Patrol information comprehensive management device based on AI intelligent analysis

Abstract

The invention belongs to the technical field of information comprehensive management, and particularly relates to a patrol information comprehensive management device based on AI intelligent analysis, which comprises an information acquisition module used for acquiring text type, image type, audio type and ledger type heterogeneous data in the whole patrol process; the preprocessing module is in communication connection with the information acquisition module and is used for carrying out localized lightweight preprocessing on the standardized intermediate data; the intelligent analysis module comprises a multi-modal fusion analysis unit and a patrol model unit; wherein the input of the multi-modal fusion analysis unit is pre-processed multi-modal data, an off-line running improved Transform model is adopted, various types of data are firstly converted into feature vectors with unified dimensions, then feature weights of all modals are calculated through a modal attention mechanism, and a unified patrol information feature vector is generated through weighted fusion; the output is a fusion feature vector; and the patrol model unit is used for inputting the fusion feature vector and a local patrol knowledge base slice, calculating the cosine similarity of the fusion feature vector and the knowledge base slice through an RAG enhanced retrieval mechanism, matching associated knowledge, and outputting a policy analysis result and a case matching result.

Description

Technical Field

[0001] This invention belongs to the field of information integrated management technology, specifically relating to an inspection information integrated management device based on AI intelligent analysis. Background Technology

[0002] Inspections include examining the inspected unit's financial data, project files, interview records, and on-site video footage. As inspection work deepens, traditional methods of information management are no longer sufficient to meet the demands of the work.

[0003] In existing technologies, inspection information management mostly adopts traditional document management systems or simple information platforms, which have the following shortcomings: First, the information collection method is singular, making it difficult to achieve comprehensive collection and unified management of diverse and heterogeneous data. Different types of data are stored in a scattered manner, resulting in serious information silos. Second, data processing mainly relies on manual operation, which is inefficient and prone to human error, making it impossible to quickly extract valuable inspection clues from massive amounts of data. Third, it lacks intelligent analysis capabilities, making it impossible to conduct in-depth analysis and trend prediction of inspection data, and difficult to predict potential corruption risks in advance. Fourth, data security is insufficient. Inspection information often involves sensitive content, and the encryption measures and access control mechanisms of existing devices are inadequate, posing a risk of information leakage. Fifth, the communication stability between modules is poor, especially in remote areas or environments with poor network signals, where data transmission is easily interrupted, affecting the normal conduct of inspection work.

[0004] Currently, most technical solutions related to inspection information management focus on managing single types of data or simple information statistics, failing to achieve deep fusion and analysis of diverse and heterogeneous data. While some solutions mention intelligent analysis, the algorithms used are relatively simple, and there are deficiencies in the design of edge data preprocessing, data security encryption, and communication stability, making it impossible to meet the high-security and intelligent management requirements of the entire inspection process. Therefore, there is an urgent need for a solution to these problems. To this end, an improved inspection information integrated management device based on AI intelligent analysis has been designed. Summary of the Invention

[0005] In view of the above-mentioned shortcomings in the existing technology, the present invention provides an AI-based intelligent analysis-based comprehensive management device for patrol information to solve the problems in the background technology.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] An AI-based intelligent analysis-based inspection information management device includes an information collection module, which is used to collect heterogeneous data such as text, images, audio, and ledgers throughout the inspection process, identify the heterogeneous data in different formats, convert them into intermediate data in a unified standard format, and complete the initial standardization process.

[0008] The preprocessing module, which is communicatively connected to the information acquisition module, is used to perform localized and lightweight preprocessing on the standardized intermediate data, and outputs a set of structured fields. The preprocessing module has a built-in data quality assessment unit and a data desensitization unit. The data quality assessment unit takes the preprocessed data as input and outputs a quality score, which is based on data integrity and accuracy. The data desensitization unit takes structured data containing sensitive information as input, processes it through a character replacement / masking algorithm, and outputs desensitized structured data.

[0009] The intelligent analysis module, which is communicatively connected to the preprocessing module, includes a multimodal fusion analysis unit and an inspection model unit. The input of the multimodal fusion analysis unit is the preprocessed multimodal data. It adopts an improved Transformer model that runs offline, first converting various types of data into feature vectors of a unified dimension, then calculating the feature weights of each modality through a modal attention mechanism, and weighted fusion to generate a unified inspection information feature vector. The output is a fused feature vector.

[0010] Inspection Model Unit: The input consists of a fused feature vector and a slice of the local inspection knowledge base. Through the RAG-enhanced retrieval mechanism, the cosine similarity between the fused feature vector and the knowledge base slice is calculated to match related knowledge. The output consists of policy analysis results and case matching results.

[0011] Furthermore, the intelligent analysis module also includes a natural language interaction unit. The natural language interaction unit takes user business scenario natural language instructions as input, converts them into structured instructions that the model can execute, and outputs standardized analysis instructions.

[0012] Furthermore, the intelligent analysis module also includes an inspection task intelligent agent unit. The input of the inspection task intelligent agent unit is standardized analysis instructions and corresponding structured data, which encapsulates special task process algorithms and outputs task processing results.

[0013] Furthermore, the graph neural network association analysis unit uses a graph convolutional neural network (GCN) to construct the inspection association graph. The input is structured entity data and association relationship data. The entity attributes are converted into node feature vectors, and the association weights are incorporated into the adjacency matrix as graph edge weights. The features of neighboring nodes are aggregated through two layers of graph convolutional layers, the node vector representation is updated, and the output is a weighted inspection association graph.

