Large model data analysis enhancement method and system based on dynamic noise filtering and multi-modal feature fusion

Abstract

The invention relates to the technical field of artificial intelligence and data processing, in particular to a large model data analysis enhancement method and system based on dynamic noise filtering and multi-modal feature fusion, and the method comprises the following steps: obtaining user historical behavior data, and carrying out data cleaning and feature extraction; constructing a user behavior feature vector, and dividing a training set and a test set; training a machine learning model based on the training set, and optimizing model parameters; verifying the model performance by using the test set, and outputting a prediction result; the method has the beneficial effects that the dynamic noise sensing module is used for performing quality evaluation on a reference file uploaded by a user, and a self-adaptive threshold cleaning mechanism is used for eliminating redundancy and conflict information; a multi-granularity feature alignment network is designed, and a document structure analysis and semantic embedding technology is utilized to realize deep association mapping between a reference file and an input task.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and data processing, and specifically provides a method and system for enhancing large model data analysis based on dynamic noise filtering and multimodal feature fusion. Background Art

[0002] In the field of artificial intelligence technology, large models (such as pre-trained language models, multimodal large models) have been widely used in data analysis tasks. However, when users provide reference files (such as documents, reports, charts, etc.) to assist analysis for large models, the following core defects exist in the existing technologies:

[0003] 1. Inefficient key information extraction: Traditional methods rely on rule matching or shallow semantic analysis, and it is difficult to accurately identify the core content related to the analysis task (such as key indicators in financial reports, diagnostic bases in medical documents) from reference files, resulting in redundant information interfering with model decision-making.

[0004] 2. Insufficient cross-modal association understanding: Reference files and data to be analyzed often exist in multimodal forms (such as text, images, tables), and existing feature fusion technologies (such as simple splicing or average pooling) cannot capture the deep semantic associations between cross-modal data, restricting the comprehensiveness of analysis results.

[0005] 3. Poor domain adaptability: For reference files in different formats (PDF / DOCX / TXT) or domains (finance / medicine / industry), the model needs to rely on a large amount of domain-specific training, lacking generality, resulting in high deployment costs and low efficiency.

[0006] 4. Weak dynamic data processing ability: Facing real-time updated reference files or noise interference (such as text typos, blurred images), existing systems are difficult to dynamically adjust processing strategies, affecting the accuracy and timeliness of analysis results.

[0007] These problems are particularly prominent in scenarios with extremely high requirements for data quality and analysis accuracy, such as financial risk control and medical diagnosis, and there is an urgent need for a technical solution that can efficiently integrate the knowledge of reference files, enhance cross-modal association analysis capabilities, and adapt to diverse inputs. Summary of the Invention

[0008] The purpose of the present invention is to provide a method and system for enhancing large model data analysis based on dynamic noise filtering and multimodal feature fusion to solve the problems mentioned in the above background art.

[0009] To achieve the above purpose, the present invention provides the following technical solution: A method for enhancing large model data analysis based on dynamic noise filtering and multimodal feature fusion, including the following steps:

[0010] Obtain the user's historical behavior data, and perform data cleaning and feature extraction;

[0011] Construct a user behavior feature vector, and divide the training set and the test set;

[0012] Train a machine learning model based on the training set and optimize the model parameters;

[0013] Use the test set to verify the model performance and output the prediction results.

[0014] Preferably, the data cleaning and feature extraction steps further include:

[0015] Remove duplicate data, fill in missing values and handle outliers;

[0016] Extract user behavior pattern features through time series analysis or clustering algorithms;

[0017] Generate a feature vector after normalizing or standardizing the features.

[0018] Preferably, the model training steps further include:

[0019] Select a deep learning model or an ensemble learning model as the prediction framework;

[0020] Adopt a cross-validation method to adjust the model hyperparameters;

[0021] Optimize the model weights based on the gradient descent algorithm or the Bayesian optimization method.

[0022] Preferably, the model verification steps further include:

[0023] Evaluate the model performance using accuracy, recall rate and F1 value;

[0024] Analyze the deviation source of the prediction results through a confusion matrix;

[0025] If the performance does not meet the standard, iterate and optimize the feature engineering or the model structure.

