Analysis system and method for AI intelligent labeling and error correction platform

Abstract

The invention discloses an analysis system and method for an AI intelligent labeling and error correction platform, and belongs to the technical field of artificial intelligence. The method comprises the following steps: splitting an image set and a text set, and selecting a seed sample to execute fusion labeling to obtain an initial labeling sample; based on the initial annotation sample, taking the image set and the text set as topological nodes, establishing initial association mapping, and forming a topological structure; in the topological structure node range, many-to-many matching is carried out on texts or images newly entering a scene; accumulating the matching frequency of each successfully matched associated pair; for each new image and text matching request, judging whether label-free is triggered or not; when no labeling is triggered, abnormity judgment is carried out through a convolutional neural network, and a labeling abnormity set is obtained; constructing an error correction training data set, and training a random forest model; the random forest model outputs a feature importance list, and locates an error source based on feature importance; and generating and executing an error correction scheme based on the error source.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to an analysis system and method for an AI intelligent labeling and error correction platform. BACKGROUND

[0002] With the deep penetration of artificial intelligence technology in vertical fields such as industrial detection, cultural heritage digitization, network equipment management, intelligent customer service, multi-modal data has become the core basis of AI model training and scene business decision-making. In such scenarios, images usually carry fixed functional attributes, and texts correspond to specific semantic information. Both of them need to be accurately associated to support subsequent labeling, analysis and application.

[0003] Traditional methods rely on static rules or single feature matching for the association of images and texts, and lack a structured organizational framework. This static nature is manifested in complex scenarios as follows: the matching error between image functional features and text semantic features accumulates with the growth of data size, and it is difficult to trace the error source. Existing technologies either rely on a large amount of manual labeling, resulting in high costs and the introduction of subjective errors, or use weak supervision learning but have limited generalization ability. Existing anomaly detection methods lack explainability and can only output abnormal results without locating the source. SUMMARY

[0004] The purpose of the present application is to provide an analysis system and method for an AI intelligent labeling and error correction platform to solve the problems existing in the prior art.

[0005] To achieve the above purpose, the present application provides the following technical solutions: In a first aspect, the present application provides an analysis method for an AI intelligent labeling and error correction platform, comprising the following steps: According to the fixed function of the images in the scene, the image set is split, and according to the semantics of the texts in the scene, the text set is split; a seed sample is selected from the scene data to perform fusion labeling to obtain an initial labeling sample; based on the initial labeling sample, the image set and the text set are taken as topological nodes to establish an initial association mapping to form a topological structure; Within the scope of the topological structure nodes, the texts or images newly entering the scene are matched in pairs; for each successfully matched association pair, its matching frequency is accumulated; Based on the scene data size and the labeling accuracy requirement, the image labeling-free threshold and the text labeling-free threshold are set; for each new image and text matching request, the matching frequency is obtained to determine whether the labeling-free is triggered; when the labeling-free is triggered, a normal sample feature library is constructed, feature extraction and abnormality calculation are performed, and abnormality determination is performed through a convolutional neural network to obtain a labeling abnormality set; Based on the labeled exception set, a correction training data set is constructed, a random forest model is trained, a feature importance list is output by the random forest model, an error source is located based on the feature importance, and a correction scheme is generated based on the error source and executed.

[0006] In combination with the first aspect, in a first implementation manner of the first aspect of the present application, the fixed function of the scene-in image is split into an image set, and the semantic of the scene-in text is split into a text set, including: According to actual application requirements, scene features are defined, including image types, text types, and business association logic of images and texts; historical labeled data and unlabeled data under the scene features are collected to form an initial data set; all images in the initial data set are extracted to extract functional features, including spatial position features, visual attribute features, and business association features, wherein the spatial position features include coordinate ranges and relative position relationships of regions in the image, the visual attribute features include size proportions, color distributions, and texture features of the regions, and the business association features represent roles of the image regions in a scene business; functional similar groups are formed by aggregating image regions that satisfy a similarity threshold based on similarity degrees of functional features of any two image regions; each functional similar group is given a unique identifier and defined as an image set; all texts in the initial data set are extracted to extract semantic features, including surface semantic features, deep semantic features, and business semantic features, wherein the surface semantic features include keywords, part-of-speech combinations, and sentence patterns, the deep semantic features include semantic vectors and semantic roles, and the business semantic features represent meanings of the texts in the scene business; semantic similar groups are formed by aggregating texts that satisfy a similarity threshold based on similarity degrees of semantic features of any two texts; each semantic similar group is given a unique identifier and defined as a text set.

[0007] In combination with the first aspect, in a second implementation manner of the first aspect of the present application, the initial association mapping is established based on the initial labeled samples, the image set and the text set are taken as topological nodes, a topological structure is formed, including: Based on the business association logic of the images and the texts, potential associated text set nodes are matched for each image set node, potential associated image set nodes are matched for each text set node, and a candidate pair is formed; seed samples of fusion labeling are called from the initial labeled samples, actual association relationships of images and texts in each seed sample are extracted; actual association frequencies of each candidate pair in the seed samples are counted, when the actual association frequency is not less than A% of the total number of the seed samples, the candidate pair is confirmed as an effective initial association pair, otherwise, the candidate pair is removed, A is a positive integer defined according to the scene; a seed sample association confidence CC of the effective initial association pair is recorded, and the formula is CC=AAF / t, wherein AAF is the actual association frequency, and t is the total number of the seed samples; The directed weighted graph is selected as a topology structure type, nodes are standardized image set nodes and text set nodes, node attributes include ID, type, function or semantic feature abstract, and seed sample quantity, edges connect nodes of effective initial association pairs, and attributes of the edges include association direction, seed sample association confidence, and association type; rules of allowed association range are formulated, specifically: only nodes with topology edges are allowed to be associated, and nodes without topology edges are prohibited from association; for node pairs with topology edges, priority of the allowed association range is divided into high, medium, and low based on CC; only nodes with completely consistent scene adaptation labels are allowed to be associated; for each node, image set node lists and text set node lists are generated according to the rules.

