Analysis system and method for ai intelligent labeling and correction platform

Abstract

The application 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 application splits an image set and a text set, selects seed samples to perform fusion labeling, and obtains initial labeling samples; based on the initial labeling samples, 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, a new text or image entering the scene is subjected to many-to-many matching; the matching frequency of each successfully matched association pair is accumulated; whether each new image and text matching request triggers exempt labeling is judged; when exempt labeling is triggered, an abnormality is determined through a convolutional neural network to obtain an abnormal labeling set; an error correction training data set is constructed to train a random forest model; a feature importance list is output by the random forest model, and an error source is located based on the feature importance; an error correction scheme is generated based on the error source and is executed.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an analysis system and method for an AI intelligent annotation and error correction platform. Background Technology

[0002] With the deep penetration of artificial intelligence technology into vertical fields such as industrial inspection, cultural heritage digitization, network equipment management, and intelligent customer service, multimodal data has become the core foundation for AI model training and scenario-based business decisions. In such scenarios, images typically carry fixed functional attributes, while text corresponds to specific semantic information. A precise correlation between the two is necessary to support subsequent annotation, analysis, and application.

[0003] Traditional methods for associating images and text often rely on static rules or single feature matching, lacking a structured framework. This static nature manifests in complex scenarios as the matching error between image functional features and text semantic features accumulates with increasing data volume, making it difficult to trace the root cause of the error. Existing technologies either rely on extensive manual annotation, leading to high costs and introducing subjective errors, or employ weakly supervised learning with limited generalization capabilities. Furthermore, existing anomaly detection methods often lack interpretability, only outputting anomaly results without pinpointing the root cause. Summary of the Invention

[0004] The purpose of this invention is to provide an analysis system and method for an AI intelligent annotation and error correction platform to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] Firstly, this application provides an analysis method for an AI-powered intelligent annotation and correction platform, comprising the following steps:

[0007] Based on the fixed function of the images in the scene, an image set is split off, and based on the semantics of the text in the scene, a text set is split off. Seed samples are selected from the scene data and fusion annotation is performed to obtain initial annotation samples. Based on the initial annotation samples, an initial association mapping is established with the image set and text set as topological nodes to form a topological structure.

[0008] Within the topological node range, perform many-to-many matching on newly entered text or images; for each successfully matched pair, accumulate its matching frequency.

[0009] Based on the scene data scale and annotation accuracy requirements, image and text annotation exemption thresholds are set; for each new image and text matching request, the matching frequency is obtained to determine whether annotation exemption is triggered; when annotation exemption is triggered, a normal sample feature library is constructed, feature extraction and anomaly calculation are performed, and anomaly determination is performed through a convolutional neural network to obtain an anomaly annotation set.

[0010] Based on the labeled anomaly set, an error correction training dataset is constructed, and a random forest model is trained. The random forest model outputs a list of feature importance, and the root cause of the error is located based on the feature importance. An error correction scheme is generated based on the root cause of the error and executed.

[0011] In conjunction with the first aspect, in the first embodiment of the first aspect of this application, the step of splitting the image set according to the fixed function of the images in the scene and splitting the text set according to the semantics of the text in the scene includes:

[0012] Based on actual application requirements, scene features are defined, including image type, text type, and business association logic between images and text. Historical labeled and unlabeled data under scene features are collected to form an initial dataset. For all images in the initial dataset, functional features are extracted, 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. Business association features indicate the role of the image region in the scene's business. The functional feature similarity between any two image regions is calculated. 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. For all text in the initial dataset, semantic features are extracted, 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. Business semantic features indicate the meaning of the text in the scene's business. The semantic feature similarity between any two texts is calculated. 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.

[0013] In conjunction with the first aspect, in the second embodiment of the first aspect of this application, the step of establishing an initial association mapping based on the initial labeled samples, using the image set and text set as topological nodes to form a topological structure, includes:

[0014] Based on the business association logic between images and text, potential associated text set nodes are matched for each image set node, and potential associated image set nodes are matched for each text set node to form candidate pairs. Seed samples with fused annotations are called from the initial labeled samples, and the actual association relationship between images and text in each seed sample is extracted. The actual association frequency of each candidate pair in the seed samples is counted. When the actual association frequency is not less than A% of the total number of seed samples, the candidate pair is confirmed as a valid initial association pair; otherwise, the candidate pair is removed. A is a positive integer defined according to the scenario. The seed sample association confidence CC of the valid initial association pair is recorded, with the formula CC=AAF / t, where AAF is the actual association frequency and t is the total number of seed samples.

[0015] A directed weighted graph is chosen as the topology type. Nodes are standardized image set nodes and text set nodes. Node attributes include ID, type, functional or semantic feature summary, and number of seed samples. Edges connect nodes of valid initial association pairs. Edge attributes include association direction, seed sample association confidence, and association type. Rules for allowing association ranges are defined as follows: associations are only allowed between nodes with topological edges; associations between nodes without topological edges are prohibited. For node pairs with topological edges, the priority of allowed association ranges is divided into high, medium, and low based on CC (Content Capability). Associations are only allowed between nodes with completely identical scene adaptation labels. For each node, an image set node list and a text set node list are generated according to the rules.

[0016] In conjunction with the first aspect, in a third embodiment of the first aspect of this application, the step of performing many-to-many matching on newly entering text or images within the topological node range includes:

[0017] Preprocess newly entered text or images in the scene; traverse all text set nodes in the topology and extract semantic feature benchmarks, specifically the mean of semantic features of all samples under that node; calculate the semantic similarity between the semantic feature set of the new text and the semantic feature benchmarks of each text set node, and select the text set node with the highest similarity to obtain the highest similarity; when the highest similarity is not less than the scene's preset attribution threshold, determine that the new text belongs to the text set node with the highest similarity; when the highest similarity is less than the scene's preset attribution threshold, temporarily store the new text in the pending confirmation text pool and trigger manual assistance attribution; traverse all image set nodes in the topology and extract functional feature benchmarks, specifically the mean of functional features of all samples under that node; calculate the functional similarity between the functional feature set of the new image and the functional feature benchmarks of each image set node, and select the image set node with the highest similarity to obtain the highest similarity; when the highest similarity is not less than the scene's preset attribution threshold, determine that the new image belongs to the image set node with the highest similarity; when the highest similarity is less than the scene's preset attribution threshold, temporarily store the new image in the pending confirmation image pool and trigger manual assistance attribution;

[0018] Iterate through each image set node in the list of image set nodes, and iterate through each text set node in the list of text set nodes. Perform many-to-many matching for each of them. A new text corresponds to multiple image set nodes, and each image set node corresponds to multiple image samples. A new image corresponds to multiple text set nodes, and each text set node corresponds to multiple text samples.

