Enterprise safety production large model data storage compression optimization method and system

Abstract

The invention relates to the technical field of data storage, and discloses an enterprise safety production large model data storage compression optimization method and system, and the method comprises the steps: carrying out the deep semantic coding of multi-modal original data in the field of enterprise safety production, and obtaining a high-dimensional semantic embedding vector; performing joint low-rank tensor decomposition on the high-dimensional semantic embedding vector to obtain a core semantic factor and an auxiliary factor; performing low-rank approximate reconstruction on the high-dimensional semantic embedding vector to obtain a low-dimensional dense vector; performing time sequence behavior analysis on the low-dimensional dense vector to obtain a data access frequency, a joint rule detection task criticality and a recent update state; performing multi-dimensional weight analysis on the low-dimensional dense vector to obtain a cache weight score; performing priority allocation on the low-dimensional dense vectors, and storing the low-dimensional dense vectors to a cache level of a multi-layer cache architecture; according to the method, the efficiency of data storage compression optimization of the enterprise safety production large model can be improved.

Description

Technical Field

[0001] This invention relates to the field of data storage technology, and in particular to a method and system for data compression and optimization of large-scale enterprise safety production model data storage. Background Technology

[0002] Enterprise safety production involves multimodal data such as text, images, videos, and structured data. This data is massive in volume and highly heterogeneous. Existing data storage and compression technologies are insufficient in mining the deep semantic relationships of multimodal data, easily leading to the loss of key semantic information during compression. Furthermore, compression algorithms are poorly adapted to high-dimensional semantic data, resulting in low compression efficiency and failing to meet the lightweight data storage requirements of large models. In addition, existing cache allocation mechanisms lack comprehensive consideration of data access characteristics, task relevance, and update status, leading to an imbalance in cache resource allocation and high latency in responding to frequently accessed and highly critical data.

[0003] Existing low-rank decomposition-based compression methods do not fully integrate the structured characteristics of multimodal data, resulting in weakened data representation capabilities after dimensionality reduction and reconstruction, and an inability to accurately retain core semantic information related to safety production. Furthermore, the single dimension of temporal behavior analysis leads to a lack of comprehensiveness and accuracy in cache weight evaluation, hindering dynamic optimization of data storage. Therefore, how to improve the data storage compression efficiency and cache resource utilization rationality of large-scale enterprise safety production models while ensuring data semantic integrity has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a method and system for data compression and optimization of large-scale enterprise safety production model data to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides a method for data compression and optimization of large-scale enterprise safety production model data, comprising:

[0006] S1. Perform deep semantic encoding on the multimodal raw data in the field of enterprise safety production to obtain a high-dimensional semantic embedding vector of the multimodal raw data;

[0007] S2. Perform joint low-rank tensor decomposition on the high-dimensional semantic embedding vector to obtain the core semantic factor and auxiliary factor of the high-dimensional semantic embedding vector;

[0008] S3. Based on the core semantic factor and the auxiliary factor, the high-dimensional semantic embedding vector is reconstructed by low-rank approximation to obtain a low-dimensional dense vector of the high-dimensional semantic embedding vector.

[0009] S4. Perform time-series behavior analysis on the low-dimensional dense vector to obtain the data access frequency, criticality of the associated compliance detection task, and recent update status of the low-dimensional dense vector;

[0010] S5. Based on the data access frequency, the criticality of the associated compliance detection task, and the recent update status, perform multi-dimensional weight analysis on the low-dimensional dense vector to obtain the cache weight score of the low-dimensional dense vector.

[0011] S6. Based on the cache weight score, prioritize the low-dimensional dense vector and store it in the cache layer of the multi-layer cache architecture.

[0012] In a preferred embodiment, the step of performing deep semantic encoding on the multimodal raw data in the field of enterprise safety production to obtain a high-dimensional semantic embedding vector of the multimodal raw data includes:

[0013] By collecting text, images, videos, and structured data in the field of enterprise safety production, multimodal raw data in the field of enterprise safety production is obtained;

[0014] Heterogeneous data normalization is performed on the original multimodal data to obtain standardized data of the original multimodal data;

[0015] Modal feature extraction is performed on the standardized data to obtain the intermediate semantic features of the standardized data;

[0016] The intermediate semantic features are projected onto a high-dimensional vector space to obtain the high-dimensional semantic embedding vector of the original multimodal data.

[0017] In a preferred embodiment, the joint low-rank tensor decomposition of the high-dimensional semantic embedding vector to obtain the core semantic factors and auxiliary factors of the high-dimensional semantic embedding vector includes:

[0018] Based on the data samples, modality types, and embedding dimensions of the high-dimensional semantic embedding vector, the high-dimensional semantic embedding vector is structurally organized to obtain the three-dimensional embedding tensor of the high-dimensional semantic embedding vector;

[0019] Tensor deconstruction is performed on the three-dimensional embedding tensor to obtain the core tensor, modality factor matrix, and eigenfactor matrix of the three-dimensional embedding tensor.

[0020] The core tensor is used as the core semantic factor of the high-dimensional semantic embedding vector;

[0021] The modality factor matrix and the feature factor matrix are coupled to obtain the auxiliary factors of the high-dimensional semantic embedding vector.

[0022] In a preferred embodiment, the step of performing low-rank approximate reconstruction of the high-dimensional semantic embedding vector based on the core semantic factor and the auxiliary factor to obtain a low-dimensional dense vector of the high-dimensional semantic embedding vector includes:

[0023] Co-factor analysis is performed on the core semantic factors and the auxiliary factors to obtain the low-rank reconstruction mapping relationship of the high-dimensional semantic embedding vector;

[0024] By applying orthogonality constraints to the low-rank reconstruction mapping relationship, a low-dimensional latent space projection operator is obtained for the high-dimensional semantic embedding vector.

[0025] Based on the low-dimensional latent space projection operator, the high-dimensional semantic embedding vector is subjected to dimensionality reduction mapping to obtain a low-dimensional dense vector of the high-dimensional semantic embedding vector.

[0026] In a preferred embodiment, the orthogonality constraint applied to the low-rank reconstruction mapping relation to obtain the low-dimensional latent space projection operator of the high-dimensional semantic embedding vector includes:

[0027] The low-rank reconstruction mapping relationship is reconstructed by matrix reconstruction to obtain the parameterized projection matrix of the low-rank reconstruction mapping relationship;

[0028] The parameterized projection matrix is ​​orthogonally decomposed to obtain the standard orthogonal basis matrix of the parameterized projection matrix;

[0029] The orthonormal basis matrix is ​​used as the low-dimensional latent space projection operator of the high-dimensional semantic embedding vector.