[0008] With reference to the first aspect, in a third implementation manner of the first aspect, the multi-to-multi matching on the text or image newly entering the scene within the range of the topology structure nodes comprises: The text or image newly entering the scene is preprocessed; all text set nodes in the topology structure are traversed to extract semantic feature references, specifically, semantic feature averages of all samples under the nodes; semantic similarity degrees of a semantic feature set of the new text and semantic feature references of the text set nodes are calculated, the text set node with the highest similarity degree is screened, and the highest similarity degree is obtained; when the highest similarity degree is not less than a scene preset attribution threshold, it is determined that the new text belongs to the text set node with the highest similarity degree; when the highest similarity degree is less than the scene preset attribution threshold, the new text is temporarily stored in a to-be-confirmed text pool, and manual assistance attribution is triggered; all image set nodes in the topology structure are traversed to extract function feature references, specifically, function feature averages of all samples under the nodes; function similarity degrees of a function feature set of the new image and function feature references of the image set nodes are calculated, the image set node with the highest similarity degree is screened, and the highest similarity degree is obtained; when the highest similarity degree is not less than the scene preset attribution threshold, it is determined that the new image belongs to the image set node with the highest similarity degree; when the highest similarity degree is less than the scene preset attribution threshold, the new image is temporarily stored in a to-be-confirmed image pool, and manual assistance attribution is triggered. Each image set node in the image set node list and each text set node in the text set node list are traversed to perform multi-to-multi matching respectively, one new text corresponds to multiple image set nodes, each image set node corresponds to multiple image samples, one new image corresponds to multiple text set nodes, and each text set node corresponds to multiple text samples.

[0009] With reference to the first aspect, in a fourth implementation manner of the first aspect, the matching frequency of each new image and text is obtained, and it is determined whether to trigger the label-free annotation, comprising: Receiving a new image and text matching request each time, obtaining the respective matching frequency; when the cumulative matching frequency of the image is not less than the image exempt from labeling threshold, triggering exempt from labeling; when the cumulative matching frequency of the text is not less than the text exempt from labeling threshold, triggering exempt from labeling; when only the image direction meets the standard, triggering exempt from labeling for the matching of the new text to the image; when only the text direction meets the standard, triggering exempt from labeling for the matching of the new image to the text; when both meet the standard, triggering exempt from labeling for both directions of matching; when neither meets the standard, not triggering exempt from labeling, and the matching request enters the manual labeling process.

[0010] In combination with the first aspect, in a fifth implementation manner of the first aspect of the present application, when exempt from labeling is triggered, a normal sample feature library is constructed, feature extraction and abnormality calculation are performed, abnormality determination is performed through a convolutional neural network, and a labeled abnormality set is obtained, including: From the historical exempt from labeling associated pairs that have been confirmed to have no labeling abnormality, image samples are extracted as a normal reference set; the images in the normal reference set are input into a pre-trained convolutional neural network, deep visual feature vectors are extracted, a normal feature cluster is generated through a K-means clustering algorithm, and the cluster center and feature distribution range are recorded; the image samples in the associated pairs that trigger exempt from labeling are standardized to be consistent with the input format of the convolutional neural network; the image samples are input into the convolutional neural network, deep visual feature vectors thereof are extracted, the distance between the feature vectors and the normal feature cluster center is calculated as an abnormality score; an abnormality determination threshold is set, when the abnormality score of the image sample exceeds the threshold, the exempt from labeling associated pair corresponding to the image is marked as a potential abnormality; for all potential abnormality associated pairs, the abnormality activation area thereof is analyzed through back propagation of the convolutional neural network, when the overlap degree between the abnormality area and the image core function feature area is greater than a set threshold, the associated pair is confirmed to be a labeled abnormality, and is added to the labeled abnormality set.

[0011] In combination with the first aspect, in a sixth implementation manner of the first aspect of the present application, based on the labeled abnormality set, a correction training data set is constructed, and a random forest model is trained, including: All associated pairs are extracted from the labeled abnormality set and are marked as error associations, correct association samples are screened from historical normal association data, the positive and negative sample ratio is controlled to be 1:1 through random oversampling, feature engineering is performed to obtain association strength features, feature matching degree features, abnormality related features and set attribute features, pre-processing is performed to obtain a training set feature matrix and a label; the random forest model is initialized and the hyperparameters are set, the training set feature matrix and the label are input, bootstrap sampling is performed to generate independent training subsets for each decision tree, each tree is recursively split based on a random feature subset, model training is performed, and model evaluation and optimization are performed after the training is completed.

[0012] In a seventh implementation manner of the first aspect, the random forest model outputs a feature importance list, and the error source is located based on the feature importance, including: a false probability threshold is set, and a high-risk association pair exceeding the threshold is screened out from the labeled anomaly set; the random forest model outputs a feature importance list, and the feature importance list is arranged in descending order of contribution to the error association determination, and specific numerical values are extracted to form a feature value matrix; the feature values of the high-risk association pair are compared with a normal range, and an abnormal feature exceeding the range is marked to locate the error source.

[0013] In an eighth implementation manner of the first aspect, the error correction scheme is generated based on the error source and executed, including: Based on the error source, the high-risk association pair is classified according to the source type, and a corresponding error correction strategy library is preset for each type of error source; for a single high-risk association pair, candidate replacement objects are screened from the corresponding node list based on the rules of the allowed association range in the topology structure; for each candidate replacement object, the association confidence and the feature matching stability of the new association pair are calculated, and when the association confidence is greater than a preset automatic execution threshold and the feature matching stability is less than an upper limit of fluctuation, the error correction is automatically executed based on the error correction strategy library, the original association pair is replaced by the new association pair, and the properties of the corresponding edge in the topology structure are updated.

[0014] In a second aspect, an analysis system for an AI intelligent labeling and error correction platform is provided, including: a topology structure construction module including a set splitting unit, a labeled text generation unit, and a topology structure construction unit; the set splitting unit splits image sets according to fixed functions of images in a scene and splits text sets according to semantics of texts in the scene; the labeled text generation unit selects seed samples from scene data to perform fusion labeling to obtain initial labeled samples; and the topology structure construction unit establishes an initial association mapping based on the initial labeled samples, takes the image sets and the text sets as topology nodes, and forms a topology structure; an association degree matching module including a many-to-many matching unit and a matching frequency accumulation unit; the many-to-many matching unit performs many-to-many matching on texts or images newly entering the scene within the range of the topology structure nodes; and the matching frequency accumulation unit accumulates the matching frequency of each successfully matched association pair; The labeling exception set generation module comprises a threshold setting unit, an exempt labeling judgment unit and a labeling exception set generation unit; wherein the threshold setting unit sets image exempt labeling thresholds and text exempt labeling thresholds based on scene data size and labeling accuracy requirements; the exempt labeling judgment unit obtains the matching frequency of each new image and text matching request and judges whether exempt labeling is triggered; the labeling exception set generation unit constructs a normal sample feature library when exempt labeling is triggered, performs feature extraction and abnormality calculation, performs abnormality determination through a convolutional neural network, and obtains a labeling exception set; The error correction scheme generation and execution module comprises a random forest model training unit, an error root location unit and an error correction scheme generation and execution unit; wherein the random forest model training unit constructs an error correction training data set based on the labeling exception set and trains a random forest model; the random forest model outputs a feature importance list, and the error root location unit locates the error root based on the feature importance; the error correction scheme generation and execution unit generates an error correction scheme based on the error root and executes it.

