Management system for digital processing of sound image archives and retrieval method thereof
By integrating intelligent methods for data acquisition and monitoring, content-aware processing, data generation and management, and hierarchical storage, the problems of unstable acquisition, low processing efficiency, and high storage costs in the digitization of audio-visual archives have been solved, achieving efficient and accurate archive management and long-term preservation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JINCHUAN GROUP CO LTD
- Filing Date
- 2025-12-26
- Publication Date
- 2026-05-12
AI Technical Summary
Traditional audio-visual archive digitization processes suffer from problems such as unstable acquisition, low processing efficiency, poor retrieval accuracy, and high storage costs, making it difficult to achieve high efficiency, accuracy, and sustainability.
The system employs a data acquisition and monitoring module that combines Kalman filtering and LSTM models for real-time quality monitoring; a content-aware processing module that uses XGBoost decision trees and CNN models for intelligent processing; a data generation and management module that constructs a knowledge graph using Whisper speech-to-text and GNN; a hierarchical storage module that implements dynamic storage based on ARIMA and MerkleTree; and a retrieval module that performs semantic retrieval using Word2Vec.
It significantly improves the stability and data quality of the acquisition process, increases processing efficiency and retrieval accuracy, reduces storage costs, and realizes intelligent management of audio-visual archives throughout the entire process.
Smart Images
Figure CN122019800A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of archival management technology, specifically to a management system and retrieval method for digitizing audio-visual archives. Background Technology
[0002] Audiovisual archives, as important carriers of historical events, cultural activities, and social changes, are widely found in institutions such as archives, radio and television stations, and universities. Their media formats include analog storage media such as videotapes, magnetic tapes, and films. With the development of information technology, the digitization of audiovisual archives has become a key path to break through the limitations of traditional storage and achieve long-term preservation and open access, as well as a core requirement for the inheritance of cultural heritage and the sharing of information resources.
[0003] However, traditional digitization methods suffer from numerous technical bottlenecks in core aspects such as acquisition, processing, management, storage, and utilization, making it difficult to meet the current comprehensive demands for efficient digitization processes, high-quality processing results, intelligent record management, and sustainable long-term preservation. Specific pain points are as follows: First, in the data acquisition and monitoring stage, existing technologies mostly rely on manual operation or equipment with fixed parameters for analog signal transcription. They lack the ability to perceive and dynamically adjust the physical state of the medium (such as magnetic tape tension and film humidity) and signal quality during the acquisition process. This makes them susceptible to signal distortion or medium damage due to medium aging or environmental fluctuations, resulting in irreversible data loss. Second, in the content processing stage, existing systems mostly use a single noise reduction algorithm to process audio and video uniformly. They cannot adaptively match processing strategies based on audio type (such as speech, music, and ambient sound), leading to the misfiltering of important voice information or incomplete noise suppression during the noise reduction process, affecting the usability of digitized archives. Third, in the archive management stage, most systems still rely on manual indexing to transcribe, classify, and tag audio and video content. This not only consumes a lot of manpower and is time-consuming, but also suffers from strong subjectivity and inconsistent indexing standards, resulting in low-structured archival metadata and difficulty in achieving accurate association and efficient retrieval. Finally, in the storage and utilization stage, existing storage solutions mostly adopt a fixed-level architecture and do not dynamically allocate storage resources based on archive access frequency, resulting in wasted storage costs. In summary, there is an urgent need for a comprehensive solution that integrates multi-source sensing, intelligent algorithms, knowledge modeling, and system collaboration to achieve efficient, accurate, and sustainable digital processing of audio-visual archives. Summary of the Invention
[0004] The purpose of this invention is to provide a management system and retrieval method for the digitization of audio-visual archives, which solves the technical problems of unstable acquisition, low processing efficiency, poor retrieval accuracy and high storage cost in traditional audio-visual archive digitization.
[0005] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, this application provides a management system for the digitization of audio-visual archives, comprising a data acquisition and monitoring module, a content perception and processing module, a data generation and management module, a hierarchical storage and preservation module, and a retrieval module connected sequentially; wherein, The data acquisition and monitoring module includes a vibration sensor, a temperature and humidity sensor, and a signal analyzer. The acquired data is fused with a Kalman filter algorithm to dynamically adjust the parameters of the acquisition equipment. The content-aware processing module includes a content-aware analysis unit, which matches audio processing strategies based on the XGBoost decision tree model and performs environmental noise reduction through a convolutional neural network (CNN). The data generation and management module combines the Whisper speech transcription model, the BERT entity label extraction model, and the graph neural network (GNN) to construct an event knowledge graph. The hierarchical storage module predicts the popularity of archives based on the ARIMA time series model and uses MerkleTree to realize file integrity verification and dynamic storage hierarchy. The retrieval module supports speech content retrieval through Word2Vec word embedding vectors.
