A method for constructing a multi-modal risk information identification library for quality and safety of consumer goods

Abstract

The application discloses a kind of consumer goods quality safety multi-modal risk information identification library construction methods, it is related to consumer goods quality safety supervision technical field, including, to quality responsibility anchor point risk factor identification record is carried out same product same batch same risk trigger chain alignment, generates anchor point cross-modal risk evidence chain set, and according to the corresponding relationship between anchor point risk trigger chain record and quality responsibility positioning point generates responsibility backhang check record, then anchor point cross-modal risk evidence chain set and responsibility backhang check record are carried out evidence consistency verification and non-quality responsibility counter-evidence isolation arbitration, form multi-modal evidence conflict arbitration account book;According to multi-modal evidence conflict arbitration account book, establish the risk index library structure with consumer goods identity normalization record as main key, and risk identification item is written into risk index library structure, obtains consumer goods quality safety multi-modal risk information identification library.The application realizes the backhang of risk phenomenon to quality responsibility positioning point.

Description

Technical Field

[0001] This invention relates to the field of consumer product quality and safety supervision technology, and in particular to a method for constructing a multimodal risk information identification database for consumer product quality and safety. Background Technology

[0002] As consumer product sales channels and regulatory data sources continue to expand, quality and safety risk identification is gradually evolving from manual complaint categorization to comprehensive analysis of multi-source data. Existing technologies typically collect complaint texts, test reports, recall notices, packaging images, usage videos, and after-sales maintenance records. Risk clues are then organized through keyword extraction, image recognition, risk labeling, and database construction, providing a data foundation for regulatory verification, corporate rectification, and recall analysis.

[0003] However, existing technologies still have shortcomings: on the one hand, risk clues are mostly collected by product name or risk keywords, making it difficult to accurately distinguish between different batches of consumer products with the same name, and also difficult to link risk phenomena back to specific quality responsibility points, resulting in risk items lacking clear responsibility orientation; on the other hand, existing technologies mostly perform simple fusion of text, images, videos and testing information, lacking alignment of risk trigger chains of the same product, batch and the same risk, and isolation of non-quality responsibility counter-evidence, which easily leads to misjudging clues formed by misuse, environmental influence or repeated dissemination as product quality risks. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for constructing a multimodal risk information identification database for consumer product quality and safety, which solves the problems in the prior art where it is difficult to assign responsibility for multi-source risk clues and where it is difficult to accurately separate non-quality responsibility interference evidence.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for constructing a multimodal risk information identification database for consumer product quality and safety. The method includes: collecting multi-source risk data on consumer product quality and safety and performing multimodal cleaning, identification, and unified calibration to form a standardized multimodal risk clue package; verifying the consumer product identity and batch normalizing the standardized multimodal risk clue package to obtain a consumer product identity normalization record; extracting risk phenomena and hazardous locations from the consumer product identity normalization record; and then linking the risk phenomena back to the quality responsibility positioning point to form a quality responsibility anchored risk element identification record; and performing same-product, same-batch verification on the quality responsibility anchored risk element identification record. Aligning with the risk trigger chain, an anchored cross-modal risk evidence chain set is generated. Based on the correspondence between the anchored risk trigger chain records and the quality responsibility positioning points, a responsibility re-attachment verification record is generated. Then, the anchored cross-modal risk evidence chain set and the responsibility re-attachment verification record are subjected to evidence consistency verification and non-quality responsibility rebuttal evidence isolation arbitration to form a multimodal evidence conflict arbitration ledger. Based on the multimodal evidence conflict arbitration ledger, a risk index database structure with the consumer product identity normalization record as the primary key is established, and risk identification entries are written into the risk index database structure to obtain a multimodal risk information identification database for consumer product quality and safety.

[0007] As a preferred embodiment of the method for constructing a multimodal risk information identification database for consumer product quality and safety described in this invention, the multi-source risk data for consumer product quality and safety includes consumer product complaint text data, consumer product packaging image data, consumer product usage process video data, consumer product testing report data, consumer product after-sales repair record data, and consumer product recall announcement data.

[0008] As a preferred embodiment of the method for constructing a multimodal risk information identification database for consumer product quality and safety according to the present invention, the specific steps for forming a standardized multimodal risk clue package are as follows: Collect multi-source risk data on the quality and safety of consumer products, and aggregate them according to source identification and carrier type to generate original multi-source risk carrier records; Multimodal cleaning and identification are performed on the original multi-source risk carrier records, and consumer product identity information and risk description information are extracted to obtain multimodal risk identification records; Multimodal risk identification records are uniformly calibrated under the same time reference and the same clue number, and the calibrated multimodal risk identification records are packaged to form a standardized multimodal risk clue package.

[0009] As a preferred embodiment of the method for constructing a multimodal risk information identification database for consumer product quality and safety according to the present invention, the specific steps for obtaining the consumer product identity normalization record are as follows: Consumer product identity information is read from the standardized multimodal risk clue package, and the consumer product identity information is processed to standardize the fields to form a set of candidate identity fields; The consumer product identity anchoring verification is performed on the candidate identity field set. First, the identity field that can uniquely point to the consumer product is locked. Then, the clues with inconsistent names are reviewed by using the packaging layout similarity relationship. Clues that are confirmed to belong to the same consumer product object are merged to obtain the consumer product identity verification record. The batch field is extracted from the consumer product identity verification record for consistency comparison. After the batch conflict clues are isolated, they are merged into clues of the same production batch to form the consumer product batch unification record. By binding the consumer product identity verification record with the consumer product batch normalization record, a consumer product identity normalization record is generated.

[0010] As a preferred embodiment of the method for constructing a multimodal risk information identification database for consumer product quality and safety according to the present invention, the specific steps for forming a quality responsibility anchored risk element identification record are as follows: Read the standardized multimodal risk clue package corresponding to the same consumer product object and the same production batch from the consumer product identity normalization record, and extract risk description statements and visible abnormal areas from the standardized multimodal risk clue package to form risk phenomenon candidate records; Risk phrase identification and abnormal area localization are performed on candidate records of risk phenomena, and risk phrases are associated with corresponding visible abnormal areas to form risk phenomenon localization records; Based on the risk phenomenon location records, a verifiable quality responsibility location table for consumer products is established, and the visible abnormal areas in the risk phenomenon location records are matched with the verifiable quality responsibility location table for consumer products to form a quality responsibility candidate location record. The strength of liability reinstatement is calculated using the candidate location records of quality responsibility. The quality responsibility positioning points are then sorted according to the strength of liability reinstatement. The quality responsibility positioning point with the highest ranking that meets the reinstatement requirements is determined as the quality responsibility anchor point. The quality responsibility anchor point is then written back to the corresponding consumer product identity normalization record to form a quality responsibility anchor point-based risk element identification record.