[0015] Compared with the prior art, the beneficial effects of the present application are: 1、The present application breaks through the limitations of traditional static association through a double-layer architecture of function and semantic set splitting and topological node mapping. On the one hand, the image fixed function and the text semantic feature are split and set, realizing fine organization of data; on the other hand, the topological structure node records the association confidence, priority and other attributes, dynamically constraining the matching range, solving the problem of semantic fragmentation of multi-modal data.

[0016] 2、The present application introduces a dynamic threshold mechanism to realize human-computer collaborative optimization of the labeling process. By setting image and text exempt labeling thresholds, high-frequency reliable association pairs are automatically exempted from manual verification, improving efficiency; for the association pairs that trigger exempt labeling, deep features are extracted and abnormality is calculated using a convolutional neural network, and precise abnormality screening is realized in combination with the normal sample feature library.

[0017] 3、The present application quantifies the error contribution degree through feature importance sorting, realizes the traceability path from feature abnormality to business logic verification, and this interpretable error correction has a significant improvement in efficiency compared with manual intervention, and through dynamic updating of the topological structure and the normal sample library, the labeling robustness is continuously optimized. BRIEF DESCRIPTION OF DRAWINGS

[0018] Fig. 1 is a step schematic diagram of the analysis method of the AI intelligent labeling and error correction platform of the present application; Fig. 2 is a system structure diagram of the analysis system of the AI intelligent labeling and error correction platform of the present application. DETAILED DESCRIPTION

[0019] 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.

[0020] Example: Figs. 1-2 As shown, the present invention provides a technical solution: like Fig. 1 As shown, this application provides an analysis method for an AI intelligent annotation and error correction platform, including the following steps: Step S100: Based on the fixed function of the images in the scene, split the image set and based on the semantics of the text in the scene, split the text set; select seed samples from the scene data and perform fusion annotation to obtain initial annotation samples; based on the initial annotation samples, establish initial association mapping with the image set and text set as topology nodes to form a topology structure. Specifically, based on actual application requirements, scenario features are defined, including image type, text type, and business association logic between images and text; historical labeled and unlabeled data under scenario features are collected to form an initial dataset; functional features are extracted from all images in the initial dataset, including spatial location features, visual attribute features, and business association features. Spatial location features include the coordinate range and relative positional relationship of each region in the image; visual attribute features include the size ratio, color distribution, and texture features of the region; and business association features represent the role of the image region in the scenario's business. The functional feature similarity between any two image regions is calculated, and image regions that meet the similarity threshold are aggregated into functionally similar groups; each functionally similar group is assigned a unique identifier and defined as an image set. Semantic features are extracted from all text in the initial dataset, including surface semantic features, deep semantic features, and business semantic features. Surface semantic features include keywords, part-of-speech combinations, and sentence structure; deep semantic features include semantic vectors and semantic roles; and business semantic features represent the meaning of the text in the scenario's business. The semantic feature similarity between any two texts is calculated, and texts that meet the similarity threshold are aggregated into semantically similar groups; each semantically similar group is assigned a unique identifier and defined as a text set.

[0021] Further, based on the image-text business association logic, a potential associated text set node is matched for each image set node, and a potential associated image set node is matched for each text set node, to form a candidate pair; a fusion-labeled seed sample is called from the initial labeled sample, and the actual association relationship between the image and the text in each seed sample is extracted; the actual association frequency of each candidate pair in the seed sample is counted, and when the actual association frequency is not less than A% of the total number of seed samples, the candidate pair is confirmed as an effective initial association pair, otherwise the candidate pair is eliminated, and A is a positive integer defined according to the scene; the seed sample association confidence CC of the effective initial association pair is recorded, and the formula is CC=AAF / t, wherein AAF is the actual association frequency, and t is the total number of seed samples; A directed weighted graph is selected as the topology structure type, the nodes are the standardized image set nodes and text set nodes, the node attributes include ID, type, function or semantic feature abstract, and the number of seed samples, the edges connect the nodes of the effective initial association pair, and the attributes of the edges include the association direction, the seed sample association confidence, and the association type; a rule of allowed association range is formulated, specifically: only the nodes with topology edges are allowed to be associated, and the nodes without topology edges are prohibited to be associated; for the node pairs with topology edges, the priority of the allowed association range is divided into high, medium, and low based on CC; only the nodes with the same scene adaptation label are allowed to be associated; for each node, the image set node list and the text set node list are generated according to the rule.

[0022] In a specific embodiment, the image type is explicitly defined as “motor core component visualization image” (including bearing, coil, rotor, etc. subtypes), the text type is “motor fault diagnosis description text” (including fault phenomenon, component position, maintenance suggestion, etc. subtypes), and the business association logic is “component fault visual feature in image → semantic information of corresponding fault description text”; historical labeled data (1200 labeled images, 800 labeled texts) and unlabeled data (3800 unlabeled images, 2200 unlabeled texts) under this scene are collected, and are merged to form an initial data set of 5000 images and 3000 texts.

[0023] Extracting functional features from images in the initial dataset: spatial position features label component coordinate range (e.g., bearing area coordinates [150, 200, 350, 400] pixels, relative position deviation from rotor area ≤50 pixels), visual attribute features calculate component size ratio (bearing inner diameter to outer diameter ratio mean 0.62), color distribution (fault wear area gray value ≤120), texture features (wear texture gray level co-occurrence matrix contrast ≥80), business association features label mapping relationships such as "bearing wear → corresponding mechanical failure", "coil discoloration → corresponding electrical failure", etc.; set the functional feature similarity threshold to 0.8, calculate the feature similarity of any two images, and aggregate to form three image sets: "motor bearing wear image set (ID: Img001, containing 820 images)", "motor coil ablation image set (ID: Img002, containing 750 images)", "motor rotor eccentricity image set (ID: Img003, containing 680 images)", and the remaining images are temporarily stored in the supplement set.

[0024] Extracting semantic features from text in the initial dataset: surface semantic feature extraction keywords (e.g., "bearing abnormal noise", "coil smoke"), word combination (noun + verb structure proportion 72%), sentence pattern (statement sentence proportion 85%), deep semantic feature generation 768-dimensional semantic vector through BERT model (cosine similarity of similar fault text vectors ≥0.85), business semantic feature label "describe bearing failure → associated mechanical failure", "describe coil failure → associated electrical failure"; set the semantic feature similarity threshold to 0.75, calculate the feature similarity of any two texts, and aggregate to form three text sets: "bearing wear failure description set (ID: Txt001, containing 630 texts)", "coil ablation failure description set (ID: Txt002, containing 580 texts)", "rotor eccentricity failure description set (ID: Txt003, containing 520 texts)", and the remaining texts are temporarily stored in the supplement set.