[0014] Furthermore, the clustering grouping unit adopts an improved K-means clustering algorithm. The input is the feature vector of inspection issues / petitions / historical cases. The optimal number of clusters is automatically determined by the elbow rule. Weighted Euclidean distance is introduced to calculate sample similarity. The cluster centers are iteratively updated until convergence. The output is the cluster label and the set of similar data.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] Through multiple acquisition sub-units and format recognition units in the information acquisition module, it is possible to comprehensively collect various types of data such as text, images, audio, and ledgers throughout the entire inspection process, and perform preliminary standardization processing, thus solving the problems of single information acquisition and information silos in existing technologies.

[0017] The edge intelligent preprocessing module performs lightweight preprocessing on data at the edge, including redundant information removal, image enhancement, audio noise reduction, and structured parsing. At the same time, it performs data quality assessment and desensitization processing, which not only reduces the bandwidth consumption and latency of data transmission, but also improves data quality and ensures information security.

[0018] The core AI intelligent analysis module integrates an improved Transformer model, graph neural network, and improved K-means clustering algorithm. It not only achieves multimodal data fusion analysis but also constructs an inspection correlation graph to uncover hidden connections, enabling the clustering and aggregation of similar issues. Combined with a dynamically updated inspection key point knowledge base and historical inspection case library, it achieves automatic identification, classification, grading, and potential risk prediction of inspection issues. At the same time, it can automatically evaluate the rectification effect, greatly reducing manual workload and improving the accuracy and efficiency of inspection work, especially enhancing the ability to identify systemic and related issues.

[0019] 4. The distributed storage and encryption module uses the national cryptographic SM4 algorithm for encrypted storage and role-based access control to ensure data storage security; the edge communication module supports 5G and Wi-Fi dual-mode communication, and has a built-in communication quality monitoring and automatic reconnection mechanism to ensure the stability and integrity of data transmission, making it suitable for different inspection environments. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described below with reference to embodiments.

[0021] Specifically, this AI-based intelligent analysis-based comprehensive management device for inspection information includes an information collection module: It collects diverse and heterogeneous data from the entire inspection process, clarifies the core attributes of various data types and standardization rules after collection, and the key algorithms are specifically applied as follows: 1) Text data: Originally in paper / electronic document form, it is converted into a standardized character stream with UTF-8 encoding after collection, accurately recording relevant text information about the inspection; 2) Image data: Originally in optical image form, it is converted into a 1920×1080 resolution two-dimensional pixel matrix after collection, with pixel grayscale values ​​ranging from 0 to 255; 3) Audio data: Originally in sound wave signal form, it is converted into a time-series amplitude sequence with a sampling rate of 16kHz after collection, with an amplitude range of [-1,1]; 4) Ledger data: Originally in heterogeneous data tables from various systems, it is converted into a standardized data table with unified fields after collection (e.g., the financial field "Amount" is "Yuan" and "Approval Time" is "Day"), achieving unified structured ledger data.

[0022] This module completes data acquisition through dedicated acquisition subunits (text: scanner / electronic import interface; images: high-definition camera; audio: high-fidelity recording equipment; ledgers: dedicated system interface). The core algorithm is a file header feature matching algorithm. Specifically, for acquired heterogeneous files (such as .pdf / jpg / mp3 / xlsx), it extracts fixed features from the file headers (e.g., .pdf file header is "%PDF-", jpg is "FFD8FF"), matches them to a preset format feature library, accurately identifies the original data format, and then performs corresponding standardized conversions (text to UTF-8 character stream, image to 1920×1080 pixel matrix, etc.), providing a unified input for subsequent preprocessing.

[0023] 2. Edge Intelligence Preprocessing Module: Communicates with the information collection module, performs lightweight processing on the standardized intermediate data to reduce the subsequent analysis load while ensuring data quality. The specific application logics of the algorithms corresponding to each data type are as follows: 1) Text data: The input is a standardized character stream. The algorithms are regular matching algorithm + TF-IDF algorithm. Specific application: First, use regular expressions to match consecutive repeated sentences / paragraphs and remove redundant character streams; then, for the text after redundancy removal, count the word frequencies of words in units of "paragraphs", calculate the inverse document frequency in combination with the inspection text corpus, and extract the top N high-weight words (N adapts to the text length, for example, take the top 20 within 500 words), form a set of "keyword - weight" key-value pairs, and retain core inspection-related information such as "special funds" and "unapproved"; 2) Image data: The input is a pixel matrix with a resolution of 1920×1080. The core algorithms are 3×3 convolution kernel Gaussian filtering algorithm + histogram equalization algorithm. Specific application: Gaussian filtering performs convolution operations on the pixel matrix through a 3×3 convolution kernel, weighted-averages the grayscale values of neighboring pixels, and corrects the blurred areas caused by jitter during on-site shooting; then, adjust the pixel grayscale value distribution through histogram equalization, stretch the contrast of low grayscale areas, make the text / details in the scanned archives and on-site images clearer, and output an enhanced standardized pixel matrix; 3) Audio data: The input is an amplitude sequence with a sampling rate of 16kHz. The core algorithms are adaptive Wiener filtering algorithm + MFCC + CTC algorithm. Specific application: Adaptive Wiener filtering first analyzes the amplitude statistical characteristics of the first 100 sampling points, constructs a background noise model (such as ambient noise), removes the noise components from the original amplitude sequence, and outputs a noise-reduced amplitude sequence; then, divide the amplitude sequence into a frame sequence of 25ms / frame through the MFCC algorithm, extract 13-dimensional Mel-frequency cepstral coefficients, and finally align the coefficient matrix with the text labels through the CTC algorithm to convert conversation recordings and meeting recordings into standardized character streams; 4) Ledger data: The input is a standardized data table. The core algorithms are regular extraction + field mapping algorithm. Specific application: Based on a preset inspection core field mapping table (such as mapping "income / expenditure / amount" for finance-related, and "approval time / approval personnel" for project-related), extract field values through regular expressions (such as "amount: ([0-9,]+) yuan") to form a set of "field name - field value" key-value pairs; The core algorithm of the data quality assessment unit is the missing rate / error rate statistical algorithm. Specific application: Calculate the number of missing fields / total number of fields (missing rate), number of incorrect fields / total number of fields (error rate), and obtain a quality score of 0-100 points through weighting. A score of ≥90 is considered qualified; The core algorithm of the data de-sensitization unit is the character replacement / masking algorithm. Specific application: Replace the middle 8 digits of the ID number (match "\\d{17}[0-9Xx]") with "*", and replace the name of the confidential unit with "[sensitive unit]", and output de-sensitized structured data;