[0006] Furthermore, the data acquisition and monitoring method of the data acquisition and monitoring module is as follows: A1. The state estimate is calculated using the Kalman filter algorithm, and the formula is as follows: Where x is the state of the medium, Kk is the Kalman gain, zk is the sensor observation, and H is the observation matrix; A2. Perform frame segmentation processing on the audio and video signals, extract spectral features, input them into the LSTM model to predict the normal signal distribution, and if the Euclidean distance between the current frame and the predicted value exceeds the threshold, trigger an anomaly warning and record the fault timestamp.
[0007] Furthermore, the processing method of the content-aware analysis unit is as follows: The speech activity detection ratio and signal-to-noise ratio are extracted from the audio, and motion vectors and color histograms are extracted from the video as input feature vectors for the XGBoost decision tree model. A dual-channel CNN model is used for audio denoising. The training data consists of noisy audio and clean audio pairs. The loss function is a weighted sum of mean squared error and perceptual loss. A generative adversarial network based on the U-Net structure is used to repair video scratches. The generator is optimized by minimizing the texture difference between the repaired area and its neighborhood.
[0008] Furthermore, the implementation method of the data generation and management module is as follows: The Whisper model with the Transformer architecture is used to perform end-to-end speech recognition on long audio files, outputting timestamped text. The transcribed text is then weighted by TF-IDF to extract keywords, and semantic analysis is performed using the BERT model to generate entity labels. Triples are extracted from metadata, event node vectors are generated through GNN embedding, and an event knowledge graph is constructed.
[0009] Furthermore, the hierarchical storage module predicts the popularity of archives based on the ARIMA time series model and dynamically allocates hot / warm / cold storage levels. The prediction model is as follows: The file is divided into blocks and hash values are calculated using MerkleTree. The current hash is compared with the original hash periodically. If they are inconsistent, the Rabin-Karp algorithm is triggered to locate the corrupted blocks and automatically repair them using RAID6 redundant storage.
[0010] Secondly, this application also provides a retrieval method for digital processing of audio-visual archives, the retrieval method comprising the following steps: The retrieval module converts the transcribed text into word embedding vectors. After the user inputs a query term, the cosine similarity is calculated, the timestamp of the matching segment is returned, and a particle filter algorithm is applied to predict the user's interaction trajectory, thereby optimizing rendering performance.
[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The audio-visual archive digitization processing and management system described in this invention employs Kalman filtering to fuse multi-source sensor data and combines an LSTM model to perform real-time quality monitoring of the acquired signals, significantly improving the stability and security of the acquisition process and effectively reducing equipment failure rates and data loss risks. The content-aware processing module uses XGBoost decision trees to achieve intelligent matching of audio processing strategies and uses a dual-channel CNN model and a U-Net-structured generative adversarial network to achieve audio noise reduction and video scratch repair, respectively, greatly improving the quality and efficiency of digitization processing. The data generation and management module integrates Whisper speech transcription, BERT semantic analysis, and GNN knowledge graph construction technologies to automatically generate structured and semantic deep metadata, realizing intelligent indexing and knowledge association of archival resources, greatly shortening manual processing time and improving retrieval accuracy. 2. The hierarchical storage module of the audio-visual archive digitization processing and management system described in this invention introduces the ARIMA time series model to predict the access frequency of archives, dynamically allocates storage levels, and combines Merkle Tree and RAID 6 to realize file integrity verification and automatic repair, ensuring the long-term safe preservation of massive audio-visual archives and significantly reducing storage costs and maintenance burdens; the retrieval module realizes semantic retrieval of "text-based audio" through Word2Vec word embedding, and by integrating five core modules of data acquisition and monitoring, content-aware processing, data generation and management, hierarchical storage and preservation, and utilization and retrieval, it realizes the intelligentization of the entire process of audio-visual archives from acquisition, processing, management to long-term preservation and intelligent utilization. Attached Figure Description
[0012] Figure 1 This is a system architecture diagram of a management system for digital processing of audio-visual archives according to the present invention.