[0025] From the initial dataset's historical annotation data, randomly select 500 "image-text" samples as seed samples (including 180 bearing, 160 coil, and 160 rotor), and perform fusion annotation by 2 senior motor engineers (cross-validation consistency ≥95%), finally obtain 485 effective initial annotation samples (eliminate 15 ambiguous annotation samples).

[0026] Based on the business association logic matching candidate pairs: matching Txt001, Txt004 (the "bearing jamming description" set in the temporary text) for Img001, matching Txt002, Txt005 (the "coil overheating description" set in the temporary text) for Img002, matching Txt003, Txt006 (the "rotor vibration description" set in the temporary text) for Img003, a total of 6 groups of candidate pairs are formed; the actual association frequency of each group of candidate pairs in the 485 initial labeling samples is counted - Img001 is associated with Txt001 for 320 times, and with Txt004 for 220 times, Img002 is associated with Txt002 for 290 times, and with Txt005 for 195 times, Img003 is associated with Txt003 for 275 times, and with Txt006 for 180 times; set A = 60 (that is, the actual association frequency needs to be ≥ 60% of the total number of seed samples), calculate Img001-Txt001 (320 / 485 ≈ 65.9% ≥ 60%), Img002-Txt002 (290 / 485 ≈ 59.8% ≈ 60%, according to the scene rule to keep the integer), Img003-Txt003 (275 / 485 ≈ 56.7% < 60%, after manual review and supplement of 5 labeling samples, it reaches 280 times, 280 / 490 ≈ 57.1%, still excluded), finally confirm Img001-Txt001, Img002-Txt002 as effective initial association pairs, respectively calculate the CC value: Img001-Txt001 CC = 320 / 485 ≈ 0.66, Img002-Txt002 CC = 290 / 485 ≈ 0.60.

[0027] A directed weighted graph topology is constructed: nodes are standardized Img001, Img002 (image set nodes) and Txt001, Txt002 (text set nodes), node attribute annotations are as follows: Img001: ID=Img001, type=image set, feature summary="bearing wear visual features (gray value≤120, texture contrast≥80)", seed sample number=485; Txt001: ID=Txt001, type=text set, feature summary="bearing wear semantics", seed sample number=485; edge attribute annotations are as follows: Img001→Txt001: association direction=image→text, CC=0.66, association type=mechanical fault-component-description; Img002→Txt002: association direction=image→text, CC=0.60, association type=electrical fault-component-description. Allowable association range rules are formulated: only the above two topological edges are reserved, and no topological edge node association is prohibited; priorities are divided according to CC (CC≥0.7 is high, 0.5-0.7 is medium, and <0.5 is low), Img001-Txt001 and Img002-Txt002 are both medium priority; it is required that node scene adaptation labels are completely consistent (for example, Img001 and Txt001 are both labeled "motor fault-mechanical class", and Img002 and Txt002 are both labeled "motor fault-electrical class"); finally, a text set list (only Txt001) is generated for Img001, a text set list (only Txt002) is generated for Img002, and corresponding image set lists are generated for Txt001 and Txt002 respectively.

[0028] Step S200: within the node range of the topology structure, a many-to-many matching is performed on the text or image newly entering the scene; for each successfully matched association pair, the matching frequency is accumulated; Specifically, the newly entered text or image in the scene is preprocessed; all text set nodes in the topology structure are traversed to extract semantic feature benchmarks, specifically the semantic feature mean of all samples under the node; the semantic similarity of the semantic feature set of the new text and the semantic feature benchmark of each text set node is calculated, the text set node with the highest similarity is selected, and the highest similarity is obtained; when the highest similarity is not less than the preset attribution threshold of the scene, it is determined that the new text belongs to the text set node with the highest similarity; when the highest similarity is less than the preset attribution threshold of the scene, the new text is temporarily stored in the to-be-confirmed text pool, and manual auxiliary attribution is triggered; all image set nodes in the topology structure are traversed to extract function feature benchmarks, specifically the function feature mean of all samples under the node; the function similarity of the function feature set of the new image and the function feature benchmark of each image set node is calculated, the image set node with the highest similarity is selected, and the highest similarity is obtained; when the highest similarity is not less than the preset attribution threshold of the scene, it is determined that the new image belongs to the image set node with the highest similarity; when the highest similarity is less than the preset attribution threshold of the scene, the new image is temporarily stored in the to-be-confirmed image pool, and manual auxiliary attribution is triggered. Each image set node in the image set node list and each text set node in the text set node list are traversed for multiple-to-multiple matching, one new text corresponds to multiple image set nodes, each image set node corresponds to multiple image samples, one new image corresponds to multiple text set nodes, and each text set node corresponds to multiple text samples.

[0029] In a specific embodiment, two newly entered texts in the scene and one newly entered image in the scene are selected as to-be-processed objects, and a preprocessing operation is first performed: the content of the new text T1 is “a persistent abnormal sound occurs during the operation of the motor bearing, accompanied by slight jamming”, the preprocessing includes Chinese word segmentation (extracting keywords such as “motor bearing”, “persistent abnormal sound”, and “slight jamming”), and removing meaningless words such as “time” and “accompanied by”, and generating a 768-dimensional semantic feature vector through a BERT model; the content of the new text T2 is “a blackened mark appears on the surface of the motor coil, and the temperature rises sharply after being electrified”, and the corresponding semantic feature vector is generated after preprocessing; the new image I1 is a “motor bearing area visualization image (pixel size 512*512)”, the preprocessing includes grayscale (eliminating color interference), Gaussian noise removal (smoothing noise pixels), and pixel value normalization (mapping to the 0-1 interval), and then spatial position features (bearing area coordinates [145, 195, 348, 398]), visual attribute features (gray value mean 115, texture contrast 82), and business association features (labeled “abnormal visual performance of bearing”) are extracted to form a function feature set.

[0030] Performing ownership determination on new text: traversing 2 text collection nodes in topology structure (Txt001: bearing wear fault description, Txt002: coil ablation fault description), extracting semantic feature reference of each node - reference of Txt001 is mean of 630 text semantic vectors of the collection (768-dimensional vector, each dimension mean range 0.12-0.85), reference of Txt002 is mean of 580 text semantic vectors (each dimension mean range 0.15-0.88). Calculating semantic similarity of new text and reference: T1 similarity with Txt001 0.82, similarity with Txt002 0.35; T2 similarity with Txt001 0.38, similarity with Txt002 0.81. Scene preset ownership threshold is 0.7, therefore determining T1 belongs to Txt001, T2 belongs to Txt002, without triggering manual assistance.