[0019] In conjunction with the first aspect, in the fourth embodiment of the first aspect of this application, the step of obtaining the matching frequency for each new image and text matching request and determining whether to trigger annotation exemption includes:

[0020] Each new image and text matching request is received, and their respective matching frequencies are obtained. When the cumulative matching frequency of an image is not less than the image annotation exemption threshold, annotation exemption is triggered. When the cumulative matching frequency of text is not less than the text annotation exemption threshold, annotation exemption is triggered. When only the image orientation meets the standard, annotation exemption is triggered for the matching of new text to that image. When only the text orientation meets the standard, annotation exemption is triggered for the matching of new image to that text. When both meet the standard, annotation exemption is triggered for both bidirectional matching. When neither meets the standard, annotation exemption is not triggered, and the matching request enters the manual annotation process.

[0021] In conjunction with the first aspect, in the fifth embodiment of the first aspect of this application, when the annotation exemption is triggered, constructing a normal sample feature library, performing feature extraction and anomaly calculation, and determining anomalies through a convolutional neural network to obtain an anomaly set includes:

[0022] Image samples are extracted from historically confirmed unlabeled association pairs as a normal baseline set. Images from this normal baseline set are input into a pre-trained convolutional neural network (CNN) to extract deep visual feature vectors. Normal feature clusters are generated using a K-means clustering algorithm, and the cluster centers and feature distribution ranges are recorded. Image samples from association pairs triggering unlabeling are standardized to match the input format of the CNN. Image samples are then input into the CNN to extract their deep visual feature vectors. The distance between this feature vector and the center of the normal feature cluster is calculated as an anomaly score. An anomaly threshold is set; when the anomaly score of an image sample exceeds the threshold, the corresponding unlabeled association pair is marked as a potential anomaly. For all potential anomaly association pairs, their anomalous activation regions are analyzed using backpropagation within the CNN. When the overlap between the anomalous region and the core functional feature region of the image exceeds a set threshold, the association pair is confirmed as a labeled anomaly and added to the labeled anomaly set.

[0023] In conjunction with the first aspect, in the sixth embodiment of the first aspect of this application, the step of constructing an error correction training dataset based on the labeled anomaly set and training a random forest model includes:

[0024] All association pairs are extracted from the anomaly set and marked as erroneous associations. Samples with correct associations are selected from historical normal association data. By random oversampling, the ratio of positive to negative samples is controlled at 1:1. Feature engineering is performed to obtain association strength features, feature matching degree features, anomaly relevance features, and set attribute features. Preprocessing is then performed to obtain the training set feature matrix and labels. The random forest model is initialized and hyperparameters are set. The training set feature matrix and labels are input, and bootstrap sampling is used to generate independent training subsets for each decision tree. Each tree is recursively split based on random feature subsets for model training. After training, the model is evaluated and optimized.

[0025] In conjunction with the first aspect, in the seventh embodiment of the first aspect of this application, the random forest model outputs a list of feature importance, and locates the root cause of errors based on feature importance, including:

[0026] An error probability threshold is set, and association pairs exceeding the threshold are selected from the labeled anomaly set as high-risk association pairs. The random forest model outputs a list of feature importance, which is sorted in descending order of the feature's contribution to the erroneous association determination. Specific values ​​are extracted to form a feature value matrix. The feature values ​​of high-risk association pairs are compared with the normal range, and abnormal features that exceed the range are marked to locate the root cause of the error.

[0027] In conjunction with the first aspect, in the eighth embodiment of the first aspect of this application, the step of generating and executing an error correction scheme based on the root cause of the error includes:

[0028] Based on the root causes of errors, high-risk association pairs are categorized by root cause type, and a corresponding error correction strategy library is preset for each type of error root cause. For a single high-risk association pair, candidate replacement objects are selected from the corresponding node list based on the rules of the allowed association range in the topology. For each candidate replacement object, the association confidence and feature matching stability of the new association pair are calculated. When the association confidence is greater than the preset automatic execution threshold and the feature matching stability is less than the fluctuation limit, the error correction is automatically executed based on the error correction strategy library, replacing the original association pair with the new association pair and updating the attributes of the corresponding edge in the topology.

[0029] Secondly, this application provides an analysis system for an AI-powered intelligent annotation and error correction platform, including:

[0030] The topology construction module includes: a set splitting unit, a labeled text generation unit, and a topology construction unit. The set splitting unit splits images into image sets based on their fixed functions within the scene, and text sets based on the semantics of the text within the scene. The labeled text generation unit selects seed samples from the scene data and performs fusion labeling to obtain initial labeled samples. The topology construction unit, based on the initial labeled samples, establishes initial association mappings using the image sets and text sets as topology nodes to form the topology.

[0031] The correlation matching module includes a many-to-many matching unit and a matching frequency accumulation unit. The many-to-many matching unit performs many-to-many matching on newly entered text or images within the topology node range. The matching frequency accumulation unit accumulates the matching frequency for each successfully matched correlation pair.

[0032] The annotation anomaly set generation module includes: a threshold setting unit, an exemption judgment unit, and an annotation anomaly set generation unit. The threshold setting unit sets image and text exemption thresholds based on the scene data scale and annotation accuracy requirements. The exemption judgment unit obtains the matching frequency for each new image and text matching request and determines whether exemption is triggered. When exemption is triggered, the annotation anomaly set generation unit constructs a normal sample feature library, performs feature extraction and anomaly calculation, and uses a convolutional neural network to determine anomalies, thus obtaining the annotation anomaly set.