[0013] Figure 2 This is a flowchart of the LSTM signal anomaly detection process in a management system for digital processing of audio-visual archives according to the present invention.
[0014] Figure 3 This is a structural diagram of a CNN audio noise reduction model for a management system for digital processing of audio-visual archives according to the present invention.
[0015] Figure 4 This is a flowchart of the Merkle Tree file verification process for a management system for digitizing audio-visual archives, as described in this invention. Detailed Implementation
[0016] Combined with appendix Figure 1 A management system for digitizing audio-visual archives includes a data acquisition and monitoring module, a content-aware processing module, a data generation and management module, a hierarchical storage and preservation module, and a retrieval module connected in sequence. The data acquisition and monitoring module includes a vibration sensor, a temperature and humidity sensor, and a signal analyzer. The acquired data is fused with a Kalman filter algorithm to dynamically adjust the parameters of the acquisition equipment. The content-aware processing module includes a content-aware analysis unit that matches audio processing strategies based on an XGBoost decision tree model and uses a convolutional neural network (CNN) for environmental noise reduction. The data generation and management module combines a Whisper speech-to-text model, a BERT entity tag extraction model, and a graph neural network (GNN) to construct an event knowledge graph. The hierarchical storage and preservation module predicts archive popularity based on an ARIMA time series model and uses a MerkleTree to verify file integrity and dynamically classify storage. The retrieval module uses SLAM technology and optical flow to achieve VR / AR interaction and supports voice content retrieval through Word2Vec word embedding vectors.
[0017] Combined with appendix Figure 2The data acquisition and monitoring methods of the data acquisition and monitoring module are as follows: A1. The state estimate is calculated using the Kalman filter algorithm, and the formula is as follows: Where x is the state of the medium, Kk is the Kalman gain, zk is the sensor observation, and H is the observation matrix; A2. Perform frame segmentation processing on the audio and video signals, extract spectral features, input them into the LSTM model to predict the normal signal distribution, and if the Euclidean distance between the current frame and the predicted value exceeds the threshold, trigger an anomaly warning and record the fault timestamp.
[0018] Combined with appendix Figure 3 The content-aware analysis unit processes the following methods: extracting the speech activity detection ratio and signal-to-noise ratio from the audio, and extracting motion vectors and color histograms from the video as input feature vectors for the XGBoost decision tree model. A dual-channel CNN model is used for audio denoising. The training data consists of pairs of noisy and clean audio. The loss function is a weighted sum of mean square error and perceptual loss. A generative adversarial network based on the U-Net structure is used to repair video scratches. The generator is optimized by minimizing the texture difference between the repaired area and its neighborhood.
[0019] Combined with appendix Figure 4 The data generation and management module is implemented as follows: using the Whisper model of the Transformer architecture to perform end-to-end speech recognition on long audio, outputting text with timestamps, extracting keywords by TF-IDF weighting of the transcribed text, and performing semantic analysis in conjunction with the BERT model to generate entity labels, extracting triples from metadata, generating event node vectors through GNN embedding, and constructing an event knowledge graph.
[0020] The tiered storage module predicts the popularity of archives based on the ARIMA time series model and dynamically allocates hot / warm / cold storage levels. The prediction model is as follows: The file is divided into blocks and hash values are calculated using MerkleTree. The current hash is compared with the original hash periodically. If they are inconsistent, the Rabin-Karp algorithm is triggered to locate the corrupted blocks and automatically repair them using RAID6 redundant storage.
[0021] A retrieval method for digital processing of audio-visual archives includes the following steps: converting transcribed text into word embedding vectors through a retrieval module; calculating cosine similarity after the user inputs query terms; returning the timestamps of matching segments; and applying a particle filter algorithm to predict user interaction trajectories and optimize rendering performance.
[0022] In its implementation, the audio-visual archive digitization management system of this invention first achieves safe and stable acquisition of original audio-visual media (such as magnetic tape and film) through a data acquisition and monitoring module. During the acquisition process, the system deploys vibration sensors to monitor magnetic tape tension changes in real time, temperature and humidity sensors to collect storage environment parameters, and a signal analyzer to acquire audio waveforms and video synchronization signals. These three types of data are input into a Kalman filter algorithm for fusion processing, dynamically estimating the media state and adjusting playback speed, magnetic head pressure, and gain parameters in real time to prevent breakage or magnetic head wear caused by abnormal tension. Simultaneously, the acquired audio and video signals are processed in 50ms frames to extract features such as MFCC and spectral energy, which are then input into a pre-trained LSTM model to predict normal signal distribution. When the Euclidean distance between a frame signal and the predicted value exceeds the mean plus three standard deviations (μ+3σ), the system immediately triggers an anomaly warning, suspends acquisition, and records the fault timestamp to ensure the safety of the original archives.