[0026] Preferably, the prediction result output steps further include:

[0027] Map the prediction results to user behavior categories or probability distributions;

[0028] Set a threshold in combination with the business scenario and generate an interpretable report;

[0029] Display the user behavior trends and potential risks through a visualization tool.

[0030] A system for an enhanced method of big model data analysis based on dynamic noise filtering and multi-modal feature fusion, comprising:

[0031] The multi-modal data preprocessing module is used to perform format parsing, semantic normalization, and domain-adaptive word segmentation on text, image, and table data;

[0032] The cross-modal association enhancement module is used to achieve cross-modal feature fusion through a multi-modal encoder, a similarity graph convolutional network, and a dynamic routing mechanism;

[0033] The interpretability enhancement module is used to provide decision-making basis based on the attribution analysis of the knowledge graph and the counterfactual explanation generator;

[0034] The dynamic noise filtering module is used to suppress data noise through a noise-sensitive Transformer and an adaptive loss function;

[0035] The multi-modal feature enhancement module is used to optimize the feature representation through cross-modal semantic anchor matching and contrastive learning.

[0036] Preferably, the multi-modal data preprocessing module further includes:

[0037] The format parsing sub-module is used to: extract text from PDF files using Apache PDFBox and detect tables in combination with PyMuPDF; parse text from DOCX files by calling python-docx and recognize embedded pictures through OpenCV; extract text from PPT files using python-pptx and perform OCR recognition through Tesseract;

[0038] The domain-adaptive word segmentation sub-module is used to: construct a dynamic vocabulary based on dependency syntactic analysis or the UMLS terminology library; train domain-specific word vectors through the FastText model, and use the BiLSTM-CRF model to handle professional term ambiguities.

[0039] Preferably, the cross-modal association enhancement module further includes:

[0040] The multi-modal encoder sub-module is used to: perform 768-dimensional semantic encoding on text through the RoBERTa model and introduce a domain adapter to adjust the feature distribution; extract image region features through Faster R-CNN and fuse global and local features in combination with an attention mechanism;

[0041] The dynamic routing mechanism sub-module is used to: identify the task type based on the BERT classifier and dynamically adjust the fusion weight of text and image features; pre-define three routing strategies: text-first, image-first, and balanced mode, and select the optimal path according to the task type.

[0042] Preferably, the dynamic noise filtering module further includes:

[0043] A noise detection sub-module, configured to: calculate noise confidence through a noise-sensitive Transformer, and locate text spelling mistakes, grammar errors, and image blurry regions; analyze the attention weight differences between the noise regions and the normal regions based on the multi-head attention matrix;

[0044] An adaptive loss optimization sub-module, configured to: dynamically adjust the loss weight according to the noise confidence, and add a penalty term to the samples in the noise regions; balance noise suppression and the model convergence speed through a joint loss function.

[0045] Preferably, the multi-modal feature enhancement module further includes:

[0046] A cross-modal semantic anchor matching sub-module, configured to: perform OCR recognition through PyTesseract, and calculate the semantic similarity between the text and the image regions based on the BERT model; combine the spatial distance and the semantic distance to achieve the joint matching of the text keywords and the image ROIs;

[0047] A contrastive learning optimization sub-module, configured to: construct a triple contrastive loss function to minimize the distance between the cross-modal positive sample pairs and maximize the distance between the negative sample pairs; verify the cross-modal feature clustering effect by visualizing the feature space through t-SNE dimensionality reduction.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] The large model data analysis enhancement method and system based on dynamic noise filtering and multi-modal feature fusion proposed by the present invention perform quality assessment on the reference files uploaded by users through the dynamic noise perception module, and adopt an adaptive threshold cleaning mechanism to eliminate redundant and conflicting information; design a multi-granularity feature alignment network, and use document structure parsing and semantic embedding technologies to achieve a deep association mapping between the reference files and the input tasks. The system includes a reference file dynamic parsing module, a context enhancement representation module, and a multi-source knowledge fusion decision-making module, and guides the large model to accurately capture the effective knowledge in the reference files by constructing an interpretable attention weight allocation strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is the mind map of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] In order to describe the objectives, technical solutions of the present invention clearly, completely, and make the advantages clearer, the following further details the embodiments of the present invention with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are part of the embodiments of the present invention, rather than all of the embodiments, and are only used to explain the embodiments of the present invention, not to limit the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