[0033] The error correction scheme generation and execution module includes: a random forest model training unit, an error root cause localization unit, and an error correction scheme generation and execution unit. The random forest model training unit constructs an error correction training dataset based on a labeled anomaly set and trains the random forest model. The random forest model outputs a list of feature importance, and the error root cause localization unit locates the error root cause based on the feature importance. The error correction scheme generation and execution unit generates and executes an error correction scheme based on the error root cause.

[0034] Compared with the prior art, the beneficial effects of the present invention are:

[0035] 1. This invention overcomes the limitations of traditional static associations through a two-layer architecture of functional and semantic set decomposition and topological node mapping. On the one hand, it decomposes the set according to fixed image functions and text semantic features to achieve refined data organization; on the other hand, it uses topological structure nodes to record attributes such as association confidence and priority, dynamically constraining the matching range and solving the problem of semantic fragmentation in multimodal data.

[0036] 2. This invention introduces a dynamic threshold mechanism to achieve human-machine collaborative optimization of the annotation process. By setting the exemption threshold for image and text annotation, high-frequency reliable association pairs are automatically exempted from manual verification, improving efficiency. For association pairs that trigger exemption from annotation, a convolutional neural network is used to extract deep features and calculate the anomaly degree, which is combined with a normal sample feature library to achieve accurate anomaly screening.

[0037] 3. This invention quantifies the contribution of errors by ranking the importance of features, and realizes the source path from feature anomalies to business logic verification. This interpretable error correction is significantly more efficient than manual intervention, and continuously optimizes the robustness of annotation by dynamically updating the topology and normal sample library. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the steps of the analysis method of the present invention for an AI intelligent annotation and error correction platform;

[0039] Figure 2 This is a system structure diagram of the analysis system of the AI ​​intelligent annotation and error correction platform of the present invention. Detailed Implementation

[0040] 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.

[0041] Example: Figures 1-2 As shown, the present invention provides a technical solution:

[0042] like Figure 1 As shown, this application provides an analysis method for an AI intelligent annotation and error correction platform, including the following steps:

[0043] 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.

[0044] 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.

[0045] Furthermore, based on the business association logic between images and text, potential associated text set nodes are matched for each image set node, and potential associated image set nodes are matched for each text set node to form candidate pairs. Seed samples with fused annotations are called from the initial labeled samples, and the actual association relationship between images and text in each seed sample is extracted. The actual association frequency of each candidate pair in the seed samples is counted. When the actual association frequency is not less than A% of the total number of seed samples, the candidate pair is confirmed as a valid initial association pair; otherwise, the candidate pair is removed. A is a positive integer defined according to the scenario. The seed sample association confidence CC of the valid initial association pair is recorded, with the formula CC=AAF / t, where AAF is the actual association frequency and t is the total number of seed samples.

[0046] A directed weighted graph is chosen as the topology type. Nodes are standardized image set nodes and text set nodes. Node attributes include ID, type, functional or semantic feature summary, and number of seed samples. Edges connect nodes of valid initial association pairs. Edge attributes include association direction, seed sample association confidence, and association type. Rules for allowing association ranges are defined as follows: associations are only allowed between nodes with topological edges; associations between nodes without topological edges are prohibited. For node pairs with topological edges, the priority of allowed association ranges is divided into high, medium, and low based on CC (Content Capability). Associations are only allowed between nodes with completely identical scene adaptation labels. For each node, an image set node list and a text set node list are generated according to the rules.

[0047] In one specific embodiment, the image type is defined as "visualized images of core motor components" (including subtypes such as bearings, coils, and rotors), the text type is defined as "description text for motor fault diagnosis" (including subtypes such as fault phenomena, component locations, and maintenance suggestions), and the business association logic is "visual features of component faults in the image → semantic information of the corresponding fault description text". Historical labeled data (1200 labeled images and 800 labeled texts) and unlabeled data (3800 unlabeled images and 2200 unlabeled texts) in this scenario are collected and merged to form an initial dataset of 5000 images and 3000 texts.

[0048] Functional features are extracted from the images in the initial dataset: spatial location features are used to label the coordinate range of each component (e.g., the bearing area coordinates are [150, 200, 350, 400] pixels, with a relative position deviation of ≤50 pixels from the rotor area); visual attribute features are used to calculate the component size ratio (the average ratio of the bearing inner diameter to the outer diameter is 0.62), color distribution (grayscale value of the fault wear area is ≤120), and texture features (the contrast of the grayscale co-occurrence matrix of the wear texture is ≥80); business association features are used to label the mapping relationship such as "bearing wear → corresponding mechanical fault" and "coil discoloration → corresponding electrical fault"; a functional feature similarity threshold of 0.8 is set, and after calculating the feature similarity between any two image areas, three image sets are aggregated: "motor bearing wear image set (ID: Img001, containing 820 images)", "motor coil ablation image set (ID: Img002, containing 750 images)", and "motor rotor eccentricity image set (ID: Img003, containing 680 images)". The remaining images are temporarily stored to be supplemented into the set.

[0049] Semantic features were extracted from the text in the initial dataset: surface semantic features included keywords (such as "bearing noise" and "coil smoke"), part-of-speech combinations (72% of which were noun + verb structures) and sentence structures (85% were declarative sentences). Deep semantic features were generated using the BERT model to produce 768-dimensional semantic vectors (cosine similarity of text vectors of the same type of fault ≥ 0.85). Business semantic features were labeled "describe bearing fault → associated with mechanical faults" and "describe coil fault → associated with electrical faults". A semantic feature similarity threshold of 0.75 was set, and after calculating the feature similarity between any two texts, three text sets were aggregated: "bearing wear fault description set (ID: Txt001, containing 630 texts)", "coil burn-out fault description set (ID: Txt002, containing 580 texts)" and "rotor eccentricity fault description set (ID: Txt003, containing 520 texts)". The remaining texts were temporarily stored to be added to the set.

[0050] From the historical labeled data of the initial dataset, 500 "image-text" samples were randomly selected as seed samples (including 180 bearings, 160 coils, and 160 rotors). Two senior motor engineers performed fusion annotation (cross-validation consistency ≥95%), and finally obtained 485 valid initial labeled samples (15 ambiguous labeled samples were removed).

