior knowledge features of the product category involves retrieving common hazardous substances and compliance history of the corresponding product category from a historical database based on the preset category label of the product to be screened, and statistically obtaining the prior probability distribution vector of the substances in the category, which serves as the prior knowledge feature vector of the product category. ; The extraction of the contextual semantic features of the component list involves extracting several explicit components adjacent to the fuzzy component item from the component list, and performing a pooling operation on the semantic vectors of each explicit component to generate a contextual semantic feature vector for the component list. ; The extraction of supplier historical submission features involves querying the historical submission records of suppliers providing the ingredient list, statistically analyzing the frequency distribution of all fuzzy descriptions submitted by the suppliers that were ultimately verified as belonging to each standard substance, and converting this distribution into a probability vector, which serves as the supplier historical submission feature vector. .

[0012] Preferably, the four types of features are used to output a candidate list and its estimated confidence level through intelligent weighted fusion, including the following steps: S3201: For the four types of feature vectors extracted , , and Calculate the corresponding uncertainty scores for each. , , and ; S3202: The four types of feature vectors and their corresponding uncertainty scores are input together into an improved cross-attention layer to generate a fused context vector. The improved cross-attention layer specifically refers to: Through the weight matrix Perform a linear transformation to generate a query vector. ; four types of feature vectors Through the weight matrix respectively and Perform a linear transformation to generate the corresponding key vector. AND value vector ; Calculate the first Attention contribution weights of class features ,in, ,in, For unnormalized attention scores, Let be the dimension of the key vector. The predetermined uncertainty penalty coefficient, For the first Uncertainty scores of class features; final fused context vector ; Using the context vector, an original matching score is calculated for each candidate standard substance; simultaneously, a confidence correction for counterfactual intervention is constructed.

[0013] Preferably, the confidence correction for constructing counterfactual intervention includes the following steps: The context vector With candidate substances semantic embedding vector Perform similarity calculation to obtain a matching score. ; Creating counterfactual feature vectors Specifically, this involves: extracting the original supplier's historical submission feature vector. Replace with an all-zero vector or a uniformly distributed probability vector; Using the counterfactual feature vector Replace the original Repeat step S3202 to obtain the counterfactual fusion context vector. To calculate candidate substances Counterfactual matching score ; Calculate the correction amount and generate the final confidence level, where the corrected estimated confidence level is specifically as follows: ,in, The preset correction strength coefficient; Based on the corrected estimated confidence level, all candidate standard substances are sorted in descending order, and the top M substances and their corresponding values ​​are output. The values ​​are used as the final candidate list.

[0014] Preferably, in step S4, the risk level assessment model adopts a risk decision-making framework based on multi-agent negotiation and evidence theory, including the following steps: Four intelligent agents with different risk dimensions are deployed, including regulatory compliance intelligent agents, health hazard intelligent agents, environmental accumulation intelligent agents, and supply chain exposure intelligent agents. Each intelligent agent encapsulates an analysis model and knowledge base for its risk dimension. Each agent independently calculates and analyzes the input standardized substance list and its concentrations, combining a real-time regulatory database and a professional risk knowledge base, and outputs its basic reliability allocation for the product's risk level in this dimension. , , , and These represent the level of trust in four risk levels: low, medium, high, and severe. Assign an ID to the agent; calculate the confidence anchor value for each agent. ,in, The peak probability predicted by the agent model. To score the quality of the input data, This represents the historical accuracy of the intelligent agent compared to similar products. , and , where is the weighting coefficient; an improved evidence fusion rule is used to perform pairwise fusion of the basic confidence assignments of the agent's output.

[0015] Preferably, the step of pairwise fusion of the basic confidence assignment of the agent's output using improved evidence fusion rules includes the following steps: When fused intelligent agents and At that time, calculate the conflict metric. ,in For a certain risk level, , For the first Individual agents assess risk levels Trust level value, For the first Individual agents assess risk levels Trust level value; if > If a preset conflict threshold is set, a conflict buffer will be created. ,in, The preset buffer coefficient, For intelligent agents The confidence anchor value, For intelligent agents The confidence anchor value; Calculate agent fusion weights and ; Will and The initial weighted fusion of the individual agents yields the following results: The conflict buffer is redistributed according to a preset risk compensation coefficient to obtain the final fusion confidence allocation. ,in, Risk level The compensation coefficient; for Normalization is performed, and the fusion result is regarded as the output of the new agent. The preliminary weighted fusion calculation is repeated until all agent evidence is fused to obtain the comprehensive confidence allocation. Based on the final comprehensive confidence level assignment, the risk level with the highest confidence level is selected as the final judgment, and a traceable decision report is generated. The report includes the original confidence level assignment of each agent, confidence anchor value, conflict metric value, and intermediate results in the fusion process.

[0016] The beneficial effects of this invention are: This invention constructs a multi-task text regularization model, especially by introducing a feature selection gating mechanism that controls the confidence of error locations. This effectively corrects spelling errors, resolves fuzzy structures, and automatically transforms non-standard natural language descriptions into structured "substance-concentration-function" information. This invention constructs a multi-scale similarity fusion algorithm, which combines global semantics and domain knowledge, to more accurately distinguish semantically similar business names and standard substances. At the same time, based on uncertainty-guided attention fusion and counterfactual intervention correction mechanism, when fusing text, category, context and supplier features, it can quantify the uncertainty of each evidence source and intelligently suppress the influence of low-quality or biased evidence, making the matching reasoning process more reliable. This invention constructs a multi-agent risk decision-making framework and improves evidence fusion rules, decomposing risk assessment into multiple professional dimensions (regulation, health, environment, and supply chain). It also innovatively introduces mechanisms such as "confidence anchors," "conflict buffers," and "risk compensation coefficients," which can not only effectively quantify and resolve judgment conflicts between different agents, but also dynamically adjust risk preferences according to product type. Finally, it outputs a comprehensive risk assessment result based on consensus and with traceable process, making the decision-making process more transparent and reasonable. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 As shown, this invention is an intelligent screening method for hazardous substances in consumer products based on artificial intelligence, comprising the following steps: S1: Receive the ingredient list of the product to be screened, the ingredient list containing vague ingredient items described in natural language or commercial names; S2: Preprocess the ingredient list using a multi-task text regularization model, and then classify and archive it. S3: Analyze the fuzzy component terms based on an artificial intelligence model: S31: Utilize a deep learning industry terminology mapping library with an embedded semantic similarity calculation model for accurate matching, and output standard substance identifiers and their confidence scores; S32: If the matching fails, the improved attention mechanism multi-evidence fusion inference unit is activated to extract four types of features, including fuzzy component text features, product category prior knowledge features, component list context semantic features and supplier historical submission features. The candidate list and its inferred confidence are output through intelligent weighted fusion. S33: Submit the output results to the user for verification, and use the feedback samples to dynamically update the mapping library and model parameters; S4: Use artificial intelligence search algorithms to query real-time synchronized hazardous substance regulatory databases, use risk level assessment models to conduct compliance screening and risk assessment, and output standardized screening reports.