[0030] In a preferred embodiment, the step of performing time-series behavior analysis on the low-dimensional dense vector to obtain the data access frequency, criticality of the associated compliance detection task, and recent update status of the low-dimensional dense vector includes:

[0031] Collect the historical access records of the low-dimensional dense vector;

[0032] Perform sliding window statistical analysis on the historical access records to obtain the data access frequency of the low-dimensional dense vector;

[0033] The correlation between the low-dimensional dense vector and the core detection tasks in the preset compliance rule base is evaluated to obtain the criticality of the low-dimensional dense vector in relation to the compliance detection tasks.

[0034] The recent update status of the low-dimensional dense vector is obtained by comparing the last modification timestamp with the current timestamp.

[0035] In a preferred embodiment, the step of performing multi-dimensional weight analysis on the low-dimensional dense vector based on the data access frequency, the criticality of the associated compliance detection task, and the recent update status to obtain the cache weight score of the low-dimensional dense vector includes:

[0036] The data access frequency, the criticality of the associated compliance detection task, and the recent update status are normalized to obtain the frequency parameter, criticality parameter, and status parameter of the low-dimensional dense vector.

[0037] The weighting factor analysis is performed on the frequency parameter, the criticality parameter, and the state parameter to obtain the weight coefficients of the low-dimensional dense vector;

[0038] Based on the weighting coefficients, the frequency parameter, the criticality parameter, and the state parameter are linearly weighted to obtain the cache weight score of the low-dimensional dense vector.

[0039] In a preferred embodiment, the formula for calculating the cache weight score is as follows:

[0040] ;

[0041] In the formula, Indicates the first The cache weight score of a low-dimensional dense vector. This represents the frequency parameter. This refers to the criticality parameter. This represents the state parameter. This represents the frequency weighting coefficient. This represents the keyness weight coefficient. Represents the state weight coefficient. This represents the preset smallest positive number.

[0042] In a preferred embodiment, the step of prioritizing the low-dimensional dense vector based on the cache weight score and storing it in the cache layer of a multi-layer cache architecture includes:

[0043] Based on the cache weight score, all the low-dimensional dense vectors are sorted in descending order to obtain a global priority queue of the low-dimensional dense vectors.

[0044] Based on the storage capacity and performance indicators of different levels in the multi-layer caching architecture, the global priority queue is segmented with capacity awareness to obtain the vector delineation of the hierarchical allocation boundary of the global priority queue.

[0045] Based on the vector, the hierarchical allocation boundary is defined, and the low-dimensional dense vector is allocated across hierarchical levels for storage.

[0046] To address the aforementioned problems, this invention also provides a data storage compression and optimization system for a large-scale enterprise safety production model, the system comprising:

[0047] The multimodal semantic encoding module is used to perform deep semantic encoding on multimodal raw data in the field of enterprise safety production to obtain a high-dimensional semantic embedding vector of the multimodal raw data;

[0048] The joint tensor decomposition module is used to perform joint low-rank tensor decomposition on the high-dimensional semantic embedding vector to obtain the core semantic factors and auxiliary factors of the high-dimensional semantic embedding vector.

[0049] The low-rank approximation reconstruction module is used to perform low-rank approximation reconstruction on the high-dimensional semantic embedding vector based on the core semantic factor and the auxiliary factor, so as to obtain a low-dimensional dense vector of the high-dimensional semantic embedding vector.

[0050] The vector time series analysis module is used to perform time series behavior analysis on the low-dimensional dense vector to obtain the data access frequency, criticality of the associated compliance detection task, and recent update status of the low-dimensional dense vector.

[0051] The cache weight parsing module is used to perform multi-dimensional weight parsing on the low-dimensional dense vector based on the data access frequency, the criticality of the associated compliance detection task, and the recent update status, so as to obtain the cache weight score of the low-dimensional dense vector.

[0052] The intelligent cache allocation module is used to prioritize the low-dimensional dense vector based on the cache weight score and store it in the cache layer of the multi-layer cache architecture.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] 1. This invention performs deep semantic encoding on multimodal raw data in the field of enterprise safety production, and combines joint low-rank tensor decomposition and low-rank approximate reconstruction techniques to accurately extract core semantic factors and auxiliary factors, efficiently realize the transformation of high-dimensional data into low-dimensional dense vectors, and completely preserve core semantic information while reducing storage overhead, thereby improving the efficiency and accuracy of data storage compression.

[0055] 2. This invention uses multi-dimensional time-series behavior analysis and cache weight parsing to generate cache weight scores based on data access frequency, task criticality, and update status. This enables intelligent allocation of data priorities and multi-layer cache adaptation storage, optimizes cache resource configuration, accelerates the response speed of high-value data access, and provides reliable data support for the stable and efficient operation of enterprise safety production models. Attached Figure Description

[0056] Figure 1A flowchart illustrating a method for data compression and optimization of a large-scale enterprise safety production model, as provided in an embodiment of the present invention;

[0057] Figure 2 A functional block diagram of a large-scale data storage compression and optimization system for enterprise safety production provided in an embodiment of the present invention;

[0058] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0059] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0060] This application provides a method for compressing and optimizing data storage of a large-scale enterprise safety production model. The executing entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0061] Reference Figure 1 The diagram shown is a flowchart illustrating a method for compressing and optimizing data storage of a large-scale enterprise safety production model according to an embodiment of the present invention. In this embodiment, the method includes:

[0062] S1. Perform deep semantic encoding on the multimodal raw data in the field of enterprise safety production to obtain a high-dimensional semantic embedding vector of the multimodal raw data;

[0063] In this embodiment of the invention, the step of performing deep semantic encoding on multimodal raw data in the field of enterprise safety production to obtain a high-dimensional semantic embedding vector of the multimodal raw data includes:

[0064] By collecting text, images, videos, and structured data in the field of enterprise safety production, multimodal raw data in the field of enterprise safety production is obtained;

[0065] Heterogeneous data normalization is performed on the original multimodal data to obtain standardized data of the original multimodal data;

[0066] Modal feature extraction is performed on the standardized data to obtain the intermediate semantic features of the standardized data;

[0067] The intermediate semantic features are projected onto a high-dimensional vector space to obtain the high-dimensional semantic embedding vector of the original multimodal data.

[0068] This process collects various types of data related to enterprise safety production. Text data includes safety regulations, operation manuals, accident reports, and hazard investigation records. Image data includes equipment inspection photos, safety protection facility pictures, production site environment pictures, and safety sign pictures. Video data includes production process monitoring videos, equipment operation status videos, safety drill recordings, and hazardous operation process recordings. Structured data includes equipment models, operating parameters, maintenance records, employee safety training files, and work permit approval information. All these different types of data are collected from various storage locations, such as the enterprise's database file server, monitoring equipment storage units, and office terminals, forming a complete multimodal raw data set for enterprise safety production.