[0051] Based on business association logic, candidate pairs were matched: Img001 was matched with Txt001 and Txt004 (the set of "bearing jamming descriptions" in the temporary text), Img002 was matched with Txt002 and Txt005 (the set of "coil overheating descriptions" in the temporary text), and Img003 was matched with Txt003 and Txt006 (the set of "rotor vibration descriptions" in the temporary text), forming a total of 6 candidate pairs; the actual association frequency of each candidate pair in 485 initially labeled samples was counted—Img001 was associated with Txt001 320 times and with Txt004 220 times, Img002 was associated with Txt002 290 times and with Txt005 195 times, Img003 was associated with Txt003 275 times and with Txt006 180 times; A=6 was set. 0 (meaning the actual association frequency must be ≥ 60% of the total number of seed samples), calculated as follows: Img001-Txt001 (320 / 485≈65.9%≥60%), Img002-Txt002 (290 / 485≈59.8%≈60%, rounded according to scenario rules), Img003-Txt003 (275 / 485≈56.7%<60%, manually verified and supplemented). After 280 iterations with 5 labeled samples (280 / 490≈57.1%, still discarded), Img001-Txt001 and Img002-Txt002 were finally confirmed as valid initial association pairs. Their CC values ​​were calculated as follows: CC of Img001-Txt001 = 320 / 485≈0.66, and CC of Img002-Txt002 = 290 / 485≈0.60.

[0052] Construct a directed weighted graph topology: Nodes are standardized Img001, Img002 (image set nodes) and Txt001, Txt002 (text set nodes). Node attribute annotations: Img001: ID=Img001, Type=Image set, Feature summary=“Bearing wear visual features (grayscale 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: Img001→Txt001: Association direction=Image→Text, CC=0.66, Association type=Mechanical fault-part-description; Img002→Txt002: Association direction=Image→Text, CC=0.60, Association type=Electrical fault-part-description. Establish rules for allowed association ranges: only retain the two topological edges mentioned above, and prohibit association of nodes without topological edges; prioritize nodes according to CC (CC≥0.7 is high, 0.5-0.7 is medium, <0.5 is low), with Img001-Txt001 and Img002-Txt002 being medium priority; require nodes to have completely consistent scene adaptation labels (e.g., both Img001 and Txt001 are labeled "Motor Fault - Mechanical", and both Img002 and Txt002 are labeled "Motor Fault - Electrical"); finally, generate a text set list for Img001 (only Txt001), generate a text set list for Img002 (only Txt002), and generate corresponding image set lists for Txt001 and Txt002 respectively.

[0053] Step S200: Within the scope of the topology nodes, perform many-to-many matching on newly entered text or images in the scene; for each successfully matched pair, accumulate its matching frequency;

[0054] Specifically, the process involves preprocessing newly entered text or images; traversing all text set nodes in the topology to extract semantic feature benchmarks, specifically the mean semantic features of all samples under that node; calculating the semantic similarity between the semantic feature set of the new text and the semantic feature benchmarks of each text set node, selecting the text set node with the highest similarity, and obtaining the highest similarity score; when the highest similarity score is not less than the scene's preset attribution threshold, the new text is determined to belong to the text set node with the highest similarity score; when the highest similarity score is less than the scene's preset attribution threshold, the new text is temporarily stored in the pending confirmation text pool, triggering manual assistance in attribution; traversing all image set nodes in the topology to extract functional feature benchmarks, specifically the mean functional features of all samples under that node; calculating the functional similarity between the functional feature set of the new image and the functional feature benchmarks of each image set node, selecting the image set node with the highest similarity, and obtaining the highest similarity score; when the highest similarity score is not less than the scene's preset attribution threshold, the new image is determined to belong to the image set node with the highest similarity score; when the highest similarity score is less than the scene's preset attribution threshold, the new image is temporarily stored in the pending confirmation image pool, triggering manual assistance in attribution;

[0055] Iterate through each image set node in the list of image set nodes, and iterate through each text set node in the list of text set nodes. Perform many-to-many matching for each of them. A new text corresponds to multiple image set nodes, and each image set node corresponds to multiple image samples. A new image corresponds to multiple text set nodes, and each text set node corresponds to multiple text samples.

[0056] In one specific embodiment, two newly entered texts and one newly entered image are selected as the objects to be processed. Preprocessing is performed first: the new text T1 contains the text "The motor bearing makes a continuous abnormal noise, accompanied by a slight stuttering," and preprocessing includes Chinese word segmentation (extracting keywords such as "motor bearing," "continuous abnormal noise," and "slight stuttering"), stop word removal (removing meaningless words such as "when" and "accompanied"), and generating a 768-dimensional semantic feature vector using the BERT model; the new text T2 contains the text "The surface of the motor coil shows charred marks, and the temperature rises after energizing." The degree of "sudden increase" is generated after preprocessing to produce corresponding semantic feature vectors; the new image I1 is "visualized image of motor bearing area (pixel size 512×512)", and the preprocessing includes grayscale conversion (eliminating color interference), Gaussian denoising (smoothing noisy pixels), and pixel value normalization (mapping to the 0-1 interval). Then, spatial location features (bearing area coordinates [145,195,348,398]), visual attribute features (mean grayscale value 115, texture contrast 82) and business-related features (labeled "abnormal visual performance of bearing") are extracted to form a functional feature set.

[0057] To determine the attribution of new text: Traverse the two text set nodes in the topology (Txt001: bearing wear fault description, Txt002: coil burn-out fault description), extracting the semantic feature benchmarks for each node—the benchmark for Txt001 is the mean of the 630 text semantic vectors in that set (mean values ​​for each dimension of the 768-dimensional vector range from 0.12 to 0.85), and the benchmark for Txt002 is the mean of the 580 text semantic vectors (mean values ​​for each dimension range from 0.15 to 0.88). Calculate the semantic similarity between the new text and the benchmarks: T1 has a similarity of 0.82 with Txt001 and 0.35 with Txt002; T2 has a similarity of 0.38 with Txt001 and 0.81 with Txt002. The preset attribution threshold is 0.7, therefore T1 is assigned to Txt001 and T2 to Txt002, without requiring manual intervention.