[0020] Specifically, the system receives a list of product ingredients to be screened from enterprise ERP systems, supply chain platforms, or manually uploaded documents. This list is typically a structured or semi-structured document, characterized by a large number of ingredient items provided by suppliers and described in natural language or internal commercial names, such as "antioxidant BHT" or "XX company's special fabric softener." The list is then input into a pre-trained multi-task text regularization model, which performs the following core operations: character-level error correction: automatically detects and corrects spelling errors and non-standard abbreviations; word and unit analysis: performs word segmentation, identifies and separates substance names, concentration values, and units of measurement; and semantic structure analysis: identifies the "substance name," "concentration / content," and "functional description" for each ingredient item. After processing the three standardized fields, the system archives the cleaned structured data according to product category, supplier, and other information to prepare for subsequent parsing. Step S3 aims to map non-standard component descriptions to a standard substance database: First, the preprocessed component "substance name" field is input into an embedded semantic similarity calculation model. This model converts it into a high-dimensional semantic vector (query vector) and compares it with the semantic vectors of all standard substance names and their aliases in a deep learning industry terminology mapping library. The matching degree is calculated through a multi-scale similarity fusion algorithm. If the highest matching confidence exceeds a preset high threshold, the corresponding standard substance identifier and the confidence level are directly output, indicating successful parsing. If S31 fails to match (highest confidence is below the threshold), the system will automatically retrieve the matching information. The system then activates a multi-evidence fusion inference unit with an improved attention mechanism. This unit extracts four types of features in parallel. The system intelligently fuses these features through a weighted network that integrates uncertainty quantification and attention mechanisms, inferring one or more most likely standard substance candidates and providing an inferred confidence level to form a candidate list. The output of S31 or S32 (standard substance candidates and confidence levels) is provided to the user (e.g., a quality inspector) for final verification or selection. The system automatically sends the user-confirmed correct matching pairs (fuzzy description - standard substance) back to the background learning engine as high-quality training samples for incrementally updating the industry terminology mapping library, optimizing the semantic similarity calculation model, and the parameters of the multi-evidence fusion inference unit, thus achieving continuous self-optimization of the system. Based on the standardized substance list and concentrations parsed by S3, the final assessment stage begins: Using an intelligent retrieval algorithm, the system queries a real-time synchronized hazardous substance regulatory database to obtain applicable limits, prohibitions, or restrictions for each substance in different target markets. The substance information and regulatory requirements are then input into a risk level assessment model. This model employs a multi-agent framework, independently assessing and fusing evidence from multiple dimensions, including regulatory compliance, health hazards, environmental risks, and supply chain stability, ultimately generating a comprehensive risk level. The system automatically integrates all intermediate results and the final judgment to generate a clearly structured, evidence-complete standardized screening report, typically including substance compliance status, risk level, specific violations, and improvement suggestions for decision-making. In this embodiment, the system refers to a software system or hardware / software integrated platform that implements the method described in this invention.

[0021] In one embodiment of the present invention, step S2, the construction and training of the multi-task text regularization model, includes the following steps: The model comprises a shared character-level feature extraction layer and three cascaded task sub-networks, the task sub-networks including: Character correction and confidence generation subnet: Taking the character feature sequence output by the shared layer as input, based on a one-dimensional convolutional neural network and a conditional random field, the output is: the corrected character sequence and the error position confidence vector of the same length as the input sequence, where the confidence value of each position represents the probability that the character at that position has a spelling error; Vocabulary and Unit Parsing Subnetwork: Taking the corrected character sequence as input, the subnetwork performs word segmentation and part-of-speech tagging tasks based on a bidirectional long short-term memory network model. When calculating the hidden state at each time step, the subnetwork introduces a feature selection gating mechanism controlled by the error position confidence vector. The final output is: the segmented word sequence and the part-of-speech tag vector corresponding to each word. Semantic Structure Parsing Subnet: Taking the fusion result of word sequence and part-of-speech tag vector corresponding to each word as input, based on Transformer encoder and pointer network, it identifies and extracts three fields of structured substance name, concentration content and functional description; The model is trained end-to-end using a multi-task learning framework, and the overall loss function is obtained by weighting the loss of the three sub-network tasks and the auxiliary loss term.

[0022] Specifically, the model uses a shared character-level feature extraction layer as the input interface. This layer receives the original text character sequence that may contain spelling errors, converts it into a continuous character feature vector sequence, and then the model performs collaborative processing through three cascaded task sub-networks to form a hierarchical parsing pipeline of "character-vocabulary-semantics". A character correction and confidence generation subnet is constructed. This subnet takes the character feature sequence output by the shared layer as input. Its core consists of a one-dimensional convolutional neural network (1D-CNN) and a conditional random field (CRF) layer. The 1D-CNN is used to capture the local contextual patterns of characters to initially judge the correctness of characters at each position. The CRF layer optimizes the coherence of character prediction from the perspective of the whole sequence. For example, it ensures that the corrected character sequence can form a reasonable word. This subnet finally outputs two results in parallel: the corrected character sequence and the confidence vector of the error position. This is a vector with the same length as the input sequence. The value of each position (between 0 and 1) represents the probability that the model judges that the original character at that position has a spelling error. For example, a confidence value of 0.9 indicates that the character is very likely to be wrong. A vocabulary and unit parsing subnet is constructed. This subnet takes the corrected character sequence output by the character error correction subnet as input. Its backbone is a bidirectional long short-term memory network (Bi-LSTM) used to perform word segmentation and part-of-speech tagging tasks. The subnet finally outputs the segmented word sequence and the part-of-speech tag vector corresponding to each word (e.g., noun-substance, number-concentration, noun-unit, etc.). A semantic structure parsing subnet is constructed. The input of this subnet is the fusion result of the word sequence output by the lexical parsing subnet and its corresponding part-of-speech tag vector. This fusion is usually achieved by concatenating the word embedding vector with its part-of-speech tag vector. The subnet uses a Transformer encoder to model the global dependencies between words and understand the syntactic structure of the entire constituent description. Subsequently, a pointer network is connected. This network is trained to "point" to and extract the start and end positions of text fragments belonging to the three fields of "substance name", "concentration / content" and "functional description" from the input sequence, thereby completing the final structured information extraction. The three sub-networks mentioned above are placed within a multi-task learning framework for end-to-end joint training. The training data consists of a large number of manually labeled component lists, each labeled with the correct character sequence, word segmentation boundaries, part-of-speech tags, and the boundaries of the three fields. The overall loss function of the model is... ,in, , and The cross-entropy losses are calculated for the character error correction and confidence generation subnet, the vocabulary and unit parsing subnet, and the semantic structure parsing subnet, respectively. , and The weighting coefficients are determined using the analytic hierarchy process (AHP). In this embodiment, , and The values ​​correspond to 0.4, 0.4 and 0.2 respectively.