[0011] As a preferred embodiment of the method for constructing a multimodal risk information identification database for consumer product quality and safety according to the present invention, the specific steps for generating an anchored cross-modal risk evidence chain set are as follows: Extract the risk trigger sequence from the risk element identification record anchored by quality responsibility, and organize the risk trigger sequence according to the quality responsibility anchor to form an anchored risk trigger chain record; Based on the anchored risk trigger chain records, a responsibility anchor sequence diagram is constructed, and anchored risk trigger chain records that do not belong to the same consumer product identity and the same production batch are isolated to obtain candidate pairs of trigger chains for the same product and the same batch. For candidate pairs of trigger chains of the same product and batch, sequential matching is performed. The risk trigger nodes in one clue are compared with the corresponding nodes in another clue in turn, and it is verified whether each node is linked back to the same quality responsibility anchor point. If the trigger order of the two clues can be continuously matched and the quality responsibility anchor points are consistent, the two clues are merged into the same risk evidence and continue to be aggregated under the same quality responsibility anchor point to generate an anchored cross-modal risk evidence chain set.

[0012] As a preferred embodiment of the method for constructing a multimodal risk information identification database for consumer product quality and safety according to the present invention, the specific steps for generating the responsibility re-verification record are as follows: The anchored risk trigger chain record and the quality responsibility location point are read by using the anchored cross-modal risk evidence chain set. After establishing the responsibility back-hanging verification task, the counterfactual bypass verification is performed. After shielding the quality responsibility location support, it is determined whether the anchored risk trigger chain record can still be closed, and the counterfactual back-hanging support record is obtained. Based on the counterfactual back-hanging support record, the degree of back-hanging support of the anchored risk trigger chain record to the quality responsibility positioning point is calculated, and a responsibility back-hanging judgment record is formed according to the degree of back-hanging support. Then, the responsibility back-hanging judgment record is bound to the corresponding anchored cross-modal risk evidence chain set to generate a responsibility back-hanging verification record.

[0013] As a preferred embodiment of the method for constructing a multimodal risk information identification database for consumer product quality and safety according to the present invention, the specific steps for forming a multimodal evidence conflict arbitration ledger are as follows: Risk evidence under the same quality responsibility anchor point is read from the anchored cross-modal risk evidence chain set, and valid back-up evidence and candidate evidence of rebuttal are distinguished according to the responsibility back-up verification record to form a positive and negative evidence slot record; The valid back-hanging evidence in the positive and negative evidence slotting record is verified for evidence consistency. Evidence that points to the same quality responsibility positioning point is grouped into the responsibility consistency evidence group to obtain the evidence consistency verification record. Based on the evidence consistency verification record, non-quality responsibility counter-evidence is identified for candidate counter-evidence, and evidence that cannot support the re-attachment of quality responsibility positioning points is isolated, generating non-quality responsibility counter-evidence isolation records. Arbitrate the anchored cross-modal risk evidence chain set based on the non-quality liability rebuttal isolation record, and bind the conflict sources and isolation basis in the anchored cross-modal risk evidence chain set to form a multimodal evidence conflict arbitration ledger.

[0014] As a preferred embodiment of the method for constructing a multimodal risk information identification database for consumer product quality and safety described in this invention, the specific steps for establishing a risk index database structure with consumer product identity normalization records as the primary key based on a multimodal evidence conflict arbitration ledger are as follows: The multimodal evidence conflict arbitration ledger is used to map the anchored cross-modal risk evidence chain set with the corresponding consumer product identity normalization record. After determining the consumer product identity normalization record as the primary key, the anchored cross-modal risk evidence chain set under the same primary key is merged and verified to form the consumer product risk primary key record. The primary key records for consumer product risks are subjected to an inbound stability verification. The primary key records for consumer product risks that pass the verification are organized into risk identification entries, and the risk identification entries are linked to the corresponding primary keys. Then, an index relationship is established between the primary keys, risk identification entries, and the multimodal evidence conflict arbitration ledger to form a risk index database structure.

[0015] As a preferred embodiment of the method for constructing a multimodal risk information identification database for consumer product quality and safety according to the present invention, the specific steps for obtaining the multimodal risk information identification database for consumer product quality and safety are as follows: Read the consumer product identity normalization record primary key from the risk index database structure, associate and confirm the risk identification entries under the consumer product identity normalization record primary key with the anchored cross-modal risk evidence chain set, and verify the correspondence between the risk identification entries, the consumer product identity normalization record primary key and the multimodal evidence conflict arbitration ledger to form a closed loop write verification record. Based on the closed-loop write verification record, the verified risk identification entries are written into the primary key of the consumer product identity normalization record, and the index relationship between the risk identification entries, the anchored cross-modal risk evidence chain set, and the multimodal evidence conflict arbitration ledger is solidified to obtain the consumer product quality and safety multimodal risk information identification library.

[0016] The beneficial effects of this invention are as follows: By linking the abnormal areas corresponding to risk phenomena back to the quality responsibility positioning points, a quality responsibility anchored risk element identification record is formed, realizing the linking of risk phenomena to quality responsibility positioning points, thereby improving the directionality of risk responsibility and reducing the mis-aggregation of similar risks; by aligning the same product, batch, and risk trigger chain and isolating non-quality responsibility counter-evidence, a multimodal evidence conflict arbitration ledger is formed, realizing the credible aggregation of multi-source risk clues under the same consumer product, batch, risk trigger chain, and quality responsibility anchor point, thereby reducing misjudgment of non-quality responsibility risks and enhancing the traceability of the risk information identification database. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating the method for constructing a multimodal risk information identification database for consumer product quality and safety.

[0019] Figure 2 A flowchart for creating a standardized multimodal risk clue package.

[0020] Figure 3 A flowchart for unifying the identity of consumer products and anchoring quality responsibility.

[0021] Figure 4 A flowchart for generating an anchored cross-modal risk evidence chain and liability re-attachment verification. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a method for constructing a multimodal risk information identification database for consumer product quality and safety, comprising the following steps: S1. Collect multi-source risk data on the quality and safety of consumer products and perform multi-modal cleaning, identification, and unified calibration to form a standardized multi-modal risk clue package.

[0026] Collect multi-source risk data on the quality and safety of consumer products, and aggregate them according to source identification and carrier type to generate original multi-source risk carrier records.

[0027] The specific process includes: extracting the complaint text, complaint time, and name of the complained consumer product from the complaint acceptance record to form consumer product complaint text data; extracting front, side, and label images of the packaging from complaint attachments or evidence materials to form consumer product packaging image data; extracting video files recording changes in the consumer product's usage status from complaint attachments or on-site evidence materials to form consumer product usage video data; extracting sample name, batch description, test items, and test conclusions from test report files to form consumer product test report data; extracting repair phenomena, repair locations, and repair times from after-sales repair records to form consumer product after-sales repair record data; extracting the name of the recalled consumer product, recall scope, and defect description from recall notice files to form consumer product recall notice data; generating source identifiers according to the respective sources of consumer product complaint text data, consumer product packaging image data, consumer product usage video data, consumer product test report data, consumer product after-sales repair record data, and consumer product recall notice data; determining the carrier type according to the content of text, images, videos, and documents; and binding and aggregating the source identifiers, carrier types, and corresponding data to generate original multi-source risk carrier records.

[0028] Multimodal cleaning and identification are performed on the original multi-source risk carrier records to extract consumer product identity information and risk description information, resulting in multimodal risk identification records.

