Cloud storage data deduplication and access acceleration methods for multimedia teaching

Abstract

This invention relates to a method for deduplication and accelerated access of cloud storage data for multimedia teaching, aiming to solve the problem of low efficiency in semantic deduplication and popularity filtering of teaching resources. Its core solution includes: extracting metadata such as courseware titles, automatic transcription of teaching videos, and teacher voice recordings; generating a business popularity fingerprint through multimodal semantic feature parsing and normalization fusion; dynamically setting a popularity threshold based on cloud storage load to filter high-popularity resources; and improving deduplication and indexing efficiency by using local fingerprint index reuse and full content hashing for precise comparison. Real-time optimization of popularity fingerprint weights is achieved by collecting user behavior data, enabling adaptive popularity judgment and dynamic iteration of cloud configuration. This solution effectively improves the deduplication accuracy, data processing efficiency, and system adaptability of teaching resource management.

Description

Technical Field

[0001] This invention relates to the field of cloud storage data deduplication and access acceleration technology, and in particular to cloud storage data deduplication and access acceleration methods for multimedia teaching. Background Technology

[0002] In the current cloud storage environment for multimedia teaching resources, data deduplication and access acceleration technologies have become crucial for improving resource management efficiency and user experience. Cloud storage platforms typically need to handle massive concurrent uploads, distributions, and retrievals of multimodal resources such as teaching materials, videos, and audio. Mainstream technologies often employ content fingerprint comparison for deduplication, commonly including full hash fingerprint matching, partial segment content analysis, and machine learning model priority prediction. Furthermore, some systems introduce mechanisms such as hot and cold data queues, priority scheduling, and cache preheating to optimize response and resource utilization in high-concurrency scenarios.

[0003] Existing technologies are primarily based on the following core features: Firstly, they use global content hash fingerprints as the uniqueness criterion for data, enabling duplicate detection and storage optimization of teaching resources. Secondly, to address high-concurrency access pressure, they employ methods such as dynamic task queues, real-time priority weights, and adaptive load scheduling to distribute resources according to the request stream processing order and popularity, thereby improving the response speed of frequently accessed data. Representative technologies include distributed fingerprint databases, local indexing of content fragments, dynamic sorting of concurrent queues, and cache hit prediction, which are widely used in teaching scenarios such as educational cloud platforms, online classrooms, and smart campuses. These technical solutions can effectively support resource deduplication and access acceleration, but they largely rely on byte-level feature comparison or external prediction and scheduling logic.

[0004] However, in multimedia teaching cloud storage applications, existing fingerprint comparison processes generally suffer from the following technical defects: First, current fingerprint comparison adopts a uniform processing strategy, performing full fingerprint generation and comparison in a fixed process regardless of resource popularity or teaching content semantics, failing to fully reflect the dynamic changes in the business attention of different resources in teaching scenarios; Second, the use of global hashing or machine learning-based priority control results in high computational load and high resource consumption of real-time inference algorithms, making it difficult to quickly adapt to peak access periods, leading to both delayed response to popular resources and wasted computation for unpopular resources; Third, existing hot and cold queues or weight function models essentially treat resource popularity as an external auxiliary scheduling parameter rather than an embedded structural attribute of the content fingerprint, causing the fingerprint generation and scheduling processes to be separated, making it difficult to achieve accurate content layered screening.

[0005] In response to the aforementioned technological status quo and problems, there is an urgent need in this field for a technical solution that can directly embed the semantic popularity of teaching resources into a fingerprint structure, enabling dynamic hierarchical comparison at the resource level. This requirement aims to break down the disconnect between traditional fingerprint comparison and scheduling collaboration. By proactively embedding content semantic features into the fingerprint generation process, the system can automatically identify and prioritize popular courses and key knowledge segments in high-concurrency upload scenarios. This achieves rapid popularity determination, segment reuse, and accurate deduplication, reducing resource waste and ensuring the responsiveness of important teaching content. Consequently, it can significantly improve the data processing capabilities and resource utilization of cloud storage platforms, better meeting the needs of rapid updates and efficient retrieval in teaching scenarios, laying the foundation for subsequent invention points and technical effects. Summary of the Invention

[0006] This application provides a cloud storage data deduplication and access acceleration method for multimedia teaching, aiming to solve one of the problems or issues of the existing technology mentioned in the background.

[0007] The cloud storage data deduplication and access acceleration method for multimedia teaching provided in this application specifically includes: S1: Obtain the original teaching metadata from the multimedia teaching resource upload request. The original teaching metadata includes the courseware title text, the video automatic speech recognition and transcription result, and the teacher's speech waveform signal. S2: Perform multi-dimensional semantic feature parsing on the original teaching metadata, calculate the inverse document frequency weight of the courseware title keywords, the frequency concentration of teaching terms on the time axis, and the statistical regularity of semantic pause intervals in the speech waveform, and generate a normalized one-dimensional semantic feature vector. S3: Perform a weighted concatenation operation based on the normalized one-dimensional semantic feature vector to generate a semantic heat fingerprint with comparable order; S4: Dynamically set the heat threshold based on the current cloud storage system load status, and input the semantic heat fingerprint into the heat threshold filter for comparison and judgment. If the semantic heat fingerprint is lower than the heat threshold, the subsequent process is terminated; if it is higher, a high heat admission instruction is generated. S5: Establish a distributed fingerprint database including a fragment fingerprint index table. Based on the high-heat admission instruction, perform local matching of the semantic heat fingerprint according to the fragment fingerprint index table. If the matching is successful, generate a fragment reuse identifier. S6: For high-popularity admission instructions that have not generated the fragment reuse identifier, use a hash algorithm to generate the final content fingerprint, and perform a precise comparison operation in the distributed fingerprint database to generate a precise deduplication determination result.

[0008] S7: Collect cloud storage data from multimedia teaching resources, update the inverse document frequency weight based on the cloud storage data, and generate adaptively evolving semantic heat fingerprint generation parameters. S8: Synchronize the adaptively evolved semantic heat fingerprint generation parameters to the edge node configuration library, and replace the original inverse document frequency weight to complete the dynamic iterative optimization of the semantic heat fingerprint generation logic.

[0009] The cloud storage data deduplication and access acceleration method for multimedia teaching provided in this application has the following beneficial effects: (1) To address the problems in existing technologies where teaching resource fingerprint comparison generally relies on external scheduling parameters and requires frequent full-content hash calculations, resulting in high response latency and large system resource overhead in high-concurrency scenarios, this solution proposes a technical path that directly internalizes teaching business semantics into fingerprint structural features. By generating a "semantic heat fingerprint," it achieves an integrated expression of content perception and heat judgment. This fingerprint is not used for precise deduplication, but rather generates a normalized vector with comparable order based on multimodal signals such as courseware title hierarchy, keyword weight distribution, speech-to-text term frequency, and semantic pause patterns, enabling resources to complete a coarse heat screening during the access phase. Compared to the traditional method that requires a serial process of first extracting a complete fingerprint and then prioritizing it, this solution introduces a semantic-driven pre-filtering mechanism at the source, so that a large number of low-heat resources do not need to enter the subsequent complex calculation stage, significantly reducing the invalid computation load and greatly improving the system throughput and response sensitivity. It is especially suitable for typical scenarios where multimedia teaching resources are continuously uploaded at high frequency.

[0010] (2) To further optimize comparison efficiency and reduce redundant calculations, this solution designs a three-level hierarchical comparison engine, realizing for the first time a progressive decision-making logic from "whether to compare" to "how to compare" and then to "whether it can be reused". The first-level heat threshold filter dynamically adjusts the admission threshold according to the real-time system load, ensuring that core computing resources are always focused on high-value teaching content; the second-level segment fingerprint index reuse mechanism establishes a local fingerprint index table for typical knowledge point segments in popular courses. After semantic heat matching, new resources can directly query existing segment results, avoiding repeated time-consuming full-file hash operations on the same teaching content; the last level only triggers standard full comparison for a few resources that do not match the first two levels, greatly compressing the actual scale of data involved in accurate deduplication. This hierarchical architecture effectively breaks the rigid process constraint of "every resource must be compared" in traditional solutions. Under the premise of ensuring deduplication accuracy, more than 90% of upload requests can be processed in milliseconds, truly realizing positive synergy between resource heat and processing efficiency, and generating a virtuous feedback mechanism of "the more popular, the faster; the less popular, the more economical".

[0011] (3) To improve the system's adaptability and timeliness, this solution also introduces an online evolution mechanism for semantic fingerprints and a stateless edge deployment mode. By continuously collecting real teaching interaction data such as key markings and reusable behaviors on the teacher's end and peak review times and concentrated note areas on the student's end, the weighting coefficients of each semantic dimension are regularly updated to ensure that the semantic heat fingerprint can dynamically adapt to the changing trends of teaching practice, thereby enhancing the accuracy of recommendation and management. At the same time, all semantic fingerprint generation and hierarchical judgment are completed locally on the edge nodes, and only the summary information that needs to be entered into the database for comparison is synchronized to the central database, which greatly reduces the network transmission pressure and the computational burden of the central node, and supports horizontal expansion and high availability deployment. The overall technical path abandons the heavy control logic that relies on multi-dimensional monitoring modeling, dynamic queue scheduling and complex weight fitting, and instead reconstructs the fingerprint ontology in a lightweight and structured semantic embedding manner. Without sacrificing accuracy, it achieves a system-level synergistic improvement in comparison efficiency, resource utilization and system scalability.