[0023] In one embodiment of the present invention, the introduction of a feature selection gating mechanism controlled by an error location confidence vector includes the following steps: For the first in the input sequence Character feature vectors at each position The corresponding error location confidence level is The feature vectors that are actually input to the bidirectional long short-term memory network unit are then... It is obtained through the gating calculation formula, specifically: ;in, The feature vectors actually input into the network after being adjusted by the gating mechanism. The index of the character's position in the sequence. For the first Error location confidence for each position For the first The original character feature vector at each position, and The weight parameter matrix, For the first Character feature vectors at each position, For the first Character feature vectors at each position.

[0024] Specifically, for the corrected character sequence to be processed, the model has obtained: for each position The original character feature vector Extracted from the upstream network, each location Corresponding error location confidence Provided by the character error correction and confidence generation subnet, in which, The closer the value is to 1, the higher the probability that the original character at that position is incorrect, and the less reliable the feature information it provides; for the th in the sequence At this position, the model does not directly... Instead of inputting into the Bi-LSTM unit, an adjusted feature vector is first calculated according to the following gating formula. , , as the actual input of the Bi-LSTM at this time step, where, This represents the component that preserves the original features of the current location. The smaller the value (higher confidence, more reliable characters), the larger the proportion of this item, indicating that the model primarily trusts it. ,and This represents the compensating component for contextual features. The larger the value (low confidence level, suspicious characters), the larger the proportion of this item. and These are two trainable weight parameter matrices, whose function is to combine the character feature vectors of the left and right adjacent positions. and Perform a linear transformation to extract useful information applicable to the current context. This constitutes a contextual substitution feature based on the left and right adjacent characters; The working principle of this gating mechanism can be compared to an adaptive switch: when When, it represents a character Basically correct, the formula simplifies to: The model directly uses the character's own features for subsequent lexical analysis; when When, it represents a character It is highly likely to be incorrect; the formula simplifies to... The model will ignore this unreliable character feature and instead infer the word information that should be present at the current position based on the features of its left and right neighbors. The calculated The sequence is used as input and fed into the subsequent Bi-LSTM network to perform standard forward propagation to complete word segmentation and part-of-speech tagging tasks. During model training, the weight matrix... and Along with other model parameters, it is learned and optimized through backpropagation, a mechanism that enables the entire model to learn how to respond to "error alerts" given from upstream (…). By dynamically adjusting the dependency strategy on different positional features, the system's robustness in accurately parsing words in texts with spelling noise is significantly enhanced.

[0025] In one embodiment of the present invention, step S31 includes the following steps: The preprocessed fuzzy component text is input into the embedded semantic similarity calculation model, and the model outputs a high-dimensional semantic vector, denoted as the query vector. ; In parallel, the semantic vectors of the standard names and aliases of all standard substances are retrieved from the deep learning industry terminology mapping library to form a candidate vector set. ; The query vector is calculated using a multi-scale similarity fusion algorithm. With each candidate vector Match confidence between ; Select matching confidence The top K candidate standard substances with the highest confidence level were identified, and the highest confidence level was determined. Does it exceed the preset first threshold? The matching decision is executed as follows: like If the match is successful, the corresponding standard substance identifier will be output. As its confidence level; if Then, the top K candidate standard substances and their confidence levels are output as the fuzzy matching results, where, For the preset second threshold; if If the match fails, step S32 is triggered directly.

[0026] Specifically, the "substance name" text (e.g., "antioxidant BHT") from the fuzzy component items preprocessed and structured in step S2 is input into a pre-trained embedding semantic similarity calculation model. This model is typically fine-tuned based on pre-trained language models such as BERT or Sentence-BERT, and can encode the input text into a fixed-dimensional high-dimensional semantic vector, which is denoted as the query vector. The system represents the overall semantics of the input text; the system accesses a deep learning industry terminology mapping library in parallel, which stores the official names, common names, and common trade names of all known standard substances and their corresponding pre-calculated semantic vectors. The system retrieves the semantic vectors corresponding to all N candidate substances from this library, forming a candidate vector set; the system uses a multi-scale similarity fusion algorithm to calculate the matching degree between the query vector and each candidate vector; the system sorts all candidate substances in descending order according to their matching confidence, takes the top K (e.g., K=5) as high-probability candidates, and performs matching decisions, specifically: 1. Exact matching: determine the highest confidence level. ,like (Based on historical data analysis), if the match is considered to have extremely high certainty, it is determined to be a perfect match, and the system directly outputs the corresponding standard substance identifier and... As its confidence level, the process ends; 2. Fuzzy matching to be inferred: if If multiple possibilities exist or the matching evidence is insufficient, the system outputs the top K candidate standard substances and their confidence levels as the fuzzy matching result. This result is then used as one of the important inputs to the multi-evidence fusion inference unit in subsequent step S32 to initiate more complex reasoning. 3. Matching failure: A second threshold is preset (again, obtained by analyzing historical data); If the current semantic database does not contain enough similar candidates, the matching will fail. The system will skip the current step and directly trigger step S32, relying entirely on multi-evidence fusion for inference. Through this hierarchical decision-making mechanism, the system achieves a balance between efficiency and accuracy: it responds quickly to explicit matches, initiates deep reasoning for ambiguous cases, and avoids unnecessary calculations for cases without matches.