[0024] 3. AI Intelligent Analysis Core Module: As the core module of the device, it communicates and connects with the edge intelligent preprocessing module to perform deep intelligent analysis on the preprocessed structured data. It focuses on mining data correlations and similar characteristics through graph neural networks and clustering algorithms. The detailed data processing logic of each unit is as follows:

[0025] Multimodal fusion analysis unit: The core algorithm is an improved Transformer model. The specific application logic is as follows: Input preprocessed multimodal structured data (text keyword set, image enhanced pixel matrix, audio-to-text conversion, ledger field set), processing flow: ① Single-modal feature extraction: For text / audio text, input the "keyword-weight" key-value pairs into the pre-trained Word2Vec word embedding layer, converting them into 512-dimensional word vectors, focusing on strengthening the vector representation of sensitive words such as "violation" and "misappropriation"; For the 1920×1080 image pixel matrix, extract visual features (such as the texture of violation documents and the contour of abnormal areas on site) through 4 layers of 3×3 convolutional layers, converting them into 512-dimensional word vectors. ① Image feature vector; For the set of ledger fields, the key-value pairs of "field name - field value" are converted into 512-dimensional ledger feature vectors through the field embedding layer, highlighting the feature weights of quantitative fields such as "amount" and "approval time"; ② Modal fusion: A modal attention mechanism is introduced to calculate the correlation weights of each modal feature with the inspection task, ranging from [0,1]. For example, when judging financial issues, the modal weight of ledger / text is increased, and when judging on-site issues, the modal weight of image is increased. The three types of 512-dimensional feature vectors are fused into a 512-dimensional unified inspection information feature vector through weighted summation; The output is a fused feature vector, which physically means to realize the unified representation of multi-source heterogeneous data, eliminate modal differences, and provide standardized input for subsequent problem judgment;

[0026] The intelligent assessment unit for inspection issues uses a CNN-RNN hybrid model (CNN + bidirectional LSTM) as its core algorithm. Its specific application enables automatic identification, classification, and severity assessment of inspection issues. The process is as follows: 1) Data input and feature adaptation: The input is a 512-dimensional fusion feature vector (integrating core features from text, images, audio, and ledgers) output by the multimodal fusion analysis unit. The auxiliary input is a "key point-weight" key-value pair from the inspection key point knowledge base (e.g., "misappropriation of special funds - 0.9"), which is converted into a 128-dimensional key point feature vector through a word embedding layer; 2) Model structure and data processing: ① Feature enhancement layer: The fusion feature vector input is fed into a fully connected layer with a ReLU activation function, compressed into a 256-dimensional enhanced feature. ① Vector, focusing on core inspection information; ② CNN branch (spatial feature processing): For spatial features originating from the 1920×1080 image in the fused features, a 3-layer convolutional layer (3×3 / 3×3 / 5×5 kernels, 64 / 128 / 256 output channels) + 2-layer 2×2 max pooling layers are used to extract spatial features such as text positions on violation documents and abnormal areas at the construction site in the image, outputting a 256-dimensional image spatial feature vector; ③ RNN branch (sequence feature processing): For sequence features originating from text / audio text in the fused features, keywords are embedded into a vector sequence (512 dimensions) and input into a 2-layer bidirectional LSTM (256 hidden units per layer) to capture "unpublished - unauthorized changes" and "no approval" information. - The temporal correlation of illegal expressions such as "illegal expenditure" is used to output a 256-dimensional text sequence feature vector; 3) Feature fusion and task calculation: ① Multi-feature fusion: Concatenate CNN spatial features (256-dimensional), RNN sequence features (256-dimensional), and key point features (128-dimensional) to obtain a 640-dimensional comprehensive feature vector, which is then normalized by a BatchNorm layer to eliminate dimensional differences; ② Classification task (problem identification): The comprehensive feature vector is input into a 2-layer ReLU fully connected layer (512 / 128 hidden units), and the softmax output is used to determine the probability distribution of 8 types of inspection problems (such as "illegal fund management" and "illegal project approval"). The one with the highest probability is the identification result. The input data is a fusion of ledger amount and text violation table. The feature vectors described above are directly matched to the corresponding problem categories; ③ Severity rating task: An ordered regression layer is introduced, which takes into account the comprehensive feature vector and the ledger quantitative data (such as the amount of violation in "yuan" and the duration in "days"), and outputs a 3-level label (1=general, 2=severe, 3=serious) through sigmoid. Specific application: If the comprehensive feature matches the key point of "misappropriation of special funds", and the "amount" in the ledger field is ≥100,000 yuan, the model outputs label 3 (serious); 4) Model optimization: The "512-dimensional fused feature vector-problem category-severity level" sample pairs (such as 10,000+ records) from the historical inspection case library are used for joint training through cross-entropy loss (classification) + mean squared error loss (grading), with a learning rate of 0.001. After 50 iterations, an attention mechanism is introduced during training to automatically focus on strongly correlated features such as "amount" and "unapproved," improving the accuracy of the judgment. The accompanying RAG enhanced retrieval algorithm is specifically applied by calculating the cosine similarity between the 512-dimensional fused feature vector and the slice vector of the local inspection knowledge base, matching the top 5 similar policy provisions / historical cases to assist in verifying the judgment results.