[0012] The aforementioned technical means jointly generate a lightweight and efficient fingerprint scheduling system oriented towards teaching semantics. This not only significantly improves the speed and stability of resource processing in high-concurrency environments, but also achieves the organic unity of business needs and underlying mechanisms through the deep integration of semantics and structure. It has good interpretability and engineering implementation value, and is particularly suitable for application scenarios in campus and off-campus teaching resource sharing platforms that have stringent requirements for real-time response, resource diversity and system economy. Attached Figure Description

[0013] Figure 1 This is the main flowchart of a cloud storage data deduplication and access acceleration method for multimedia teaching.

[0014] Figure 2 This is a sub-flowchart of cloud storage data deduplication and access acceleration methods for multimedia teaching.

[0015] Figure 3 This is another sub-flowchart for cloud storage data deduplication and access acceleration methods for multimedia teaching. Detailed Implementation

[0016] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0017] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0018] like Figure 1 As shown, this application provides a cloud storage data deduplication and access acceleration method for multimedia teaching, specifically including: S1: Obtain the original teaching metadata from the multimedia teaching resource upload request. The original teaching metadata includes the courseware title text, the video automatic speech recognition and transcription result, and the teacher's speech waveform signal. S2: Perform multi-dimensional semantic feature parsing on the original teaching metadata, calculate the inverse document frequency weight of the courseware title keywords, the frequency concentration of teaching terms on the time axis, and the statistical regularity of semantic pause intervals in the speech waveform, and generate a normalized one-dimensional semantic feature vector. S3: Perform a weighted concatenation operation based on the normalized one-dimensional semantic feature vector to generate a semantic heat fingerprint with comparable order; S4: Dynamically set the heat threshold based on the current cloud storage system load status, and input the semantic heat fingerprint into the heat threshold filter for comparison and judgment. If the semantic heat fingerprint is lower than the heat threshold, the subsequent process is terminated; if it is higher, a high heat admission instruction is generated. S5: Establish a distributed fingerprint database including a fragment fingerprint index table. Based on the high-heat admission instruction, perform local matching of the semantic heat fingerprint according to the fragment fingerprint index table. If the matching is successful, generate a fragment reuse identifier. S6: For high-popularity admission instructions that have not generated the fragment reuse identifier, use a hash algorithm to generate the final content fingerprint, and perform a precise comparison operation in the distributed fingerprint database to generate a precise deduplication determination result.

[0019] S7: Collect cloud storage data from multimedia teaching resources, update the inverse document frequency weight based on the cloud storage data, and generate adaptively evolving semantic heat fingerprint generation parameters. S8: Synchronize the adaptively evolved semantic heat fingerprint generation parameters to the edge node configuration library, and replace the original inverse document frequency weight to complete the dynamic iterative optimization of the semantic heat fingerprint generation logic.

[0020] Step S1: Obtain the original teaching metadata from the multimedia teaching resource upload request. The original teaching metadata includes the courseware title text, the video automatic speech recognition and transcription result, and the teacher's speech waveform signal. Specifically, it includes: S1.1: Perform protocol parsing on the multimedia teaching resource upload request to extract the original teaching metadata set containing the courseware title text string, video streaming data packet, and teacher's voice audio waveform signal.

[0021] For multimedia teaching resource upload request messages received by the access layer, the protocol identifier extraction operation is first performed at the transmission channel terminal. Based on the protocol version number and data type field in the message header, the protocol type of the upload request is determined to clarify the decoding rules and semantic mapping table required for subsequent parsing. Next, the multimodal data splitting function corresponding to the protocol type is called to extract the message payload into courseware title text field segments, video streaming media data segments, and teacher voice audio data segments according to preset segmentation boundaries. During the extraction process, a checksum verification mechanism is used to check the integrity and validity of each segment of data. Subsequently, character encoding format recognition and unification processing is performed on the courseware title text field segment. Different encoding formats (such as UTF-8, GBK, etc.) are uniformly converted to the internal standard encoding format to eliminate cross-platform display differences, and a unique resource identifier is assigned to this text segment to ensure fast indexing in subsequent feature parsing. Then, a container format parsing operation is performed on the video streaming media data segment to read streaming media metadata containing encoder type, frame rate, resolution, and timestamp information, and the original compressed data packet set is extracted as the input source for automatic speech recognition processing. Subsequently, the audio format header of the teacher's voice audio data segment is parsed to lock the sampling rate, sampling bit depth, number of channels, and timestamp information. The undecompressed waveform data stream is then extracted as the raw signal input for subsequent endpoint detection and silence removal. Through the above chain-like parsing process, the heterogeneous data segments in the upload request message are transformed into a raw teaching metadata set containing the courseware title text string, video streaming media data packets, and the teacher's voice audio waveform signal. This achieves decoupling and structuring of different modal data and provides a unified input condition for downstream automatic speech recognition, word segmentation weight labeling, and audio pause analysis.

[0022] For example, in a multimedia teaching resource upload scenario on an online education platform, the access layer receives an upload request message based on the WebSocket protocol. The protocol version number in the message header is 2.1, and the data type field is marked as a trimodal composite type. After determining the protocol type, the trimodal splitting function is called to segment the payload into a 128-byte courseware title segment, a continuous video frame stream segment, and an audio waveform segment. When extracting the courseware title segment, the encoding format is detected as GBK, which is uniformly converted to UTF-8 and a resource identifier "RES-202406150001" is generated. The video segment container format is parsed to show an encoder of H.264, a frame rate of 30fps, a resolution of 1920×1080, and a timestamp interval of 33ms. The continuous compressed video frame data packets are extracted as input for the subsequent ASR model. The audio segment format header is parsed to show a sampling rate of 44.1kHz, a bit depth of 16bit, 1 channel, and a timestamp precision of 10ms. The original PCM waveform stream is extracted for subsequent pause analysis. In the data integrity verification, the CRC check value of all data segments matches the attached checksum to ensure no transmission errors. The raw teaching metadata set processed in this way includes a UTF-8 encoded courseware title text string, a structured set of video streaming data packets, and a structured audio waveform signal stream. This provides a stable and high-quality input source for subsequent speech recognition, keyword segmentation, and pause interval analysis, significantly improving the parsing accuracy and execution efficiency of the subsequent feature extraction module.

[0023] S1.2: Perform automatic speech recognition and conversion processing based on the video streaming data packets in the original teaching metadata set to generate a sequence of transcribed teaching content with timestamp alignment information.

[0024] Based on the video streaming data packets in the original teaching metadata set output from the preceding steps, the automatic speech recognition access module is invoked to demultiplex the compressed video data packets to separate the audio track signals, generating the original speech data stream for subsequent recognition. Sampling rate recalibration and quantization accuracy standardization are performed on the original speech data stream to ensure that the input signal is consistent with the training parameters of the speech recognition acoustic model, obtaining a standardized audio signal sequence. Endpoint detection and segmentation are performed on the standardized audio signal sequence, dividing the continuous speech into a set of independent speech segments containing start and end timestamps, laying the segmentation foundation for the generation of time alignment information. Feature extraction processing is performed on the speech segment set, using various acoustic features such as Mel-frequency cepstral coefficients (MFCC), short-time energy, and harmonic structure to construct a high-dimensional feature vector matrix, which serves as the input to the acoustic model. The high-dimensional feature vector matrix is ​​input into a pre-set combination structure of the multimedia teaching domain automatic speech recognition acoustic model and language model, generating corresponding text transcription results and appending corresponding timestamp information next to each word or phrase, generating a sequence of transcribed teaching content with timestamp alignment information. Through the above-mentioned automatic speech recognition and conversion process, the video streaming data packets extracted in the previous step are transformed into a structured text sequence of teaching content with time indexing capabilities, thereby achieving the technical effects of time axis clustering and frequency concentration calculation in subsequent semantic feature analysis.

[0025] For example, in the multimedia teaching video data packet input scenario, the video encoding format is H.264, the audio encoding format is AAC, the audio sampling rate is configured to 16000Hz, and the quantization precision is 16bit. After performing demultiplexing, the length of the output raw speech stream is 1800 seconds. During endpoint detection, the energy threshold is set to 0.02, the silence duration threshold is set to 500ms, and approximately 450 speech segments are obtained after segmentation. MFCC feature extraction is performed on each speech segment, with a filter bank of 40, cepstral coefficients of 13 dimensions, and first-order and second-order difference coefficients are calculated to generate a 39-dimensional feature vector. In the acoustic model recognition stage, a deep neural network (DNN) based acoustic model is used, with 1024 hidden layer neurons and 6 layers; the language model is a ternary grammar model, and the training corpus covers 100,000 teaching sentences. In the recognized output transcribed text sequence of teaching content, each term is appended with start and end timestamps accurate to 0.01 seconds. For example, the identified term "Newton's Second Law" has a start timestamp of 352.15 seconds and an end timestamp of 355.42 seconds in the sequence. This time alignment information significantly improves the accuracy of subsequent calculations of the frequency concentration of teaching terms.