[0027] In one embodiment of the present invention, the query vector is calculated using a multi-scale similarity fusion algorithm. With each candidate vector Match confidence between This includes the following steps: By calculating the query vector With each candidate vector Cosine similarity between them is used as global semantic similarity. ,in, Let L2 be the norm of the vector; Computational domain enhanced similarity Constructing the domain subspace projection matrix The column vectors of the projection matrix are formed by orthogonalizing the semantic vectors of K pre-selected high-frequency core terms in the field of consumer product chemistry; the query vector With candidate vectors Projecting each vector onto the respective neighborhood subspace yields a projection vector. and ,in, Projection matrix Transpose of the vector; calculate the normalized similarity between the two projected vectors. ,in, Projection vector and The Euclidean distance between them These are the preset scaling parameters; Dynamically determine the fusion weight based on the characteristics of the query text. The matching confidence level is obtained. Among them, fusion weight The length and term density features of the query text are extracted and then calculated using a preset nonlinear mapping function. Based on the calculated matching confidence Sets, perform matching decisions and output results.

[0028] Specifically, for a query vector and any candidate vector, their global semantic similarity is first calculated. This is achieved by calculating the cosine similarity between the two vectors, where the global semantic similarity ranges from [-1, 1]. In this application scenario, vectors encoded by the semantic model are usually located in the positive value region. The closer the value is to 1, the more similar the overall meaning of the two texts is in the general semantic space. To enhance the semantic discrimination capability within a professional domain, this invention introduces domain-enhanced similarity, the calculation process of which is as follows: K consumer chemical... In this embodiment, "high-frequency core terms" are defined as professional terms that appear in the top 5% of the industry corpus and are strongly related to hazardous substances, material properties, and safety attributes (e.g., "phthalates," "flame retardants," "volatile organic compounds," and "migration"). These terms are input into a semantic model to obtain their semantic vectors, which are then subjected to Schmitt orthogonalization to form a set of orthogonal basis vectors. These basis vectors are used as columns to construct a domain subspace projection matrix. The query vector and candidate vectors are projected onto the domain subspace to obtain the projection vectors. and This operation is equivalent to retaining only the semantic components related to the core concepts of the domain in the original vectors; the normalized similarity between the two projected vectors is calculated using a distance-based similarity transformation formula, where, The preset scaling parameter is used to control the sensitivity of distance to the influence of similarity. The formula uses a variant of Cauchy distribution or inverse distance weight to ensure that the domain similarity approaches 1 as the projected distance approaches 0. Among them, the fusion weights are determined dynamically. Specifically, for the input fuzzy component text, two key features are extracted: length feature. (Number of characters in the text) and term density features (The proportion of words in the text that belong to the predefined "high-frequency core terms" set to the total number of words), will be used to construct the feature vector. Input a predefined non-linear mapping function, which can be a simple lookup table or a lightweight neural network (e.g., a two-layer perceptron), designed to satisfy the following property: when the query text is short ( Small) or dense with professional terms ( When (large), the output is smaller. The value is higher because the amount of information being queried is limited or highly specialized, and therefore more reliance should be placed on domain knowledge; conversely, a larger value should be output. The value depends more on global semantics. The value is constrained to ; Finally, the matching confidence scores of the query vector and the candidate vectors are obtained by weighted summation. The system performs the above calculation on all candidate vectors in the candidate set to obtain the matching confidence score set. Subsequently, the system makes a decision on this set based on a preset threshold, outputting an exact matching result, a fuzzy matching candidate list, or triggering a multi-evidence fusion inference.

[0029] In one embodiment of the present invention, step S32 involves extracting four types of features, including the following steps: The extraction of the fuzzy component text features involves inputting the fuzzy component into a pre-trained language model encoder to obtain its context-dependent semantic representation, which serves as the fuzzy component text feature vector. ; The extraction of prior knowledge features of the product category involves retrieving common hazardous substances and compliance history of the corresponding product category from a historical database based on the preset category label of the product to be screened, and statistically obtaining the prior probability distribution vector of the substances in the category, which serves as the prior knowledge feature vector of the product category. ; The extraction of the contextual semantic features of the component list involves extracting several explicit components adjacent to the fuzzy component item from the component list, and performing a pooling operation on the semantic vectors of each explicit component to generate a contextual semantic feature vector for the component list. ; The extraction of supplier historical submission features involves querying the historical submission records of suppliers providing the ingredient list, statistically analyzing the frequency distribution of all fuzzy descriptions submitted by the suppliers that were ultimately verified as belonging to each standard substance, and converting this distribution into a probability vector, which serves as the supplier historical submission feature vector. .