[0069] The collected multimodal raw data underwent standardized format and numerical processing. Text data was uniformly converted to UTF-8 encoding, and the length of each text was adjusted to 5000 characters. Any text shorter than 5000 characters was padded with meaningless characters, while any text longer than 5000 characters was truncated to 5000 characters from the end. Image data was uniformly adjusted to a resolution of 1024×768 pixels, and all pixel values ​​were converted to integers between 0 and 255. Video data was extracted at a frame rate of 30 frames per second, and each frame was processed according to the same standards as individual image data, while recording the corresponding video timestamp. Numerical data in the structured data was uniformly converted to decimal floating-point numbers, and non-numerical data such as employee names and equipment numbers were converted to 64-bit binary encoding. This processing yielded standardized multimodal raw data.

[0070] Features were extracted for different modalities of standardized data. For standardized text data, the text content was analyzed sentence by sentence to identify keywords related to safety production, such as key phrases like "hazardous operations," "equipment malfunctions," and "safety protection," as well as phrases like "high-altitude work permits" and "hot work approvals." Semantic relationships between words were also analyzed, such as the connection between equipment malfunctions and maintenance records. This information was integrated to form intermediate semantic features for the text modality. For standardized image data, the pixel distribution, color composition ratio, object outlines, and texture details were analyzed to identify specific equipment components, the location of safety protection facilities, and employee operating postures. These features directly related to safety production were extracted as intermediate semantic features for the image modality. For visual data… For standardized video data, features extracted from each frame are used to track the motion trajectories of objects between frames, such as equipment operation and employee movement paths. Scene changes are analyzed, such as the progress of production processes and the emergence and elimination of hazardous scenes. These dynamic features are extracted as intermediate semantic features for the video modality. For standardized structured data, the relationships between various data fields are analyzed, such as the correspondence between equipment operating parameters and fault records. The patterns of numerical changes, such as the trend of equipment temperature changes, are analyzed. Key indicators reflecting the state of safe production, such as continuous fault-free operation time of equipment and employee training completion rate, are extracted as intermediate semantic features for the structured data modality. Finally, the features of all modalities are integrated to obtain the intermediate semantic features of the standardized data.

[0071] A 1000-dimensional high-dimensional vector space is pre-constructed. Each dimension of this space corresponds to a specific semantic attribute related to safety production. For example, the first dimension corresponds to the normal operation status of equipment, the second dimension corresponds to the type of high-altitude operation, and the third dimension corresponds to the integrity of fire-fighting facilities. A total of 1000 different semantic attributes are set, each corresponding to one of the 1000 dimensions of the vector space. The previously extracted intermediate semantic features of each modality are assigned to specific dimensions of the high-dimensional vector space according to the preset correspondence rules. Each intermediate semantic feature is assigned a specific value in a specific dimension of the vector space based on its specific meaning and importance in the safety production scenario. For example, the semantic feature of "normal equipment operation" corresponds to the first dimension and is assigned a value of 0.9, and the semantic feature of "high-altitude operation" corresponds to the second dimension and is assigned a value of 0.8. After all the intermediate semantic features of all modalities are assigned in the high-dimensional vector space, a vector containing 1000 dimension values ​​is formed. This vector is the high-dimensional semantic embedding vector of the multimodal raw data.

[0072] The beneficial effects are as follows: by comprehensively collecting various multimodal data in the field of enterprise safety production, the integrity and comprehensiveness of the data are ensured, providing a rich data source for subsequent processing. Heterogeneous data normalization processing unifies the format and numerical range of different modal data, eliminates the processing obstacles caused by data heterogeneity, and ensures the consistency and accuracy of subsequent feature extraction. Modal feature extraction can accurately capture the core semantic information related to safety production in various types of data, providing a solid foundation for semantic expression. Projecting intermediate semantic features onto a preset 1000-dimensional high-dimensional vector space, the resulting high-dimensional semantic embedding vector completely retains the semantic association and core information of the original multimodal data, providing high-quality data support for subsequent steps such as joint low-rank tensor decomposition. At the same time, each step of the entire process has clear operating standards and specific requirements, ensuring the reproducibility of the technical solution and enabling the stable and reliable acquisition of high-dimensional semantic embedding vectors.

[0073] S2. Perform joint low-rank tensor decomposition on the high-dimensional semantic embedding vector to obtain the core semantic factor and auxiliary factor of the high-dimensional semantic embedding vector;

[0074] In this embodiment of the invention, the step of performing joint low-rank tensor decomposition on the high-dimensional semantic embedding vector to obtain the core semantic factors and auxiliary factors of the high-dimensional semantic embedding vector includes:

[0075] Based on the data samples, modality types, and embedding dimensions of the high-dimensional semantic embedding vector, the high-dimensional semantic embedding vector is structurally organized to obtain the three-dimensional embedding tensor of the high-dimensional semantic embedding vector;

[0076] Tensor deconstruction is performed on the three-dimensional embedding tensor to obtain the core tensor, modality factor matrix, and eigenfactor matrix of the three-dimensional embedding tensor.

[0077] The core tensor is used as the core semantic factor of the high-dimensional semantic embedding vector;

[0078] The modality factor matrix and the feature factor matrix are coupled to obtain the auxiliary factors of the high-dimensional semantic embedding vector.

[0079] Each data sample for the high-dimensional semantic embedding vector is defined as an independent vector instance, with the total number of samples denoted as N, where N is a positive integer determined based on the actual amount of multimodal raw data collected. Modality types are categorized into four types: text, image, video, and structured data, uniquely identified by the numbers 1, 2, 3, and 4, respectively. The embedding dimension is a preset 1000 dimensions, corresponding to indices 1 to 1000. The first dimension corresponds to the data sample indices, increasing sequentially from 1 to N; the second dimension corresponds to the modality type identifier, with 1 to 4 matching the four modalities respectively; the third dimension corresponds to the embedding dimension indices, using a three-dimensional structure of 1 to 1000 as the framework. The 1000 dimensional values ​​of each high-dimensional semantic embedding vector are then filled into the corresponding positions in the three-dimensional structure according to the correspondence between their data sample indices, corresponding modality type identifiers, and embedding dimension indices, forming a complete three-dimensional embedding tensor of the high-dimensional semantic embedding vector.

[0080] For the constructed 3D embedding tensor, its data distribution characteristics are first analyzed as a whole to identify the key data regions that carry the core semantic associations of all data samples in each modality. The data in these regions can directly reflect the core semantic information in the field of enterprise safety production. These key data regions are extracted completely to form the core tensor of the 3D structure. Then, the data parts that describe only the inherent attributes of each modality, such as the semantic expression attributes of text and the visual feature attributes of images, are separated from the 3D embedding tensor. They are then arranged in a format where rows correspond to modality type identifiers and columns correspond to modality inherent attribute feature items to form a modality factor matrix. At the same time, the data parts that describe only the feature associations of each embedding dimension are separated from the 3D embedding tensor. They are then arranged in a format where rows correspond to embedding dimension indices and columns correspond to dimension feature parameters to form a feature factor matrix.