[0031] Performing ownership determination on new image: traversing 2 image collection nodes in topology structure (Img001: bearing wear image, Img002: coil ablation image), extracting functional feature reference of each node - reference of Img001 is mean of 820 image functional features (spatial position coordinate mean [152, 202, 352, 402], gray value mean 118, texture contrast 80), reference of Img002 is mean of 750 image functional features (spatial position coordinate mean [210, 260, 410, 460], gray value mean 95, texture contrast 75). Calculating functional similarity of I1 and reference: similarity with Img001 0.85, similarity with Img002 0.32, higher than ownership threshold 0.7, determining I1 belongs to Img001.

[0032] Performing many-to-many matching and frequency accumulation: based on node list (Txt001 corresponds to image collection list Img001, Txt002 corresponds to image collection list Img002; Img001 corresponds to text collection list Txt001), traversing nodes for matching: T1 (belonging to Txt001) matches with 10 image samples under Img001, calculating semantic-functional cross similarity, 8 samples similarity ≥0.75, successfully matching 8 times, increasing cumulative matching frequency of associated pair (Img001-Txt001) from 325 times (value after S100 ends) to 333 times; T2 (belonging to Txt002) matches with 8 image samples under Img002, 6 samples similarity ≥0.75, successfully matching 6 times, cumulative matching frequency of associated pair (Img002-Txt002) increases from 290 times to 296 times; I1 (belonging to Img001) matches with 12 text samples under Txt001, 9 samples similarity ≥0.75, successfully matching 9 times, cumulative matching frequency of (Img001-Txt001) further increases to 342 times.

[0033] Step S300: based on the scene data size and the labeling accuracy requirement, set the image labeling exemption threshold and the text labeling exemption threshold; for each new image and text matching request, obtain its matching frequency, and determine whether to trigger labeling exemption; when labeling exemption is triggered, construct a normal sample feature library, perform feature extraction and abnormality calculation, perform abnormality determination through a convolutional neural network, and obtain a labeling abnormality set; Specifically, receive each new image and text matching request, and obtain the respective matching frequencies thereof; when the cumulative matching frequency of the image is not less than the image labeling exemption threshold, labeling exemption is triggered; when the cumulative matching frequency of the text is not less than the text labeling exemption threshold, labeling exemption is triggered; when only the image direction meets the requirement, labeling exemption is triggered for the matching of the new text to the image; when only the text direction meets the requirement, labeling exemption is triggered for the matching of the new image to the text; when both meet the requirement, labeling exemption is triggered for both directions of matching; when neither meets the requirement, labeling exemption is not triggered, and the matching request enters the manual labeling process.

[0034] Further, from the labeling exemption associated pairs that have been confirmed to have no labeling abnormality, image samples are extracted as a normal reference set; the images in the normal reference set are input into a pre-trained convolutional neural network, deep visual feature vectors are extracted, a normal feature cluster is generated through a K-means clustering algorithm, and the cluster center and feature distribution range are recorded; the image samples in the labeling exemption associated pairs that are triggered are standardized to be consistent with the input format of the convolutional neural network; the image samples are input into the convolutional neural network, deep visual feature vectors thereof are extracted, the distance between the feature vectors and the normal feature cluster center is calculated as an abnormality score; an abnormality determination threshold is set, and when the abnormality score of the image sample exceeds the threshold, the labeling exemption associated pair corresponding to the image is marked as a potential abnormality; for all potential abnormality associated pairs, the convolutional neural network is used for backward propagation analysis of the abnormality activation area thereof, and when the overlap degree between the abnormality area and the image core function feature area is greater than a set threshold, the associated pair is confirmed to be a labeling abnormality and is added to the labeling abnormality set.

[0035] In a specific embodiment, in combination with the current data size of the scene (the image set Img001 contains 820 images, and the image set Img002 contains 750 images; the text set Txt001 contains 630 texts, and the text set Txt002 contains 580 texts) and the labeling accuracy requirement (the mislabeling rate needs to be ≤3%), the labeling exemption threshold is set: the image labeling exemption threshold is 80 (that is, the cumulative matching frequency of the image set to which the image belongs is ≥80 to trigger), and the text labeling exemption threshold is 70 (that is, the cumulative matching frequency of the text set to which the text belongs is ≥70 to trigger).

[0036] Receive a new matching request: the matching request of a new image I2 (belonging to Img001, the current cumulative matching frequency of Img001 is 342 times) and a new text T3 (belonging to Txt001, the current cumulative matching frequency of Txt001 is 335 times). After obtaining the frequencies of both parties, it is judged that Img001 frequency 342 ≥ 80, Txt001 frequency 335 ≥ 70, both of which meet the standard, and the bidirectional matching (I2→T3, T3→I2) is triggered without labeling.

[0037] Construct a normal sample feature library: randomly extract 200 image samples from the historical confirmed non-abnormal Img001-Txt001 association pair as a normal reference set; input these samples into the pre-trained ResNet50 convolutional neural network (CNN), extract the 2048-dimensional deep visual feature vector output by the last fully connected layer; generate a normal feature cluster by K-means clustering (cluster number k = 1), record the cluster center vector V0 (each dimension mean range 0.08-0.92), and the feature distribution range is standard deviation σ = 0.12 (each dimension standard deviation mean).

[0038] Preprocess the image I2 in the association pair (I2-T3) that triggers the free labeling: uniformly scale to 224x224 pixels (matching ResNet50 input format), perform normalization (pixel value mapping to [-1,1] interval); input I2 into ResNet50, extract 2048-dimensional visual feature vector V1; calculate the Euclidean distance between V1 and V0, get distance value 0.35 (abnormality score). Set the abnormality judgment threshold to 0.3 (based on the 3σ principle, 0.12x2.5 = 0.3), since 0.35 > 0.3, mark the association pair (I2-T3) as potential abnormality.

[0039] Generate a heat map of I2 through CNN back propagation to analyze the abnormal activation area: the abnormal area is concentrated in the bearing outer ring edge (pixel range [180,220,310,350]), while the core functional feature area of the image is the bearing wear area (pixel range [150,200,320,370]), and the overlap degree is 65% (number of overlapping pixels / total number of core area pixels); set the overlap degree threshold to 50%, since 65% > 50%, confirm that the association pair (I2-T3) is a labeled abnormality and add it to the labeled abnormality set.

[0040] Subsequently, another matching request is received (new image I3 belongs to Img002, frequency 296 < 80; new text T4 belongs to Txt002, frequency 285 ≥ 70), only the text direction meets the standard, and the matching of I3→T4 triggers free labeling; the distance between the feature vector of I3 and the center of the Img002 normal feature cluster is 0.22 < 0.3, it is judged that there is no abnormality, and it is not added to the labeled abnormality set.