[0058] To determine the attribution of a new image: Traverse the two image set nodes in the topology (Img001: bearing wear image, Img002: coil ablation image), extracting the functional feature benchmarks for each node—Img001's benchmark is the mean of functional features from 820 images (mean spatial coordinates [152, 202, 352, 402], mean grayscale value 118, texture contrast 80); Img002's benchmark is the mean of functional features from 750 images (mean spatial coordinates [210, 260, 410, 460], mean grayscale value 95, texture contrast 75). Calculate the functional similarity between I1 and the benchmarks: similarity with Img001 is 0.85, similarity with Img002 is 0.32, both exceeding the attribution threshold of 0.7, therefore I1 is assigned to Img001.

[0059] Perform many-to-many matching and frequency accumulation: Based on the node list (Txt001 corresponds to the image set list Img001, Txt002 corresponds to the image set list Img002; Img001 corresponds to the text set list Txt001), traverse the nodes for matching: T1 (belonging to Txt001) is matched with 10 image samples under Img001, and the semantic-functional cross-similarity is calculated. Among them, 8 samples have a similarity ≥ 0.75, and 8 matches are successfully performed. The cumulative matching frequency of the association pair (Img001-Txt001) is increased from 3. The number of matches increased from 25 (values ​​after S100) to 333; T2 (belonging to Txt002) matched with 8 image samples under Img002, with 6 samples having a similarity ≥0.75, resulting in 6 successful matches, and the cumulative matching frequency of the association pair (Img002-Txt002) increased from 290 to 296; I1 (belonging to Img001) matched with 12 text samples under Txt001, with 9 samples having a similarity ≥0.75, resulting in 9 successful matches, and the cumulative matching frequency of (Img001-Txt001) further increased to 342.

[0060] Step S300: Based on the scene data scale and annotation accuracy requirements, set the image annotation exemption threshold and text annotation exemption threshold; for each new image and text matching request, obtain its matching frequency and determine whether annotation exemption is triggered; when annotation exemption is triggered, construct a normal sample feature library, perform feature extraction and anomaly calculation, and use a convolutional neural network to determine anomalies and obtain an anomaly annotation set.

[0061] Specifically, each new image and text matching request is received, and their respective matching frequencies are obtained. When the cumulative matching frequency of an image is not less than the image annotation exemption threshold, annotation exemption is triggered. When the cumulative matching frequency of text is not less than the text annotation exemption threshold, annotation exemption is triggered. When only the image orientation meets the standard, annotation exemption is triggered for the matching of new text to that image. When only the text orientation meets the standard, annotation exemption is triggered for the matching of new image to that text. When both meet the standard, annotation exemption is triggered for both bidirectional matching. When neither meets the standard, annotation exemption is not triggered, and the matching request enters the manual annotation process.

[0062] Furthermore, image samples are extracted from historically confirmed unlabeled association pairs as a normal baseline set. Images from this normal baseline set are input into a pre-trained convolutional neural network to extract deep visual feature vectors. Normal feature clusters are generated using a K-means clustering algorithm, and the cluster centers and feature distribution ranges are recorded. Image samples from association pairs triggering unlabeling are standardized to match the input format of the convolutional neural network. Image samples are input into the convolutional neural network to extract their deep visual feature vectors. The distance between this feature vector and the center of the normal feature cluster is calculated as an anomaly score. An anomaly determination threshold is set. When the anomaly score of an image sample exceeds the threshold, the corresponding unlabeled association pair is marked as a potential anomaly. For all potential anomaly association pairs, their anomalous activation regions are analyzed through backpropagation using a convolutional neural network. When the overlap between the anomalous region and the core functional feature region of the image is greater than a set threshold, the association pair is confirmed as a labeled anomaly and added to the labeled anomaly set.

[0063] In one specific embodiment, considering the current data scale of the scene (image set Img001 contains 820 images, Img002 contains 750 images; text set Txt001 contains 630 texts, Txt002 contains 580 texts) and the annotation accuracy requirements (error rate must be ≤3%), the following annotation exemption thresholds are set: the image annotation exemption threshold is 80 (i.e., triggered when the cumulative matching frequency of the image set is ≥80), and the text annotation exemption threshold is 70 (triggered when the cumulative matching frequency of the text set is ≥70).

[0064] Receive a new matching request: a matching request between a new image I2 (belonging to Img001, Img001 currently has a cumulative matching frequency of 342 times) and a new text T3 (belonging to Txt001, Txt001 currently has a cumulative matching frequency of 335 times). After obtaining the frequencies of both, determine if: Img001 frequency 342 ≥ 80, Txt001 frequency 335 ≥ 70. Both meet the criteria, triggering annotation-free matching for both bidirectional matches (I2→T3, T3→I2).

[0065] Construct a normal sample feature library: Randomly extract 200 image samples from the historically confirmed Img001-Txt001 association pairs without anomalies as a normal baseline set; Input these samples into a pre-trained ResNet50 convolutional neural network (CNN) to extract the 2048-dimensional deep visual feature vector output from the last fully connected layer; Generate normal feature clusters through K-means clustering (number of clusters k=1), and record the cluster center vector V0 (mean of each dimension ranges from 0.08 to 0.92), with the feature distribution range being the standard deviation σ=0.12 (mean of the standard deviation of each dimension).

[0066] The image I2 in the association pair (I2-T3) that triggered the label-free process was preprocessed: it was uniformly scaled to 224×224 pixels (matching the ResNet50 input format) and normalized (pixel values ​​were mapped to the [-1,1] interval); I2 was input into ResNet50 to extract a 2048-dimensional visual feature vector V1; the Euclidean distance between V1 and V0 was calculated, yielding a distance value of 0.35 (anomaly score). The anomaly detection threshold was set to 0.3 (based on the 3σ principle, 0.12×2.5=0.3). Since 0.35 > 0.3, the association pair (I2-T3) was labeled as a potential anomaly.