[0023] After data acquisition, the digitized files enter the content-aware processing module for intelligent repair and enhancement. The content-aware analysis unit extracts the speech activity detection (VAD) ratio and signal-to-noise ratio (SNR) from the audio and motion vectors and color histograms from the video, forming feature vectors that are input into the XGBoost decision tree model to automatically match the optimal processing strategy (such as "voice-first noise reduction" or "high dynamic range video enhancement"). For audio, a dual-channel convolutional neural network (CNN) model is used to process the time-frequency map and the original waveform separately. Noisy audio and corresponding clean audio are used as training pairs, and a loss function weighted by mean squared error (MSE) and perceptual loss is used for training to achieve end-to-end noise reduction. For video, a generative adversarial network (GAN) based on the U-Net structure is used to identify and repair scratched areas. The generator output is optimized by minimizing the difference in texture and edge consistency between the repaired area and its surrounding neighborhood (such as the SSIM index), significantly improving image quality.
[0024] Subsequently, the system enters the data generation and management module, where it performs deep metadata extraction and knowledge organization on the processed audio and video. The Whisper model, based on the Transformer architecture, is used to perform high-precision speech-to-text transcription of long audio files, outputting timestamped text. TF-IDF weighted analysis is then performed on this text to extract keywords, and the BERT model is used for semantic understanding to identify entity tags such as people, events, and locations. Furthermore, "subject-relationship-object" triples (e.g., "1980 school anniversary - includes - principal's speech") are extracted from the metadata. These triples are then embedded into low-dimensional vectors using a graph neural network (GNN) to construct an event knowledge graph, achieving semantic association and structured storage of the archive content.
[0025] In the tiered storage module, the system models historical access logs based on the ARIMA time series model to predict the access frequency of each file in the future. Combining this with the value level, the system calculates the cost-effectiveness ratio and dynamically allocates the data to different levels of storage media, such as SSD (hot data), HDD (warm data), or tape library (cold data). Simultaneously, a Merkle Tree is used to divide each file into blocks according to its type (10MB / block for audio, 50MB / block for video). The hash value of each block is calculated and a hash tree is constructed. The hash is periodically compared with the original hash. If an inconsistency is found, the Rabin-Karp algorithm is used to quickly locate the corrupted data block and automatically repair it using the redundancy mechanism of RAID 6, ensuring the long-term integrity and accessibility of digital archives.
[0026] Finally, in the retrieval module, the system supports intelligent retrieval and immersive display. The transcribed text generated by Whisper is converted into Word2Vec word embedding vectors. After the user inputs a query term (such as "opening ceremony"), the system calculates the cosine similarity and returns the matching audio clips and their timestamps, achieving semantic retrieval of "text-to-sound". For visualization, SLAM technology (such as ORB-SLAM) is integrated to track the user's viewpoint position in the VR / AR environment in real time. Combined with optical flow, historical image backgrounds are dynamically matched, and particle filtering algorithms are used to predict user interaction trajectories, optimizing 3D scene rendering performance and providing a low-latency, highly immersive historical scene restoration experience.
[0027] Through the coordinated operation of the above modules, this system achieves full automation and intelligence in the process of audio-visual archives from acquisition, processing, management to storage and utilization, significantly improving digitization efficiency, data quality and long-term preservation reliability. It is suitable for large-scale audio-visual resource management scenarios such as universities, archives, and broadcasting institutions.
[0028] The following detailed implementation demonstration of this system will be provided with reference to specific embodiments. Example: Digitization of Enterprise Historical Audio Recordings A company archive holds a large number of analog magnetic tapes from the 1980s, containing important historical recordings such as research forums, production scenes, and speeches by leaders. Due to severe tape aging, issues such as magnetic powder shedding, signal attenuation, and high background noise exist, making traditional digitization methods inefficient and of poor quality. After deployment, this system will automate and intelligently process these recordings according to the following workflow.
[0029] 1. Implementation of the data acquisition and monitoring module After the system is started, the tape playback device is connected first, and then the vibration sensor (installed on the tape guide), temperature and humidity sensor (placed in the equipment compartment) and digital signal analyzer (acquiring audio output waveforms) are connected.