[0030] Specifically, the text of the fuzzy component to be parsed (e.g., "high-temperature resistant plasticizer") is input into a pre-trained language model encoder (e.g., the encoder part of BERT or RoBERTa). Based on its knowledge learned on large-scale corpora, the encoder generates a vector representation containing contextual information for each input token. Typically, the vector corresponding to the [CLS] token output by the encoder is taken, or the output vectors of all tokens in the sequence are averaged and pooled to obtain the global semantic representation of the fuzzy component. This representation is a high-dimensional dense vector, denoted as the text feature vector. It encapsulates the core semantics of the description itself; The system queries the historical compliance screening database based on the preset category labels of the products to be screened. This database stores the categories, components, and final compliance judgments of all previously screened products. The system retrieves all historical product records belonging to the same category, calculates the frequency of each standard substance appearing in these historical products, and calculates the proportion of substances that have been judged as "risk substances" (e.g., exceeding limits, restricted, or prohibited) in historical records. The system combines these statistical results into a probability distribution vector. For example, each dimension of the vector corresponds to a common standard substance, and its value can be the prior probability of that substance appearing in this type of product, or a comprehensive score combining frequency and risk weight. This vector serves as the prior knowledge feature vector of the product category. It reflects domain experience regarding "what these types of products typically use, and which of them are more likely to be risky"; From the current list of components to be screened, the focus is on the surrounding environment of the ambiguous component to be parsed: The ambiguous component is located in the list, and its preceding m and following n (e.g., m=n=2) clearly identified or structurally distinct adjacent components are extracted. These adjacent clearly identified components (e.g., "calcium carbonate," "titanium dioxide," etc.) are also converted into their respective semantic vectors using a pre-trained language model encoder. Then, pooling operations (max pooling or average pooling) are performed on the semantic vectors of these adjacent items. Max pooling captures the most prominent semantic features in the context, while average pooling reflects the overall semantic atmosphere of the context. The combined vector obtained after the pooling operation is the semantic feature vector of the component list context. It characterizes the "chemical environment" in which the fuzzy term is located. For example, if the surrounding area is full of pigment components, then the probability that the fuzzy term is a certain pigment additive is greatly increased. The system tracks and learns each supplier's data submission habits: Based on the supplier identifier providing this list, it queries all past ingredient list history submitted by that supplier. From this history, it extracts all records that were previously marked as "fuzzy descriptions" and subsequently verified manually (or determined by the system with high confidence) to correspond to a standard substance. It counts the number of times each specific fuzzy description used by the supplier (e.g., "our company's special additive A") is ultimately verified as various different standard substances. For the currently unparsed fuzzy ingredient item, it checks whether the supplier has used the same or very similar descriptions in the past. If so, it forms a probability vector based on the distribution of historical verification results, where each dimension represents the historical conditional probability of the fuzzy description being verified as a certain standard substance. If no completely identical descriptions are found, the overall distribution of all fuzzy description verification results can be used, or a weighted statistical analysis based on the semantic similarity of the descriptions can be used to obtain a vector reflecting the supplier's overall word choice preferences and accuracy. This vector is the supplier's historical submission feature vector. It encodes habitual information from specific data sources; for example, a supplier might habitually use "stabilizer" to refer to "organotin stabilizer".

[0031] In one embodiment of the present invention, the four types of features are used to output a candidate list and its estimated confidence level through intelligent weighted fusion, including the following steps: S3201: For the four types of feature vectors extracted , , and Calculate the corresponding uncertainty scores for each. , , and ; S3202: The four types of feature vectors and their corresponding uncertainty scores are input together into an improved cross-attention layer to generate a fused context vector. The improved cross-attention layer specifically refers to: Through the weight matrix Perform a linear transformation to generate a query vector. ; four types of feature vectors Through the weight matrix respectively and Perform a linear transformation to generate the corresponding key vector. AND value vector ; Calculate the first Attention contribution weights of class features ,in, ,in, For unnormalized attention scores, Let be the dimension of the key vector. The predetermined uncertainty penalty coefficient, For the first Uncertainty scores of class features; final fused context vector ; Using the context vector, an original matching score is calculated for each candidate standard substance; simultaneously, a confidence correction for counterfactual intervention is constructed.

[0032] Specifically, the process uses the extracted four types of feature vectors as input: before fusion, the credibility of each type of feature is evaluated, and for each type of feature vector... Calculate its corresponding uncertainty score. The uncertainty score: for features extracted by the neural network The uncertainty score is calculated by applying random Dropout during multiple forward propagations and calculating the variance of the feature vectors; for statistically extracted features... , and The information entropy of its probability distribution vector or the width of the confidence interval calculated based on historical data is used as the uncertainty score; The four types of feature vectors and their uncertainty scores are input into an improved cross-attention layer to generate a comprehensive fused context vector with text features. As the "focus" of the current query, through a trainable weight matrix Generate query vectors: For four types of feature vectors ,in, ; respectively through trainable weight matrices and Perform a linear transformation to generate the corresponding key vector. AND value vector ; Calculate the query vector and each key vector Unnormalized attention score ,in, It is the square root of the dimension of the key vector (scaling factor). The preset uncertainty penalty coefficient is used to control the degree to which uncertainty inhibits attention allocation; The larger (the more uncertain), The lower the score, the more likely it is to be penalized; for four The scores are then Softmax normalized to obtain the attention contribution weights for each class of features. The final fusion context vector is obtained by weighted summation of the value vectors. This vector integrates information from all features, and the contribution of features with high uncertainty in the fusion is adaptively reduced.

[0033] In one embodiment of the present invention, the confidence correction for constructing counterfactual intervention includes the following steps: The context vector With candidate substances semantic embedding vector Perform similarity calculation to obtain a matching score. ; Creating counterfactual feature vectors Specifically, this involves: extracting the original supplier's historical submission feature vector. Replace with an all-zero vector or a uniformly distributed probability vector; Using the counterfactual feature vector Replace the original Repeat step S3202 to obtain the counterfactual fusion context vector. To calculate candidate substances Counterfactual matching score ; Calculate the correction amount and generate the final confidence level, where the corrected estimated confidence level is specifically as follows: ,in, The preset correction strength coefficient; Based on the corrected estimated confidence level, all candidate standard substances are sorted in descending order, and the top M substances and their corresponding values ​​are output. The values ​​are used as the final candidate list.

[0034] Specifically, the input consists of the original fusion context vector, the semantic embedding vectors of each candidate substance, and the original features; for each candidate standard substance to be evaluated... , and its pre-computed semantic embedding vector (From the industry terminology mapping library) and the original fusion context vector generated in step S3202 The original matching score is obtained by performing similarity calculation. In this embodiment, cosine similarity is used for similarity calculation, and the original matching score is used. This represents the candidate material supported by the fusion of multiple pieces of evidence (including supplier characteristics). The original propensity score matched with the current fuzzy component; in order to quantify the supplier's historical characteristics. To determine the independent impact of the current matching decision, a counterfactual scenario needs to be constructed, and the system creates a counterfactual feature vector. Specifically, in this embodiment, a typical implementation is to... Setting it to an all-zero vector simulates the scenario of "having absolutely no prior knowledge of the supplier's historical habits." Another implementation is to replace it with a neutral prior vector, such as a uniformly distributed probability vector, representing no prior preference for the supplier's behavior. The construction of this system cuts off the current inference's reliance on historical data from specific suppliers; Using the built Completely replace the original Meanwhile, keeping the other three types of features unchanged, and then repeating step S3202 (uncertainty-guided cross-attention fusion), specifically: using , , and As input, the uncertainty of the four types of features is recalculated, where, because As a neutral vector, its uncertainty is typically set to 0 or a baseline value; the uncertainty-guided cross-attention calculation is re-performed to generate a counterfactual fusion context vector. This vector represents the comprehensive semantic representation without considering specific supplier biases. Same candidate position semantic embedding vector The similarity is calculated again to obtain the counterfactual matching score. ; Compare the original matching score with the counterfactual matching score, and the difference between them. Here, the difference is denoted as This can be approximated as an "intervention" of the supplier's historical characteristics on candidate substances. The "causal effect" of matching scores: if This indicates that the presence of supplier characteristics boosted the substance's score, which could stem from reasonable supplier practices or some form of data bias; if This indicates that supplier characteristics contribute neither positively nor negatively to the matching of the substance. To be conservative and to mitigate potential overestimation of confidence due to supplier data bias, the system uses the following correction formula to calculate the final estimated confidence level. , This means that correction is only considered necessary if the supplier's characteristics positively improve the score; if the improvement is negative or zero, no deduction is made. The preset correction strength coefficient has a value range of [value range missing]. , This indicates that the original result is completely trusted and no correction is made. This indicates that all positive gains from supplier characteristics are subtracted; typically, an intermediate value (e.g., 0.3-0.6) is taken to achieve a moderate reduction of bias rather than a complete rejection; the corrected confidence level retains support from other reliable evidence (text, categories, context) while suppressing unreasonable gains that may be introduced by a single supplier data source. After calculating the corrected presumed confidence scores for all candidate standard substances, the system sorts them in descending order of the values. The top M substances (e.g., M=3 or 5), along with their corrected presumed confidence scores, are selected as the final candidate list. This list represents a more reliable inference result after multi-evidence fusion and anti-bias correction, for users to use for final verification or decision-making.