[0027] Trend Prediction Unit: The core algorithm is an LSTM time-series prediction model, specifically applied as follows: The input data consists of two parts: first, historical inspection time-series data (structured data of "inspection problem type - quantity - amount involved" for each quarter of the past 5 years, with dimensions: number of problems / quarter, amount / quarter); second, current inspection progress data (statistical data of problems that have been inspected this year). Model processing logic: The historical data is divided into time series by "quarter". Using "data from the previous 3 quarters" as the input window and "distribution of problem types / amount scale in the 4th quarter" as the prediction target, a 3-layer LSTM model (256 / 128 / 64 hidden units) is constructed. Through time-series modeling, the annual / quarterly variation patterns of inspection problems are captured (e.g., violations of special funds are mostly concentrated at the end of the year). The output is the prediction results of potential risk points for each quarter in the next year (e.g., "the risk of project approval violations will increase by 30% in the next quarter"), generating a risk trend curve. Quantitative indicators such as "amount involved" and "number of problems" in the input data directly affect the prediction weights, providing data support for adjusting the focus of inspections.

[0028] Graph Neural Network Association Analysis Unit and Clustering Grouping Unit: 1) Graph Neural Network Association Analysis Unit: The core algorithm is the GCN (Graph Convolutional Neural Network) model. Its specific application in conjunction with the entity data in this application is as follows: The input consists of preprocessed entity data and association data. Entity data is a key-value pair of "entity ID - entity attribute" (e.g., inspection object: "ID001 - Unit A / Education Field"; inspection issue: "ID101 - Misappropriation of Special Funds / Amount 500,000 RMB / March 2024"). Association data is a triple of "entity pair - association type - weight" (e.g., "ID001 - ID101 - Existence - 0.8", where weight is calculated from entity co-occurrence frequency). Among these, inspection issues / petitions... The entity attributes corresponding to each case / historical case are ultimately converted into a 256-dimensional feature vector, specifically in the following form: Inspection Problem Feature Vector (256 dimensions): [Text Keyword Features (128 dimensions) + Quantitative Indicator Features (32 dimensions) + Time Features (16 dimensions) + Problem Type Encoding Features (64 dimensions) + Auxiliary Features (16 dimensions)]; Explanation of each sub-dimension: ① Text Keyword Features (128 dimensions): derived from the TF-IDF high-weight keywords (first 128) of the preprocessed text, each keyword corresponding to a 1-dimensional word embedding vector (generated by the Word2Vec model); ② Quantitative Indicator Features (32 dimensions): includes 32 items such as the amount involved (standardized), the duration of the violation (standardized), and the number of people involved. Core quantitative fields; ③ Time features (16 dimensions): a standardized vector after the discovery time is converted to a timestamp; ④ Problem type coding features (64 dimensions): one-hot coding for 8 types of problems such as "illegal fund management" and "illegal project approval" (8 types correspond to 8 dimensions, extended to 64 dimensions to adapt to the model input); ⑤ Auxiliary features (16 dimensions): including inspection stage coding, source channel coding, etc.; Petition feature vector (256 dimensions): [text keyword features (128 dimensions) + demand quantitative features (32 dimensions) + time features (16 dimensions) + complaint type coding features (64 dimensions) + auxiliary features (16 dimensions)]; Explanation of each sub-dimension: ① Text keyword features (128 dimensions): TF-ID of the petition description text F. High-weight keyword embedding vector; ② Quantitative features of demands (32 dimensions): including number of complaints, number of people involved, amount of demand (if any); ③ Time features (16 dimensions): standardized vectors of complaint time and first response time; ④ Complaint type coding features (64 dimensions): unique hot coding of 16 types of complaints such as "noise pollution" and "rights disputes" (extended to 64 dimensions); ⑤ Auxiliary features (16 dimensions): including source of petition (telephone / email / visit), code of the area involved, etc.; Historical case feature vector (256 dimensions): [Text keyword features (128 dimensions) + Case quantitative features (32 dimensions) + Time features (16 dimensions) + Case type coding features (64 dimensions) + Rectification effect features (16 dimensions)];Sub-dimension Descriptions: ① Text Keyword Features (128 dimensions): High-weight keyword embedding vectors for case descriptions and rectification measures; ② Case Quantitative Features (32 dimensions): Includes amount involved in the case, amount invested in rectification, duration of violations, etc.; ③ Time Features (16 dimensions): Standardized vectors for case occurrence time and rectification completion time; ④ Case Type Encoding Features (64 dimensions): Historical cases are encoded into 8 categories such as "funding / project / work style" (extended to 64 dimensions); ⑤ Rectification Effect Features (16 dimensions): Standardized values ​​such as rectification completion rate and review score; Model Specific Processing: ① Node Construction: Integrate the attributes of each entity (text attributes converted to word embedding, standardized quantitative attributes) into the above 256-dimensional node feature vector (), with each entity corresponding to 1 node; ② Edge Construction: Use association relationships as graph edges and weights as edge weights to construct an adjacency matrix (elements are edge weights); ③ Graph Convolution Calculation: Use 2 layers of graph convolution (input 256 dimensions → hidden 128 dimensions → output 64 dimensions, ReL) U activation), using adjacency matrix to aggregate neighborhood node features (e.g., the "Unit A" node aggregates its associated "Special Fund Misappropriation" and "Supervisor B" node features), and updates the node vector; ④ Association mining: calculate