[0029] The specific process includes: retrieving corresponding content from the original multi-source risk carrier records based on source identification and carrier type; deleting irrelevant characters, duplicate sentences, and non-quality and safety descriptions from the consumer product complaint text data, while retaining the name of the complained consumer product and the risk description content; performing image orientation correction and text recognition on the consumer product packaging image data, extracting the consumer product name, brand, model, batch number, and label description content from the front, side, and label images; and extracting key frames from the video data of the consumer product's usage process, identifying the content before and after the anomaly occurred from the video files recording changes in the consumer product's usage status. It extracts risk description information that can characterize damage, deformation, detachment, leakage, and ablation; it reads fields from consumer product testing report data, consumer product after-sales repair record data, and consumer product recall notice data to extract sample name, batch description, test conclusion, repair phenomenon, repair part, recalled consumer product name, recall scope, and defect description; it incorporates the consumer product name, brand, model, batch, and label description extracted from the original multi-source risk carrier records into the consumer product identity information, and incorporates complaint descriptions, abnormal screen content, test conclusions, repair phenomena, and defect descriptions into the risk description information to obtain multimodal risk identification records.

[0030] Multimodal risk identification records are uniformly calibrated under the same time reference and the same clue number, and the calibrated multimodal risk identification records are packaged to form a standardized multimodal risk clue package.

[0031] The specific process includes: reading source identifiers, carrier types, consumer product identity information, risk description information, and corresponding time information from multimodal risk identification records; converting complaint time, video recording time, test report time, repair time, and recall announcement time into a unified time format; merging multimodal risk identification records according to the same consumer product identity information and the same risk description information; assigning the same clue number to the merged multimodal risk identification records; correcting the chronological relationship between different carriers using the same time benchmark; and correspondingly labeling the text content, image content, video content, test conclusion content, repair phenomenon content, and defect description content under the same clue number; and encapsulating the multimodal risk identification records that have completed time labeling and clue number labeling according to source identifiers, carrier types, consumer product identity information, and risk description information to form a standardized multimodal risk clue package.

[0032] S2. Perform consumer product identity verification and batch normalization on the standardized multimodal risk clue package to obtain a consumer product identity normalization record. Extract risk phenomena and dangerous parts from the consumer product identity normalization record, and then link the risk phenomena back to the quality responsibility positioning point to form a quality responsibility anchored risk element identification record.

[0033] Consumer product identity information is read from the standardized multimodal risk clue package, and the fields of the consumer product identity information are normalized to form a set of candidate identity fields.

[0034] The specific process includes: retrieving consumer product identity information from a standardized multimodal risk clue package according to the same clue number; extracting the name of the complained consumer product from the consumer product complaint text data, the consumer product name, brand, model, batch number, and label description from the consumer product packaging image data, the sample name and batch description from the consumer product test report data, the name associated with the repair part from the consumer product after-sales repair record data, and the name of the recalled consumer product and the recall scope from the consumer product recall announcement data as fields to be standardized; removing spaces, standardizing punctuation, standardizing full-width and half-width characters, standardizing capitalization, merging common aliases, and deleting invalid modifiers from the fields to be standardized; merging fields with the same meaning but different spellings into the same field name; and retaining the field source identifier, carrier type, and the same clue number to form a set of candidate identity fields.

[0035] The candidate identity field set is subjected to consumer product identity anchoring verification. First, the identity field that can uniquely point to the consumer product is locked. Then, the clues with inconsistent names are reviewed by using the packaging layout similarity relationship. Clues that are confirmed to belong to the same consumer product object are merged to obtain consumer product identity verification records. The batch field is extracted from the consumer product identity verification records for consistency comparison. After batch conflict clues are isolated, they are merged into clues of the same production batch to form a consumer product batch unification record.

[0036] The specific process includes: extracting the consumer product name, brand, model, and label description under the same clue number from the identity candidate field set; first, grouping clues with the same brand and model into the same consumer product object; for clues with abbreviations, aliases, or inconsistent spellings from different sources for the consumer product name, then retrieving the packaging image data of the packaging front text position, label area position, and main identification content to verify the packaging layout similarity; if the packaging layout is similar and the brand, model, or label description can be matched, then the clues with inconsistent names are grouped into the same consumer product object, forming a consumer product identity verification record; reading the batch field, batch description, and recall scope from the consumer product identity verification record, grouping clues with the same batch field under the same consumer product object into the same production batch, adding clues with missing batch fields but corresponding batch descriptions and recall scopes into the same production batch, and isolating clues with conflicting batch fields from the current production batch, forming a consumer product batch unification record.

[0037] By binding the consumer product identity verification record with the consumer product batch normalization record, a consumer product identity normalization record is generated.

[0038] The specific process includes: using clues confirmed to belong to the same consumer product in the consumer product identity verification record as the basis for identity normalization; reading the consumer product name, brand, model, label description, and the same clue number from the consumer product identity verification record; then reading the production batch clues already merged under the same consumer product object from the consumer product batch normalization record; and binding the consumer product identity verification record and the consumer product batch normalization record to correspond to the same consumer product object; for clues that have matched the same production batch, classifying the corresponding batch field, batch description, and recall scope into the batch level under the same consumer product object; for clues that have isolated the current production batch, retaining them in the batch conflict marker under the same consumer product object to avoid mixing with confirmed production batches; and uniformly associating the same consumer product object, the same production batch, the same clue number, source identifier, and carrier type to generate a consumer product identity normalization record that can simultaneously represent the consumer product identity and production batch affiliation.

[0039] The standardized multimodal risk clue package corresponding to the same consumer product object and the same production batch is read from the consumer product identity normalization record, and risk description statements and visible abnormal areas are extracted from the standardized multimodal risk clue package to form risk phenomenon candidate records.

[0040] The specific process includes: searching for the same clue number from the consumer product identity normalization record according to the same consumer product object and the same production batch; retrieving risk description information, text content, image content, video content, test conclusion content, maintenance phenomenon content, and defect description content from the standardized multimodal risk clue package based on the same clue number; extracting statements that can characterize quality and safety anomalies from the text content, test conclusion content, maintenance phenomenon content, and defect description content as risk description statements; reading abnormal scene content from the image content and video content, and locating the visible abnormal area corresponding to the risk description statement; and then associating the risk description statement, visible abnormal area, same consumer product object, same production batch, and same clue number to form a risk phenomenon candidate record.

[0041] Risk phrase identification and abnormal area localization are performed on candidate records of risk phenomena, and risk phrases are associated with corresponding visible abnormal areas to form risk phenomenon localization records.

[0042] The specific process includes: reading risk description statements, visible abnormal areas, the same clue number, the same consumer product object, and the same production batch from the candidate records of risk phenomena; performing word segmentation and semantic relationship organization on the risk description statements, retaining words indicating abnormal states, and combining words indicating abnormal states with adjacent descriptions of dangerous parts to form risk phrases; finding the visible abnormal areas corresponding to the risk phrases based on the same clue number, confirming the location boundaries of the visible abnormal areas in packaging images and usage process videos, and marking the visible abnormal areas that can correspond to the risk phrases as the areas corresponding to the risk phrases; and binding the risk phrases, descriptions of dangerous parts, areas corresponding to the risk phrases, the same consumer product object, and the same production batch to form a risk phenomenon location record.

[0043] A verifiable quality responsibility location table for consumer products is established based on the risk phenomenon location records. Visible abnormal areas in the risk phenomenon location records are then matched with the verifiable quality responsibility location table for consumer products to form a record of candidate quality responsibility locations.