[0026] S1.3: The courseware title text string is segmented and weighted using a pre-set domain keyword dictionary to output a structured title feature vector carrying inverse document frequency feature values.

[0027] The input data is the courseware title text string extracted through protocol parsing. It needs to be segmented and weighted using a pre-defined domain keyword dictionary to output a structured title feature vector carrying inverse document frequency (IDF) features. The courseware title text string is scanned using a domain keyword dictionary, and the matched keywords are used as segmentation boundaries to generate a title word sequence segmented by semantic units. For this title word sequence, a character regularization module is called to unify word forms and eliminate punctuation and capitalization differences to ensure consistency in subsequent calculations. The word frequency statistics and inverse document frequency (IDF) reference values ​​are searched one by one in the domain keyword dictionary for the regularized word sequence, constructing a word frequency mapping table containing word indices and initial IDF values. The TF-IDF algorithm is used to calculate the weight w for each word, as shown in the following formula: Here, TF is the number of times the word appears in the title divided by the total number of words in the title, and IDF is the inverse document frequency (IDF) value in the pre-built dictionary. The calculated weight values ​​are bound to the corresponding word indices to generate a preliminary word vector set. Vector length standardization is performed on the preliminary word vector set to ensure that feature vectors under different title lengths are comparable on the same numerical scale. The standardized word vector set is then positionally encoded according to the original word order to generate a structured title feature vector carrying the IDF feature value. Through the above processing, the title text string from the previous step is transformed into a quantitative indicator that can be directly used for multi-dimensional semantic feature analysis, achieving accurate weight expression of the courseware title text in semantic popularity fingerprint modeling.

[0028] S1.4: Perform endpoint detection and silence segment removal processing on the teacher's voice audio waveform signal to obtain a set of clean voice waveform segments that retain effective semantic pause interval information.

[0029] S1.5: Perform spatiotemporal alignment encapsulation processing on the structured title feature vector, the transcribed text sequence of teaching content, and the set of clean speech waveform fragments to generate a standardized multidimensional original teaching metadata object for subsequent feature parsing.

[0030] Step S2: Perform multi-dimensional semantic feature parsing on the original teaching metadata, calculating the inverse document frequency weight of courseware title keywords, the frequency concentration of teaching terms on the time axis, and the statistical regularity of semantic pause intervals in the speech waveform, to generate a normalized one-dimensional semantic feature vector. Specifically, this includes: S2.1: Perform word segmentation and inverse document frequency weight calculation on the courseware title text to extract the inverse document frequency weight of the courseware title keywords. This distribution quantifies the distinguishability and importance of each word in the title in the set of teaching resources, serving as the initial expression of the text semantic features.

[0031] The structured title feature vector generated by step S1.3 is read and used as the initial input object for text semantic feature parsing. A pre-set domain keyword dictionary is loaded and a word frequency mapping table from keywords to the global document set is established.

[0032] Based on the mapping table, precise matching and word segmentation are performed to split the title text into independent terms and mark the number of times each term appears in the document collection and the number of times it appears in the title, generating a term statistics matrix.

[0033] The inverse document frequency calculation module is called on the term statistics matrix to calculate the inverse document frequency value for each term. The calculation is completed using the following formula: in, This represents the total number of documents in the teaching resource collection. The ratio of the number of title documents containing the term to the number of titles used to characterize the rarity of the term.

[0034] The inverse document frequency value is multiplied by the word frequency of the term obtained from word segmentation in the current title to generate a sequence of TF-IDF weight values ​​for the term, which serves as the quantitative result of the importance of the title keywords.

[0035] Normalize the TF-IDF weight value sequence to map all weights to a uniform interval to eliminate scale differences caused by different title lengths and term densities, and output the inverse document frequency weight vector of courseware title keywords.

[0036] Through the above processing method, the structured title feature vector of the previous step is transformed into data indicators that can quantify the distinguishability and importance of words in the entire set of teaching resources, thereby achieving the initial accurate expression of the semantic features of the text.

[0037] S2.2: Based on the video automatic speech recognition and transcription results, perform teaching term recognition and temporal clustering analysis to generate the concentration of the frequency of teaching terms on the time axis. This concentration reflects the distribution density and clustering characteristics of core knowledge points in the video stream, serving as a key indicator of the semantics of the video content.

[0038] Based on the sequence of transcribed teaching content and its corresponding timestamp information output by the automatic speech recognition and transcription module, domain terminology dictionary matching and location index extraction are performed on the transcribed text to generate a set of term entries and their location coordinates on the global timeline.

[0039] Perform time series aggregation on the set of location coordinates, calculate the number of occurrences of each term on the time axis and the interval between adjacent occurrences, and generate a term distribution time series matrix as the input source for subsequent cluster analysis.

[0040] A time-series similarity measurement method based on dynamic time warping (DTW) is adopted to perform multiple distance calculations on the term distribution time series matrix, and the occurrence records of terms with high similarity and physical time intervals less than a preset threshold are divided into the same time cluster.

[0041] The concentration index is calculated using the following formula, taking into account the mean and standard deviation of the frequency of clustering terms at each time point: Where σ is the standard deviation of the frequency of occurrence of the term within the cluster.

[0042] The concentration vector of the frequency of occurrence of teaching terms over time is generated based on the concentration index. The difference in dimensions caused by different cluster sizes is eliminated by the maximum value normalization process, so as to ensure that the concentration vector can be directly used in the fusion of multimodal semantic features.

[0043] Through the aforementioned dynamic temporal clustering and centralized quantification processing methods, the sequence of teaching terms transcribed by automatic speech recognition is transformed into measurable temporal distribution density data, achieving a precise mapping from speech content to video semantic structure features.

[0044] S2.3: Zero-crossing rate detection and short-time energy envelope analysis are performed using the teacher's speech waveform signal to statistically analyze the semantic pause intervals in the speech waveform. This statistical regularity characterizes the logical breakpoints and emphasis intervals in the teacher's teaching rhythm, serving as a quantitative basis for audio behavioral semantics.

[0045] S2.4: Perform multimodal feature fusion and maximum / minimum value normalization on the inverse document frequency weight of the courseware title keywords, the concentration of the frequency of teaching terms on the time axis, and the statistical regularity of semantic pause intervals in the speech waveform to generate a normalized one-dimensional semantic feature vector. This vector eliminates the dimensional differences of different modal data and unifies the feature space, directly supporting the subsequent weighted splicing operation of semantic heat fingerprint.

[0046] like Figure 2 As shown, step S3 involves performing a weighted concatenation operation based on the normalized one-dimensional semantic feature vector to generate a semantic heat fingerprint with comparable order. Specifically, this includes: S3.1: Obtain the inverse document frequency weight vector of courseware title keywords, the frequency concentration vector of teaching terms on the time axis, and the statistical regularity vector of semantic pause intervals in speech waveforms output from the previous steps. Perform dimension alignment processing on the three normalized one-dimensional semantic feature vectors to eliminate the dimensional differences of different modal data in the feature space and generate a standardized multimodal semantic feature matrix.

[0047] The input conditions are three types of normalized one-dimensional semantic feature vectors output by the previous step S2.4, namely, the inverse document frequency weight vector of courseware title keywords, the frequency concentration vector of teaching terms on the time axis, and the statistical regularity vector of semantic pause intervals in speech waveforms. These vectors have eliminated the differences in their internal dimensions during the original generation process, but there are still differences in dimensionality, resolution, and index space when comparing across modalities.

[0048] For three types of normalized one-dimensional semantic feature vectors, a dimension benchmark acquisition process is performed. By scanning the dimension length, value range and index step size of each type of vector, a unified feature alignment configuration table is generated, providing a clear benchmark for subsequent mapping according to the same coordinate system.

[0049] Interpolation completion is performed based on the feature alignment configuration table. For vectors with a dimension length lower than the global benchmark, linear interpolation or spline interpolation methods are used to increase the number of sampling points, so that each feature position of the same modality has a directly comparable correspondence in the vector index.

[0050] The interpolated vectors are subjected to scale normalization mapping, and the standardization function is called to unify the value range of each modality to the closed interval [0,1], so as to ensure that the comparison of the numerical values ​​of different modal data is not affected by the difference in units.

[0051] An index mapping matrix-based rearrangement algorithm is adopted to align the normalized vectors of different modalities by feature index position, generating a standardized multimodal semantic feature matrix in which rows represent modalities and columns represent the same feature position, thereby achieving cross-modal spatial alignment and data comparability.

[0052] Through the above processing method, the three types of normalized one-dimensional semantic feature vectors from the previous step are transformed into standardized multimodal semantic feature matrices with unified dimension, unified value range, and unified index structure, realizing the alignment of cross-modal semantic features and providing a consistent input basis for subsequent weighted mapping operations.