[0081] The core tensor concentrates the most critical and essential semantic association information in the high-dimensional semantic embedding vector. This information is the key part that reflects the core meaning of data in the field of enterprise safety production. Without additional data processing or transformation, the core tensor can be directly determined as the core semantic factor of the high-dimensional semantic embedding vector to ensure that the core semantic information is not lost or distorted.

[0082] First, semantic association matching is performed between the modality intrinsic attribute feature terms represented by the columns of the modality factor matrix and the dimensional feature parameters corresponding to the embedding dimension indices represented by the rows of the feature factor matrix. This ensures that the intrinsic attribute feature terms of each modality can form an accurate correspondence with the corresponding embedding dimension feature parameters. Next, the corresponding matching elements in the modality factor matrix and the feature factor matrix are associated and integrated. Following the format that rows correspond to modality type identifiers and columns correspond to embedding dimension indices, the integrated values ​​are filled into the corresponding positions one by one to form a new matrix. This matrix serves as the auxiliary factor for the high-dimensional semantic embedding vector, supplementing auxiliary semantic information related to modalities and embedding dimensions not covered by the core semantic factors.

[0083] S3. Based on the core semantic factor and the auxiliary factor, the high-dimensional semantic embedding vector is reconstructed by low-rank approximation to obtain a low-dimensional dense vector of the high-dimensional semantic embedding vector.

[0084] In this embodiment of the invention, the step of performing low-rank approximate reconstruction of the high-dimensional semantic embedding vector based on the core semantic factor and the auxiliary factor to obtain a low-dimensional dense vector of the high-dimensional semantic embedding vector includes:

[0085] Co-factor analysis is performed on the core semantic factors and the auxiliary factors to obtain the low-rank reconstruction mapping relationship of the high-dimensional semantic embedding vector;

[0086] By applying orthogonality constraints to the low-rank reconstruction mapping relationship, a low-dimensional latent space projection operator is obtained for the high-dimensional semantic embedding vector.

[0087] Based on the low-dimensional latent space projection operator, the high-dimensional semantic embedding vector is subjected to dimensionality reduction mapping to obtain a low-dimensional dense vector of the high-dimensional semantic embedding vector.

[0088] The orthogonality constraint applied to the low-rank reconstruction mapping relationship to obtain the low-dimensional latent space projection operator of the high-dimensional semantic embedding vector includes:

[0089] The low-rank reconstruction mapping relationship is reconstructed by matrix reconstruction to obtain the parameterized projection matrix of the low-rank reconstruction mapping relationship;

[0090] The parameterized projection matrix is ​​orthogonally decomposed to obtain the standard orthogonal basis matrix of the parameterized projection matrix;

[0091] The orthonormal basis matrix is ​​used as the low-dimensional latent space projection operator of the high-dimensional semantic embedding vector.

[0092] The core semantic factors, i.e., the core semantic associations of enterprise safety production carried in the core tensor, are analyzed. Specifically, these include the precise semantic correspondence between parameters such as temperature, vibration, and pressure in equipment operation status and fault detection such as overheating faults, mechanical wear faults, and pressure exceeding limits; the strict correlation between approval steps, operating procedures, and compliance requirements for specific operations such as hot work, high-altitude work, and confined space work in the safe operation process, including work permit processing, implementation of protective measures, and on-site supervision requirements; the matching of the quantity, location, and integrity of facilities such as fire hydrants, fire extinguishers, guardrails, and emergency passages in the configuration of protective facilities with low-risk, medium-risk, and high-risk areas in the risk level; and the correspondence between employee safety training records and work qualification certification, as well as the correlation between equipment maintenance cycles and failure rates. Simultaneously, the correspondence between the inherent attributes of each modality in the auxiliary factors and the features of the embedding dimension is accurately matched. The inherent attributes of the text modality are reflected in the accurate expression of safety terminology and the coherence of sentence logic, corresponding to the features of the semantic description dimension in the embedding dimension; the inherent attributes of the image modality are reflected in the morphological features, color distribution, and spatial layout of equipment components, corresponding to the features of the visual feature dimension in the embedding dimension; the inherent attributes of the video modality are reflected in the continuous actions of the work process and the dynamic changes of the scene, corresponding to the features of the dynamic feature dimension in the embedding dimension; the inherent attributes of structured data are reflected in the logical relationship between parameters and the change law of values, corresponding to the features of the numerical relationship dimension in the embedding dimension. Further identification of overlapping semantic information between core semantic factors and auxiliary factors is conducted. For example, the term "wearing protective equipment for high-altitude operations" in the text modality aligns with the core semantic concept of safe operating procedures; the morphological features of intact equipment guardrails in the image modality align with the core semantic concept of protective facility configuration. Simultaneously, complementary relationships are clarified. For instance, detailed safety operation instructions in the text modality supplement the specific execution requirements of the operating procedures in the core semantic concept; subtle wear features of equipment in the image modality supplement the potential hazard information for equipment fault detection in the core semantic concept; and detailed employee operation actions in the video modality supplement the criteria for determining operational compliance in the core semantic concept. Based on these clear relationships, a correspondence rule is established between 1000 dimensions of data in the high-dimensional semantic embedding vector and 200 dimensions in the low-dimensional space. This clarifies the specific dimensions in the low-dimensional space corresponding to different categories of dimensions in the high-dimensional data, such as equipment status, operating procedures, protective facilities, risk levels, and training certification, forming a low-rank reconstruction mapping relationship that can accurately guide the transformation of high-dimensional data to low-dimensional data.

[0093] The semantic correspondence rules and dimensional association logic contained in the low-rank reconstruction mapping relationship are comprehensively transformed into a structured matrix form. The number of rows in this matrix is ​​strictly set to the number of dimensions of the high-dimensional semantic embedding vector, 1000, corresponding to 1000 specific dimensions in the high-dimensional space. The number of columns is set to the preset number of dimensions of the low-dimensional space, 200, corresponding to 200 core semantic categories in the low-dimensional space. The matrix is ​​filled according to the association strength between each high-dimensional and low-dimensional dimension in the mapping relationship. The association strength is determined by the core semantic matching metric. The calculation of the core semantic matching degree is based on the degree of fit between the semantic information carried by the high-dimensional dimension and the core semantic category corresponding to the low-dimensional dimension. If the degree of fit between the high-dimensional semantic and the low-dimensional core semantic category reaches 80% or more, for example, the high-dimensional dimension "equipment operating temperature ≤60℃" and the low-dimensional category "equipment temperature safety status" are completely matched, with a matching degree of 85%, then the association strength is judged as high, and the corresponding matrix position is filled. The matrix is ​​filled with a value of 0.9. If the core semantic matching degree is between 50% and 79%, for example, the high-dimensional dimension "equipment operating vibration frequency 10-20Hz" partially matches the low-dimensional category "equipment mechanical safety status", with a matching degree of 65%, then the association strength is judged as medium, and the corresponding matrix position is filled with a value of 0.5. If the core semantic matching degree is less than 50%, for example, the high-dimensional dimension "equipment manufacturer number" has a low matching degree with the low-dimensional category "equipment safety status", with a matching degree of 30%, then the association strength is judged as low, and the corresponding matrix position is filled with a value of 0.1. According to this explicit rule, each position of the 1000 rows × 200 columns of the matrix is ​​filled with a specific value between 0 and 1 to ensure that each value in the matrix has a corresponding judgment basis, and finally a complete and accurate parameterized projection matrix of low-rank reconstruction mapping relationship is formed.