the cosine similarity of node vectors and mine hidden associations (e.g., "Unit A's Special Fund Misappropriation" and "Unit C's Irregular Approval" share the hidden association of "Supervisor B"); the output is an inspection association graph containing 64-dimensional node features and edge weights. When new inspection data is added, the node / edge information is automatically recalculated to adapt to the data update of the entire inspection process; 2) Clustering grouping unit: the core algorithm is the improved K-means clustering algorithm, which is specifically applied in combination with inspection issues / petitions / historical case data as follows: the input is the above 256-dimensional feature vector (generated by conversion of preprocessed text keywords, ledger fields, time information, etc., and the specific dimension composition is as described above); algorithm implementation steps: ① Data standardization: standardize the 256-dimensional features dimension by dimension according to the Z-score formula, formula: ; in The j-th dimension of the i-th sample contains the original features (such as "number of complaints", "amount involved", "keyword embedding value", etc.). Let j be the mean of the j-th dimension. ① Standard deviation is used to eliminate the difference in the numerical range of "amount (ten thousand yuan)" and "number of times" and "word embedding value"; ② Determination of the optimal K value: Candidate K = 2~10 (matching the number of common inspection problems), perform preliminary clustering calculation for each K to calculate SSE (sum of squared errors), draw the "SSE-K" curve to find the elbow point, and then calculate the average profile coefficient (≥0.7 is qualified) to determine the optimal K (e.g., K=5, corresponding to 5 types of core inspection problems); ③ Weighted clustering iteration: a. Initial center selection: Select K initial centers from the standardized feature matrix using the "farthest priority method" (first randomly select 1 petition sample, then select the sample farthest from it, and so on); b. Weighted distance calculation: Use weighted Euclidean distance, assign a weight of 1.2 to the core feature dimensions such as "problem type code", "amount involved", and "complaint type code", and assign a weight of 0.8 to the secondary features such as "time feature" and "source channel", formula: ,in c. Sample assignment: Assign each sample to the cluster with the smallest noise pollution distance (e.g., "noise pollution" type petitions are assigned to cluster 0); d. Center update: Calculate the feature mean of each sample as the new center; e. Convergence judgment: Stop when the maximum change in the center between adjacent iterations Δ≤0.001; ④ Label generation: Assign integer labels of 0~K-1 to the K clusters, count the core noise pollution of each type (e.g., core feature of cluster 0 - noise pollution), generate a "label-semantic" mapping table, output the label of each sample and the set of similar types (e.g., cluster 0 contains all noise pollution type petitions), and realize the centralized analysis of similar problems;

[0029] 4. Distributed Storage and Encryption Module: Communicating with the core AI intelligent analysis module, this module employs a distributed storage architecture to store the collected raw data, preprocessed data, and AI analysis results. This architecture features high scalability and reliability, allowing for flexible expansion of storage capacity as data volume grows. Multiple data copies are used for backup to prevent data loss. This module uses the national cryptographic algorithm SM4 to encrypt sensitive inspection data, ensuring data security during storage. A built-in access control unit assigns different data access permissions based on roles; for example, inspection team leaders can access all data, while ordinary inspection personnel can only access data related to their assigned inspection tasks, preventing data leakage.

[0030] 5. Intelligent Interaction and Visualization Module: It communicates with the AI ​​intelligent analysis core module and the distributed storage and encryption module respectively, providing a user-friendly human-computer interaction interface. Inspection personnel can use this interface to perform operations such as data query, viewing analysis results, and issuing inspection tasks. Meanwhile, this module visualizes the AI ​​analysis results in the form of charts, heatmaps, time-series curves, and graph neural network correlation maps. The graph neural network correlation map uses a combination of hierarchical and force-oriented layouts. Different types of nodes (inspection targets, inspection problems, rectification measures, and responsible entities) are distinguished by different colors and icons. For example, inspection target nodes use blue circular icons, inspection problem nodes use red triangle icons, rectification measure nodes use green square icons, and responsible entity nodes use yellow diamond icons. The node size is adaptively adjusted according to the amount of associated data, and the thickness of the correlation edges corresponds to the edge weight, with higher weights resulting in thicker edges. The module supports mouse wheel zooming and dragging to adjust the layout of nodes. Clicking on a node displays detailed attribute information (such as the problem description, amount involved, and discovery time for inspection problem nodes), and double-clicking an edge displays the correlation details. The module provides a node filtering function, which can filter and display target nodes and related paths by node type, time range, problem severity, etc. It supports correlation path tracing. By inputting the start node and end node, all correlation paths between the two are automatically highlighted, helping inspection personnel quickly locate the root cause and scope of the problem. In addition, the module has a built-in inspection task management unit, which is used to generate inspection task lists, assign inspection responsibilities to personnel, track the progress of inspection tasks, and link the task progress with AI analysis results. When an anomaly occurs at a certain task node, it will automatically remind relevant personnel to achieve closed-loop management of inspection work.