[0053] S3.2: Read the adaptive evolution weight coefficient set stored in the edge node configuration library. The adaptive evolution weight coefficient set includes text keyword weight factors, terminology time sequence weight factors, and speech pause weight factors. Based on the adaptive evolution weight coefficient set, perform element-wise weighted mapping operation on the standardized multimodal semantic feature matrix to quantify the contribution of each semantic dimension to the overall business popularity and generate a weighted semantic feature component sequence.

[0054] The edge node configuration library read interface is called to obtain the current version of the adaptive evolution weight coefficient set for the standardized multimodal semantic feature matrix, which is then used as the control parameter input for this step.

[0055] The adaptive evolution weight coefficient set is structured and parsed, and the text keyword weight factor, term temporal weight factor and speech pause weight factor are mapped to the corresponding feature dimensions of the matrix, respectively. An element operation correspondence table is established based on the dimension index.

[0056] Element-wise operations are performed on the standardized multimodal semantic feature matrix. The weighted product mapping method is used to multiply each matrix element with its corresponding weight factor to generate a weighted feature matrix. This process maintains the consistency of the position between elements under the condition of matrix row and column index alignment.

[0057] For each column of the weighted feature matrix, the feature contribution is calculated, and the relative influence of that column in the overall popularity index is quantified using the normalized cumulative sum method, calculated by the following formula: Where C is the contribution index, x is the feature value, and w is the corresponding weight.

[0058] The contribution indexes calculated from each column are output as a weighted semantic feature component sequence in the order of the matrix columns, ensuring that the sequence has sortability and numerical stability in the subsequent concatenation of heat vectors.

[0059] By using element-wise weighted mapping and contribution calculation, the standardized multimodal semantic feature matrix obtained in the previous step is transformed into a sequence of weighted semantic feature components that can quantitatively reflect the contribution ratio of each semantic dimension to the overall business popularity, thus realizing the structured mapping from multimodal feature vectors to popularity representation components.

[0060] S3.3: Perform a serial splicing and aggregation operation on the weighted semantic feature component sequence. According to the preset fixed arrangement order of text first, time sequence second, and speech last, integrate the discrete weighted semantic feature component sequence into a continuous high-dimensional original heat vector to generate an initial data structure that can completely represent the multi-dimensional semantic attributes of teaching resources and generate a high-dimensional original heat vector.

[0061] Based on the weighted semantic feature component sequence generated in the previous step, a preset multimodal concatenation control strategy is invoked to serialize and organize the components according to a fixed order of text first, time sequence second, and speech last, ensuring the consistency of the order of features in the storage structure. Data type unification processing is performed on each serialized component, converting numerical components to floating-point format and retaining decimal precision to four decimal places to eliminate encoding differences caused by different calculation paths. An index mapping table is used to assign a unique position identifier to each component. A continuous feature record stream is generated by combining the position identifier and the component value, thus supporting fast access and position locking during subsequent projection transformations. After the feature record stream is generated, a continuous memory block encapsulation operation is performed, using a high-dimensional array structure to transform discrete components into a continuous storage layout, ensuring that all feature components can be loaded at once in matrix operations, reducing the overhead of multiple memory accesses. Through the above processing, the result of the previous step is transformed into a high-dimensional original heat vector data structure, achieving a complete representation of the multidimensional semantic attributes of teaching resources and supporting comparability.

[0062] S3.4: The high-dimensional original heat vector is subjected to feature compression processing using a linear dimensionality reduction algorithm. While preserving the relative distance relationship between vectors, i.e., the monotonicity of Euclidean distance, the high-dimensional original heat vector is mapped to a low-dimensional dense vector space to reduce the computational complexity of subsequent comparison processes and improve storage efficiency, thereby generating low-dimensional dense heat candidate vectors.

[0063] The object of feature compression processing on the high-dimensional original heat vector generated in the preceding step S3.3 is a continuous high-dimensional vector structure composed of three weighted semantic feature components of text, time sequence, and speech with a fixed arrangement order. For this high-dimensional original heat vector, mean centering is first performed on each component to eliminate the offset of each feature in terms of numerical scale, thereby ensuring that the covariance matrix in the dimensionality reduction process can accurately reflect the correlation between features. Then, variance standardization is performed on the centered vector to scale the standard deviation of each component to a uniform numerical range, so as to avoid the high variance component causing a dominant bias to the low variance component during the projection process. Based on the standardization result, a feature covariance matrix is ​​generated, and its eigenvector set is obtained through eigenvalue decomposition. The eigenvectors are sorted by the size of the eigenvalues ​​to determine the descending order of feature contribution. The eigenvectors corresponding to the first few eigenvalues ​​are selected to form the projection transformation matrix. The number of selected eigenvectors is determined based on the constraint of preserving the monotonicity of Euclidean distance and the preset cumulative contribution rate threshold. Matrix multiplication is performed between the original high-dimensional heat vector and the projection transformation matrix to achieve a linear mapping from the high-dimensional space to the low-dimensional dense vector space. Simultaneously, the monotonic correspondence between the Euclidean distances between the mapped candidate vectors and the original high-dimensional vectors is verified to ensure the stability of the alignment and ranking. Through this dimensionality reduction projection process, the results of the previous step are transformed into low-dimensional dense heat candidate vectors, significantly reducing the computational complexity and improving storage efficiency of subsequent alignment processes.

[0064] S3.5: Perform normalized scalar encoding processing on the low-dimensional dense heat candidate vector, map the vector values ​​to a standardized sortable interval, and give it a clear numerical comparison attribute, thereby completing the transformation from multi-dimensional semantic features to a single business heat index, and finally generating a semantic heat fingerprint with comparable order.

[0065] like Figure 3 As shown, step S4 involves dynamically setting a heat threshold based on the current cloud storage system load status, and inputting the semantic heat fingerprint into the heat threshold filter for comparison. If the semantic heat fingerprint is lower than the heat threshold, the subsequent process is terminated; if it is higher, a high-heat admission instruction is generated. Specifically, this includes: S4.1: Perform multi-dimensional indicator collection and processing on the monitoring data stream of the current operating status of the cloud storage system to extract a real-time load feature vector containing the CPU utilization, memory usage, and network bandwidth throughput of the central node. This real-time load feature vector quantifies the current resource stress of the system and serves as the benchmark input data for dynamic threshold calculation.

[0066] The continuous data stream from the cloud storage system operation status monitoring module is divided into time windows and processed by signal separation. The index data, including CPU utilization, memory usage and network bandwidth throughput, are extracted into independent measurement sequences.

[0067] Sampling noise is filtered for each measurement sequence, and abnormal peak values ​​are removed using median filtering to ensure the stability of subsequent feature calculations.

[0068] Peak interval identification and average value calculation are performed on the CPU utilization measurement sequence. The average value is recorded as the CPU load parameter, and the fluctuation characteristics of processor resources are evaluated in combination with variance statistics.

[0069] The memory usage measurement sequence is normalized using the maximum and minimum values, and the normalized memory usage metric is calculated using the following formula: in, This is the current memory usage measurement. and These represent the maximum and minimum values ​​within the statistics window, respectively.

[0070] Perform a unit-time traffic integration operation on the network bandwidth throughput measurement sequence to obtain the total bandwidth usage, and combine it with the maximum theoretical bandwidth value to calculate the network load percentage.

[0071] The CPU load parameters, normalized memory usage indicators, and network load percentages mentioned above are combined in a preset order to generate a real-time load feature vector, which serves as the baseline input for dynamic calculation of the heat threshold.

[0072] By using a chained processing approach, the semantic heat fingerprint generation result from the previous stage is linked with the real-time load feature vector to establish a data association, thereby enabling the quantitative support of the operating status required for subsequent adaptive threshold setting.

[0073] S4.2: Based on the real-time load feature vector, perform nonlinear mapping function operation to generate a dynamic adaptive heat threshold. This dynamic adaptive heat threshold establishes an inverse adjustment relationship between system load pressure and resource access standards, ensuring that the access threshold is automatically raised under high load conditions to protect core computing power.

[0074] Based on the real-time load feature vector output by the preceding step S4.1, normalization mapping is performed on each load indicator to transform the original percentage values ​​of CPU utilization, memory usage and network bandwidth throughput to a standard range of 0 to 1, so as to eliminate the interference of different units on subsequent calculations.

[0075] The normalized multidimensional load feature vector is processed by calling the weight allocation module. Based on the resource bottleneck sensitivity coefficient preset during system architecture design, the weighted contribution ratio of each indicator is calculated, and a comprehensive load pressure indicator vector for threshold calculation is generated.

[0076] The dynamically adaptive heat threshold obtained by mapping is subjected to interval pruning processing with the absolute minimum and maximum threshold boundary values ​​set by the business to ensure that the threshold does not exceed the upper and lower limits that the system can withstand under extreme load conditions.