[0094] The 200 column vectors in the parameterized projection matrix are checked and adjusted one by one. First, the dot product between any two different column vectors is calculated. The dot product is calculated by multiplying the values ​​of the corresponding row positions of the two vectors respectively, and then summing all the multiplication results. For example, if we select the first and second column vectors, the values ​​in the first row of the first column vector are 0.9 and 0.5, and the values ​​in the second row of the second column vector are 0.1 and 0.5. The dot product of the two is 0.9 × 0.1 + 0.5 × 0.5 = 0.09 + 0.25 = 0.34. If the dot product is not 0, the adjustment ratio is determined based on the difference between the dot product and 0. The larger the difference, the larger the adjustment ratio. All values ​​of one of the vectors are adjusted proportionally, and the dot product is recalculated until the dot product of the two vectors is 0. In the example above, the difference is 0.34. All values ​​of the second column vector are multiplied by 0.6, and the dot product is recalculated as 0.9 × 0.06 + 0.5 × 0.3 = 0.054 + 0.15 = 0.204. The adjustment continues until the dot product is 0. Next, the dot product of each column vector is calculated, which is the sum of the squares of the values ​​at all rows of the vector. If the result is not 1, for example, if the dot product of a column vector is 2.25, then a scaling factor is calculated: 1 divided by the square root of the vector's dot product, i.e., 1 ÷ 1.5 ≈ 0.667. All values ​​of the vector are scaled proportionally by multiplying each value by 0.667, making the vector's dot product equal to 1. Through this process of checking and adjusting the cross-dot product and the dot product of each column vector one by one, the orthogonal decomposition of the parametric projection matrix is ​​completed, ultimately obtaining an orthogonal basis matrix that satisfies the condition that the dot product of any two column vectors is 0 and the dot product of a single column vector is 1.

[0095] Using the orthonormal basis matrix as the low-dimensional latent space projection operator for the high-dimensional semantic embedding vector, the association operation process between the high-dimensional vector and the projection operator is initiated. Each dimension value of the 1000-dimensional high-dimensional semantic embedding vector is matched one by one with the 200 column values ​​of the corresponding row of the projection operator. Specifically, the value of a certain dimension of the high-dimensional vector is multiplied by each column value of the corresponding row of the projection operator, and then all the multiplication results are summed to obtain the corresponding low-dimensional value for that column. For example, if the value of the 50th dimension of the high-dimensional semantic embedding vector is 0.8, corresponding to a sufficient number of fire hydrants, and the values ​​of the projection operator in the 50th row, 30th column are 0.9, the 50th row, 31st column are 0.05, and the remaining columns in the 50th row are 0.01 or lower, the sum of these values ​​is 0.8 × 0.9 + 0.8 × 0.05 + 0.8 × remaining column values ​​≈ 0.72 + 0.04 + 0.008 = 0.768, which is used as the value of the 30th column in the low-dimensional dimension. After the calculation is completed, valid values ​​are filtered according to preset standards. Values ​​greater than 0.05 that are directly related to the core semantics of enterprise safety production are judged as valid values, while values ​​less than or equal to 0.05 that have no actual semantic connection are judged as meaningless redundant values ​​and are removed. Through this series of precise calculations and filtering, the original 1000-dimensional high-dimensional vector is finally transformed into a 200-dimensional vector. This 200-dimensional vector is the low-dimensional dense vector of the high-dimensional semantic embedding vector, which fully retains all the core semantic information related to enterprise safety production, such as equipment safety status, operation compliance, risk prevention and control, protective facility configuration, and employee training and certification. Each dimension corresponds to clear and key semantic content, with no redundant or invalid data.

[0096] S4. Perform time-series behavior analysis on the low-dimensional dense vector to obtain the data access frequency, criticality of the associated compliance detection task, and recent update status of the low-dimensional dense vector;

[0097] In this embodiment of the invention, the step of performing time-series behavior analysis on the low-dimensional dense vector to obtain the data access frequency, criticality of the associated compliance detection task, and recent update status of the low-dimensional dense vector includes:

[0098] Collect the historical access records of the low-dimensional dense vector;

[0099] Perform sliding window statistical analysis on the historical access records to obtain the data access frequency of the low-dimensional dense vector;

[0100] The correlation between the low-dimensional dense vector and the core detection tasks in the preset compliance rule base is evaluated to obtain the criticality of the low-dimensional dense vector in relation to the compliance detection tasks.

[0101] The recent update status of the low-dimensional dense vector is obtained by comparing the last modification timestamp with the current timestamp.

[0102] Historical access records of low-dimensional dense vectors are collected. The records include the access time, access subject, and access purpose of each low-dimensional dense vector. The scope of access record collection covers all access behaviors within the past 90 days. All access records related to low-dimensional dense vectors are extracted from storage locations such as the access log database, cache call record files, and application interaction logs of the enterprise data management platform through data synchronous reading, ensuring that the collected historical access records completely cover all access scenarios without omissions or missing information.

[0103] The sliding window duration is set to 7 days, and the window sliding step size is 1 day. The collected historical access records are arranged in chronological order of access time. Starting from the first sliding window, the windows are slid forward sequentially. The total number of times the low-dimensional dense vector is accessed is counted in each window. For example, if a window records 28 accesses, then the access count for that window is 28. After traversing all sliding windows, the average number of accesses for all windows is calculated. This average value is the data access frequency of the low-dimensional dense vector. If the total duration of the historical access records is less than 7 days, the actual duration is used as a single sliding window for statistics, and the access count of that window is directly used as the data access frequency.

[0104] The pre-defined compliance rule base includes core testing tasks in the field of enterprise safety production, specifically covering five categories: equipment operation safety testing, work process compliance testing, protective facility configuration compliance testing, employee operation standard testing, and emergency response process compliance testing. Each core testing task is clearly marked with key semantic features and testing indicators. For example, the key semantic features for equipment operation safety testing include equipment temperature parameters, vibration frequency, and pressure values. The 200-dimensional core semantic features of the low-dimensional dense vector are compared one by one with the key semantic features of each core testing task. The number of successfully matched semantic features is counted, and the proportion of the number of matches to the total number of key semantic features for that core testing task is calculated. The highest proportion among all core testing tasks is taken as the relevance matching value. If the relevance matching value is ≥80%, the criticality of the associated compliance testing task is judged as high; if the relevance matching value is between 60% and 79%, it is judged as medium; and if the relevance matching value is below 60%, it is judged as low. This judgment result is the criticality of the low-dimensional dense vector associated with the compliance testing task.