[0067] A heatmap of I2 was generated using backpropagation via CNN, and the abnormal activation regions were analyzed: the abnormal regions were concentrated on the outer edge of the bearing (pixel range [180,220,310,350]), while the core functional feature region of the image was the bearing wear area (pixel range [150,200,320,370]). The overlap between the two was calculated to be 65% (number of overlapping pixels / total number of pixels in the core region). The overlap threshold was set to 50%. Since 65% > 50%, the association pair (I2-T3) was confirmed as an anomaly and added to the anomaly annotation set.

[0068] 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 the exemption from annotation. The distance between the extracted feature vector of I3 and the center of the normal feature cluster of Img002 is 0.22 < 0.3, and it is determined that there is no abnormality, so it is not added to the annotation abnormal set.

[0069] Step S400: Based on the labeled anomaly set, construct the error correction training dataset and train the random forest model; the random forest model outputs a list of feature importance, and the root cause of the error is located based on the feature importance; an error correction scheme is generated based on the root cause of the error and executed.

[0070] Specifically, all association pairs are extracted from the anomaly set and marked as erroneous associations. Samples with correct associations are selected from historical normal association data. Through random oversampling, the ratio of positive to negative samples is controlled at 1:1. Feature engineering is performed to obtain association strength features, feature matching degree features, anomaly correlation features, and set attribute features. Preprocessing is then performed to obtain the training set feature matrix and labels. The random forest model is initialized and hyperparameters are set. The training set feature matrix and labels are input, and bootstrap sampling is used to generate independent training subsets for each decision tree. Each tree is recursively split based on random feature subsets for model training. After training, the model is evaluated and optimized.

[0071] Furthermore, an error probability threshold is set, and association pairs exceeding the threshold are selected from the labeled anomaly set as high-risk association pairs. The random forest model outputs a list of feature importance, which is sorted in descending order of the contribution of features to the erroneous association determination. Specific values ​​are extracted to form a feature value matrix. The feature values ​​of high-risk association pairs are compared with the normal range, and abnormal features that exceed the range are marked to locate the root cause of the error.

[0072] Furthermore, based on the root cause of the error, high-risk association pairs are categorized by root cause type, and a corresponding error correction strategy library is preset for each type of error root cause. For a single high-risk association pair, candidate replacement objects are selected from the corresponding node list based on the rules of the allowed association range in the topology. For each candidate replacement object, the association confidence and feature matching stability of the new association pair are calculated. When the association confidence is greater than the preset automatic execution threshold and the feature matching stability is less than the fluctuation limit, the error correction is automatically executed based on the error correction strategy library, replacing the original association pair with the new association pair and updating the attributes of the corresponding edge in the topology.

[0073] In one specific embodiment, 120 erroneous association samples (such as the I2-T3 association pair mentioned above) are extracted from the labeled anomaly set, and 120 correct association samples (all from the Img001-Txt001 and Img002-Txt002 valid association pairs) are selected from the historical normal association data. The erroneous samples are expanded to 120 through random oversampling to ensure a 1:1 ratio of positive to negative samples. Feature engineering is performed: the association strength feature is "cumulative matching frequency (e.g., 342 times for I2-T3) and seed sample association confidence CC (0.66)"; the feature matching degree feature is "image-text cross-similarity (0.35 for I2-T3)"; the anomaly correlation feature is "anomaly score (0.35)"; and the set attribute feature is "functional feature stability of Img001 (0.15)". Normalize the features (e.g., map the frequency 342 to 0.72 and the CC value 0.66 to 0.68), and divide them into a training set (168 samples) and a validation set (72 samples) in a 7:3 ratio to form a feature matrix and labels (error = 1, correct = 0).

[0074] Initialize the random forest model: Set the number of decision trees to 150, maximum depth to 12, minimum leaf node sample size to 6, and use the Gini coefficient as the splitting criterion. Input the training set feature matrix and labels, and generate independent subsets for each tree through bootstrap sampling (sampling rate 70%). Randomly select 3 features for splitting each tree; during training, record that the OOB error decreases from the initial 0.28 to 0.14 and then stabilizes, and terminate training early. Evaluate on the validation set: F1 score 0.86, precision 0.88 (error association recognition accuracy), recall 0.84 (outlier set coverage), meeting the scene accuracy requirements, and complete model training.

[0075] With an error probability threshold of 0.7, 120 association pairs from the labeled anomaly set were input into the model, and the error probability of each association pair was output. 85 high-risk association pairs (error probability ≥ 0.7) were then selected. The model output a list of feature importance (in descending order): cross-similarity (0.32), anomaly score (0.28), cumulative match frequency (0.21), and CC value (0.19). Feature values ​​of high-risk association pairs I2-T3 were extracted: cross-similarity 0.35, anomaly score 0.35, cumulative match frequency 342, and CC value 0.66. Compared with the feature distribution of normal samples (cross-similarity normal range 0.5-0.9, anomaly score ≤0.3, cumulative match frequency 200-400, CC value 0.6-0.8), "cross-similarity 0.35 < 0.5" and "anomaly score 0.35 > 0.3" were marked as anomalous features. Since cross-similarity is a Top 1 important feature, the root cause of the error was located as "the functional features of image I2 do not match the semantic features of text T3".

[0076] The high-risk association pair is classified as "feature mismatch" and the "re-matching of text nodes with high cross-similarity" strategy from the preset strategy library is invoked. Based on the topological structure rule (Img001 is only allowed to associate with nodes under Txt001), candidate objects Txt001-1 (including texts of the "bearing noise with stuttering" category, with an average cross-similarity of 0.78 between semantic feature benchmark and Img001 functional feature) are selected from the sub-text set of Txt001. The metrics for the new association pair I2-Txt001-1 are calculated as follows: association confidence 0.88 (weighted calculation based on seed sample association frequency and cross similarity), feature matching stability 0.05 (feature value fluctuation coefficient); the preset automatic execution threshold is 0.85 and the fluctuation limit is 0.1. Since 0.88≥0.85 and 0.05≤0.1, automatic error correction is performed—replacing the original association pair I2-T3 with I2-Txt001-1, and updating the attributes of the edge Img001→Txt001-1 in the topology: the association confidence CC is corrected from 0.66 to 0.72, and the cumulative matching frequency increases by 12 times (from 342 times to 354 times).