[0077] Through the above chained processing method, the real-time load feature vector is transformed into a dynamic adaptive heat threshold with reverse adjustment characteristics, so as to achieve the expected technical effect of automatically adjusting the admission criteria according to the system resource pressure to protect the core computing power.

[0078] S4.3: Perform a scalar numerical comparison operation between the semantic heat fingerprint with comparability generated in the previous steps and the dynamic adaptive heat threshold, and generate a binary heat judgment result signal. This binary heat judgment result signal clarifies whether the current teaching resource meets the business heat conditions for entering the deep comparison process.

[0079] The input semantic heat fingerprint with comparable order and the dynamic adaptive heat threshold are subjected to scalar numerical comparison processing, and the normalized business heat fingerprint value output from the previous step is read into the high-speed register buffer unit.

[0080] A precision consistency calibration process is performed on the dynamic adaptive heat threshold, mapping the threshold to the same numerical scale range as the heat fingerprint, thus eliminating scale mismatch caused by different calculation paths.

[0081] The calculated decision signal is processed by logical encoding, and the decision result is mapped to a binary status code of a high-heat admission branch or a low-heat buffer branch, generating decision data that can directly drive the downstream process control.

[0082] The generated binary status code is written into the flow control register buffer and the status broadcast mechanism is triggered, so that the two processing paths of high-heat admission and low-heat cache can receive the judgment information at the same time, thereby realizing the parallel preparation of scheduling instructions for subsequent processing modules.

[0083] By using scalar comparison and logical encoding, the semantic heat fingerprint results from the previous step are transformed into clear judgment signals, enabling real-time decision-making on whether teaching resources should enter the deep comparison process.

[0084] S4.4: Based on the termination of the execution process and the mark encapsulation process of the negative judgment branch in the binary heat judgment result signal, a low heat pending cache mark carrying timestamp information is generated. The low heat pending cache mark indicates that the system will directly transfer the corresponding teaching resources into the asynchronous cache queue and skip the subsequent fragment index query and full hash comparison.

[0085] S4.5: Based on the positive decision branch in the binarized heat determination result signal, execute the admission instruction generation and routing distribution process to generate a high heat admission instruction carrying a unique request identifier. This high heat admission instruction triggers the downstream segment fingerprint index table query module to start a parallel retrieval mechanism to reuse historical deduplication results.

[0086] The input conditions include the binarized heat determination result signal obtained through sub-step S4.3, the positive determination branch of which indicates that the current teaching resource has the business heat conditions to enter the deep comparison process, the dynamic adaptive heat threshold calculated in sub-step S4.2, and the semantic heat fingerprint with comparability generated in sub-step S3.5.

[0087] Based on the positive judgment result signal, the unique judgment status bit and the associated heat value index field are parsed to generate the original information carrier of the admission instruction.

[0088] The edge node task scheduling interface is called using the original information carrier of the admission instruction to generate an instruction data object carrying a unique request identifier. This unique request identifier is generated by concatenating the timestamp and the resource hash digest bit by bit, ensuring that the downstream processing module can accurately associate with the corresponding resource.

[0089] The routing distribution path is calculated for the input instruction data object. Based on the fragment fingerprint index table in the distributed storage system and the network topology delay parameters, the optimal set of nodes is determined by the shortest path first and bandwidth utilization balancing method.

[0090] The path calculation result is embedded into the control field of the admission instruction data object, and the admission instruction is distributed to the fragment fingerprint index table query module of the target node through a distributed message queue transmission mechanism, triggering the start signal of the parallel retrieval mechanism.

[0091] Through the above chain-like derivation process, the positive judgment result signal is transformed into a high-heat admission instruction with a unique request identifier and optimal routing information, thereby enabling parallel retrieval triggering of the downstream segment fingerprint index table and rapid reuse of historical deduplication results.

[0092] Step S5: Establish a distributed fingerprint database including a fragment fingerprint index table. Based on the high-heat admission instruction, perform local matching of the semantic heat fingerprint according to the fragment fingerprint index table. If the match is successful, generate a fragment reuse identifier. This includes: The specific steps for establishing a fragment fingerprint index table are as follows: Step 1: Historical Popular Segment Data Screening and Preprocessing: Screen course resources uploaded in the system in the past that meet the "high popularity" standard, and extract typical teaching segments; extract semantic popularity fingerprint key feature vectors for each segment, and perform integrity verification (remove missing values ​​and outliers) to provide standardized data for subsequent index construction.

[0093] Step 2: Local retrieval key generation: For the preprocessed fragment fingerprint key feature vector, perform spatial hash mapping processing that is completely consistent with S5.1 (including spatial block mapping, local projection transformation, hash value generation, collision detection and incremental offset correction) to generate a unique local retrieval key value, which serves as the core data of the first-level index area of ​​the index table, ensuring that the key value of the resource to be processed can be accurately matched with the key value of the index table.

[0094] Step 3: Fingerprint cluster aggregation and metadata association: The fragment fingerprints that generate local search key values ​​are classified and aggregated according to "semantic popularity range + core knowledge unit" to form candidate similar fragment fingerprint clusters, which are stored in the secondary fingerprint cluster area; at the same time, the historical deduplication judgment metadata (rights confirmation status, timestamp, etc.) of the fragment are associated to complete the filling of the tertiary metadata association area, ensuring that each fingerprint cluster can correspond to reusable historical metadata.

[0095] Step 4: Index Structure Construction and Optimization: The index is constructed using a B+ tree structure, associating local search key values ​​with the storage offset of fingerprint clusters, dividing the storage into multiple partitions, each corresponding to a physical node, supporting multi-way concurrent lookups; at the same time, auxiliary indexes of semantic tags and timestamps are established to optimize the efficiency of pattern matching and intersection operations in S5.2; the index table is updated regularly (e.g., daily), adding the latest high-frequency fragment fingerprints and updating the reference count value to ensure the timeliness and completeness of the index table data.

[0096] The specific steps for establishing a distributed fingerprint database are as follows: Step 1: Distributed Cluster Deployment: Deploy a multi-node distributed storage cluster, dividing the storage into shards according to "semantic popularity intervals" (such as high, medium, and low popularity shards). Each shard is assigned an independent physical node, supporting concurrent read and write operations and data redundancy backup to ensure data security and query efficiency.

[0097] Step 2: Basic Data Entry: Completely enter the historical high-popularity teaching segment data selected in Step 1 (Index Table Creation) into the corresponding shard of the distributed cluster. This includes complete semantic popularity fingerprints of the segments, segment association information, deduplication history records, and data indexing. This ensures that each segment's data can be quickly retrieved through fingerprint summary and course ID, providing underlying data support for the index table.

[0098] Step 3: Link and bind with the fragment fingerprint index table: Establish a mapping relationship between the distributed fingerprint database and the fragment fingerprint index table, and synchronize the storage address and fragment information of each fragment fingerprint in the database to the third-level metadata association area of ​​the index table. This ensures that after S5.2 locates the candidate fingerprint cluster through the index table, it can quickly retrieve the complete fingerprint data and metadata from the distributed fingerprint database, achieving a seamless connection between "index addressing and data retrieval".

[0099] Step 4: Data Standardization and Validation: Standardize all data in the database to ensure the feature vector format, ownership status identifier, and timestamp format of the fragment fingerprint. Perform data validation regularly to remove abnormal data and supplement missing metadata to ensure data accuracy and avoid affecting matching and reuse performance.

[0100] Step 5: Dynamic Updates and Maintenance: Establish a dynamic update mechanism. When the system adds new high-popularity course segments, the deduplication status of historical segments changes, or the reference count is updated, the corresponding data in the distributed fingerprint database is updated synchronously, and the segment fingerprint index table is updated in conjunction with the database to ensure data consistency between the database and the index table. At the same time, data backup and cleanup are performed to delete expired and low-popularity segment data, optimize storage resources, and ensure the efficiency of S5 queries.

[0101] S5.1: Based on the high-popularity access instruction, extract the semantic popularity fingerprint key feature vector of the teaching resource to be processed, perform spatial hash mapping processing on the semantic popularity fingerprint key feature vector, and generate a local retrieval key value for fast location. The local retrieval key value serves as the sole addressing basis for accessing the fragment fingerprint index table.

[0102] Based on the semantic popularity fingerprint key feature vector of the teaching resources to be processed, which is input by the high popularity admission command, a multi-dimensional data integrity check is performed on the feature vector to remove missing values ​​and outliers to ensure the accuracy of subsequent mapping.

[0103] The verified feature vector is input into the preset spatial block mapping algorithm module, and the mapping unit index range is determined based on the Euclidean distance metric of the feature space.

[0104] A local projection transformation is performed on the mapping unit index range to compress the continuous numerical field to a fixed range in order to improve the stability of the hash operation.

[0105] The spatial hash function is invoked to perform hash value generation operations on each element of the feature vector after local projection transformation, resulting in the initial hash sequence.

[0106] The initial hash sequence is subjected to collision detection processing. A reversible shift correction is added to the hash values ​​that collide using the incremental offset method to ensure that a unique local retrieval key value is generated.