[0105] The last modification timestamp, accurate to the second, is extracted from the metadata information of the low-dimensional dense vector. The format is year-month-day hour:minute:second. Simultaneously, the current system timestamp is obtained, with the same format as the last modification timestamp. The time difference between the two timestamps is calculated by subtracting the last modification timestamp from the current timestamp, resulting in a value in hours. If the time difference is ≤24 hours, the recent update status is determined as updated; if the time difference is between 25 and 72 hours, the recent update status is determined as partially updated; if the time difference exceeds 72 hours, the recent update status is determined as not updated. This determination result is the recent update status of the low-dimensional dense vector.

[0106] S5. Based on the data access frequency, the criticality of the associated compliance detection task, and the recent update status, perform multi-dimensional weight analysis on the low-dimensional dense vector to obtain the cache weight score of the low-dimensional dense vector.

[0107] In this embodiment of the invention, the step of performing multi-dimensional weight analysis on the low-dimensional dense vector based on the data access frequency, the criticality of the associated compliance detection task, and the recent update status to obtain the cache weight score of the low-dimensional dense vector includes:

[0108] The data access frequency, the criticality of the associated compliance detection task, and the recent update status are normalized to obtain the frequency parameter, criticality parameter, and status parameter of the low-dimensional dense vector.

[0109] The weighting factor analysis is performed on the frequency parameter, the criticality parameter, and the state parameter to obtain the weight coefficients of the low-dimensional dense vector;

[0110] Based on the weighting coefficients, the frequency parameter, the criticality parameter, and the state parameter are linearly weighted to obtain the cache weight score of the low-dimensional dense vector.

[0111] The formula for calculating the cache weight score is as follows:

[0112] ;

[0113] In the formula, Indicates the first The cache weight score of a low-dimensional dense vector. This represents the frequency parameter. This refers to the criticality parameter. This represents the state parameter. This represents the frequency weighting coefficient. This represents the keyness weight coefficient. Represents the state weight coefficient. This represents the preset smallest positive number.

[0114] The data access frequency, criticality of the compliance detection task, and recent update status are normalized. Data access frequency is the average number of accesses over the past 90 days, obtained through a sliding window statistical method. During normalization, the access frequency of a single low-dimensional dense vector is divided by the maximum access frequency of all low-dimensional dense vectors; the result is the frequency parameter, with a value limited to 0 to 1. The criticality of the compliance detection task was previously categorized into three levels (high, medium, and low) based on relevance assessment, corresponding to initial values ​​of 0.8 to 1.0, 0.6 to 0.79, and 0 to 0.59, respectively. During normalization, the initial criticality value of each low-dimensional dense vector is divided by 1.0, mapping the result to the 0-1 range; the result is the criticality parameter. The recent update status was previously categorized into three states (updated, partially updated, and not updated) based on time difference comparison, corresponding to initial values ​​of 1.0, 0.5, and 0.1, respectively. These initial values ​​are already within the 0-1 range and require no further mapping; they are directly used as the status parameter.

[0115] Weighted factor analysis was performed on frequency, criticality, and status parameters. This analysis was based on the cache resource allocation requirements of the enterprise's large-scale safety production model. Core compliance testing tasks are crucial for ensuring enterprise safety production; therefore, criticality parameters closely related to these core tasks have the greatest impact on cache allocation. Access frequency reflects the activity level of data usage and has a secondary impact, while data update status reflects timeliness and has a relatively smaller impact. Based on this ranking of impact, a weight coefficient of 0.4 was set for criticality, 0.35 for frequency, and 0.25 for status, with a total weight coefficient of 1.0. This ensures that the weight allocation aligns with the core requirements of optimized cache resource configuration. These three determined values ​​are respectively used as the frequency weight coefficient, criticality weight coefficient, and status weight coefficient.

[0116] A very small, fixed positive constant, 0.0001, is preset. This value prevents the denominator from being zero in subsequent calculations without affecting the validity of the overall result. When calculating the cache weight score, three division operations are performed: first, the frequency weight coefficient is divided by the sum of the frequency parameter and the very small positive constant, yielding the first intermediate result; second, the criticality weight coefficient is divided by the sum of the criticality parameter and the very small positive constant, yielding the second intermediate result; third, the state weight coefficient is divided by the sum of the state parameter and the very small positive constant, yielding the third intermediate result. These three intermediate results are then added together to obtain a total value. Finally, 3 is divided by this total value, resulting in the final cache weight score of the low-dimensional dense vector. This score is limited to the range of 0 to 3, with a higher score indicating a higher cache priority for the low-dimensional dense vector.

[0117] The beneficial effects are as follows: By standardizing and normalizing data access frequency, criticality of associated compliance detection tasks, and recent update status, indicators of different dimensions and magnitudes are uniformly mapped to the same numerical range, ensuring that the three indicators are comparable and calculable. The weighted factor analysis process clarifies the influence weight of each indicator based on the core needs of enterprise safety production, highlighting the importance of criticality indicators. This aligns with the configuration logic of tilting cache resources towards high-value data. The calculation process of cache weight scores avoids calculation anomalies by pre-setting minimal normal numbers. At the same time, the comprehensive calculation of multi-dimensional indicators achieves accurate quantification of the value of low-dimensional dense vector cache. The obtained cache weight scores can comprehensively and selectively reflect the data access activity, task association importance, and timeliness, providing an objective, unified, and reproducible judgment standard for subsequent priority allocation, and ensuring the rationality and scientific nature of cache resource allocation.

[0118] S6. Based on the cache weight score, prioritize the low-dimensional dense vector and store it in the cache layer of the multi-layer cache architecture.

[0119] In this embodiment of the invention, the step of prioritizing the low-dimensional dense vector based on the cache weight score and storing it in the cache layer of a multi-layer cache architecture includes:

[0120] Based on the cache weight score, all the low-dimensional dense vectors are sorted in descending order to obtain a global priority queue of the low-dimensional dense vectors.

[0121] Based on the storage capacity and performance indicators of different levels in the multi-layer caching architecture, the global priority queue is segmented with capacity awareness to obtain the vector delineation of the hierarchical allocation boundary of the global priority queue.

[0122] Based on the vector, the hierarchical allocation boundary is defined, and the low-dimensional dense vector is allocated across hierarchical levels for storage.