[0077] like Figure 2 As shown, this application provides an analysis system for an AI intelligent annotation and error correction platform, including:

[0078] The topology construction module includes: a set splitting unit, a labeled text generation unit, and a topology construction unit. The set splitting unit splits images into image sets based on their fixed functions within the scene, and text sets based on the semantics of the text within the scene. The labeled text generation unit selects seed samples from the scene data and performs fusion labeling to obtain initial labeled samples. The topology construction unit, based on the initial labeled samples, establishes initial association mappings using the image sets and text sets as topology nodes to form the topology.

[0079] The correlation matching module includes a many-to-many matching unit and a matching frequency accumulation unit. The many-to-many matching unit performs many-to-many matching on newly entered text or images within the topology node range. The matching frequency accumulation unit accumulates the matching frequency for each successfully matched correlation pair.

[0080] The annotation anomaly set generation module includes: a threshold setting unit, an exemption judgment unit, and an annotation anomaly set generation unit. The threshold setting unit sets image and text exemption thresholds based on the scene data scale and annotation accuracy requirements. The exemption judgment unit obtains the matching frequency for each new image and text matching request and determines whether exemption is triggered. When exemption is triggered, the annotation anomaly set generation unit constructs a normal sample feature library, performs feature extraction and anomaly calculation, and uses a convolutional neural network to determine anomalies, thus obtaining the annotation anomaly set.

[0081] The error correction scheme generation and execution module includes: a random forest model training unit, an error root cause localization unit, and an error correction scheme generation and execution unit. The random forest model training unit constructs an error correction training dataset based on a labeled anomaly set and trains the random forest model. The random forest model outputs a list of feature importance, and the error root cause localization unit locates the error root cause based on the feature importance. The error correction scheme generation and execution unit generates and executes an error correction scheme based on the error root cause.

[0082] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An analysis method for an AI intelligent annotation and error correction platform, characterized in that, Includes the following steps: Based on the fixed function of the images in the scene, an image set is split off; based on the semantics of the text in the scene, a text set is split off; seed samples are selected from the scene data and fusion annotation is performed to obtain initial annotation samples. Based on the initial labeled samples, an initial association mapping is established using the image set and text set as topological nodes to form a topological structure; Define scenario characteristics based on actual application requirements, including image type, text type, and business relationship logic between images and text; Collect historical labeled and unlabeled data under scene features to form an initial dataset; For all images in the initial dataset, functional features are extracted, including spatial location features, visual attribute features, and business-related 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. Business-related features represent the role of the image region in the scene's business context. The functional feature similarity between any two image regions is calculated. 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. For all text in the initial dataset, semantic features are extracted, 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. Business semantic features represent the meaning of the text in the scene's business context. The semantic feature similarity between any two texts is calculated. 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. Samples containing images and text are selected from the historical labeled data of the initial dataset as seed samples. Professionals label the seed samples and verify the consistency of the labels through cross-validation. Seed samples with ambiguous labels are removed, and the remaining samples that meet the standards are used as the initial labeled samples. Based on the business association logic between images and text, potential associated text set nodes are matched for each image set node, and potential associated image set nodes are matched for each text set node to form candidate pairs. Seed samples with fused annotations are called from the initial labeled samples, and the actual association relationship between images and text in each seed sample is extracted. The actual association frequency of each candidate pair in the seed samples is counted. When the actual association frequency is not less than A% of the total number of seed samples, the candidate pair is confirmed as a valid initial association pair; otherwise, the candidate pair is removed. A is a positive integer defined according to the scenario. The seed sample association confidence CC of the valid initial association pair is recorded, with the formula CC=AAF / t, where AAF is the actual association frequency and t is the total number of seed samples. A directed weighted graph is chosen as the topology type. Nodes are standardized image set nodes and text set nodes. Node attributes include ID, type, functional or semantic feature summary, and number of seed samples. Edges connect nodes of valid initial association pairs, and edge attributes include association direction, seed sample association confidence, and association type. Rules for allowing association ranges are defined as follows: associations are only allowed between nodes with topological edges; associations between nodes without topological edges are prohibited. For node pairs with topological edges, the priority of allowed association ranges is divided into high, medium, and low based on CC (Content Capability). Associations are only allowed between nodes with completely identical scene adaptation labels. For each node, an image set node list and a text set node list are generated according to the rules. Within the topological node range, perform many-to-many matching on newly entered text or images; for each successfully matched pair, accumulate its matching frequency. Based on the scene data scale and annotation accuracy requirements, image and text annotation exemption thresholds are set; for each new image and text matching request, the matching frequency is obtained to determine whether annotation exemption is triggered; when annotation exemption is triggered, a normal sample feature library is constructed, feature extraction and anomaly calculation are performed, and anomaly determination is performed through a convolutional neural network to obtain an anomaly annotation set. Image samples are extracted from historically confirmed unlabeled association pairs as a normal baseline set. Images from this normal baseline set are input into a pre-trained convolutional neural network (CNN) to extract deep visual feature vectors. Normal feature clusters are generated using a K-means clustering algorithm, and the cluster centers and feature distribution ranges are recorded. Image samples from association pairs triggering unlabeling are standardized to match the input format of the CNN. Image samples are then input into the CNN to extract their deep visual feature vectors. The distance between this feature vector and the center of the normal feature cluster is calculated as an anomaly score. An anomaly threshold is set; when the anomaly score of an image sample exceeds the threshold, the corresponding unlabeled association pair is marked as a potential anomaly. For all potential anomaly association pairs, their anomalous activation regions are analyzed using backpropagation within the CNN. When the overlap between the anomalous region and the core functional feature region of the image exceeds a set threshold, the association pair is confirmed as a labeled anomaly and added to the labeled anomaly set. Based on the labeled anomaly set, an error correction training dataset is constructed, and a random forest model is trained. The random forest model outputs a list of feature importance, and the root cause of the error is located based on the feature importance. An error correction scheme is generated based on the root cause of the error and executed. All association pairs are extracted from the anomaly set and marked as erroneous associations. Samples with correct associations are selected from historical normal association data. By random oversampling, the ratio of positive to negative samples is controlled at 1:

1. Feature engineering is performed to obtain association strength features, feature matching degree features, anomaly relevance features, and set attribute features. Preprocessing is then performed to obtain the training set feature matrix and labels. The random forest model is initialized and hyperparameters are set. The training set feature matrix and labels are input, and bootstrap sampling is used to generate independent training subsets for each decision tree. Each tree is recursively split based on random feature subsets for model training. After training, the model is evaluated and optimized.