[0107] By using spatial hash mapping, the semantic heat fingerprint key feature vector from the previous step is transformed into a local retrieval key value with a single addressing attribute, thereby achieving the expected technical effect of quickly locating the fragment fingerprint index table.

[0108] This key is used to access the fragment fingerprint index table, which can locate a set of historical popular course fragments that are highly similar to the semantic features of the current resource with only one address, significantly improving query speed and reducing computational load.

[0109] S5.2: Using the local retrieval key value, perform a multi-way concurrent search operation in the pre-built fragment fingerprint index table to filter out candidate similar fragment fingerprint clusters from the stored set of local content fingerprints of typical teaching fragments of historical popular course resources. The candidate similar fragment fingerprint clusters contain historical fragment data that have potential semantic overlap with the current teaching resources.

[0110] Upon receiving the local retrieval key value generated in the previous step, this key value is used as the sole addressing input to access the pre-built fragment fingerprint index table. A multi-way search task queue is established through a distributed concurrent retrieval scheduling module to ensure that retrieval tasks start synchronously on different physical nodes and different storage partitions to reduce query latency. For each search task, the address offset of its storage partition is calculated based on the hash mapping result of the local retrieval key value, and this offset is used to quickly locate potential matching candidate regions from the stored set of local content fingerprints of typical teaching fragments of historical popular course resources. Pattern matching is performed on the local content fingerprint data within the candidate regions. By comparing the consistency between the local retrieval key value and the fingerprint data index field, fragment fingerprints that do not have the same semantic popularity range are eliminated, generating a preliminary candidate set. For each fragment fingerprint in the preliminary candidate set, an intersection operation is performed based on the semantic tags and timestamp distribution characteristics pre-stored in the index table. Fingerprint records that overlap with the current teaching resource semantic popularity fingerprint in terms of core knowledge units and temporal clustering are retained, and candidate similar fragment fingerprint clusters are output. Through the above multi-path concurrent search and layer-by-layer filtering process, the potential matching range corresponding to the local search key value is converged to historical fragment data with potential semantic overlap, thereby achieving efficient filtering and candidate aggregation in the local matching stage.

[0111] S5.3: Perform fine-grained vector similarity measurement calculation on each historical segment data in the candidate similar segment fingerprint cluster, and compare the calculated similarity value with the preset dynamic confidence threshold to identify the target matching segment fingerprint that meets the reuse conditions. The target matching segment fingerprint represents the high consistency between the current resource and the historical resource in the core teaching content.

[0112] S5.4: Based on the historical deduplication judgment result data associated with the target matching fragment fingerprint, perform metadata inheritance and mapping transformation processing, directly assign the ownership status of the historical resources to the currently uploaded teaching resources, and generate an initialized fragment reuse decision signal. The fragment reuse decision signal indicates that the current resource does not need to enter the subsequent full comparison process.

[0113] Based on the historical deduplication results associated with the target matching fragment fingerprint, a complete metadata record set containing historical resource ownership status, entry timestamp, citation count, and content fingerprint summary is received as the initial processing object. Field parsing and data type validation are performed on this metadata record set, converting the ownership status field into a system-wide universal binary authoritative status identifier, and confirming that its data integrity meets the assignment conditions. After standardizing the format of the ownership status, a mapping table is generated, associating the historical resource ownership status identifier with the unique request identifier of the currently uploaded teaching resource using key-value pairs, realizing the direct mapping and transfer of ownership information from historical fragments to new resources. Conflict detection is performed on the assignment of ownership status values, comparing the current resource's record status in the transaction log with the historical ownership status. If the consistency check passes, an initial fragment reuse decision signal is generated. Through the above metadata inheritance and mapping transformation processing, the historical ownership status corresponding to the target matching fragment fingerprint selected in the previous step is directly assigned to the currently uploaded resource, generating a fragment reuse decision signal that clearly indicates the current resource can skip the full comparison process, achieving the expected technical effect of quickly terminating the comparison process for high-frequency resources.

[0114] S5.5: Based on the fragment reuse decision signal, perform a standardized encapsulation operation to generate a fragment reuse identifier data packet carrying a unique traceability identifier, and write the fragment reuse identifier data packet into the system transaction log to complete the early termination of the deduplication process. The fragment reuse identifier data packet serves as the final output result and triggers the reference count update mechanism of the storage system.

[0115] Step S6: For high-popularity admission instructions that have not generated the fragment reuse identifier, a hash algorithm is used to generate the final content fingerprint, and a precise comparison operation is performed in the distributed fingerprint database to generate a precise deduplication determination result. Specifically, this includes: S6.1: Perform binary stream reassembly processing on the original teaching metadata set associated with high-popularity access instructions that have not generated fragment reuse identifiers, so as to restore the spatiotemporally aligned and encapsulated multidimensional original teaching metadata objects into a continuous byte sequence stream, which serves as the direct input source for the standard content definition hash algorithm.

[0116] For the original teaching metadata set associated with high-popularity access instructions that lack segment reuse identifiers, binary stream reassembly processing needs to be performed based on the structured information of multi-dimensional encapsulated objects to ensure the continuity and integrity of the input source for subsequent hash algorithms. This involves obtaining the multi-dimensional original teaching metadata object from the high-popularity access instructions, parsing the spatiotemporal alignment encapsulation format of the courseware title text, video transcription data, and audio waveform segments, and extracting the byte offset index and arrangement order information of each modality's data within the encapsulation. The byte sequence reassembly control module is then invoked to establish a modal data reassembly mapping table based on the offset index and arrangement order. This converts the courseware title text into binary segments of the corresponding encoding format, maps the video transcription data into encoded blocks containing timestamps and text, and converts audio waveform segments into continuous byte streams in PCM or a specified encoding format. Finally, a buffer queue serial splicing operation is performed on all modal binary segments, merging them into a single byte sequence stream according to the mapping table order, and performing byte alignment padding during the splicing process to eliminate storage misalignments caused by varying lengths across modalities. Integrity verification is performed on the serially concatenated byte sequence stream. A CRC checksum is calculated and compared with a preset verification benchmark to ensure that no byte ordering or omissions are introduced during the reassembly process. The continuous byte sequence stream that passes integrity verification is encapsulated into a standard content-defined hash algorithm input object for subsequent full feature mapping operations. Through modal data binary reassembly and continuation processing based on an offset mapping table, the multi-dimensional encapsulated object from the previous step is transformed into a structurally consistent, non-cross-modal misaligned continuous byte sequence stream, achieving format uniformity and data integrity assurance for the full hash input source.

[0117] S6.2: Based on the continuous byte sequence stream, call the standard content definition hash algorithm to perform a full feature mapping operation to calculate a final content fingerprint digest with a fixed length and anti-collision properties. This digest uniquely represents the complete binary content features of the current teaching resource.

[0118] S6.3: Generate a distributed query request packet using the final content fingerprint digest, and distribute the distributed query request packet to the corresponding distributed fingerprint database storage node through a consistent hash routing strategy to locate a specific physical shard area that stores the historical full fingerprint data.

[0119] For the final content fingerprint digest generated in the previous steps, structured encapsulation processing is performed based on the fixed-length binary feature value of the digest to generate a distributed query request packet with a distributed addressing field, a digest verification field, and a load metadata field. A consistent hash calculation is performed on the addressing field of the distributed query request packet, mapping the digest value to a unique location point on the hash ring, serving as the location basis for the distributed fingerprint database node. Node allocation processing is performed using the location point on the consistent hash ring, selecting the first storage node identifier that meets the load balancing condition according to the clockwise traversal rule of the hash ring, and embedding this identifier into the routing information field of the query request packet. Distributed routing transmission control is performed on the routing information field of the query request packet, distributing the packet data to the receiving buffer of the corresponding distributed fingerprint database storage node via a dedicated data channel. The fragmentation positioning mechanism of this receiving buffer is triggered, accurately locating the specific physical fragment region storing the historical full fingerprint data based on the digest value and fragmentation index table in the request packet, to support the subsequent bit-by-bit precise matching and comparison. By employing a consistent hashing routing strategy and a sharding location mechanism, the final fingerprint digest from the previous step is transformed into a query request packet with target node addressing and sharding location attributes, thereby enabling high-speed location and load balancing of the fingerprint digest in the distributed fingerprint database.

[0120] S6.4: Perform a bit-by-bit precise matching and comparison operation on the final content fingerprint digest within the specific physical fragmentation area to determine whether the current resource has a completely consistent bit sequence with the resources already stored in the library, thereby generating a precise deduplication determination result signal containing the existence or uniqueness status.

[0121] Within a specific physical partition area located in the distributed fingerprint database storage node, the final content fingerprint digest generated by the preceding step S6.3 is used as the comparison object. All historical full fingerprint data within the physical partition area is loaded and bit-by-bit data reading processing is performed to generate a comparison data block sequence of the same length as the current fingerprint digest.

[0122] The XOR operation is performed on each bit in the data block sequence, the bit difference count value of the XOR output is recorded, and the bit difference count value is accumulated to generate a full difference metric, which serves as the basic metric parameter for determining content consistency.