[0041] Step S400: based on the labeled exception set, build a correction training dataset, train a random forest model; the random forest model outputs a feature importance list, locates the error source based on the feature importance; generates a correction scheme based on the error source and executes it.

[0042] Specifically, all associated pairs are extracted from the labeled exception set and marked as error associations. Correct association samples are selected from historical normal association data. Through random oversampling, the positive and negative sample ratio is controlled at 1:1. Feature engineering is performed to obtain association strength features, feature matching degree features, abnormal correlation features, and set attribute features. Preprocessing is performed to obtain a training set feature matrix and labels. Model initialization and hyperparameter setting are performed on the random forest model. The training set feature matrix and labels are input. Bootstrap sampling is used to generate independent training subsets for each decision tree. Each tree is recursively split based on a random feature subset. Model training is performed. After training, model evaluation and tuning are performed.

[0043] Further, set an error probability threshold, and select association pairs exceeding the threshold from the labeled exception set as high-risk association pairs. The random forest model outputs a feature importance list, which is arranged in descending order of contribution to error association determination. The specific values are extracted to form a feature value matrix. Compare the feature values of the high-risk association pairs with the normal range, mark the abnormal features that exceed the range, and locate the error source.

[0044] Further, based on the error source, classify the high-risk association pairs by source type, and preset a corresponding correction strategy library for each type of error source. For a single high-risk association pair, based on the rules of the allowed association range in the topology structure, select candidate replacement objects from the corresponding node list. For each candidate replacement object, calculate the association confidence and feature matching stability of the new association pair. When the association confidence is greater than the preset automatic execution threshold and the feature matching stability is less than the upper limit of fluctuation, automatically execute the correction based on the correction strategy library, replace the original association pair with a new association pair, and update the properties of the corresponding edge in the topology structure.

[0045] In a specific embodiment, 120 false association samples (such as the I2-T3 association pair in the foregoing) are extracted from the labeled abnormal set, 120 correct association samples (all from the valid association pairs of Img001-Txt001 and Img002-Txt002) are screened from the historical normal association data, and the false samples are expanded to 120 through random oversampling to ensure the ratio of positive and negative samples to be 1:1. Feature engineering is performed: the association strength feature takes the "cumulative matching frequency (such as 342 times for I2-T3), seed sample association confidence CC (0.66)"; the feature matching degree feature takes the "image-text cross similarity (0.35 for I2-T3)"; the abnormal correlation feature takes the "abnormal degree score (0.35)"; and the set attribute feature takes the "function feature stability of Img001 (0.15)". Normalization is performed on the features (such as mapping the frequency 342 to 0.72, and mapping the CC value 0.66 to 0.68), and the training set (168 samples) and the verification set (72 samples) are divided according to the ratio of 7:3 to form the feature matrix and the label (false = 1, correct = 0).

[0046] Initialize the random forest model: set the number of decision trees to 150, the maximum depth to 12, and the minimum leaf node sample size to 6, and use the Gini coefficient as the splitting criterion. Input the training set feature matrix and label, and generate independent subsets for each tree through bootstrap sampling (sampling rate 70%), and randomly select 3 features for each tree to split; record the OOB error during training from the initial 0.28 to 0.14, and terminate the training in advance. Evaluation on the validation set: F1 score 0.86, precision 0.88 (false association recognition accuracy), recall 0.84 (abnormal set coverage), which meets the scene accuracy requirement, and completes the model training.

[0047] Set the false probability threshold to 0.7, input the 120 association pairs in the labeled abnormal set into the model, output the false probability of each association pair, and select 85 high-risk association pairs (false probability ≥ 0.7). The model outputs a feature importance list (descending order): cross similarity (0.32), abnormal degree score (0.28), cumulative matching frequency (0.21), and CC value (0.19). Extract the feature values of the high-risk association pair I2-T3: cross similarity 0.35, abnormal degree score 0.35, cumulative matching frequency 342, and CC value 0.66; compare the normal sample feature distribution (cross similarity normal range 0.5-0.9, abnormal degree score ≤ 0.3, cumulative matching frequency 200-400, and CC value 0.6-0.8), and mark "cross similarity 0.35 < 0.5" and "abnormal degree score 0.35 > 0.3" as abnormal features; since the cross similarity is the Top1 important feature, the error root is located as "the function feature of image I2 does not match the semantic feature of text T3".

[0048] The high-risk association is classified as a "feature mismatch class", and the strategy of "re-matching high cross-similarity text node" in the preset strategy library is called. Based on the topological structure rule (Img001 is only allowed to associate with the nodes under Txt001), the candidate object Txt001-1 (containing "bearing abnormal sound with jam" class text, the average cross-similarity of semantic feature benchmark and Img001 function feature is 0.78) is screened from the subtext set of Txt001. Calculate the indicators of the new association pair I2-Txt001-1: association confidence 0.88 (based on seed sample association frequency and cross-similarity weighted calculation), feature matching stability 0.05 (feature value fluctuation coefficient); the preset automatic execution threshold is 0.85, and the fluctuation upper limit is 0.1. Because 0.88≥0.85 and 0.05≤0.1, the error correction is automatically executed - replacing the original association pair I2-T3 with I2-Txt001-1, updating the properties of the Img001→Txt001-1 edge in the topological structure: the association confidence CC is corrected from 0.66 to 0.72, and the cumulative matching frequency is increased by 12 times (original 342 times→354 times).

[0049] As shown in Fig. 2 The present application provides an analysis system for an AI intelligent labeling and error correction platform, which comprises: A topological structure construction module, comprising: a set splitting unit, a labeled text generation unit, and a topological structure construction unit; wherein the set splitting unit splits out an image set according to the fixed functions of the images within the scene, and splits out a text set according to the semantics of the texts within the scene; the labeled text generation unit selects seed samples from the scene data to perform fusion labeling to obtain initial labeled samples; the topological structure construction unit establishes an initial association mapping based on the initial labeled samples, taking the image set and the text set as topological nodes, to form a topological structure; An association degree matching module, comprising: a many-to-many matching unit and a matching frequency accumulation unit; wherein the many-to-many matching unit performs many-to-many matching on new texts or images entering the scene within the scope of the topological structure nodes; the matching frequency accumulation unit accumulates the matching frequency of each successfully matched association pair; A labeled anomaly set generation module, comprising: a threshold setting unit, an exempt labeling judgment unit, and a labeled anomaly set generation unit; wherein the threshold setting unit sets image exempt labeling thresholds and text exempt labeling thresholds based on the scene data size and labeling accuracy requirements; the exempt labeling judgment unit obtains the matching frequency of each new image and text matching request and judges whether exempt labeling is triggered; the labeled anomaly set generation unit constructs a normal sample feature library when exempt labeling is triggered, performs feature extraction and anomaly degree calculation, and makes an abnormality judgment through a convolutional neural network to obtain a labeled anomaly set; The error correction scheme generation and execution module comprises a random forest model training unit, an error source positioning unit and an error correction scheme generation and execution unit; the random forest model training unit constructs an error correction training data set based on the labeled anomaly set, and trains a random forest model; the random forest model outputs a feature importance list, and the error source positioning unit positions an error source based on the feature importance; and the error correction scheme generation and execution unit generates an error correction scheme based on the error source and executes the error correction scheme.