2. The analysis method for an AI intelligent annotation and error correction platform according to claim 1, characterized in that, Within the scope of the topological structure nodes, performing many-to-many matching on newly entering text or images includes: Preprocess newly entered text or images in the scene; traverse all text set nodes in the topology and extract semantic feature benchmarks, specifically the mean of semantic features of all samples under that node; calculate the semantic similarity between the semantic feature set of the new text and the semantic feature benchmarks of each text set node, and select the text set node with the highest similarity to obtain the highest similarity; when the highest similarity is not less than the scene's preset attribution threshold, determine that the new text belongs to the text set node with the highest similarity; when the highest similarity is less than the scene's preset attribution threshold, temporarily store the new text in the pending confirmation text pool and trigger manual assistance attribution; traverse all image set nodes in the topology and extract functional feature benchmarks, specifically the mean of functional features of all samples under that node; calculate the functional similarity between the functional feature set of the new image and the functional feature benchmarks of each image set node, and select the image set node with the highest similarity to obtain the highest similarity; when the highest similarity is not less than the scene's preset attribution threshold, determine that the new image belongs to the image set node with the highest similarity; when the highest similarity is less than the scene's preset attribution threshold, temporarily store the new image in the pending confirmation image pool and trigger manual assistance attribution; Iterate through each image set node in the list of image set nodes, and iterate through each text set node in the list of text set nodes. Perform many-to-many matching for each of them. A new text corresponds to multiple image set nodes, and each image set node corresponds to multiple image samples. A new image corresponds to multiple text set nodes, and each text set node corresponds to multiple text samples.

3. The analysis method for an AI intelligent annotation and error correction platform according to claim 1, characterized in that, For each new image and text matching request, the matching frequency is obtained, and it is determined whether annotation exemption is triggered, including: Each new image and text matching request is received, and their respective matching frequencies are obtained. When the cumulative matching frequency of an image is not less than the image annotation exemption threshold, annotation exemption is triggered. When the cumulative matching frequency of text is not less than the text annotation exemption threshold, annotation exemption is triggered. When only the image orientation meets the standard, annotation exemption is triggered for the matching of new text to that image. When only the text orientation meets the standard, annotation exemption is triggered for the matching of new image to that text. When both meet the standard, annotation exemption is triggered for both bidirectional matching. When neither meets the standard, annotation exemption is not triggered, and the matching request enters the manual annotation process.

4. The analysis method for an AI intelligent annotation and error correction platform according to claim 1, characterized in that, The random forest model outputs a list of feature importance, which is used to locate the root cause of errors, including: An error probability threshold is set, and association pairs exceeding the threshold are selected from the labeled anomaly set as high-risk association pairs. The random forest model outputs a list of feature importance, which is sorted in descending order of the feature's contribution to the erroneous association determination. Specific values ​​are extracted to form a feature value matrix. The feature values ​​of high-risk association pairs are compared with the normal range, and abnormal features that exceed the range are marked to locate the root cause of the error.

5. The analysis method for an AI intelligent annotation and error correction platform according to claim 1, characterized in that, The generation and execution of the error correction scheme based on the root cause of the error includes: Based on the root causes of errors, high-risk association pairs are categorized by root cause type, and a corresponding error correction strategy library is preset for each type of error root cause. For a single high-risk association pair, candidate replacement objects are selected from the corresponding node list based on the rules of the allowed association range in the topology. For each candidate replacement object, the association confidence and feature matching stability of the new association pair are calculated. When the association confidence is greater than the preset automatic execution threshold and the feature matching stability is less than the fluctuation limit, the error correction is automatically executed based on the error correction strategy library, replacing the original association pair with the new association pair and updating the attributes of the corresponding edge in the topology.

6. An analysis system for an AI intelligent annotation and error correction platform, using the analysis method for an AI intelligent annotation and error correction platform as described in any one of claims 1-5, characterized in that, include: The topology construction module includes: a set splitting unit, a labeled text generation unit, and a topology construction unit. The set splitting unit splits images into image sets based on their fixed functions within the scene, and text sets based on the semantics of the text within the scene. The labeled text generation unit selects seed samples from the scene data and performs fusion labeling to obtain initial labeled samples. The topology construction unit, based on the initial labeled samples, establishes initial association mappings using the image sets and text sets as topology nodes to form the topology. The correlation matching module includes a many-to-many matching unit and a matching frequency accumulation unit. The many-to-many matching unit performs many-to-many matching on newly entered text or images within the topology node range. The matching frequency accumulation unit accumulates the matching frequency for each successfully matched correlation pair. The annotation anomaly set generation module includes: a threshold setting unit, an exemption judgment unit, and an annotation anomaly set generation unit. The threshold setting unit sets image and text exemption thresholds based on the scene data scale and annotation accuracy requirements. The exemption judgment unit obtains the matching frequency for each new image and text matching request and determines whether exemption is triggered. When exemption is triggered, the annotation anomaly set generation unit constructs a normal sample feature library, performs feature extraction and anomaly calculation, and uses a convolutional neural network to determine anomalies, thus obtaining the annotation anomaly set. The error correction scheme generation and execution module includes: a random forest model training unit, an error root cause localization unit, and an error correction scheme generation and execution unit. The random forest model training unit constructs an error correction training dataset based on a labeled anomaly set and trains the random forest model. The random forest model outputs a list of feature importance, and the error root cause localization unit locates the error root cause based on the feature importance. The error correction scheme generation and execution unit generates and executes an error correction scheme based on the error root cause.

Images