[0123] The positive or negative determination signal is encapsulated into a precise deduplication determination result signal data packet, and a comparison timestamp and fragment region identifier are attached for S6.5 to perform subsequent storage status update processing.

[0124] By precisely matching and comparing each bit and calculating the bit difference ratio, the final content fingerprint digest of the previous step is transformed into a precise deduplication result with clear existence or uniqueness attributes, thus achieving the technical effect of ensuring deduplication accuracy in high-concurrency scenarios.

[0125] S6.5: Based on the accurate deduplication determination result signal, perform storage state update processing. If it is determined to be a duplicate, generate a reference count increment instruction and return the existing resource pointer. If it is determined to be a new resource, generate an inbound write instruction and persist the final content fingerprint digest to the distributed fingerprint database, thus completing the final closed loop of the deduplication process.

[0126] Step S7: Collect cloud storage data from multimedia teaching resources, update the inverse document frequency weights based on the cloud storage data, and generate adaptively evolving semantic heat fingerprint generation parameters. Specifically, this includes: S7.1: Perform multi-source heterogeneous data collection and protocol parsing on the key chapter marking operation logs generated by the teacher's teaching courseware editing interface and the playback timeline click event stream generated by the student's video player, in order to extract the original user interaction behavior feature set containing the marked chapter index identifier, the marked frequency statistics value, and the start and end timestamps of the playback time period, as the initial input basis for updating the weight coefficients.

[0127] S7.2: Based on the marked chapter index identifier in the original user interaction behavior feature set, perform chapter content mapping and association processing, aggregate discrete marking operations to the corresponding courseware title keyword inverse document frequency weight vector dimension, and perform video segment temporal overlap calculation based on the start and end timestamps of the playback time period, map continuous playback events to the teaching term time axis frequency concentration vector dimension, and generate a modal alignment enhancement feature matrix carrying user attention quantification indicators.

[0128] S7.3: Utilize the modality alignment enhancement feature matrix to perform multidimensional attention weight derivation operations, calculate the tag density normalization factor in the text keyword dimension, the replay peak clustering factor in the video terminology dimension, and the implicit logic emphasis inference factor in the audio pause dimension, respectively, in order to quantify the relative importance change trend of each semantic modality in the current teaching cycle and generate a dynamic weight adjustment increment sequence to be updated.

[0129] Attention weight vectorization is performed on each modality dimension of the modality alignment enhancement feature matrix to establish a multidimensional feature increment mapping relationship between text, video, and audio semantic modalities within the current teaching cycle. Label density is calculated for the feature components of the text keyword semantic dimension. The frequency of feature values ​​corresponding to key chapter labels on the teacher's end within the title keyword vector dimension is counted, and a label density normalization factor is obtained based on the total number of global keywords. These three factors are sequentially combined into a dynamic weight adjustment increment sequence, retaining the contribution ratio of each modality in the multidimensional feature fusion process. Through multidimensional weight derivation calculations, the enhanced feature matrix is ​​transformed into quantified incremental data that can be used to update weight coefficients, achieving adaptive correction of the semantic heat fingerprint generation parameters.

[0130] S7.4: Perform exponential moving average smoothing filtering on the dynamic weight adjustment increment sequence, and perform linear fusion iterative calculation in combination with the historically stored adaptive evolution weight coefficient set to eliminate noise interference caused by single interaction behavior and retain the long-term popularity evolution pattern, and generate a new version of text keyword weight factor, terminology time sequence weight factor and speech pause weight factor combination with time continuity and stability.

[0131] For the incremental sequence of dynamic weight adjustment to be updated, the historically stored adaptive evolution weight coefficient set is loaded as the benchmark coefficient vector for smoothing, and a time series matrix is ​​generated to ensure that the increment of each dimension corresponds to the historical coefficient.

[0132] An exponential moving average filter is applied to each dimension of the time series matrix. The decay factor is set based on the teaching cycle length and the rate of change of resource popularity. The incremental sequence and historical weights are merged in a decreasing proportion according to the time weight to reduce the interference of short-term fluctuations on the long-term trend.

[0133] By using exponential moving average smoothing and linear fusion iterative calculation, the dynamic weight adjustment increment sequence is transformed into a new version of text keyword weight factor, terminology time sequence weight factor and speech pause weight factor combination with time continuity and stability, achieving the technical effect of long-term trend preservation and short-term noise suppression.

[0134] S7.5: Perform structured encapsulation and version verification processing on the combination of text keyword weight factors, terminology time sequence weight factors and speech pause weight factors in the new version, and generate an adaptively evolving semantic heat fingerprint generation parameter data package containing complete weight configuration information and effective time window, which serves as the sole authoritative input source for the edge node configuration library to dynamically iteratively optimize the fingerprint generation logic.

[0135] The new version of text keyword weight factors, terminology time sequence weight factors, and speech pause weight factors generated by multi-objective optimization are split into fields. The values ​​of each weight factor, the scope of the weight, and the source of the calculation are recorded in an independent data structure to maintain the traceability of the parameters.

[0136] Meta-information injection processing is performed on the split data structure to embed the generation batch number, algorithm version number and historical iteration sequence number associated with the weight factor combination into the unified metadata header, so as to ensure accurate matching capability in the scenario of multi-version weight update of edge nodes.

[0137] The effective time window is calculated based on the unified metadata header. The current system time and the teaching cycle reference time axis are used to perform interval mapping. The effective start and end times are written into the control field of the weight factor combination in the form of timestamps to ensure that the configuration is effective within the specified teaching cycle.

[0138] A digital signature algorithm is used to generate a signature verification code for a combination of weight factors containing complete weight configuration information and an effective time window. Elliptic curve digital signatures are used to ensure that data packets are not tampered with during distribution to edge nodes.

[0139] The version consistency detection module is invoked to perform version verification processing on the weight factor combination. The currently generated version sequence number is compared with the version number of the edge node configuration inventory to generate an update legality judgment signal, which serves as the sole authoritative input source for the dynamic iterative optimization of the fingerprint generation logic of the edge node configuration library.

[0140] By using structured encapsulation and version verification, the result of the previous step is transformed into an adaptive evolutionary semantic heat fingerprint generation parameter data package containing complete weight configuration information and effective time window, thereby achieving parameter iteration controllability of the edge node configuration library in the heat fingerprint generation logic.

[0141] For example, for a new version of text generated within an online course cycle, a combination of keyword weight factor 0.35, terminology temporal weight factor 0.45, and speech pause weight factor 0.20 is used. Field-based splitting is performed, and the sources are recorded as TF-IDF term frequency analysis, temporal clustering analysis, and pause statistical pattern optimization, respectively. The generation batch number of the injected metadata is 15, the algorithm version number is 3.2, and the historical iteration sequence number is 58. When calculating the effective time window, the start time is mapped to 1672531200 and the end time is mapped to 1675123200 (seconds) based on the teaching cycle timeline. A 256-bit digital signature is generated using the ECDSA algorithm to ensure the integrity of this parameter combination during transmission. In the version consistency check, the currently generated version serial number is 60, and the edge node configuration inventory version number is 59, which is determined to be a legitimate update. The final output data packet includes the weight factor value, scope, generation batch number, algorithm version number, historical serial number, effective time window, and signature verification code. After being encapsulated, it is written into the edge node configuration library, realizing the dynamic optimization of the weight coefficient in the multimedia teaching resource fingerprint generation logic.

[0142] Step S8: Synchronize the adaptively evolved semantic heat fingerprint generation parameters to the edge node configuration library, replacing the original inverse document frequency weights to complete the dynamic iterative optimization of the semantic heat fingerprint generation logic. Specifically, this includes: S8.1: Perform structured data encapsulation processing on the combination of text keyword weight factors, terminology time sequence weight factors and speech pause weight factors in the new version to generate an adaptively evolving semantic heat fingerprint generation parameter data package containing complete weight configuration information, effective time window identifier and digital signature verification code, which serves as the authoritative input source for edge node configuration updates.

[0143] S8.2: Based on the adaptively evolved semantic heat fingerprint, the parameter data packet is generated to perform integrity verification and version conflict detection processing. By comparing the existing version number in the current edge node configuration library with the version sequence number in the data packet, a version verification pass signal or version rollback instruction indicating the legality of the configuration update is generated.

[0144] S8.3: Based on the version verification pass signal, the distributed configuration center communication protocol is used to perform parameter synchronization transmission processing, and the adaptively evolved semantic heat fingerprint generation parameter data packet is distributed to the local configuration receiving buffer of each edge node through an encrypted channel to generate a temporary configuration cache object to be written.

[0145] S8.4: Perform an atomic replacement write operation on the temporary configuration cache object to be written, and replace the old set of weight coefficients stored in the edge node configuration library with the new version of the combination of text keyword weight factor, terminology time sequence weight factor and speech pause weight factor to generate an effective dynamic iterative optimization configuration record.