[0031] 6. Edge Communication Module: This module communicates with the edge intelligent preprocessing module, AI intelligent analysis core module, distributed storage and encryption module, and intelligent interaction and visualization module to enable data transmission between these modules. It supports dual-mode 5G and Wi-Fi communication, automatically switching modes based on network conditions. Wi-Fi is used in environments with good network signal to reduce communication costs, while 5G is used in remote areas or environments with poor network signal to ensure stable and efficient data transmission. Simultaneously, the module incorporates a communication quality monitoring unit to monitor data transmission rate and stability in real time. In the event of communication anomalies, it automatically reconnects and retransmits data to ensure data integrity.

[0032] 7. Power Management Module: Provides stable power to all the above modules, adopts a wide voltage input design to adapt to different power supply environments. The module has a built-in backup power unit (such as a lithium battery pack). In the event of a sudden power outage, the backup power unit automatically starts to ensure that the device can continue to work for a period of time, ensuring the safe storage of data and the proper completion of unfinished tasks.

[0033] Example 1

[0034] This embodiment of an AI-based intelligent analysis-based patrol information integrated management device includes an information acquisition module, an edge intelligent preprocessing module, an AI intelligent analysis core module, a distributed storage and encryption module, an intelligent interaction and visualization module, an edge communication module, and a power management module.

[0035] 1. Information Acquisition Module: The text acquisition subunit uses an EPSONDS-870 scanner to scan paper inspection plans, reports, etc., and connects to the main unit via a USB interface. It also includes an electronic document import interface supporting doc and pdf files. The image acquisition subunit uses a Hikvision DS-2CD3T46WDA3-L high-definition camera to capture on-site inspection images and a DJI Phantom 4 RTK drone to capture outdoor wide-area scene images. The audio acquisition subunit uses a ZoomH1n high-fidelity voice recorder to capture conversation and meeting recordings, with the audio files transmitted to the device via Bluetooth. The ledger acquisition subunit interfaces with the inspected unit's UFIDA financial system through a dedicated financial data interface and with the OA project management system through a project management system interface, automatically collecting financial statements and project approval documents. The format recognition unit uses the Tika toolkit to recognize different file formats and converts similar data into a unified format, such as converting all text data to txt format, image data to jpg format, and audio data to mp3 format.

[0036] 2. Edge Intelligence Preprocessing Module: The NVIDIA Jetson Xavier NX edge computing module serves as the core processing unit for preprocessing the collected data. For text data, the jieba word segmentation tool is used for keyword extraction, and redundant information is removed using regular expressions. For image data, Gaussian filtering in the OpenCV library is used for deblurring, and histogram equalization is used for image enhancement. For audio data, Audacity is used for noise reduction, and audio is converted to text using the Baidu speech recognition API. For ledger data, the pandas library in Python is used for structured parsing to extract key fields. The data quality assessment unit sets pre-defined evaluation indicators of data integrity ≥90%, accuracy ≥95%, and image clarity ≥80 points (out of 100). The preprocessed data is scored; qualified data proceeds to subsequent analysis, while unqualified data is returned to the acquisition module for re-collection. The data anonymization unit uses character replacement, replacing the middle 8 digits of the ID number with "*" and the name of sensitive units with "[Sensitive Unit]".

[0037] 3. AI Intelligent Analysis Core Module: Utilizing an Intel Core i9-12900K processor as the core computing unit, and equipped with the TensorFlow / PyTorch Geometric deep learning framework, the specific implementation and data flow of the core algorithms in each unit are as follows: 1) Multimodal Fusion Analysis Unit: Based on an improved Transformer model (implemented in PyTorch). Specific applications: The word embedding layer uses a pre-trained Word2Vec model (adapted to the patrol text corpus), the convolutional layer uses a 3×3 convolutional kernel, and the modal attention weights are calculated through a fully connected layer; the input is pre-processed multimodal structured data (text keyword set, image enhancement pixel matrix, etc.), and the output is a 512-dimensional fusion feature vector (dimensionless); 2) Patrol Model Unit: The offline deployed general-purpose large model and the patrol-specific large model are loaded and run through local model files. The RAG enhanced retrieval algorithm is specifically implemented by calculating the cosine similarity (dimensionless) between the fusion feature vector and the knowledge base slice vector using PyTorch, and matching related knowledge. 1) Output structured parsing / matching results; 2) Graph Neural Network Association Analysis Unit: Implements a GCN model based on the PyTorchGeometric framework. Specific application: Sets 2 layers of graph convolutional layers (input dimension 256, hidden layer 128, output 64), ReLU activation function, learning rate 0.001, and optimizes parameters through backpropagation; the input is entity key-value pairs and association triples, and the output is a 64-dimensional node feature vector and updated edge weights; 3) Clustering Grouping Unit: Implements an improved K-means algorithm based on the Scikit-learn framework. Specific application: Determines the cluster number range of 2-10 through the elbow rule, the silhouette coefficient threshold is 0.7, and the weight coefficients of the weighted Euclidean distance are determined through training with 300 manually labeled inspection problem samples; the input is a standardized 256-dimensional feature vector, and the output is cluster labels and similar sets; 4) LSTM Trend Prediction Model: Implements a 3-layer LSTM structure based on TensorFlow. The input is the inspection time series data of the past 5 years, and the output is the risk trend prediction results for the next year;