[0052] Please refer to Figure 1 , the present invention provides a technical solution: a large model data analysis enhancement method and system based on dynamic noise filtering and multi-modal feature fusion, including:

[0053] I. Reference document preprocessing module

[0054] 1. Multi-modal data normalization

[0055] Operation process:

[0056] Format parsing:

[0057] PDF file: Use Apache PDFBox to extract text, combine with PyMuPDF for table detection, and identify mathematical formulas through regular expressions (such as E = mc 2 );

[0058] DOCX file: Call the python-docx library to parse the text, identify embedded pictures through OpenCV and generate OCR text;

[0059] PPT file: Use python-pptx to extract the text of each page, simulate screenshots through PyAutoGUI and call Tesseract for picture text recognition.

[0060] 2. Hybrid parsing engine

[0061] Regular expression filtering: Remove redundant content such as headers, footers, and repeated titles;

[0062] Deep learning text completion: For texts misrecognized by OCR (such as "EBITDA" misrecognized as "EBITDA"), use the pre-trained BERT model for error correction.

[0063] 3. Unified format output

[0064] Generate a JSON format with semantic tags:

[0065]

[0066]

[0067] Domain adaptive tokenizer

[0068] Operation process:

[0069] Dynamic vocabulary construction

[0070] Domain term extraction:

[0071] Financial domain: Identify terms such as "asset - liability ratio" and "price - earnings ratio" through dependency syntactic analysis; Medical domain: Match professional vocabulary such as "pathological section" and "CT scan" based on UMLS (Unified Medical Language System).

[0072] Word vector training:

[0073] Use more than 100,000 documents in the field to train the FastText model to generate domain - specific word vectors (e.g., the cosine similarity between the vector of "EBITDA" and "profit" reaches 0.82).

[0074] Word segmentation optimization

[0075] Ambiguity resolution:

[0076] Adopt the BiLSTM - CRF model to handle the ambiguity of professional terms (e.g., "apple" refers to a company in financial news and a fruit in medical documents);

[0077] Entity linking:

[0078] Link the word - segmented results to Wikipedia entities (e.g., "apple" is linked to "Apple Inc." or "Malus domestica").

[0079] II. Cross - modal association enhancement module

[0080] 1. Low - level alignment: Multimodal encoder

[0081] Operation process:

[0082] Text encoding

[0083] Use the RoBERTa model to encode the text content to generate 768 - dimensional semantic vectors;

[0084] Introduce the Domain Adapter:

[0085]

[0086] For image encoding, use Faster R - CNN to extract the region of interest (ROI) features of the image, and each ROI generates a 2048 - dimensional vector;

[0087] Fuse the global image features and local ROI features through the attention mechanism: \img_vec = ∑iαi·roi_veci+β·global_vec

[0088] 2. Middle - level fusion: Cross - modal similarity graph

[0089] Operation process:

[0090] Similarity matrix calculation

[0091] Cosine similarity between text and image features: \Stext-img

[0092] = cos_sim(text_vec, img_vec)

[0093] Construct a bipartite graph: Nodes are divided into text blocks (Text Node) and image regions (Image Node), and the edge weights are similarity values.

[0094] Graph Convolutional Network Training

[0095] Use GCN for 3-layer feature propagation: \H(l+1) = σ

[0096] (D^-1 / 2A^D^-1 / 2H(l)W(l))

[0097] where A^ = A + I is the adjacency matrix with self-loops, and D^ is the degree matrix.