[0146] For the temporary configuration cache object to be written generated and distributed to the local configuration receive buffer, the atomic write interface of the edge node configuration library is called to establish an overwrite transaction to ensure that the entire set of weight coefficients is completely replaced in a single indivisible unit of operation. Block-level mapping verification is performed on the combination of text keyword weight factors, terminology timing weight factors, and speech pause weight factors in the temporary configuration cache object to be written, mapping them to their respective storage slots in the configuration library, and verifying the matching between slot capacity and data precision. Based on the data blocks that have passed the mapping verification, pointer redirection and redundancy cleanup operations are performed, pointing the pointers of the old version weight coefficient set to the new version data blocks, releasing the storage space occupied by the old version, and synchronously updating the version number field of the index table. A pre-write logging mechanism is used to record version number changes, new weight coefficient values, and effective time windows in persistent storage media, ensuring that a consistent state can be restored through log replay in the event of any interruption. After logging is complete, a commit trigger signal is sent to notify the upper-layer configuration management module that the replacement transaction has been completed and the new version weight coefficient combination is in effect, generating an effective dynamic iterative optimization configuration record. By using the atomic replacement write and transaction control processing methods described above, the temporary cache object from the previous step is transformed into a persistent and immediately applicable configuration record, achieving the expected technical effect of synchronously updating the global weight coefficient of the edge node configuration library without interruption or version aliasing.

[0147] For example, on the edge node of the multimedia teaching cloud platform, the temporary configuration cache object contains a text keyword weight factor of 0.42, a terminology timing weight factor of 0.35, and a speech pause weight factor of 0.23. The new version needs to replace the old version parameters 0.38, 0.31, and 0.31 within the effective time window. The write operation is called through the configuration library interface, and the block ID mapping is used to verify that the slots corresponding to the three weight coefficients are all floating-point numbers with a precision of no less than two decimal places. The old version parameter block length is 96 bytes, and the new version length is also 96 bytes; since they are identical, no new slots need to be allocated, only pointer redirection is performed. The pre-write log recording format includes the version number changing from V3.4 to V3.5, the effective time window Unix timestamp range [1698816000, 1698902400], and the replaced parameter values, ensuring that the replacement process can be replayed and recovered. The commit trigger signal is sent to the configuration management module when the transaction is committed. After the module confirms the replacement, the system log shows that the configuration library V3.5 is successfully loaded. The subsequent semantic heat fingerprint generation process has been tested and the response time has been greatly reduced under high concurrency requests, indicating that the dynamic iterative optimization of configuration records has taken effect and is being applied stably.

[0148] S8.5: Based on the already effective dynamic iterative optimization configuration record, trigger the hot loading mechanism of the fingerprint generation engine, re-initialize the weighted mapping operation logic in the semantic feature parsing and processing module, so that the subsequently generated semantic heat fingerprints can directly apply the updated weight coefficients, and complete the closed-loop dynamic iterative optimization of the semantic heat fingerprint generation logic.

[0149] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0150] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0151] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. Cloud storage data deduplication and access acceleration methods for multimedia teaching, specifically including: S1: Obtain the original teaching metadata from the multimedia teaching resource upload request. The original teaching metadata includes the courseware title text, the video automatic speech recognition and transcription result, and the teacher's speech waveform signal. S2: Perform multi-dimensional semantic feature parsing on the original teaching metadata, calculate the inverse document frequency weight of the courseware title keywords, the frequency concentration of teaching terms on the time axis, and the statistical regularity of semantic pause intervals in the speech waveform, and generate a normalized one-dimensional semantic feature vector. S3: Perform a weighted concatenation operation based on the normalized one-dimensional semantic feature vector to generate a semantic heat fingerprint with comparable order; S4: Dynamically set the heat threshold based on the current cloud storage system load status, and input the semantic heat fingerprint into the heat threshold filter for comparison and judgment. If the semantic heat fingerprint is lower than the heat threshold, the subsequent process is terminated; if it is higher, a high heat admission instruction is generated. S5: Establish a distributed fingerprint database including a fragment fingerprint index table. Based on the high-heat admission instruction, perform local matching of the semantic heat fingerprint according to the fragment fingerprint index table. If the matching is successful, generate a fragment reuse identifier. S6: For high-popularity admission instructions that have not generated the fragment reuse identifier, use a hash algorithm to generate the final content fingerprint, and perform a precise comparison operation in the distributed fingerprint database to generate a precise deduplication determination result.

2. The cloud storage data deduplication and access acceleration method for multimedia teaching according to claim 1, characterized in that, Step S6 is followed by: S7: Collect cloud storage data from multimedia teaching resources, update the inverse document frequency weight based on the cloud storage data, and generate adaptively evolving semantic heat fingerprint generation parameters. S8: Synchronize the adaptively evolved semantic heat fingerprint generation parameters to the edge node configuration library, and replace the original inverse document frequency weight to complete the dynamic iterative optimization of the semantic heat fingerprint generation logic.

3. The cloud storage data deduplication and access acceleration method for multimedia teaching according to claim 2, characterized in that, The cloud storage data includes behavioral data on teachers marking key chapters and interactive data on students during peak review periods.

4. The cloud storage data deduplication and access acceleration method for multimedia teaching according to claim 1, characterized in that, Step S3 specifically includes: Obtain the inverse document frequency weight vector of courseware title keywords, the frequency concentration vector of teaching terms on the time axis, and the statistical regularity vector of semantic pause intervals in speech waveforms output from step S2, and perform dimension alignment processing to obtain a standardized multimodal semantic feature matrix. Read the set of adaptive evolution weight coefficients stored in the edge node configuration library, and perform element-wise weighted mapping operation on the standardized multimodal semantic feature matrix based on the set of adaptive evolution weight coefficients to quantify the contribution of each semantic dimension to the overall business popularity and generate a weighted semantic feature component sequence. Perform a serial concatenation and aggregation operation on the weighted semantic feature component sequence to generate a high-dimensional original heat vector; The high-dimensional original heat vector is compressed using a linear dimensionality reduction algorithm, and then mapped to a low-dimensional dense vector space to generate a low-dimensional dense heat candidate vector. Normalized scalar encoding is performed on the low-dimensional dense heat candidate vectors to map the vector values ​​to a standardized sortable interval, giving them explicit numerical comparison attributes, and finally generating semantic heat fingerprints with comparable order.

5. The cloud storage data deduplication and access acceleration method for multimedia teaching according to claim 4, characterized in that, The adaptive evolution weight coefficient set includes text keyword weight factors, terminology temporal weight factors, and speech pause weight factors.

6. The cloud storage data deduplication and access acceleration method for multimedia teaching according to claim 4, characterized in that, The weighted semantic feature component sequence is subjected to a serial splicing and aggregation operation, which includes integrating the discrete weighted semantic feature component sequence into a continuous high-dimensional original heat vector by arranging the weighted semantic feature component sequence in a fixed order of text priority, time sequence second, and speech last.

7. The cloud storage data deduplication and access acceleration method for multimedia teaching according to claim 1, characterized in that, Step S4 specifically includes: Multi-dimensional indicator collection and processing of monitoring data streams of the current operating status of cloud storage systems; Based on the real-time load feature vector, a nonlinear mapping function is performed to generate a dynamic adaptive heat threshold. A scalar numerical comparison operation is performed between the generated semantic heat fingerprint with comparable order and the dynamic adaptive heat threshold to generate a binarized heat determination result signal. Based on the negative decision branch in the binarized heat determination result signal, the execution process is terminated and the mark encapsulation process is completed to generate a low heat mark to be cached carrying timestamp information; Based on the positive decision branch in the binarized heat determination result signal, the admission instruction generation and routing distribution process is performed to generate a high-heat admission instruction carrying a unique request identifier. This high-heat admission instruction triggers the downstream segment fingerprint index table query module to start a parallel retrieval mechanism to reuse historical deduplication results.

8. The cloud storage data deduplication and access acceleration method for multimedia teaching according to claim 7, characterized in that, The multidimensional indicator acquisition and processing includes extracting a real-time load feature vector containing the central node's CPU utilization, memory usage, and network bandwidth throughput.

9. The cloud storage data deduplication and access acceleration method for multimedia teaching according to claim 1, characterized in that, The step of calculating the inverse document frequency weights of the courseware title keywords, the frequency concentration of teaching terms on the time axis, and the statistical regularity of semantic pause intervals in the speech waveform to generate a normalized one-dimensional semantic feature vector includes performing multimodal feature fusion and maximum / minimum value normalization processing on the inverse document frequency weights of the courseware title keywords, the frequency concentration of teaching terms on the time axis, and the statistical regularity of semantic pause intervals in the speech waveform to generate a normalized one-dimensional semantic feature vector.

10. The cloud storage data deduplication and access acceleration method for multimedia teaching according to claim 1, characterized in that, The step of using a pre-built fragment fingerprint index table to perform local matching queries on the semantic heat fingerprint includes performing spatial hash mapping processing on the semantic heat fingerprint to generate a local retrieval key value for fast location, and the local retrieval key value serves as the sole addressing basis for accessing the fragment fingerprint index table.

Images