[0044] The specific process includes: reading risk phrases, descriptions of hazardous parts, corresponding areas of risk phrases, the same consumer product object, and the same production batch from the risk phenomenon location record; using the consumer product part pointed to by the description of hazardous parts as the verifiable location name; using the location boundary of the area corresponding to the risk phrase as the verifiable location range; and organizing the verifiable location names and verifiable location ranges according to the same consumer product object to establish a verifiable quality responsibility location table for consumer products; comparing the visible abnormal areas in the risk phenomenon location record with the verifiable location ranges in the verifiable quality responsibility location table for consumer products; retaining records where the visible abnormal areas fall into or are adjacent to the corresponding verifiable location ranges as candidate responsibility locations; and associating the risk phrase, description of hazardous parts, visible abnormal areas, verifiable location names, the same consumer product object, and the same production batch to form a quality responsibility candidate location record.

[0045] The strength of liability reinstatement is calculated using the candidate location records of quality responsibility. The quality responsibility positioning points are then sorted according to the strength of liability reinstatement. The quality responsibility positioning point with the highest ranking that meets the reinstatement requirements is determined as the quality responsibility anchor point. The quality responsibility anchor point is then written back to the corresponding consumer product identity normalization record to form a quality responsibility anchor point-based risk element identification record.

[0046] The specific process includes: retrieving risk phrases, risk phenomena, visible abnormal areas, quality responsibility positioning points, and supporting test results from the quality responsibility candidate location record for the same consumer product and the same production batch; mapping the content indicating abnormal status in the risk phrases to the name of the quality responsibility positioning point to obtain a semantic matching value; verifying the overlap between the location boundary of the visible abnormal area and the location range corresponding to the quality responsibility positioning point to obtain a location matching value; obtaining a time-fitting value based on the time interval between the occurrence time of the risk phenomenon and the occurrence time of the abnormality at the quality responsibility positioning point; and determining whether there are supporting test results for the abnormality at the quality responsibility positioning point. The content is regularly analyzed to obtain supporting values. Simultaneously, inconsistencies are checked between the location of the dangerous area pointed to by the risk phrase, the location of the visible abnormal area, and the supporting information of the detection conclusion to obtain the conflict intensity. Based on the semantic matching value, location matching value, time fit value, and supporting value strengthening the back-hanging relationship while the conflict intensity weakens the back-hanging relationship, a responsibility back-hanging intensity is generated. The quality responsibility positioning points corresponding to the same risk phenomenon are then ranked according to the responsibility back-hanging intensity. The quality responsibility positioning point with the highest ranking and meeting the back-hanging requirements is determined as the quality responsibility anchor point. The quality responsibility anchor point is then written back to the corresponding consumer product identity normalization record, forming a quality responsibility anchor point-based risk element identification record.

[0047] It should be noted that the quality responsibility location point is the specific responsible part of the consumer product to which the risk phenomenon can be attributed. It is determined by establishing a verifiable quality responsibility location table for consumer products based on the description of the dangerous parts and the corresponding visible abnormal areas in the risk phenomenon location record. The attribute attribute requirement is that the strength of the attribute attribute must reach the judgment condition that a stable correspondence exists between the risk phenomenon and the quality responsibility location point. It is determined by the distribution of the strength of the attribute attribute of the confirmed risk clues under the same consumer product object.

[0048] The expression for calculating the liability backing strength is: ; in, Indicates the first The risk phenomenon was reverted to the first The responsibility reinforcing strength of each quality responsibility positioning point; The number indicating the risk phenomenon; Indicates the number of the quality responsibility location point; Indicates the first The risk phrase corresponding to the first risk phenomenon and the first Semantic matching values ​​between quality responsibility location points; Indicates the first The visible abnormal area corresponding to the first risk phenomenon is the same as the first... Location matching values ​​between quality responsibility positioning points; Indicates the first The first risk phenomenon and the first The intensity of conflict between individual quality responsibility positioning points; Indicates the first The timing of the occurrence of the first risk phenomenon and the first The time overlap value between the occurrence times of anomalies at each quality responsibility location point; Indicates the first The risk phenomenon was reverted to the first Supporting values ​​for each quality responsibility positioning point.

[0049] It should be noted that, Through the first The risk phrase for the first risk phenomenon and the first The names and descriptions of each quality responsibility location point are vectorized into text, and then normalized after taking the cosine similarity between the two. By reading the first The risk phenomenon corresponds to the coordinates of the visible abnormal area and the first... The coordinates of the quality responsibility location points are obtained by taking the intersection and comparison of the two coordinates and then normalizing them; Through statistics The first risk phenomenon and the first The results were obtained by normalizing the number of inconsistencies in location, description, and evidence among the quality responsibility positioning points, and then normalizing them according to the total amount of evidence. By reading the first The timing of the occurrence of the first risk phenomenon and the first The time of anomaly occurrence at each quality responsibility location point was calculated by normalizing the time difference between the two points according to the same clue time span and then converting them in reverse. By reading the corresponding number in the test report The test results of the quality responsibility positioning point, and in accordance with the test results of the first quality responsibility positioning point, and the conclusions of the test results of the first quality responsibility positioning point. The degree of correspondence and the degree of qualified anomaly of each risk phenomenon are obtained by normalization.

[0050] S3. Align the quality responsibility anchored risk element identification records with the same product, batch, and risk trigger chain to generate an anchored cross-modal risk evidence chain set. Based on the correspondence between the anchored risk trigger chain records and the quality responsibility positioning points, generate responsibility back-hanging verification records. Then, verify the consistency of evidence and arbitrate non-quality responsibility counter-evidence by separating and arbitrating the anchored cross-modal risk evidence chain set and the responsibility back-hanging verification records to form a multimodal evidence conflict arbitration ledger.

[0051] Extract the risk trigger sequence from the risk element identification record anchored by quality responsibility, and organize the risk trigger sequence according to the quality responsibility anchor to form an anchored risk trigger chain record.

[0052] The specific process includes: extracting risk phrases, corresponding regions, quality responsibility anchors, clue numbers, and temporal relationships from the risk element identification record for the same consumer product and production batch; determining the triggering order of risk phrases based on the appearance time of their corresponding regions in packaging images and usage videos; correcting the triggering order by combining the description order of risk phrases in text content, inspection conclusions, repair phenomena, and defect descriptions; arranging risk phrases that are linked back to the same quality responsibility anchor and can be continuously connected in time; supplementing the connection relationship of risk phrases with sequential breaks under the same quality responsibility anchor based on the image and video content corresponding to the same clue number; removing risk phrases that cannot be linked back to the same quality responsibility anchor from the current triggering order; and associating the retained risk phrases, their corresponding regions, triggering order, and quality responsibility anchors to form an anchored risk triggering chain record.

[0053] Based on the anchored risk trigger chain records, a responsibility anchor sequence diagram is constructed, and anchored risk trigger chain records that do not belong to the same consumer product identity and the same production batch are isolated to obtain candidate pairs of trigger chains for the same product and the same batch.