[0035] In one embodiment of the present invention, step S4, the risk level assessment model, adopts a risk decision-making framework based on multi-agent negotiation and evidence theory, and includes the following steps: Four intelligent agents with different risk dimensions are deployed, including regulatory compliance intelligent agents, health hazard intelligent agents, environmental accumulation intelligent agents, and supply chain exposure intelligent agents. Each intelligent agent encapsulates an analysis model and knowledge base for its risk dimension. Each agent independently calculates and analyzes the input standardized substance list and its concentrations, combining a real-time regulatory database and a professional risk knowledge base, and outputs its basic reliability allocation for the product's risk level in this dimension. , , , and These represent the level of trust in four risk levels: low, medium, high, and severe. Assign an ID to the agent; calculate the confidence anchor value for each agent. ,in, The peak probability predicted by the agent model. To score the quality of the input data, This represents the historical accuracy of the intelligent agent compared to similar products. , and , where is the weighting coefficient; an improved evidence fusion rule is used to perform pairwise fusion of the basic confidence assignments of the agent's output.

[0036] Specifically, the system deploys four independent agents, each focusing on a different risk dimension: Regulatory Compliance Agent: This agent encapsulates a matching and logical judgment model based on a rule engine and a real-time regulatory database, focusing on determining whether a substance's concentration violates the specific regulatory limits of the target market; Health Hazard Agent: This agent encapsulates an assessment module based on a toxicology database (e.g., GHS classification) and a dose-response model, focusing on assessing the potential health hazards (carcinogenicity, reproductive toxicity) of a substance; Environmental Accumulation Agent: This agent encapsulates an analysis model based on substance degradation, bioaccumulation (BCF), and ecotoxicity data, focusing on assessing the long-term environmental impact of a substance; Supply Chain Exposure Agent: This agent encapsulates an analysis model based on product usage scenarios (e.g., chewing on children's toys, evaporation in enclosed spaces of car air fresheners), supply chain stability, and historical recall data, focusing on assessing actual exposure risks and supply chain controllability. Each agent maintains its own dedicated knowledge base and trained analysis model. Each agent receives the same input: a standardized list of substances and their corresponding concentration information obtained from step S3. Each agent independently runs its internal analysis model, combining a real-time synchronized regulatory database with its own specialized risk knowledge base. Output format: Each agent... (in, The output of each of the four agents (corresponding to the four agents mentioned above) is a basic confidence assignment vector. ,in, , , and These represent the level of trust (reliability scores) for four risk levels: low, medium, high, and severe. These reliability scores satisfy... For example, a health hazard agent might provide a response based on a highly toxic substance. and Reliability allocation; Not all agents' judgments are equally reliable; the system calculates a dynamic confidence anchor value for each agent. This is used to quantify the overall credibility of its judgment, among which, The highest probability value or the confidence level of the most certain category output by the agent's internal model reflects the inherent certainty of its judgment in this instance, i.e., the probability peak value. The quality score is determined by the input data (substance concentration, product information, etc.) upon which this risk assessment is based. The more complete and accurate the data, the higher the score. This represents the historical accuracy of the intelligent agent compared to similar products. , and The weighting coefficients are determined using the analytic hierarchy process (AHP). In this embodiment, , and The values ​​were 0.5, 0.3, and 0.2, respectively; Obtain the basic confidence assignments of the four agents. and its corresponding confidence anchor points Then, the system uses improved evidence fusion rules to fuse these pieces of evidence in pairs.

[0037] In one embodiment of the present invention, the step of performing pairwise fusion of the basic confidence assignment of the agent's output using improved evidence fusion rules includes the following steps: When fused intelligent agents and At that time, calculate the conflict metric. ,in For a certain risk level, , For the first Individual agents assess risk levels Trust level value, For the first Individual agents assess risk levels Trust level value; if > If a preset conflict threshold is set, a conflict buffer will be created. ,in, The preset buffer coefficient, For intelligent agents The confidence anchor value, For intelligent agents The confidence anchor value; Calculate agent fusion weights and ; Will and The initial weighted fusion of the individual agents yields the following results: The conflict buffer is redistributed according to a preset risk compensation coefficient to obtain the final fusion confidence allocation. ,in, Risk level The compensation coefficient; for Normalization is performed, and the fusion result is regarded as the output of the new agent. The preliminary weighted fusion calculation is repeated until all agent evidence is fused to obtain the comprehensive confidence allocation. Based on the final comprehensive confidence level assignment, the risk level with the highest confidence level is selected as the final judgment, and a traceable decision report is generated. The report includes the original confidence level assignment of each agent, confidence anchor value, conflict metric value, and intermediate results in the fusion process.