[0038] 4. Distributed Storage and Encryption Module: Utilizing a Hadoop distributed storage architecture, this module features three data nodes, each configured with a 10TB hard drive, enabling multi-copy data backup. Sensitive data is encrypted using the national standard SM4 algorithm, with encryption keys stored via a hardware encryption chip to ensure key security. The access control unit categorizes user roles into patrol team leaders, deputy team leaders, regular patrol personnel, and system administrators, assigning different access permissions to each role. For example, regular patrol personnel can only access data related to their assigned patrol tasks and cannot access data from other tasks.

[0039] 5. Intelligent Interaction and Visualization Module: A 15.6-inch touchscreen display is used as the interactive interface. Human-computer interaction software is developed based on the Qt framework to realize functions such as data query, viewing analysis results, and task assignment. Visualization is implemented using the ECharts chart library combined with the D3.js visualization framework. The visualization of the graph neural network correlation graph is developed using the ECharts graph component. Specific configurations are as follows: The layout is set to switchable between "force" (force-oriented layout) and "hierarchy" (hierarchical layout). The default is to use the force-oriented layout to display global relationships. When switching to the hierarchical layout, it displays relationships according to the logical hierarchy of "inspection object - inspection problem - responsible entity - rectification measures". In the node style configuration, the fill color and border color of different types of nodes are set through itemStyle, the corresponding icon is set through the symbol attribute, and symbolSize is set to [30, 60]. The adaptive range dynamically calculates its size based on the amount of data associated with each node. Edge styles are set via `lineStyle`, with `lineWidth` binding edge weight parameters, mapping the weight range to a line width of 1-5px. Interactive functions include a node / edge details pop-up display using ECharts' tooltip component, graph zooming via the `dataZoom` component, and drag-and-drop functionality via the `roam` property. Node filtering and path tracing functions process the graph data output by the graph neural network association analysis unit through front-end logic. Filtering involves filtering the node dataset based on user input criteria and re-rendering the graph. Path tracing uses a depth-first search algorithm to traverse the graph data, extracting all paths between the starting and ending nodes and highlighting them. Other visualizations include a heatmap showing the distribution of inspection issues, a line chart showing risk trends, and a pie chart showing issue classification statistics. The task management unit generates an inspection task list, clearly defining the task name, responsible personnel, and completion deadline. A progress bar displays the task completion status in real time, and an automatic audio-visual reminder is issued when a task is not completed on time.

[0040] 6. Edge Communication Module: Dual-mode communication is achieved using a Huawei ME909S-8215G module and an Intel AX200 Wi-Fi 6 module, with automatic switching between communication modes implemented through embedded software. A communication quality monitoring unit monitors the data transmission rate in real time; when the Wi-Fi communication rate drops below 1Mbps, it automatically switches to 5G communication; in the event of a communication interruption, it automatically reconnects and retransmits any data that was not successfully transmitted.

[0041] 7. Power Management Module: It adopts a wide voltage input module (input voltage range: 12V-24V) to adapt to different power supply environments; it has two built-in 12V / 10Ah lithium battery packs as backup power. In the event of a sudden power outage, the backup power can support the device to work continuously for 2 hours, ensuring safe data storage and work completion.

[0042] The application process of the device in this embodiment during actual inspection work is as follows:

[0043] (1) Data collection: Through the various collection sub-units of the information collection module, the inspection plan, the report materials of the inspected unit, the on-site video, the recording of the conversation, the financial statements and other data are collected in a comprehensive manner. The format recognition unit recognizes and performs preliminary standardization processing on the collected data.

[0044] (2) Edge preprocessing: The edge intelligent preprocessing module performs redundancy removal, image enhancement, audio noise reduction, and structured parsing on the collected data. The data quality assessment unit selects qualified data, and the data desensitization unit desensitizes sensitive information.

[0045] (3) AI intelligent analysis: The multimodal fusion analysis unit of the core module of AI intelligent analysis performs feature fusion on the preprocessed data to generate feature vectors; the intelligent judgment unit of inspection problems combines the inspection key points knowledge base and historical case base to realize the automatic identification, classification and rating of inspection problems; the trend prediction unit predicts potential risk points; and the rectification effect evaluation unit evaluates the rectification effect of the inspected unit.

[0046] (4) Data storage and interaction: Distributed storage and encryption modules encrypt and store various types of data. Inspectors can view the analysis results, issue inspection tasks, and track task progress through intelligent interaction and visualization modules, thereby realizing full-process management of inspection work.

[0047] (5) Communication assurance: The edge communication module automatically switches between 5G / Wi-Fi communication modes according to the network environment to ensure the stability and integrity of data transmission between modules; the power management module provides stable power supply, and the backup power unit ensures data security in the event of a sudden power outage.

[0048] The above are merely embodiments of the present invention. The circuits, electronic components, and modules involved are all prior art, fully achievable by those skilled in the art, and require no further explanation. The scope of protection in this application does not involve improvements to the software and methods. Commonly known structures and characteristics in the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all prior art in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the application.