[0098] 3. High-level decision-making: Dynamic routing mechanism

[0099] Operation process:

[0100] Task type recognition

[0101] Identify the current task type (such as "financial risk control", "medical diagnosis") through the BERT classifier;

[0102] Concatenate the task type vector and the fusion feature:

[0103] \final_vec = concat(fusion_vec, task_vec)

[0104] Routing path selection

[0105] Pre-define three routing strategies:

[0106] Text first: Suitable for contract clause analysis (weight: text 60% + image 40%);

[0107] Image first: Suitable for trend chart interpretation (weight: text 30% + image 70%);

[0108] Balanced mode: Default scenario (weight: text 50% + image 50%);

[0109] Dynamically adjust the weights according to the task type: \αtask = softmax(W·task_vec)

[0110] III. Interpretability Enhancement Module

[0111] 1. Knowledge graph-driven attribution analysis

[0112] Operation Process:

[0113] Entity Relationship Extraction

[0114] Extract key information using SPO (Subject-Predicate-Object) triples:

[0115] Text: "The company's EBITDA increased by 20% in 2024" → (Company, has, EBITDA increased by 20%);

[0116] Image: "Revenue growth curve" in the trend chart → (Company, shows, revenue growth trend).

[0117] Inference Path Generation

[0118] Path search based on the domain knowledge graph: \The starting node of the knowledge graph is the contract clause, and the target node is the approval decision

[0119] Visualization tool: Use Cytoscape.js to generate an interactive graph, supporting path highlighting and weight display.

[0120] 2. Counterfactual Explanation Generator

[0121] Operation Process:

[0122] Feature Perturbation Strategy

[0123] Text perturbation: Randomly replace 10% of the vocabulary with synonyms (e.g., "growth" → "rise");

[0124] Image perturbation: Add Gaussian blur (σ = 2) to the ROI region.

[0125] Contribution Quantification

[0126] Calculate feature importance using SHAP values:

[0127] Generate counterfactual samples:

[0128]

[0129]

[0130] IV. Dynamic Noise Filtering Algorithm

[0131] 1. Confidence-Guided Noise Detection Mechanism

[0132] Implementation Steps:

[0133] Training of noise-sensitive Transformer

[0134] Data Annotation:

[0135] Text: Manually mark spelling mistakes (e.g., "EBITDA" → "EBITDA") and grammar mistakes (subject-verb disagreement).

[0136] Image: Use the COCO-Text dataset to mark blurred regions (e.g., low-contrast text).

[0137]

[0138]

[0139] Attention weight analysis

[0140] Calculate the multi-head attention matrix: \Attention(Q, K, V) = softmax(\frac{QK^T}{\sqrt{d_k}})V

[0141] Noise region localization: The attention weight of the noise-sensitive Transformer to the noise region is significantly higher than that of the normal region (experiments show that the average weight of the noise region is 0.72 vs. 0.35 for the normal region).

[0142] Adaptive loss function optimization

[0143] Mathematical derivation:

[0144] Noise confidence calculation

[0145] Output probability based on the noise-sensitive Transformer: \c_i = noise\_detector(x_i) (\c_i \in [0, 1])

[0146] Noise region weight: \alpha_i = \frac{1}{1 + exp(-\beta(\c_i - \theta))}

[0147] Among them, \beta = 5 is the temperature parameter, and \theta = 0.6 is the confidence threshold.

[0148] Dynamic weight adjustment

[0149] During the training process, increase the loss weight for samples in the noise region (\c_i > \theta) to suppress the impact of noise on the model: \L_{noisy} = \sum_{i} \alpha_i \cdot L(y_i, \hat{y}_i) \L_{clean} = \sum_{i} (1 - \alpha_i) \cdot L(y_i, \hat{y}_i) \L_{total} = \lambda \cdot L_{noisy} + (1 - \lambda) \cdot L_{clean} (\lambda = 0.8)

[0150] V. Multimodal feature enhancement technology

[0151] 1. Cross-modal semantic anchor matching

[0152] Implementation steps:

[0153] Fusion of OCR and text encoding

[0154] Image preprocessing:

[0155] Use PyTesseract for OCR recognition to generate text blocks with coordinates (e.g., "EBITDA" is located at (100, 200));

[0156] Text semantic encoding:

[0157] The BERT model encodes the OCR text and the original document text respectively, and matches the anchor points through cosine similarity (e.g., the similarity between the OCR text "EBITDA" and the document text "EBITDA" reaches 0.91).