[0054] The specific process includes: taking the consumer product identity, production batch, risk phrase, trigger sequence, and quality responsibility anchor point recorded in the anchored risk trigger chain record as the processing object, determining each risk phrase as a risk trigger node, connecting adjacent risk trigger nodes according to the trigger sequence, and establishing a correlation between the risk trigger node and the corresponding quality responsibility anchor point to form a responsibility anchor point sequence diagram that can represent the sequence relationship of risk trigger nodes and the back-hanging relationship of quality responsibility anchor points; within the responsibility anchor point sequence diagram, analyzing the consumer product identity and production batch corresponding to each anchored risk trigger chain record, keeping anchored risk trigger chain records with consistent consumer product identity and consistent production batch within the same comparison range, and isolating anchored risk trigger chain records with inconsistent consumer product identity or inconsistent production batch from the same comparison range; pairing anchored risk trigger chain records retained within the same comparison range that can be compared in trigger sequence to obtain candidate pairs of trigger chains of the same product and batch.

[0055] For candidate pairs of trigger chains of the same product and batch, sequential matching is performed. The risk trigger nodes in one clue are compared with the corresponding nodes in another clue in turn, and it is verified whether each node is linked back to the same quality responsibility anchor point. If the trigger order of the two clues can be continuously matched and the quality responsibility anchor points are consistent, the two clues are merged into the same risk evidence and continue to be aggregated under the same quality responsibility anchor point to generate an anchored cross-modal risk evidence chain set.

[0056] The specific process includes: based on candidate trigger chain pairs of the same product and batch, confirming that two leads are within the same merging range according to the same consumer product identity and the same production batch; then extracting the order of risk trigger nodes from the two leads respectively, and comparing them position by position according to their arrangement in the anchored risk trigger chain record; first verifying whether the risk phrase corresponding to the risk trigger node in one lead represents the same abnormal state as the corresponding risk phrase in the other lead; then verifying whether the areas corresponding to the risk phrases of the two risk trigger nodes point to the same dangerous location; and further verifying whether both risk trigger nodes are reattached to the same quality responsibility anchor point; for two leads with an inconsistent number of risk trigger nodes... Based on the triggering order between adjacent risk triggering nodes and the image and video content under the same clue number, it is determined whether there is a continuous connection relationship. Risk triggering nodes that can be continuously connected are retained as valid corresponding nodes, while risk triggering nodes that cannot be continuously connected are not included in the merging of the same risk evidence. When the valid corresponding nodes in two clues can maintain the same triggering order and both valid corresponding nodes are attached back to the same quality responsibility anchor point, the two clues are determined to be the same risk evidence. Using the same quality responsibility anchor point as the aggregation center, the clues that have been determined to be the same risk evidence are merged with other clues that meet the same corresponding relationship in the candidate pairs of triggering chains of the same product and batch, generating an anchored cross-modal risk evidence chain set.

[0057] By using the anchored cross-modal risk evidence chain set, the correspondence between the anchored risk trigger chain record and the quality responsibility positioning point is read. After establishing the responsibility back-hanging verification task, the counterfactual bypass verification is performed. After shielding the quality responsibility position support, it is determined whether the anchored risk trigger chain record can still be closed, and the counterfactual back-hanging support record is obtained.

[0058] The specific process includes: identifying the anchored risk trigger chain record corresponding to each piece of the same risk evidence, based on the set of anchored cross-modal risk evidence chains; establishing a responsibility re-attachment verification task based on the relationships between risk trigger nodes, trigger order, quality responsibility anchor points, and quality responsibility positioning points in the anchored risk trigger chain records; the responsibility re-attachment verification task retains the order of risk trigger nodes and the connection relationship between adjacent risk trigger nodes, while temporarily removing the support relationship between quality responsibility positioning points and risk trigger nodes, and re-verifying whether risk trigger nodes can still correspond continuously according to the original trigger order, and whether risk phrases can still maintain correspondence with the corresponding areas of risk phrases; if there is a break between risk trigger nodes, or if risk trigger nodes cannot continue to be re-attached to quality responsibility anchor points, then it is confirmed that the anchored risk trigger chain record depends on the quality responsibility positioning point; if the correspondence can still be maintained through other descriptive content after removing the support of the quality responsibility positioning point, the correspondence is marked as a bypass closure relationship; binding the responsibility re-attachment verification task, the closure judgment before and after the quality responsibility positioning point is masked, the breakage of risk trigger nodes, and the bypass closure relationship, to obtain the counterfactual re-attachment support record.

[0059] Based on the counterfactual back-hanging support record, the degree of back-hanging support of the anchored risk trigger chain record to the quality responsibility positioning point is calculated, and a responsibility back-hanging judgment record is formed according to the degree of back-hanging support. Then, the responsibility back-hanging judgment record is bound to the corresponding anchored cross-modal risk evidence chain set to generate a responsibility back-hanging verification record.

[0060] The specific process includes, based on the counterfactual backtracking support records, verifying the closure of the anchored risk trigger chain records in the retained and shielded states of the quality responsibility positioning points. In the retained state, the closure strength is determined by whether the risk trigger nodes maintain their original trigger order, whether the risk phrases still correspond to their respective regions, and whether the risk trigger nodes can be continuously backtracked to the quality responsibility anchor points. In the shielded state, the closure strength is determined by whether the risk trigger nodes can still be continuously corresponded to other descriptive content. Finally, based on the risk trigger node breaks and the corresponding regions of the risk phrases observed during the counterfactual bypass verification, the closure strength is determined. The conflict intensity is determined by the inconsistency of domains and the interruption of the quality responsibility anchor point reattachment. The support consistency value is determined based on the continuous support of the quality responsibility positioning point for the risk triggering sequence. The degree of reattachment support of the anchored risk triggering chain record for the quality responsibility positioning point is obtained by comprehensively considering the closure strength when the support of the quality responsibility positioning point is retained, the strength of closure after the support of the quality responsibility positioning point is shielded, the conflict intensity, and the support consistency value. The degree of reattachment support is used to determine whether the anchored risk triggering chain record depends on the quality responsibility positioning point. The judgment content is associated with the closure judgment in the counterfactual reattachment support record to form a responsibility reattachment judgment record. The responsibility reattachment judgment record is bound to the corresponding anchored cross-modal risk evidence chain set to generate a responsibility reattachment verification record.

[0061] The expression for calculating the degree of support provided by the anchored risk trigger chain record to the quality responsibility location point is as follows: ; in, Indicates the first Anchored risk trigger chain record for the first The degree of backing support at each quality responsibility positioning point; Indicates the record number of the anchoring risk trigger chain; Indicates that the first [number] is reserved. When supporting a quality responsibility positioning point, the first The closure strength of the anchored risk triggering chain record; Indicates blocking the first After the first quality responsibility positioning point is supported, the first The strength at which the anchored risk trigger chain can still close; Indicates the first Anchored risk trigger chain record and the first The intensity of conflict between quality responsibility positioning points during counterfactual bypass verification; Indicates the first The quality responsibility positioning point for the first The anchor point-based risk trigger chain record supports a consistent value for the risk triggering sequence.

[0062] It should be noted that, By retaining the first After verifying the supporting evidence for each quality responsibility positioning point, the first... The risk triggering chain record is obtained by checking whether each triggering node corresponds continuously in the anchor point-based risk triggering chain record. By blocking the first After verifying the supporting evidence for each quality responsibility positioning point, the first... Whether each trigger node in the anchored risk trigger chain record can still be continuously obtained; Through statistical counterfactual bypass verification, the first Anchored risk trigger chain record and the first This was obtained from the situation of node breakage, missing evidence, and inconsistent reattachment between the quality responsibility positioning points; By checking the first The quality responsibility positioning point for the first The continuity of risk triggering sequence and the correspondence of nodes in the anchored risk triggering chain record are obtained.