[0123] Collect all low-dimensional dense vectors obtained through multi-dimensional weight parsing and their corresponding cache weight scores. The cache weight scores range from 0 to 3. Sort all low-dimensional dense vectors in descending order of scores. If multiple low-dimensional dense vectors have the same cache weight score, use the data access frequency in descending order as a supplementary sorting criterion. If the data access frequencies are also the same, sort them in ascending order of the last modification timestamp. After sorting, a continuous sequence is formed. The earlier the low-dimensional dense vector in the sequence, the higher its priority. This sequence is the global priority queue of low-dimensional dense vectors. Each vector in the queue corresponds to a unique priority sorting number.

[0124] The default multi-level caching architecture consists of three levels: L1 cache, L2 cache, and L3 cache. The L1 cache has a storage capacity of 100GB and performance metrics of read / write latency ≤1ms and concurrent accesses ≥10,000 per second. The L2 cache has a storage capacity of 500GB and performance metrics of read / write latency ≤5ms and concurrent accesses ≥5,000 per second. The L3 cache has a storage capacity of 1TB and performance metrics of read / write latency ≤10ms and concurrent accesses ≥2,000 per second. Each low-dimensional dense vector has a storage size of 800 bytes. Based on this, the maximum number of vectors that each level can hold is calculated: L1 cache can hold a maximum of 125,000 vectors, L2 cache can hold a maximum of 625,000 vectors, and L3 cache can hold a maximum of 1,250,000 vectors. The total number of vectors in the global priority queue is counted. The queues are segmented based on the maximum capacity of each level. The allocation boundary of the L1 cache is set as the first 125,000 vectors in the queue, the allocation boundary of the L2 cache is from the 125,001st to the 750,000th vector in the queue, and the allocation boundary of the L3 cache is from the 750,001st vector in the queue to the vector corresponding to the total number of vectors. The queue segmentation range corresponding to each level is clearly defined.

[0125] Based on vector-based hierarchical allocation boundaries, the first 125,000 low-dimensional dense vectors in the global priority queue are directly stored in the L1 cache, vectors 125,001 to 750,000 are stored in the L2 cache, and vectors from 750,001 to the total number of vectors are stored in the L3 cache. Simultaneously, cross-level backup rules are set: the first 20,000 highest-priority low-dimensional dense vectors in the L1 cache are synchronously backed up in the L2 cache, and the first 100,000 vectors in the L2 cache are synchronously backed up in the L3 cache. This ensures that high-priority data has redundant copies in lower-level caches, preventing data inaccessibility due to a single cache level failure. During storage, hierarchical identifier metadata is added to each low-dimensional dense vector to clarify its primary and backup storage levels, and a storage completion timestamp is recorded, completing the cross-level storage allocation of low-dimensional dense vectors.

[0126] The beneficial effects are as follows: by forming a global priority queue in descending order of cache weight scores, the priority of high-value data is clearly defined. Capacity-aware segmentation is performed in conjunction with the specific storage capacity and performance indicators of the three-level cache architecture, ensuring that cache resources at each level are fully utilized and accurately matched with data priorities. The setting of cross-level storage allocation and backup rules not only ensures fast access response for high-priority data, but also improves the reliability of data storage, avoids cache resource waste and high-value data access delays, and achieves optimized configuration of cache resources, providing efficient and stable storage support for data calls in the enterprise's large-scale safe production model.

[0127] like Figure 2The diagram shown is a functional block diagram of a large-scale data storage compression and optimization system for enterprise safety production provided in an embodiment of the present invention.

[0128] The enterprise safety production large-scale model data storage compression and optimization system 100 described in this invention can be installed in an electronic device. Depending on the functions implemented, the enterprise safety production large-scale model data storage compression and optimization system 100 may include a multimodal semantic coding module 101, a joint tensor decomposition module 102, a low-rank approximation reconstruction module 103, a vector temporal analysis module 104, a cache weight parsing module 105, and an intelligent cache allocation module 106. The modules described in this invention can also be referred to as units, which are a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, stored in the memory of the electronic device.

[0129] In this embodiment, the functions of each module / unit are as follows:

[0130] The multimodal semantic encoding module 101 is used to perform deep semantic encoding on multimodal raw data in the field of enterprise safety production to obtain a high-dimensional semantic embedding vector of the multimodal raw data.

[0131] The joint tensor decomposition module 102 is used to perform joint low-rank tensor decomposition on the high-dimensional semantic embedding vector to obtain the core semantic factors and auxiliary factors of the high-dimensional semantic embedding vector.

[0132] The low-rank approximation reconstruction module 103 is used to perform low-rank approximation reconstruction on the high-dimensional semantic embedding vector based on the core semantic factor and the auxiliary factor, so as to obtain a low-dimensional dense vector of the high-dimensional semantic embedding vector.

[0133] The vector time series analysis module 104 is used to perform time series behavior analysis on the low-dimensional dense vector to obtain the data access frequency, criticality of the associated compliance detection task, and recent update status of the low-dimensional dense vector.

[0134] The cache weight parsing module 105 is used to perform multi-dimensional weight parsing on the low-dimensional dense vector based on the data access frequency, the criticality of the associated compliance detection task, and the recent update status, so as to obtain the cache weight score of the low-dimensional dense vector.

[0135] The intelligent cache allocation module 106 is used to prioritize the low-dimensional dense vector based on the cache weight score and store it in the cache layer of the multi-layer cache architecture.

[0136] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0137] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0138] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0139] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0140] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for data compression and optimization of a large-scale enterprise safety production model, characterized in that, The method includes: S1. Perform deep semantic encoding on the multimodal raw data in the field of enterprise safety production to obtain a high-dimensional semantic embedding vector of the multimodal raw data; S2. Perform joint low-rank tensor decomposition on the high-dimensional semantic embedding vector to obtain the core semantic factor and auxiliary factor of the high-dimensional semantic embedding vector; S3. Based on the core semantic factor and the auxiliary factor, the high-dimensional semantic embedding vector is reconstructed by low-rank approximation to obtain a low-dimensional dense vector of the high-dimensional semantic embedding vector. S4. Perform time-series behavior analysis on the low-dimensional dense vector to obtain the data access frequency, criticality of the associated compliance detection task, and recent update status of the low-dimensional dense vector; S5. Based on the data access frequency, the criticality of the associated compliance detection task, and the recent update status, perform multi-dimensional weight analysis on the low-dimensional dense vector to obtain the cache weight score of the low-dimensional dense vector. S6. Based on the cache weight score, prioritize the low-dimensional dense vector and store it in the cache layer of the multi-layer cache architecture.