[0158] Spatio-temporal alignment algorithm

[0159] Text keyword localization:

[0160] \keyword_pos = BERT_token_positions(text, keyword)

[0161] Image ROI association:

[0162] For each ROI in the image, calculate its spatial distance from the text keyword: \dspatial = (xroi - xkeyword)2 + (yroi - ykeyword)2

[0163] Joint matching of semantic distance and spatial distance: \match_score = dspatial

[0164] + 1cos_sim(text_vec, roi_vec)

[0165] 2. Feature fusion optimized by contrastive learning

[0166] Mathematical derivation:

[0167] Triplet contrast loss function

[0168] Positive sample pairs (cross-modal data with the same semantics): \d(a, p) = ∥text_vec - img_vec∥2

[0169] Negative sample pairs (cross-modal data with different semantics): \d(a, n) = ∥

[0170] text_vec - random_img_vec∥2

[0171] Loss function: \Ltriplet = ∑imax(0, d(ai, pi) - d(ai, ni) + γ)

[0172] Among them, γ = 0.5 is the interval parameter. Experiments show that when γ = 0.5, the cross-modal feature discrimination is optimal.

[0173] Feature space visualization

[0174] t-SNE dimensionality reduction shows that after triple contrast loss training, the text and image features with the same semantics are more closely clustered in space (the intra-class distance decreases from 0.8 to 0.4).

[0175] VI. Domain-independent generalization framework

[0176] 1. Adversarial training mechanism

[0177] Implementation steps:

[0178] Domain discriminator design

[0179] Input: Cross-modal fusion feature vector (2048 dimensions); architecture

[0180]

[0181] Adversarial training process generator objective: Minimize the domain classification accuracy;

[0182] Discriminator objective: Maximize the domain classification accuracy;

[0183] Joint loss function: \Ladv = λ·Ltask - Ldomain where λ = 0.3 is the balance parameter.

[0184] 2. Meta-learning initialization

[0185] Implementation steps:

[0186] MAML algorithm process

[0187] Fast adaptation stage:

[0188] Update the model parameters on the support set:

[0189] Meta-training stage: Calculate the gradient and update the initial parameter θ on the query set:

[0190] VII. Training phase operation process

[0191] 1. Multi-task joint training

[0192] Implementation steps:

[0193] Loss function design

[0194] Noise filtering loss ($L_{noise}$): $L_{noise}=\sum_{i}\alpha_{i}\cdot BCEWithLogitsLoss(c_{i},\hat{c}_{i})$

[0195] Where $c_{i}$ is the true value of noise confidence and $\hat{c}_{i}$ is the predicted value.

[0196] Feature fusion loss ($L_{fusion}$): $L_{fusion}=\sum_{i}L_{triplet}(x_{i}^{text},x_{i}^{img})$ Combining triplet contrast loss and mean squared error (MSE): $L_{fusion}=\lambda_{triplet}\cdot L_{triplet}+\lambda_{MSE}\cdot MSE(f(x_{i}^{text}),f(x_{i}^{img}))$

[0197] Interpretability loss ($L_{explain}$): $L_{explain}=\sum_{i}CrossEntropyLoss(p_{i}^{path},\hat{p}_{i}^{path})$

[0198] Where $p_{i}^{path}$ is the true distribution of the knowledge graph path and $\hat{p}_{i}^{path}$ is the predicted distribution.