[0038] Specifically, conflict metrics This is obtained by calculating half the sum of the absolute values ​​of their trust levels across the four risk levels. This value ranges from 0 to 1; a larger value indicates a more severe disagreement between the two agents. A preset conflict threshold (obtained through historical data statistics) is set. If a conflict is detected, it is considered a significant conflict requiring special handling. In this case, the system creates a conflict buffer. ,in, The buffer is a preset buffer coefficient (a fixed value pre-set by R&D personnel or domain experts based on their experience and understanding of conflict handling in the risk decision-making process). The physical meaning of the buffer is: to reserve a portion of the confidence value from the conflict quality; its size is proportional to the degree of conflict, but limited by the relatively low confidence levels of either party. This is so that a safer redistribution can be carried out later. For each risk level Perform the following operations: Q1: Calculate the initial fusion reliability This is equivalent to a simple weighted average based on credibility; Q2: If a conflict buffer was created in step Q1 If the risk level is too high, the reliability score needs to be reallocated. The allocation principle is to favor higher risk levels to reflect the "precautionary principle" in risk assessment. Therefore, the system assigns a risk score to each risk level. Preset a risk compensation coefficient ,satisfy and For example, for general consumer goods, one could set... , , , Then, the buffer confidence is added to the initial fusion confidence according to the compensation coefficient to obtain the final fusion confidence allocation. ;Q3: If If the value is less than or equal to the preset conflict threshold, then there is no need to create a buffer; simply set the value to 0. Q4: Yes After normalization, at this point, Represents intelligent agents and The new reliability assignment after fusion will The output is considered as a new "agent" and its confidence anchor value is set to... or Repeat steps Q1-Q4 to fuse the new agent with the next agent, iterating in this way until all evidence from all four agents has been fused, yielding the final overall confidence score. ; In the final overall reliability allocation In the process, the risk level with the highest reliability score is selected as the final risk level for the product. For example, if If the value is the maximum, it is judged as a "serious" risk; Generate a traceability report: The system automatically records and generates a traceable decision report, which includes at least: the basic confidence allocation and confidence anchor value of each original agent, the conflict metric value of each round of pairwise fusion, the size of the created buffer (if it exists), the intermediate confidence allocation results after each round of fusion, and the final comprehensive confidence allocation and judgment reason.

[0039] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for intelligent screening of hazardous substances in consumer products based on artificial intelligence, characterized in that, Includes the following steps: S1: Receive the ingredient list of the product to be screened, the ingredient list containing vague ingredient items described in natural language or commercial names; S2: Preprocess the ingredient list using a multi-task text regularization model, and then classify and archive it. S3: Analyze the fuzzy component terms based on an artificial intelligence model: S31: Utilize a deep learning industry terminology mapping library with an embedded semantic similarity calculation model for accurate matching, and output standard substance identifiers and their confidence scores; S32: If the matching fails, the improved attention mechanism multi-evidence fusion inference unit is activated to extract four types of features, including fuzzy component text features, product category prior knowledge features, component list context semantic features and supplier historical submission features. The candidate list and its inferred confidence are output through intelligent weighted fusion. S33: Submit the output results to the user for verification, and use the feedback samples to dynamically update the mapping library and model parameters; S4: Use artificial intelligence search algorithms to query real-time synchronized hazardous substance regulatory databases, use risk level assessment models to conduct compliance screening and risk assessment, and output standardized screening reports.

2. The intelligent screening method for hazardous substances in consumer products based on artificial intelligence according to claim 1, characterized in that, In step S2, the construction and training of the multi-task text regularization model includes the following steps: The model comprises a shared character-level feature extraction layer and three cascaded task sub-networks, the task sub-networks including: Character correction and confidence generation subnet: Taking the character feature sequence output by the shared layer as input, based on a one-dimensional convolutional neural network and a conditional random field, the output is: the corrected character sequence and the error position confidence vector of the same length as the input sequence, where the confidence value of each position represents the probability that the character at that position has a spelling error; Vocabulary and Unit Parsing Subnetwork: Taking the corrected character sequence as input, the subnetwork performs word segmentation and part-of-speech tagging tasks based on a bidirectional long short-term memory network model. When calculating the hidden state at each time step, the subnetwork introduces a feature selection gating mechanism controlled by the error position confidence vector. The final output is: the segmented word sequence and the part-of-speech tag vector corresponding to each word. Semantic Structure Parsing Subnet: Taking the fusion result of word sequence and part-of-speech tag vector corresponding to each word as input, based on Transformer encoder and pointer network, it identifies and extracts three fields of structured substance name, concentration content and functional description; The model is trained end-to-end using a multi-task learning framework, and the overall loss function is obtained by weighting the loss of the three sub-network tasks and the auxiliary loss term.

3. The intelligent screening method for hazardous substances in consumer products based on artificial intelligence according to claim 2, characterized in that, The introduction of a feature selection gating mechanism controlled by the error location confidence vector includes the following steps: For the first in the input sequence Character feature vectors at each position The corresponding error location confidence level is The feature vectors that are actually input to the bidirectional long short-term memory network unit are then... It is obtained through the gating calculation formula, specifically: ;in, The feature vectors actually input to the network after being adjusted by the gating mechanism. The index of the character's position in the sequence. For the first Error location confidence for each position For the first The original character feature vector at each position, and The weight parameter matrix, For the first Character feature vectors at each position, For the first Character feature vectors at each position.

4. The intelligent screening method for hazardous substances in consumer products based on artificial intelligence according to claim 1, characterized in that, Step S31 includes the following steps: The preprocessed fuzzy component text is input into the embedded semantic similarity calculation model, and the model outputs a high-dimensional semantic vector, denoted as the query vector. ; In parallel, the semantic vectors of the standard names and aliases of all standard substances are retrieved from the deep learning industry terminology mapping library to form a candidate vector set. ; The query vector is calculated using a multi-scale similarity fusion algorithm. With each candidate vector Match confidence between ; Select matching confidence The top K candidate standard substances with the highest confidence level were identified, and the highest confidence level was determined. Does it exceed the preset first threshold? The matching decision is executed as follows: like If the match is successful, the corresponding standard substance identifier will be output. As its confidence level; if Then, the top K candidate standard substances and their confidence levels are output as the fuzzy matching results, where, The second threshold is preset; if If the match fails, step S32 is triggered directly.