[0030] Kalman filtering for fusing multi-source data: The sensors acquire data once per second, which is then input into the Kalman filter algorithm. The state vector x includes magnetic tape tension, ambient temperature and humidity, and signal-to-noise ratio; the observed values zk are the real-time readings of the three types of sensors; the observation matrix H maps the state to the observation space. The system recursively updates the state estimate according to the following formula: The Kalman gain Kk is dynamically calculated based on the system noise covariance. If the tension estimate exceeds the safety threshold (e.g., >5N), the system automatically reduces the playback speed and issues an alarm to prevent tape breakage.
[0031] LSTM signal anomaly detection: The audio signal is segmented into 50ms frames. The MFCC (Melbourne Frequency Cepstral Coefficients) and spectral energy of each frame are extracted as features and input into a pre-trained LSTM model. This model is trained on historical normal recording data to learn the temporal distribution patterns of the signal. The system calculates the Euclidean distance between the current frame's features and the LSTM prediction value. If the distance between three consecutive frames exceeds μ+3σ (based on historical data statistics), it is determined to be "signal distortion". The system will pause acquisition, record the fault timestamp, and prompt "head blockage or media damage". The operator can intervene in time.
[0032] 2. Implementation of the Content-Aware Processing Module After the data collection is completed, the system automatically enters the content analysis and repair phase.
[0033] XGBoost Strategy Matching: The content-aware analysis unit extracts the following features from the audio: Voice activity detection (VAD) percentage (to determine whether human voice is the dominant sound); signal-to-noise ratio (SNR); spectral flatness (to determine whether it is music or ambient sound).
[0034] The above features are used to construct a vector [0.72, 18.5, 0.45] and input into the XGBoost decision tree model. The model outputs the optimal processing strategy based on the training data (labeled as "speech", "exterior", and "mixed" categories). For example, if VAD > 0.6 and SNR < 20dB, then the strategy of "voice enhancement + broadband noise reduction" is matched.
[0035] Dual-channel CNN audio noise reduction: Employing a two-input convolutional neural network: The first channel inputs a time-frequency graph generated by the short-time Fourier transform (STFT); The second channel inputs the original waveform segment.
[0036] The network structure consists of 4 convolutional layers and 2 fully connected layers, with the loss function being: The perceptual loss, Lperceptual, extracts high-level feature differences based on a pre-trained speech recognition model, ensuring that speech intelligibility does not decrease after noise reduction. The training data consists of noisy / clean audio pairs (such as speech with added white noise or air conditioning noise).
[0037] U-Net GAN video scratch repair (for video tapes): For video files, the system uses a generative adversarial network (GAN) based on the U-Net architecture: The generator uses an encoder-decoder structure to identify scratch areas; The discriminator determines whether the repaired area looks natural; The loss function minimizes the texture difference between the repaired region and its neighborhood (such as SSIM, L1 Loss).
[0038] The training data consists of pairs of video frames with artificially added scratches and the original frames.
[0039] 3. Implementation of the Data Generation and Management Module Whisper speech-to-text transcription: The Whisper-large model, based on the Transformer architecture, performs end-to-end audio transcription, supporting multiple languages, punctuation, automatic segmentation, and outputting a timestamp for each sentence. For example: [00:02:15 - 00:02:30] General Manager XXX emphasized at the Second Session of the Eighth Staff and Workers' Congress... BERT entity tag extraction: Perform TF-IDF analysis on the transcribed text to extract keywords (such as "general manager", "construction site", "1985"), then input the data into the BERT model for Named Entity Recognition (NER) to output structured tags. {Person: "General Manager", Event: "Construction Site", Time: "1985", Location: "Flash Furnace"} GNN constructs an event knowledge graph: Extract triples from metadata, such as: (Audio A, includes, General Manager), (General Manager, time of occurrence, 1985), (1985, related event, construction site) Nodes are embedded using graph neural networks (GNNs) (such as GraphSAGE) to generate vector representations, and node importance is calculated using the PageRank algorithm to construct a visual knowledge graph.
[0040] 4. Implementation of the Hierarchical Storage Preservation Module ARIMA Heat Prediction: The system collects the access logs of the past year (such as download times, retrieval frequencies), establishes a time series {xt}, and fits an ARIMA(1,1,1) model: Predict the access volume in the next 30 days and calculate the "cost-benefit ratio": Ratio > 5: Store in SSD (hot storage); 2 - 5: HDD (warm storage); < 2: Tape library (cold storage).