[0063] Risk evidence under the same quality responsibility anchor point is read from the anchored cross-modal risk evidence chain set, and valid back-attachment evidence and candidate evidence of rebuttal are distinguished according to the responsibility back-attachment verification record to form a positive and negative evidence slot record.

[0064] The specific process includes: grouping the same risk evidence according to the quality responsibility anchor point around the anchored cross-modal risk evidence chain set; extracting the anchored risk trigger chain record corresponding to each risk evidence under the same quality responsibility anchor point; retrieving the degree of back-hanging support, closure judgment, and bypass closure relationship corresponding to each risk evidence from the responsibility back-hanging verification record; classifying risk evidence that meets the back-hanging support requirements and shows a break in the risk trigger node after shielding the support of the quality responsibility positioning point as valid back-hanging evidence; classifying risk evidence that does not meet the back-hanging support requirements or can still maintain closure through other descriptive content after shielding the support of the quality responsibility positioning point as candidate evidence for rebuttal; and associating the valid back-hanging evidence and candidate evidence for rebuttal under the same quality responsibility anchor point to the corresponding anchored cross-modal risk evidence chain set and responsibility back-hanging verification record to form a positive and negative evidence slot record.

[0065] The valid back-hanging evidence in the positive and negative evidence slotting records is verified for evidence consistency. Evidence that points to the same quality responsibility positioning point is grouped into the responsibility consistency evidence group, and evidence consistency verification records are obtained.

[0066] The specific process includes: based on the valid back-attached evidence in the positive and negative evidence slotting records, merging the valid back-attached evidence according to the same quality responsibility anchor point, and then verifying whether the risk phrase, the corresponding area of ​​the risk phrase, the anchored risk trigger chain record, and the responsibility back-attached verification record in each valid back-attached evidence all point to the same quality responsibility positioning point; when the risk phrase in the valid back-attached evidence represents the same abnormal state, the area corresponding to the risk phrase falls within the same quality responsibility positioning point, the anchored risk trigger chain record maintains the same trigger order, and the responsibility back-attached verification record confirms that the quality responsibility positioning point has a back-attached support role for the risk triggering order, the valid back-attached evidence is included in the responsibility consistency evidence group; for valid back-attached evidence with inconsistent meanings of risk phrases, risk phrase corresponding areas deviating from the quality responsibility positioning point, broken sequence of anchored risk trigger chain records, or responsibility back-attached verification records that cannot confirm the back-attached support role, it is not included in the responsibility consistency evidence group; the responsibility consistency evidence group, the commonly pointed quality responsibility positioning point, the corresponding quality responsibility anchor point, and the reason for not being included are associated to obtain the evidence consistency verification record.

[0067] Based on the evidence consistency verification record, non-quality responsibility counter-evidence is identified for candidate evidence. Evidence that cannot support the re-attachment of quality responsibility positioning points is isolated, and non-quality responsibility counter-evidence isolation records are generated.

[0068] The specific process includes: determining the abnormal state, quality responsibility positioning point, and quality responsibility anchor point of the evidence consistency group based on the evidence consistency verification record; verifying each counter-evidence candidate in the positive and negative evidence slotting record against the evidence consistency group; comparing whether the risk phrases in the counter-evidence candidate represent the same abnormal state as the risk phrases in the evidence consistency group; comparing whether the area corresponding to the risk phrase in the counter-evidence candidate falls within the scope of the quality responsibility positioning point; and further verifying whether the anchored risk trigger chain record corresponding to the counter-evidence candidate can maintain the same trigger order as the evidence consistency group. For counter-evidence candidate with deviations in the meaning of risk phrases, deviations in the area corresponding to risk phrases, broken sequence of anchored risk trigger chain records, insufficient support for back-hanging as shown in the responsibility back-hanging verification record, or still having a bypassable closure relationship after the quality responsibility positioning point is shielded, these are determined to be evidence that cannot support the back-hanging of the quality responsibility positioning point. The corresponding deviation content, closure status after shielding, back-hanging support level, and associated anchored cross-modal risk evidence chain set are isolated and bound to generate a non-quality responsibility counter-evidence isolation record.

[0069] Arbitrate the anchored cross-modal risk evidence chain set based on the non-quality liability rebuttal isolation record, and bind the conflict sources and isolation basis in the anchored cross-modal risk evidence chain set to form a multimodal evidence conflict arbitration ledger.

[0070] The specific process includes, based on the deviations in the non-quality responsibility counter-evidence isolation records, the closure status after shielding, the degree of back-hanging support, and the associated anchored cross-modal risk evidence chain set, arbitrating each piece of risk evidence within the anchored cross-modal risk evidence chain set, verifying whether the risk evidence has been included in the responsibility consistency evidence group, and further verifying whether the risk evidence has deviations in the meaning of risk phrases, deviations in the corresponding areas of risk phrases, breaks in the order of anchored risk trigger chain records, insufficient back-hanging support, and whether there are still bypass closure relationships after shielding the quality responsibility positioning point; and whether they can all point to the same quality responsibility positioning point. Furthermore, risk evidence that can support the re-attachment of quality responsibility positioning points is retained under the corresponding quality responsibility anchor point. Candidate evidence that cannot support the re-attachment of quality responsibility positioning points is separated from the evidence group of consistency of responsibility. The source of conflict is identified as deviation of the meaning of the risk phrase, deviation of the corresponding area of ​​the risk phrase, breakage of the record sequence of the anchored risk trigger chain, insufficient support for re-attachment, and the closure relationship of the bypass. The deviation content, the closure situation after shielding, and the support for re-attachment are identified as the basis for isolation. The source of conflict, the basis for isolation, and the corresponding set of anchored cross-modal risk evidence chains are bound together to form a multimodal evidence conflict arbitration ledger.

[0071] S4. Based on the multimodal evidence conflict arbitration ledger, establish a risk index database structure with the consumer product identity normalization record as the main key, and write the risk identification entries into the risk index database structure to obtain the multimodal risk information identification database for consumer product quality and safety.

[0072] The multimodal evidence conflict arbitration ledger is used to map the anchored cross-modal risk evidence chain set to the corresponding consumer product identity normalization record. After determining the consumer product identity normalization record as the primary key, the anchored cross-modal risk evidence chain set under the same primary key is merged and verified to form the consumer product risk primary key record.

[0073] The specific process includes: using the conflict sources, isolation criteria, and anchored cross-modal risk evidence chain sets in the multimodal evidence conflict arbitration ledger; organizing risk evidence retained under the same quality responsibility anchor point before warehousing; associating the anchored cross-modal risk evidence chain sets with the corresponding same consumer product object and the same production batch; and then verifying consistency with the consumer product name, brand, model, batch field, same clue number, source identifier, and carrier type in the consumer product identity normalization record; after verification, determining the consumer product identity normalization record as the primary key, and then... Anchored cross-modal risk evidence chains under the same primary key that point to the same quality responsibility positioning point, the same quality responsibility anchor point, and the same risk triggering order are merged and verified. During the merge verification, risk evidence that can jointly support the re-attachment of the quality responsibility positioning point is retained, while content marked as non-quality responsibility counter-evidence in the multimodal evidence conflict arbitration ledger is excluded. The consumer product identity normalization record, anchored cross-modal risk evidence chain set, quality responsibility anchor point, quality responsibility positioning point, conflict source, and isolation basis under the same primary key are bound to form a consumer product risk primary key record.