[0199] Hyperparameter optimization

[0200] 1. Grid search range:

[0201]

[0202]

[0203] Optimization tool: Use Optuna for Bayesian optimization, with the goal of minimizing the combined loss on the validation set: $Objective=\alpha\cdot Accuracy+\beta\cdot ExplainabilityScore$

[0204] Training code example

[0205]

[0206]

[0207] 2. Domain adaptation training

[0208] Implementation steps:

[0209] Cross-domain dataset construction

[0210] Data source:

[0211] Finance: 500,000 credit contracts and financial reports (2010 - 2024);

[0212] Medical: 300,000 electronic medical records and imaging reports (including data from 500 hospitals); Industrial: 200,000 equipment logs and maintenance manuals (covering more than 100 equipment types). Data augmentation:

[0213] Text: EDA (Easy Data Augmentation) generates synonymous sentences;

[0214] Image: Random rotation (±15°), adding salt and pepper noise (density 0.05).

[0215] Implemented through adversarial training

[0216] Domain discriminator optimization:

[0217]

[0218] VIII. Operation process in the inference stage

[0219] 1. Real-time data processing

[0220] Technical implementation:

[0221] C++ high-performance parsing engine

[0222] Text parsing:

[0223] Use RapidJSON to parse JSON format documents, with a throughput of 100MB / s;

[0224] Regular expression library (PCRE2) filters redundant information, and the rule set contains more than 200 financial /

[0225] Medical keyword matching patterns.

[0226] Image preprocessing: Use OpenCV to implement image scaling (longest side ≤ 2000px) and channel conversion (RGB → BGR);

[0227] Tesseract OCR parallel processing (multi-thread support, number of threads = number of CPU cores ×

[0228] 2).

[0229] Batch processing optimization

[0230] Dynamic batch size:

[0231] \batch_size = min(100, available_memory / 2048)

[0232] CUDA stream parallel processing:

[0233] / / CUDA stream configuration

[0234] cudaStream_t streams[4]; for (int i = 0; i < 4; ++i) {

[0235] cudaStreamCreate(&streams[i]);}

[0236] / / Parallel inference for (int i = 0; i < num_batches; ++i) {

[0237] cudaMemcpyAsync(input_buffers[i], host_data[i], size,

[0238] cudaMemcpyHostToDevice, streams[i % 4]);

[0239] model_inference(input_buffers[i], output_buffers[i],

[0240] streams[i % 4]);

[0241] }

[0242] 2. Interactive analysis interface

[0243] Implementation details:

[0244] Heatmap generation and feature importance calculation:

[0245] def compute_heatmap(input_data, model):

[0246] shap_values = shap.DeepExplainer(model, input_data).shap_values(input_data)

[0247] heatmap = np.mean(np.abs(shap_values), axis = 0)

[0248] return heatmap

[0249] Visualization tool: Use Plotly Dash to generate an interactive heatmap, supporting zooming and hovering to display specific contribution values.

[0250] Drag-and-drop correlation graph

[0251] Front-end implementation:

[0252] Cytoscape.js is used to construct a graph, where the node size is positively correlated with the feature importance;

[0253] D3.js is used to implement drag-and-drop interaction, and the edge weights are automatically updated when the nodes are moved.

[0254] Backend interface:

[0255] @app.route(' / generate_graph',

[0256] methods=['POST']) def generate_graph():

[0257] data = request.json

[0258] graph = model.generate_knowledge_graph(data['text'], data['images'])

[0259] return jsonify(graph).

[0260] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for enhancing large model data analysis based on dynamic noise filtering and multimodal feature fusion, characterized in that: It includes the following steps: Obtain the user's historical behavior data, and perform data cleaning and feature extraction; Construct a user behavior feature vector, and divide the training set and the test set; Train a machine learning model based on the training set and optimize the model parameters; Use the test set to verify the model performance and output the prediction result.

2. An enhanced method for large model data analysis based on dynamic noise filtering and multi-modal feature fusion according to claim 1, characterized in that: The data cleaning and feature extraction step further includes: Remove duplicate data, fill in missing values, and handle outliers; Extract user behavior pattern features through time series analysis or clustering algorithms; Generate feature vectors after normalizing or standardizing the features.