[0041] Merkle Tree Integrity Check: The files are chunked by type (10MB per chunk for audio, 50MB per chunk for video), calculate the SHA-256 hash for each chunk, and construct a Merkle Tree. The system performs the check once a week: Compare the current root hash with the original root hash; If they are inconsistent, use the Rabin-Karp algorithm to quickly locate the damaged chunk (based on rolling hash); Automatically repair the redundant data through RAID 6 and record the log.
[0042] 5. Implementation of the Retrieval Module Word2Vec Semantic Retrieval: After segmenting all the transcribed texts, use the Word2Vec model (Skip-gram, vector dimension 128) to train word vectors. When the user inputs "construction site", the system calculates the cosine similarity between the query word and the text of each audio segment, and returns the results with a similarity > 0.7 and their timestamps.
[0043] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.
Claims
1. A management system for digitizing audio-visual archives, characterized in that, It includes a data acquisition and monitoring module, a content-aware processing module, a data generation and management module, a hierarchical storage and preservation module, and a retrieval module, which are connected in sequence; among them, The data acquisition and monitoring module includes a vibration sensor, a temperature and humidity sensor, and a signal analyzer. The acquired data is fused with a Kalman filter algorithm to dynamically adjust the parameters of the acquisition equipment. The content-aware processing module includes a content-aware analysis unit, which matches audio processing strategies based on the XGBoost decision tree model and performs environmental noise reduction through a convolutional neural network (CNN). The data generation and management module combines the Whisper speech transcription model, the BERT entity label extraction model, and the graph neural network (GNN) to construct an event knowledge graph. The hierarchical storage module predicts the popularity of archives based on the ARIMA time series model and uses MerkleTree to realize file integrity verification and dynamic storage hierarchy. The retrieval module supports speech content retrieval through Word2Vec word embedding vectors.
2. The management system for digitizing audio-visual archives according to claim 1, characterized in that, The data acquisition and monitoring methods of the data acquisition and monitoring module are as follows: A1. The state estimate is calculated using the Kalman filter algorithm, and the formula is as follows: Where x is the state of the medium, Kk is the Kalman gain, zk is the sensor observation, and H is the observation matrix; A2. Perform frame segmentation processing on the audio and video signals, extract spectral features, input them into the LSTM model to predict the normal signal distribution, and if the Euclidean distance between the current frame and the predicted value exceeds the threshold, trigger an anomaly warning and record the fault timestamp.
3. The management system for digitizing audio-visual archives according to claim 1, characterized in that, The processing method of the content-aware analysis unit is as follows: The speech activity detection ratio and signal-to-noise ratio are extracted from the audio, and motion vectors and color histograms are extracted from the video as input feature vectors for the XGBoost decision tree model. A dual-channel CNN model is used for audio denoising. The training data consists of noisy audio and clean audio pairs. The loss function is a weighted sum of mean squared error and perceptual loss. A generative adversarial network based on the U-Net structure is used to repair video scratches. The generator is optimized by minimizing the texture difference between the repaired area and its neighborhood.
4. The management system for digitizing audio-visual archives according to claim 1, characterized in that, The implementation method of the data generation and management module is as follows: The Whisper model with the Transformer architecture is used to perform end-to-end speech recognition on long audio files, outputting timestamped text. The transcribed text is then weighted by TF-IDF to extract keywords, and semantic analysis is performed using the BERT model to generate entity labels. Triples are extracted from metadata, event node vectors are generated through GNN embedding, and an event knowledge graph is constructed.
5. The management system for digitizing audio-visual archives according to claim 1, characterized in that, The hierarchical storage module predicts the popularity of archives based on the ARIMA time series model and dynamically allocates hot / warm / cold storage levels. The prediction model is as follows: The file is divided into blocks and hash values are calculated using MerkleTree. The current hash is compared with the original hash periodically. If they are inconsistent, the Rabin-Karp algorithm is triggered to locate the corrupted blocks and automatically repair them using RAID6 redundant storage.
6. A retrieval method for digitized audio-visual archives, based on the management system for digitized audio-visual archives as described in claim 1, characterized in that, Includes the following steps: The retrieval module converts the transcribed text into word embedding vectors. After the user inputs a query term, the cosine similarity is calculated, the timestamp of the matching segment is returned, and a particle filter algorithm is applied to predict the user's interaction trajectory, thereby optimizing rendering performance.