[0074] The primary key records for consumer product risks are subjected to an inbound stability verification. The primary key records for consumer product risks that pass the verification are organized into risk identification entries, and the risk identification entries are linked to the corresponding primary keys. Then, an index relationship is established between the primary keys, risk identification entries, and the multimodal evidence conflict arbitration ledger to form a risk index database structure.

[0075] The specific process includes, for the consumer product risk primary key record, first verifying whether the consumer product identity normalization record used as the primary key is consistent with the same consumer product object, the same production batch, and the same clue number; then verifying whether the anchored cross-modal risk evidence chain set under the same primary key still points to the same quality responsibility anchor point and quality responsibility positioning point; and verifying whether the conflict sources and isolation basis recorded in the multimodal evidence conflict arbitration ledger have excluded non-quality responsibility counter-evidence content; and ensuring that there are no batch conflicts or responsibility positioning conflicts among the consumer product identity normalization record, the anchored cross-modal risk evidence chain set, and the multimodal evidence conflict arbitration ledger. When conflicting and contradictory evidence is mixed in, the consumer product risk primary key record is determined to have passed the warehousing stability verification. The consumer product risk primary key records that have passed the verification are organized according to risk phrase, corresponding region of risk phrase, quality responsibility anchor point, quality responsibility positioning point, anchored cross-modal risk evidence chain set, and conflict source to form risk identification entries. The risk identification entries are then linked to the corresponding consumer product identity normalization record primary key to establish the query relationship between the consumer product identity normalization record primary key and the risk identification entries, and the traceability relationship between the risk identification entries and the multimodal evidence conflict arbitration ledger is established to form a risk index database structure.

[0076] The consumer product identity normalization record primary key is read from the risk index database structure. The risk identification entries under the consumer product identity normalization record primary key are associated and confirmed with the anchored cross-modal risk evidence chain set. The correspondence between the risk identification entries, the consumer product identity normalization record primary key and the multimodal evidence conflict arbitration ledger is verified to form a closed loop for writing verification records.

[0077] The specific process includes: determining the primary key of the consumer product identity normalization record based on the risk index database structure; locating the risk identification entries already attached under the consumer product identity normalization record primary key; verifying whether the risk phrase, corresponding area, quality responsibility anchor point, and quality responsibility positioning point recorded in the risk identification entry are consistent with the risk evidence in the anchored cross-modal risk evidence chain set; confirming the association between the risk identification entry and the corresponding anchored cross-modal risk evidence chain set; further verifying whether the consumer product identity normalization record primary key is still consistent with the same consumer product object, the same production batch, and the same clue number; and verifying whether the conflict source and isolation basis recorded in the multimodal evidence conflict arbitration ledger can be traced back to the corresponding risk identification entry; when there is no identity misattachment, batch misattachment, evidence chain misattachment, or missing arbitration basis among the risk identification entry, consumer product identity normalization record primary key, anchored cross-modal risk evidence chain set, and multimodal evidence conflict arbitration ledger, binding the association confirmation, primary key verification, evidence chain verification, and arbitration ledger verification to form a closed loop and write the verification record.

[0078] Based on the closed-loop write verification record, the verified risk identification entries are written into the primary key of the consumer product identity normalization record, and the index relationship between the risk identification entries, the anchored cross-modal risk evidence chain set, and the multimodal evidence conflict arbitration ledger is solidified to obtain the consumer product quality and safety multimodal risk information identification library.

[0079] The specific process includes, based on the confirmation of associations, primary key verification, evidence chain verification, and arbitration ledger verification in the closed-loop verification record, screening out risk identification entries that do not have identity mislabeling, batch mislabeling, evidence chain mislabeling, or missing arbitration basis. The verified risk identification entries are then written under the corresponding consumer product identity normalization record primary key. The risk phrase, corresponding area, quality responsibility anchor point, and quality responsibility positioning point in the risk identification entry are then fixedly associated with the consumer product identity normalization record primary key. Simultaneously, the anchored cross-modal risk evidence chain set corresponding to the risk identification entry is retained as evidence tracing basis, and the conflict sources and isolation basis in the multimodal evidence conflict arbitration ledger are identified. The data is retained as the basis for arbitration. The relationships between the consumer product identity normalization record primary key, risk identification entries, anchored cross-modal risk evidence chain set, and multimodal evidence conflict arbitration ledger are re-verified to confirm consistency among the same consumer product object, the same production batch, the same quality responsibility anchor point, and the same quality responsibility positioning point. The evidence tracing relationship between the risk identification entries and the anchored cross-modal risk evidence chain set, and the arbitration tracing relationship between the risk identification entries and the multimodal evidence conflict arbitration ledger are then solidified to obtain a consumer product quality and safety multimodal risk information identification database that allows querying risk identification entries and tracing evidence chains and arbitration bases by the consumer product identity normalization record primary key.

[0080] In summary, this invention achieves the following: by linking the abnormal areas corresponding to risk phenomena back to the quality responsibility positioning point, forming a quality responsibility anchored risk element identification record, the risk phenomena are linked back to the quality responsibility positioning point, thus improving the directionality of risk responsibility and reducing the misclassification of similar risks; by aligning the same product, batch, and risk trigger chain and isolating non-quality responsibility counter-evidence, a multimodal evidence conflict arbitration ledger is formed, realizing the credible aggregation of multi-source risk clues under the same consumer product, batch, risk trigger chain, and quality responsibility anchor point, thus reducing misjudgment of non-quality responsibility risks and enhancing the traceability of the risk information identification database.

[0081] It should be noted that the above embodiments are only used to illustrate the technical solutions 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 solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for constructing a library of multi-modal risk information identification for quality and safety of consumer products, characterized in that, include: Collect multi-source risk data on consumer product quality and safety, perform multi-modal cleaning, identification, and unified calibration to form a standardized multi-modal risk clue package; The standardized multimodal risk clue package is verified for consumer product identity and batch normalized to obtain a consumer product identity normalization record. Risk phenomena and dangerous parts are extracted from the consumer product identity normalization record. The risk phenomena are then linked back to the quality responsibility positioning point to form a quality responsibility anchored risk element identification record. Align the quality responsibility anchored risk element identification records with the same product, batch, and risk trigger chain to generate an anchored cross-modal risk evidence chain set. Based on the correspondence between the anchored risk trigger chain records and the quality responsibility positioning points, generate responsibility back-hanging verification records. Then, verify the consistency of evidence and arbitrate non-quality responsibility counter-evidence by isolating the anchored cross-modal risk evidence chain set and the responsibility back-hanging verification records to form a multimodal evidence conflict arbitration ledger. Based on the multimodal evidence conflict arbitration ledger, a risk index database structure with the consumer product identity normalization record as the main key is established, and risk identification entries are written into the risk index database structure to obtain a multimodal risk information identification database for consumer product quality and safety.

2. The method of claim 1, wherein the method further comprises: The multi-source risk data on consumer product quality and safety includes consumer product complaint text data, consumer product packaging image data, consumer product usage video data, consumer product testing report data, consumer product after-sales repair record data, and consumer product recall notice data.