3. An enhanced method for large model data analysis based on dynamic noise filtering and multi-modal feature fusion according to claim 2, characterized in that: The model training step further includes: Select a deep learning model or an ensemble learning model as the prediction framework; Adopt a cross-validation method to adjust the model hyperparameters; Optimize the model weights based on the gradient descent algorithm or the Bayesian optimization method.

4. An enhanced method for large model data analysis based on dynamic noise filtering and multi-modal feature fusion according to claim 3, characterized in that: The model verification step further includes: Evaluate the model performance using accuracy, recall rate, and F1 value; Analyze the deviation source of the prediction result through a confusion matrix; If the performance does not meet the standard, iterate to optimize the feature engineering or the model structure.

5. A method for enhancing large model data analysis based on dynamic noise filtering and multi-modal feature fusion according to claim 4, characterized in that: The prediction result output step further includes: Map the prediction result to user behavior categories or probability distributions; Set a threshold in combination with the business scenario and generate an interpretable report; Display the user behavior trend and potential risks through a visualization tool.

6. A system for the method of enhancing large model data analysis based on dynamic noise filtering and multimodal feature fusion according to claim 5, characterized in that: It includes: A multi-modal data preprocessing module, which is used to perform format parsing, semantic normalization, and domain-adaptive word segmentation on text, image, and table data; A cross-modal association enhancement module, which is used to achieve cross-modal feature fusion through a multi-modal encoder, a similarity graph convolutional network, and a dynamic routing mechanism; An interpretability enhancement module, which is used to provide decision-making basis based on the attribution analysis of the knowledge graph and the counterfactual explanation generator; A dynamic noise filtering module, which is used to suppress data noise through a noise-sensitive Transformer and an adaptive loss function; A multi-modal feature enhancement module, which is used to optimize the feature representation through cross-modal semantic anchor point matching and contrast learning.

7. A system according to claim 6, wherein: The multi-modal data preprocessing module further includes: A format parsing sub-module, which is used to: extract text from PDF files using Apache PDFBox and detect tables in combination with PyMuPDF; parse text from DOCX files by calling python-docx and recognize embedded pictures through OpenCV; extract text from PPT files using python-pptx and perform OCR recognition through Tesseract; A domain-adaptive word segmentation sub-module, which is used to: construct a dynamic vocabulary based on dependency syntax analysis or the UMLS term library; train domain-specific word vectors through the FastText model, and use the BiLSTM-CRF model to handle professional term ambiguities.

8. A system according to claim 7, characterized in that: The cross-modal association enhancement module further includes: A multi-modal encoder sub-module, which is used to: perform 768-dimensional semantic encoding on text through the RoBERTa model and introduce a domain adapter to adjust the feature distribution; extract image region features through Faster R-CNN and fuse global and local features in combination with an attention mechanism; The dynamic routing mechanism sub-module is used for: identifying task types based on the BERT classifier and dynamically adjusting the fusion weights of text and image features; pre-defining three routing strategies of text priority, image priority, and balanced mode, and selecting the optimal path according to the task type.

9. A system according to claim 8, characterized in that: The dynamic noise filtering module further includes: The noise detection sub-module is used for: calculating the noise confidence through a noise-sensitive Transformer, and locating text spelling mistakes, grammar errors, and blurred image areas; analyzing the attention weight differences between the noise area and the normal area based on the multi-head attention matrix; The adaptive loss optimization sub-module is used for: dynamically adjusting the loss weight according to the noise confidence, adding a penalty term to the samples in the noise area; balancing noise suppression and model convergence speed through the joint loss function.

10. A system according to claim 9, wherein: The multi-modal feature enhancement module further includes: The cross-modal semantic anchor matching sub-module is used for: performing OCR recognition through PyTesseract, and calculating the semantic similarity between the text and the image area based on the BERT model; realizing the joint matching of text keywords and image ROIs by combining the spatial distance and the semantic distance; The contrastive learning optimization sub-module is used for: constructing a triple contrastive loss function, minimizing the distance between cross-modal positive sample pairs and maximizing the distance between negative sample pairs; visualizing the feature space through t-SNE dimensionality reduction to verify the cross-modal feature clustering effect.

Images