Claims

1. A comprehensive management device for inspection information based on AI intelligent analysis, characterized in that: It includes an information collection module, which is used to collect heterogeneous data such as text, images, audio and ledgers throughout the inspection process, and to identify the heterogeneous data in different formats, convert them into intermediate data in a unified standard format, and complete the initial standardization process. The preprocessing module, which is communicatively connected to the information acquisition module, is used to perform localized and lightweight preprocessing on the standardized intermediate data, and outputs a set of structured fields. The preprocessing module has a built-in data quality assessment unit and a data desensitization unit. The data quality assessment unit takes the preprocessed data as input and outputs a quality score, which is based on data integrity and accuracy. The data desensitization unit takes structured data containing sensitive information as input, processes it through a character replacement / masking algorithm, and outputs desensitized structured data. The intelligent analysis module, which is communicatively connected to the preprocessing module, includes a multimodal fusion analysis unit and an inspection model unit. The input of the multimodal fusion analysis unit is the preprocessed multimodal data. It adopts an improved Transformer model that runs offline, first converting various types of data into feature vectors of a unified dimension, then calculating the feature weights of each modality through a modal attention mechanism, and weighted fusion to generate a unified inspection information feature vector. The output is a fused feature vector. Inspection Model Unit: The input consists of a fused feature vector and a slice of the local inspection knowledge base. Through the RAG-enhanced retrieval mechanism, the cosine similarity between the fused feature vector and the knowledge base slice is calculated to match related knowledge. The output consists of policy analysis results and case matching results.

2. The patrol information integrated management device based on AI intelligent analysis as described in claim 1, characterized in that: The intelligent analysis module also includes a natural language interaction unit. The input of the natural language interaction unit is the natural language instructions of the user's business scenario, which are converted into structured instructions that the model can execute, and the output is standardized analysis instructions.

3. The patrol information integrated management device based on AI intelligent analysis as described in claim 1, characterized in that: The intelligent analysis module also includes an inspection task intelligent agent unit. The input of the inspection task intelligent agent unit is standardized analysis instructions and corresponding structured data, which encapsulates special task process algorithms and outputs task processing results.

4. The patrol information integrated management device based on AI intelligent analysis as described in claim 1, characterized in that: The graph neural network association analysis unit uses a graph convolutional neural network (GCN) to construct an inspection association graph. The input is structured entity data and association relationship data. The entity attributes are converted into node feature vectors, and the association weights are incorporated into the adjacency matrix as graph edge weights. The features of neighboring nodes are aggregated through two layers of graph convolutional layers, the node vector representation is updated, and the output is a weighted inspection association graph.

5. The patrol information integrated management device based on AI intelligent analysis as described in claim 1, characterized in that: The clustering grouping unit adopts an improved K-means clustering algorithm. The input is the feature vector of inspection issues / petitions / historical cases. The optimal number of clusters is automatically determined by the elbow rule. Weighted Euclidean distance is introduced to calculate sample similarity. The cluster centers are iteratively updated until convergence. The output is the cluster label and the set of similar data.

6. The patrol information integrated management device based on AI intelligent analysis as described in claim 1, characterized in that: It also includes a local encrypted storage module that communicates with the intelligent analysis module. All data is stored locally in a closed loop and is not transmitted to the outside. Sensitive inspection data is encrypted and stored. At the same time, it has a built-in role-based access control unit that assigns different data access permissions to inspection team leaders, ordinary inspection personnel, and system administrators.

7. The patrol information integrated management device based on AI intelligent analysis as described in claim 6, characterized in that: It also includes an intelligent interaction and visualization module, which is communicatively connected to the intelligent analysis module and the local encrypted storage module, including an LED touch screen and localized interactive software, to enable fully offline interaction between inspection personnel and the device, supporting natural language command input, data query, analysis result viewing, and inspection task assignment.

8. The patrol information integrated management device based on AI intelligent analysis as described in claim 7, characterized in that: It also includes a field collaborative communication module, which communicates with the preprocessing module, intelligent analysis module, local encrypted storage module, and intelligent interaction and visualization module respectively. It has a built-in network switching chip, supports only local area network communication, and does not have external network connection function to ensure data security.

9. The patrol information integrated management device based on AI intelligent analysis as described in claim 1, characterized in that: The preprocessing module includes a text processing unit, an image processing unit, an audio processing unit, and a ledger processing unit. The text processing unit incorporates a regular expression matching algorithm and a TF-IDF algorithm. The regular expression matching algorithm matches and removes duplicate character streams using a preset regular expression, while the TF-IDF algorithm extracts text keywords and generates a set of "keyword-weight" key-value pairs. The image processing unit incorporates a Gaussian filtering algorithm with a 3×3 convolution kernel and a histogram equalization algorithm. The Gaussian filtering algorithm smooths the pixel matrix, while the histogram equalization algorithm adjusts the distribution of pixel grayscale values.

10. The patrol information integrated management device based on AI intelligent analysis as described in claim 9, characterized in that: The audio processing unit incorporates an adaptive Wiener filtering algorithm, an MFCC algorithm, and a CTC algorithm. The adaptive Wiener filtering algorithm eliminates background noise, the MFCC algorithm converts the amplitude sequence into a Mel-frequency cepstral coefficient matrix, and the CTC algorithm converts the coefficient matrix into a normalized character stream. The ledger processing unit has a built-in regular expression extraction algorithm and field mapping table. It uses the regular expression extraction algorithm to extract key fields and generate a set of "field name - field value" key-value pairs.