3. The method of claim 2, wherein the method further comprises: The specific steps for forming a standardized multimodal risk clue package are as follows: Collect multi-source risk data on the quality and safety of consumer products, and aggregate them according to source identification and carrier type to generate original multi-source risk carrier records; Multimodal cleaning and identification are performed on the original multi-source risk carrier records, and consumer product identity information and risk description information are extracted to obtain multimodal risk identification records; Multimodal risk identification records are uniformly calibrated under the same time reference and the same clue number, and the calibrated multimodal risk identification records are packaged to form a standardized multimodal risk clue package.

4. The method of claim 3, wherein the method further comprises: The specific steps for obtaining the consumer product identity normalization record are as follows: Consumer product identity information is read from the standardized multimodal risk clue package, and the consumer product identity information is processed to standardize the fields to form a set of candidate identity fields; The consumer product identity anchoring verification is performed on the candidate identity field set. First, the identity field that can uniquely point to the consumer product is locked. Then, the clues with inconsistent names are reviewed by using the packaging layout similarity relationship. Clues that are confirmed to belong to the same consumer product object are merged to obtain the consumer product identity verification record. The batch field is extracted from the consumer product identity verification record for consistency comparison. After the batch conflict clues are isolated, they are merged into clues of the same production batch to form the consumer product batch unification record. By binding the consumer product identity verification record with the consumer product batch normalization record, a consumer product identity normalization record is generated.

5. The method of claim 4, wherein the method further comprises: The specific steps for forming a quality responsibility anchored risk factor identification record are as follows: Read the standardized multimodal risk clue package corresponding to the same consumer product object and the same production batch from the consumer product identity normalization record, and extract risk description statements and visible abnormal areas from the standardized multimodal risk clue package to form risk phenomenon candidate records; Risk phrase identification and abnormal area localization are performed on candidate records of risk phenomena, and risk phrases are associated with corresponding visible abnormal areas to form risk phenomenon localization records; Based on the risk phenomenon location records, a verifiable quality responsibility location table for consumer products is established, and the visible abnormal areas in the risk phenomenon location records are matched with the verifiable quality responsibility location table for consumer products to form a quality responsibility candidate location record. The strength of liability reinstatement is calculated using the candidate location records of quality responsibility. The quality responsibility positioning points are then sorted according to the strength of liability reinstatement. The quality responsibility positioning point with the highest ranking that meets the reinstatement requirements is determined as the quality responsibility anchor point. The quality responsibility anchor point is then written back to the corresponding consumer product identity normalization record to form a quality responsibility anchor point-based risk element identification record.

6. The consumer goods quality safety multi-modal risk information identification library construction method of claim 1 or 5, wherein, The specific steps for generating the anchored cross-modal risk evidence chain set are as follows: Extract the risk trigger sequence from the risk element identification record anchored by quality responsibility, and organize the risk trigger sequence according to the quality responsibility anchor to form an anchored risk trigger chain record; Based on the anchored risk trigger chain records, a responsibility anchor sequence diagram is constructed, and anchored risk trigger chain records that do not belong to the same consumer product identity and the same production batch are isolated to obtain candidate pairs of trigger chains for the same product and the same batch. For candidate pairs of trigger chains of the same product and batch, sequential matching is performed. The risk trigger nodes in one clue are compared with the corresponding nodes in another clue in turn, and it is verified whether each node is linked back to the same quality responsibility anchor point. If the trigger order of the two clues can be continuously matched and the quality responsibility anchor points are consistent, the two clues are merged into the same risk evidence and continue to be aggregated under the same quality responsibility anchor point to generate an anchored cross-modal risk evidence chain set.

7. The method of claim 1, wherein the method further comprises: The specific steps for generating the responsibility re-attachment verification record are as follows: The anchored risk trigger chain record and the quality responsibility location point are read by using the anchored cross-modal risk evidence chain set. After establishing the responsibility back-hanging verification task, the counterfactual bypass verification is performed. After shielding the quality responsibility location support, it is determined whether the anchored risk trigger chain record can still be closed, and the counterfactual back-hanging support record is obtained. Based on the counterfactual back-hanging support record, the degree of back-hanging support of the anchored risk trigger chain record to the quality responsibility positioning point is calculated, and a responsibility back-hanging judgment record is formed according to the degree of back-hanging support. Then, the responsibility back-hanging judgment record is bound to the corresponding anchored cross-modal risk evidence chain set to generate a responsibility back-hanging verification record.

8. The method for constructing a multimodal risk information identification database for consumer product quality and safety as described in claim 7, characterized in that, The specific steps for forming the multimodal evidence conflict arbitration ledger are as follows: Risk evidence under the same quality responsibility anchor point is read from the anchored cross-modal risk evidence chain set, and valid back-up evidence and candidate evidence of rebuttal are distinguished according to the responsibility back-up verification record to form a positive and negative evidence slot record; The valid back-hanging evidence in the positive and negative evidence slotting record is verified for evidence consistency. Evidence that points to the same quality responsibility positioning point is grouped into the responsibility consistency evidence group to obtain the evidence consistency verification record. Based on the evidence consistency verification record, non-quality responsibility counter-evidence is identified for candidate counter-evidence, and evidence that cannot support the re-attachment of quality responsibility positioning points is isolated, generating non-quality responsibility counter-evidence isolation records. Arbitrate the anchored cross-modal risk evidence chain set based on the non-quality liability rebuttal isolation record, and bind the conflict sources and isolation basis in the anchored cross-modal risk evidence chain set to form a multimodal evidence conflict arbitration ledger.

9. The method for constructing a multimodal risk information identification database for consumer product quality and safety as described in claim 8, characterized in that, Based on the multimodal evidence conflict arbitration ledger, a risk index database structure with consumer product identity normalization records as the primary key is established. The specific steps are as follows: The multimodal evidence conflict arbitration ledger is used to map the anchored cross-modal risk evidence chain set with the corresponding consumer product identity normalization record. After determining the consumer product identity normalization record as the primary key, the anchored cross-modal risk evidence chain set under the same primary key is merged and verified to form the consumer product risk primary key record. The primary key records for consumer product risks are subjected to an inbound stability verification. The primary key records for consumer product risks that pass the verification are organized into risk identification entries, and the risk identification entries are linked to the corresponding primary keys. Then, an index relationship is established between the primary keys, risk identification entries, and the multimodal evidence conflict arbitration ledger to form a risk index database structure.

10. The method for constructing a multimodal risk information identification database for consumer product quality and safety as described in claim 9, characterized in that, The specific steps for obtaining the multimodal risk information identification database for consumer product quality and safety are as follows: Read the consumer product identity normalization record primary key from the risk index database structure, associate and confirm the risk identification entries under the consumer product identity normalization record primary key with the anchored cross-modal risk evidence chain set, and verify the correspondence between the risk identification entries, the consumer product identity normalization record primary key and the multimodal evidence conflict arbitration ledger to form a closed loop write verification record. Based on the closed-loop write verification record, the verified risk identification entries are written into the primary key of the consumer product identity normalization record, and the index relationship between the risk identification entries, the anchored cross-modal risk evidence chain set, and the multimodal evidence conflict arbitration ledger is solidified to obtain the consumer product quality and safety multimodal risk information identification library.

Images