[0050] It will be obvious to a person skilled in the art that the application is not limited to the details of the foregoing exemplary embodiments and can be implemented in other concrete forms without departing from the spirit or essential characteristics of the application. The foregoing embodiments are therefore considered in all respects to be illustrative and not restrictive, the scope of the application being indicated by the appended claims rather than by the foregoing description, and it is intended to encompass all changes and modifications that fall within the meaning and scope of equivalents of the claims. Any reference signs in the claims should not be construed as limiting the claims concerned.

Claims

1. An analysis method for an AI intelligent labeling and correction platform, characterized in that, The method comprises the following steps: According to the fixed function of the image in the scene, the image set is split, and the text set is split according to the semantics of the text in the scene; a seed sample is selected from the scene data to perform fusion labeling to obtain an initial labeling sample; Based on the initial labeling sample, the image set and the text set are taken as topological nodes to establish an initial association mapping and form a topological structure; Within the node range of the topological structure, the text or image newly entering the scene is matched in a many-to-many manner; the matching frequency of each successfully matched association pair is accumulated; Based on the size of the scene data and the labeling accuracy requirement, the image labeling exemption threshold and the text labeling exemption threshold are set; the matching frequency of each new image and text matching request is obtained to determine whether the labeling exemption is triggered; when the labeling exemption is triggered, a normal sample feature library is constructed, feature extraction and abnormality calculation are performed, abnormality judgment is performed through a convolutional neural network, and a labeling abnormality set is obtained; Based on the labeling abnormality set, a correction training data set is constructed, and a random forest model is trained; the random forest model outputs a feature importance list, the error root is located based on the feature importance, a correction scheme is generated based on the error root, and the correction scheme is executed.

2. The analysis method for the AI intelligent labeling and correction platform according to claim 1, wherein, According to the fixed function of the image in the scene, the image set is split, and the text set is split according to the semantics of the text in the scene, comprising: According to the actual application requirement, the scene features are defined, including the image type, the text type and the business association logic of the image and the text; the historical labeling data and the unlabeled data under the scene features are collected to form an initial data set; the functional features of all images in the initial data set are extracted, including the spatial position feature, the visual attribute feature and the business association feature, wherein the spatial position feature includes the coordinate range and the relative position relationship of each region in the image, the visual attribute feature includes the size ratio, the color distribution and the texture feature of the region, and the business association feature represents the role of each region in the image in the scene business; the functional feature similarity of any two image regions is calculated, and the functional similar groups are formed by aggregating the image regions meeting the similarity threshold; each functional similar group is given a unique identifier, and is defined as an image set; the semantic features of all texts in the initial data set are extracted, including the surface semantic feature, the deep semantic feature and the business semantic feature, wherein the surface semantic feature includes the keyword, the part-of-speech combination and the sentence pattern, the deep semantic feature includes the semantic vector and the semantic role, and the business semantic feature represents the meaning of the text in the scene business; the semantic feature similarity of any two texts is calculated, and the semantic similar groups are formed by aggregating the texts meeting the similarity threshold; each semantic similar group is given a unique identifier, and is defined as a text set.

3. The analysis method for AI intelligent labeling and correction platform according to claim 1, wherein, The initial association mapping is established based on the initial labeling sample, the image set and the text set are taken as topological nodes, and the topological structure is formed, comprising: Based on the image and text business association logic, a potential associated text set node is matched for each image set node, and a potential associated image set node is matched for each text set node to form a candidate pair; a fusion-labeled seed sample is called from the initial labeled sample, and an actual association relationship between the image and the text in each seed sample is extracted; the actual association frequency of each candidate pair in the seed sample is counted, and when the actual association frequency is not less than A% of the total number of seed samples, the candidate pair is confirmed as an effective initial association pair, otherwise the candidate pair is removed, and A is a positive integer defined according to the scene; the seed sample association confidence CC of the effective initial association pair is recorded, and the formula is CC=AAF / t, wherein AAF is the actual association frequency, and t is the total number of seed samples; A directed weighted graph is selected as the topology structure type, the nodes are the standardized image set nodes and text set nodes, the node attributes include ID, type, function or semantic feature abstract and seed sample number, the edges connect the nodes of the effective initial association pair, and the attributes of the edges include the association direction, the seed sample association confidence and the association type; a rule of allowed association range is formulated, specifically: only the nodes with topology edges are allowed to be associated, and the nodes without topology edges are prohibited to be associated; for the node pairs with topology edges, the priority of the allowed association range is divided into high, medium and low based on CC; only the nodes with the same scene adaptation label are allowed to be associated; for each node, the image set node list and the text set node list are generated according to the rule.

4. The analysis method for AI intelligent labeling and correction platform according to claim 1, wherein, In the range of the topology structure nodes, a many-to-many matching is performed on the text or image newly entering the scene, including: The text or image newly entering the scene is preprocessed; all text set nodes in the topology structure are traversed to extract semantic feature benchmarks, specifically the semantic feature mean values of all samples under the node; the semantic similarity between the semantic feature set of the new text and the semantic feature benchmarks of each text set node is calculated, the text set node with the highest similarity is selected, and the highest similarity is obtained; when the highest similarity is not less than the scene preset attribution threshold, it is determined that the new text belongs to the text set node with the highest similarity; when the highest similarity is less than the scene preset attribution threshold, the new text is temporarily stored in the to-be-confirmed text pool, and manual assistance is triggered; all image set nodes in the topology structure are traversed to extract function feature benchmarks, specifically the function feature mean values of all samples under the node; the function similarity between the function feature set of the new image and the function feature benchmarks of each image set node is calculated, the image set node with the highest similarity is selected, and the highest similarity is obtained; when the highest similarity is not less than the scene preset attribution threshold, it is determined that the new image belongs to the image set node with the highest similarity; when the highest similarity is less than the scene preset attribution threshold, the new image is temporarily stored in the to-be-confirmed image pool, and manual assistance is triggered; Traverse each image set node in the image set node list, traverse each text set node in the text set node list, and perform many-to-many matching respectively, one new text corresponding to multiple image set nodes, each image set node corresponding to multiple image samples, one new image corresponding to multiple text set nodes, and each text set node corresponding to multiple text samples.