5. The intelligent screening method for hazardous substances in consumer products based on artificial intelligence according to claim 4, characterized in that, The query vector is calculated using a multi-scale similarity fusion algorithm. With each candidate vector Match confidence between This includes the following steps: By calculating the query vector With each candidate vector Cosine similarity between them is used as global semantic similarity. ,in, Let L2 be the norm of the vector; Computational domain enhanced similarity Constructing the domain subspace projection matrix The column vectors of the projection matrix are formed by orthogonalizing the semantic vectors of K pre-selected high-frequency core terms in the field of consumer product chemistry; the query vector With candidate vectors Projecting each vector onto the respective neighborhood subspace yields a projection vector. and ,in, Projection matrix Transpose of the vector; calculate the normalized similarity between the two projected vectors. ,in, Projection vector and The Euclidean distance between them These are the preset scaling parameters; Dynamically determine the fusion weight based on the characteristics of the query text. The matching confidence level is obtained. Among them, the fusion weight The length and term density features of the query text are extracted and then calculated using a preset nonlinear mapping function. Based on the calculated matching confidence Sets, perform matching decisions and output results.

6. The intelligent screening method for hazardous substances in consumer products based on artificial intelligence according to claim 1, characterized in that, In step S32, four types of features are extracted, including the following steps: The extraction of the fuzzy component text features involves inputting the fuzzy component into a pre-trained language model encoder to obtain its context-dependent semantic representation, which serves as the fuzzy component text feature vector. ; The extraction of prior knowledge features of the product category involves retrieving common hazardous substances and compliance history of the corresponding product category from a historical database based on the preset category label of the product to be screened, and statistically obtaining the prior probability distribution vector of the substances in the category, which serves as the prior knowledge feature vector of the product category. ; The extraction of the contextual semantic features of the component list involves extracting several explicit components adjacent to the fuzzy component item from the component list, and performing a pooling operation on the semantic vectors of each explicit component to generate a contextual semantic feature vector for the component list. ; The extraction of supplier historical submission features involves querying the historical submission records of suppliers providing the ingredient list, statistically analyzing the frequency distribution of all fuzzy descriptions submitted by the suppliers that were ultimately verified as belonging to each standard substance, and converting this distribution into a probability vector, which serves as the supplier historical submission feature vector. .

7. The intelligent screening method for hazardous substances in consumer products based on artificial intelligence according to claim 6, characterized in that, The four types of features are used to output a candidate list and its estimated confidence level through intelligent weighted fusion, including the following steps: S3201: For the four types of feature vectors extracted , , and Calculate the corresponding uncertainty scores for each. , , and ; S3202: The four types of feature vectors and their corresponding uncertainty scores are input together into an improved cross-attention layer to generate a fused context vector. The improved cross-attention layer specifically refers to: Through the weight matrix Perform a linear transformation to generate a query vector. ; four types of feature vectors Through the weight matrix respectively and Perform a linear transformation to generate the corresponding key vector. AND value vector ; Calculate the first Attention contribution weights of class features ,in, ,in, For unnormalized attention scores, Let be the dimension of the key vector. The predetermined uncertainty penalty coefficient, For the first Uncertainty scores of class features; final fused context vector ; Using the context vector, an original matching score is calculated for each candidate standard substance; simultaneously, a confidence correction for counterfactual intervention is constructed.

8. The intelligent screening method for hazardous substances in consumer products based on artificial intelligence according to claim 7, characterized in that, The confidence correction for constructing counterfactual interventions includes the following steps: The context vector With candidate materials semantic embedding vector Perform similarity calculation to obtain a matching score. ; Creating counterfactual feature vectors Specifically, this involves: extracting the original supplier's historical submission feature vector. Replace with an all-zero vector or a uniformly distributed probability vector; Using the counterfactual feature vector Replace the original Repeat step S3202 to obtain the counterfactual fusion context vector. To calculate candidate substances Counterfactual matching score ; Calculate the correction amount and generate the final confidence level, where the corrected estimated confidence level is specifically as follows: ,in, The preset correction strength coefficient; Based on the corrected estimated confidence level, all candidate standard substances are sorted in descending order, and the top M substances and their corresponding values ​​are output. The values ​​are used as the final candidate list.

9. The intelligent screening method for hazardous substances in consumer products based on artificial intelligence according to claim 1, characterized in that, In step S4, the risk level assessment model adopts a risk decision-making framework based on multi-agent negotiation and evidence theory, including the following steps: Four intelligent agents with different risk dimensions are deployed, including regulatory compliance intelligent agents, health hazard intelligent agents, environmental accumulation intelligent agents, and supply chain exposure intelligent agents. Each intelligent agent encapsulates an analysis model and knowledge base for its risk dimension. Each agent independently calculates and analyzes the input standardized substance list and its concentrations, combining a real-time regulatory database and a professional risk knowledge base, and outputs its basic reliability allocation for the product's risk level in this dimension. , , , and These represent the level of trust in four risk levels: low, medium, high, and severe. Assign an ID to the agent; calculate the confidence anchor value for each agent. ,in, The peak probability predicted by the agent model. To score the quality of the input data, This represents the historical accuracy of the intelligent agent compared to similar products. , and , where is the weighting coefficient; an improved evidence fusion rule is used to perform pairwise fusion of the basic confidence assignments of the agent's output.

10. The intelligent screening method for hazardous substances in consumer products based on artificial intelligence according to claim 9, characterized in that, The method of pairwise fusing the basic confidence assignment of the agent's output using improved evidence fusion rules includes the following steps: When fused intelligent agents and At that time, calculate the conflict metric. ,in For a certain risk level, , For the first Individual agents assess risk levels Trust level value, For the first Individual agents assess risk levels Trust level value; if > If a preset conflict threshold is set, a conflict buffer will be created. ,in, The preset buffer coefficient, For intelligent agents The confidence anchor value, For intelligent agents The confidence anchor value; Calculate agent fusion weights and ; Will and The initial weighted fusion of the individual agents yields the following results: The conflict buffer is redistributed according to a preset risk compensation coefficient to obtain the final fusion confidence allocation. ,in, Risk level The compensation coefficient; for Normalization is performed, and the fusion result is regarded as the output of the new agent. The preliminary weighted fusion calculation is repeated until all agent evidence is fused to obtain the comprehensive confidence allocation. Based on the final comprehensive confidence level assignment, the risk level with the highest confidence level is selected as the final judgment, and a traceable decision report is generated. The report includes the original confidence level assignment of each agent, confidence anchor value, conflict metric value, and intermediate results in the fusion process.

Images