2. The data storage compression and optimization method for a large-scale enterprise safety production model as described in claim 1, characterized in that, The process of performing deep semantic encoding on multimodal raw data in the field of enterprise safety production to obtain a high-dimensional semantic embedding vector of the multimodal raw data includes: By collecting text, images, videos, and structured data in the field of enterprise safety production, multimodal raw data in the field of enterprise safety production is obtained; Heterogeneous data normalization is performed on the original multimodal data to obtain standardized data of the original multimodal data; Modal feature extraction is performed on the standardized data to obtain the intermediate semantic features of the standardized data; The intermediate semantic features are projected onto a high-dimensional vector space to obtain the high-dimensional semantic embedding vector of the original multimodal data.

3. The data storage compression and optimization method for a large-scale enterprise safety production model as described in claim 1, characterized in that, The joint low-rank tensor decomposition of the high-dimensional semantic embedding vector yields the core semantic factors and auxiliary factors of the high-dimensional semantic embedding vector, including: Based on the data samples, modality types, and embedding dimensions of the high-dimensional semantic embedding vector, the high-dimensional semantic embedding vector is structurally organized to obtain the three-dimensional embedding tensor of the high-dimensional semantic embedding vector; Tensor deconstruction is performed on the three-dimensional embedding tensor to obtain the core tensor, modality factor matrix, and eigenfactor matrix of the three-dimensional embedding tensor. The core tensor is used as the core semantic factor of the high-dimensional semantic embedding vector; The modality factor matrix and the feature factor matrix are coupled to obtain the auxiliary factors of the high-dimensional semantic embedding vector.

4. The data storage compression and optimization method for a large-scale enterprise safety production model as described in claim 1, characterized in that, The step of performing low-rank approximate reconstruction of the high-dimensional semantic embedding vector based on the core semantic factor and the auxiliary factor to obtain a low-dimensional dense vector of the high-dimensional semantic embedding vector includes: Co-factor analysis is performed on the core semantic factors and the auxiliary factors to obtain the low-rank reconstruction mapping relationship of the high-dimensional semantic embedding vector; By applying orthogonality constraints to the low-rank reconstruction mapping relationship, a low-dimensional latent space projection operator is obtained for the high-dimensional semantic embedding vector. Based on the low-dimensional latent space projection operator, the high-dimensional semantic embedding vector is subjected to dimensionality reduction mapping to obtain a low-dimensional dense vector of the high-dimensional semantic embedding vector.

5. The data storage compression and optimization method for a large-scale enterprise safety production model as described in claim 4, characterized in that, The orthogonality constraint applied to the low-rank reconstruction mapping relationship to obtain the low-dimensional latent space projection operator of the high-dimensional semantic embedding vector includes: The low-rank reconstruction mapping relationship is reconstructed by matrix reconstruction to obtain the parameterized projection matrix of the low-rank reconstruction mapping relationship; The parameterized projection matrix is ​​orthogonally decomposed to obtain the standard orthogonal basis matrix of the parameterized projection matrix; The orthonormal basis matrix is ​​used as the low-dimensional latent space projection operator of the high-dimensional semantic embedding vector.

6. The data storage compression and optimization method for a large-scale enterprise safety production model as described in claim 1, characterized in that, The temporal behavior analysis of the low-dimensional dense vector yields the data access frequency, criticality of the associated compliance detection task, and recent update status of the low-dimensional dense vector, including: Collect the historical access records of the low-dimensional dense vector; Perform sliding window statistical analysis on the historical access records to obtain the data access frequency of the low-dimensional dense vector; The correlation between the low-dimensional dense vector and the core detection tasks in the preset compliance rule base is evaluated to obtain the criticality of the low-dimensional dense vector in relation to the compliance detection tasks. The recent update status of the low-dimensional dense vector is obtained by comparing the last modification timestamp with the current timestamp.

7. The data storage compression and optimization method for a large-scale enterprise safety production model as described in claim 1, characterized in that, The method involves performing multi-dimensional weight analysis on the low-dimensional dense vector based on the data access frequency, the criticality of the associated compliance detection task, and the recent update status to obtain the cache weight score of the low-dimensional dense vector, including: The data access frequency, the criticality of the associated compliance detection task, and the recent update status are normalized to obtain the frequency parameter, criticality parameter, and status parameter of the low-dimensional dense vector. The weighting factor analysis is performed on the frequency parameter, the criticality parameter, and the state parameter to obtain the weight coefficients of the low-dimensional dense vector; Based on the weighting coefficients, the frequency parameter, the criticality parameter, and the state parameter are linearly weighted to obtain the cache weight score of the low-dimensional dense vector.

8. The data storage compression and optimization method for a large-scale enterprise safety production model as described in claim 7, characterized in that, The formula for calculating the cache weight score is as follows: ； In the formula, Indicates the first The cache weight score of a low-dimensional dense vector. This represents the frequency parameter. This refers to the criticality parameter. This represents the state parameter. This represents the frequency weighting coefficient. This represents the keyness weight coefficient. Represents the state weight coefficient. This represents the preset smallest positive number.

9. The data storage compression and optimization method for a large-scale enterprise safety production model as described in claim 1, characterized in that, The step of prioritizing the low-dimensional dense vector based on the cache weight score and storing it in the cache layer of the multi-layer cache architecture includes: Based on the cache weight score, all the low-dimensional dense vectors are sorted in descending order to obtain a global priority queue of the low-dimensional dense vectors. Based on the storage capacity and performance indicators of different levels in the multi-layer caching architecture, the global priority queue is segmented with capacity awareness to obtain the vector delineation of the hierarchical allocation boundary of the global priority queue. Based on the vector, the hierarchical allocation boundary is defined, and the low-dimensional dense vector is allocated across hierarchical levels for storage.

10. A data storage compression and optimization system for a large-scale enterprise safety production model, characterized in that, The system for implementing the enterprise safety production large-scale model data storage compression and optimization method as described in claim 1 includes: The multimodal semantic encoding module is used to perform deep semantic encoding on multimodal raw data in the field of enterprise safety production to obtain a high-dimensional semantic embedding vector of the multimodal raw data; The joint tensor decomposition module is used to perform joint low-rank tensor decomposition on the high-dimensional semantic embedding vector to obtain the core semantic factors and auxiliary factors of the high-dimensional semantic embedding vector. The low-rank approximation reconstruction module is used to perform low-rank approximation reconstruction on the high-dimensional semantic embedding vector based on the core semantic factor and the auxiliary factor, so as to obtain a low-dimensional dense vector of the high-dimensional semantic embedding vector. The vector time series analysis module is used to perform time series behavior analysis on the low-dimensional dense vector to obtain the data access frequency, criticality of the associated compliance detection task, and recent update status of the low-dimensional dense vector. The cache weight parsing module is used to perform multi-dimensional weight parsing on the low-dimensional dense vector based on the data access frequency, the criticality of the associated compliance detection task, and the recent update status, so as to obtain the cache weight score of the low-dimensional dense vector. The intelligent cache allocation module is used to prioritize the low-dimensional dense vector based on the cache weight score and store it in the cache layer of the multi-layer cache architecture.

Images