5. The analysis method for AI intelligent labeling and correction platform according to claim 1, wherein, The matching frequency of each new image and text is obtained, and it is judged whether the labeling is triggered or not, including: Receiving each new image and text matching request, obtaining its respective matching frequency; when the cumulative matching frequency of the image is not less than the image labeling threshold, the labeling is triggered; when the cumulative matching frequency of the text is not less than the text labeling threshold, the labeling is triggered; when only the image direction meets the standard, the labeling is triggered for the matching of the new text to the image; when only the text direction meets the standard, the labeling is triggered for the matching of the new image to the text; when both meet the standard, the labeling is triggered for both directions; when neither meets the standard, the labeling is not triggered, and the matching request enters the manual labeling process.

6. The analysis method for AI intelligent labeling and correction platform according to claim 1, wherein, When the labeling is triggered, a normal sample feature library is constructed, feature extraction and abnormality calculation are performed, abnormality judgment is performed through a convolutional neural network, and a labeled abnormality set is obtained, including: From the history of confirmed no labeling abnormality labeling association pair, the image sample is extracted as a normal reference set; the image in the normal reference set is input into the pre-trained convolutional neural network to extract the deep visual feature vector, and the K-means clustering algorithm is used to generate a normal feature cluster, and the cluster center and feature distribution range are recorded; the image sample in the association pair triggering the labeling is standardized to be consistent with the input format of the convolutional neural network; the image sample is input into the convolutional neural network to extract its deep visual feature vector, and the distance between the feature vector and the normal feature cluster center is calculated as an abnormality score; set an abnormality judgment threshold, when the abnormality score of the image sample exceeds the threshold, mark the labeling association pair corresponding to the image as a potential abnormality; for all potential abnormal association pairs, analyze the abnormal activation area through the convolutional neural network, and when the overlap degree between the abnormal area and the image core function feature area is greater than a set threshold, confirm that the association pair is a labeled abnormality and add it to the labeled abnormality set.

7. The analysis method for AI intelligent labeling and correction platform according to claim 1, wherein, Based on the labeled abnormality set, a correction training data set is constructed, and a random forest model is trained, including: All association pairs are extracted from the labeled abnormality set and marked as error associations, correct association samples are selected from historical normal association data, the positive and negative sample ratio is controlled to be 1:1 through random oversampling, feature engineering is performed to obtain association strength features, feature matching degree features, abnormality related features and set attribute features, preprocessing is performed to obtain a training set feature matrix and a label; the random forest model is initialized and the hyperparameters are set, the training set feature matrix and the label are input, bootstrap sampling is used to generate independent training subsets for each decision tree, each tree is recursively split based on a random feature subset, the model is trained, and the model is evaluated and optimized after training. Based on the labeled abnormality set, a correction training data set is constructed, and a random forest model is trained, including: All association pairs are extracted from the labeled abnormality set and marked as error associations, correct association samples are selected from historical normal association data, the positive and negative sample ratio is controlled to be 1:1 through random oversampling, feature engineering is performed to obtain association strength features, feature matching degree features, abnormality related features and set attribute features, preprocessing is performed to obtain a training set feature matrix and a label; the random forest model is initialized and the hyperparameters are set, the training set feature matrix and the label are input, bootstrap sampling is used to generate independent training subsets for each decision tree, each tree is recursively split based on a random feature subset, the model is trained, and the model is evaluated and optimized after training.

8. The analysis method for AI intelligent labeling and correction platform according to claim 1, wherein, The random forest model outputs a feature importance list, and the error root is located based on the feature importance, including: Set an error probability threshold, filter out the associated pairs exceeding the threshold from the labeled abnormal set as high-risk associated pairs; the random forest model outputs a feature importance list, which is arranged in descending order of the contribution of the feature to the error association judgment, the specific numerical value is extracted, and a feature value matrix is formed; compare the feature values of the high-risk associated pairs with the normal range, mark the abnormal features that exceed the range, and locate the error root.

9. The analysis method for AI intelligent labeling and correction platform according to claim 1, wherein, The error correction scheme is generated and executed based on the error root, including: Based on the error root, the high-risk associated pairs are classified by root type, and a corresponding error correction strategy library is preset for each type of error root; for a single high-risk associated pair, based on the rules of the allowed association range in the topology structure, candidate replacement objects are selected from the corresponding node list; for each candidate replacement object, the association confidence and feature matching stability of the new associated pair are calculated, when the association confidence is greater than the preset automatic execution threshold, and the feature matching stability is less than the upper limit of fluctuation, the error correction is automatically executed based on the error correction strategy library, the original associated pair is replaced by the new associated pair, and the properties of the corresponding edge in the topology structure are updated.

10. The analysis system for AI intelligent annotation and correction platform, using the analysis method for AI intelligent annotation and correction platform of any one of claims 1-9, characterized in that, Including: The topology structure construction module includes: a set splitting unit, a labeled text generation unit, and a topology structure construction unit; wherein the set splitting unit splits the image set according to the fixed function of the image in the scene, and splits the text set according to the semantics of the text in the scene; the labeled text generation unit selects seed samples from the scene data to perform fusion labeling to obtain initial labeled samples; the topology structure construction unit establishes an initial association mapping based on the initial labeled samples, taking the image set and the text set as the topology nodes, to form a topology structure; The association degree matching module includes: a many-to-many matching unit and a matching frequency accumulation unit; wherein the many-to-many matching unit performs many-to-many matching on the text or image newly entering the scene within the scope of the topology structure nodes; the matching frequency accumulation unit accumulates the matching frequency of each successfully matched associated pair; The labeled abnormal set generation module includes: a threshold setting unit, an exempt labeling judgment unit, and a labeled abnormal set generation unit; wherein the threshold setting unit sets the image exempt labeling threshold and the text exempt labeling threshold based on the scene data size and the labeling accuracy requirement; the exempt labeling judgment unit obtains the matching frequency of each new image and text matching request, and judges whether the exempt labeling is triggered; the labeled abnormal set generation unit constructs a normal sample feature library when the exempt labeling is triggered, performs feature extraction and abnormality calculation, and determines the abnormality through a convolutional neural network to obtain a labeled abnormal set; The error correction scheme generation and execution module includes: a random forest model training unit, an error root positioning unit, and an error correction scheme generation and execution unit; wherein the random forest model training unit constructs an error correction training data set based on the labeled abnormal set, and trains a random forest model; the random forest model outputs a feature importance list, and the error root positioning unit locates the error root based on the feature importance; the error correction scheme generation and execution unit generates and executes the error correction scheme based on the error root.

Images