Standardized processing and intelligent storage architecture method for construction task data

Abstract

The invention relates to the technical field of constructional engineering informatization and artificial intelligence fusion, in particular to a construction task data standardization processing and intelligent storage architecture method which comprises the steps that multi-source construction task data from a construction site is collected to be subjected to unified format packaging and structure initialization processing; establishing a plurality of candidate structure models based on a minimum description length principle, and performing structure reconstruction on the multi-source construction task data; the standardized construction task data are embedded into a multi-scale information representation space; estimating local information entropy and entropy gradient density for the embedded construction task data; and according to the time change of the construction task data, constructing a self-supervised driven embedding updating mechanism, and completing iterative optimization of an embedding function. According to the method, unified model expression is carried out on the multi-source heterogeneous construction data through a structure modeling mode based on the minimum description length principle, and the technical effects of reducing redundant information and enhancing the data structure definition are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of the integration of construction project informatization and artificial intelligence, and specifically to a method for standardizing the processing and intelligent storage architecture of construction task data. Background Art

[0002] Under the trend of the accelerating informatization transformation of the construction industry, the task data involved in the construction site is becoming increasingly complex and dynamic. Especially in large-scale engineering projects, the construction tasks are not only diverse in types, but also the data sources show diverse characteristics, including text task descriptions, schedule plans, process records, material tracking data, sensor information, etc. These data have the typical characteristics of strong timeliness, inconsistent structures, and dense semantic intersections.

[0003] Some existing construction management systems or engineering information platforms have introduced the basic concept of structuring and hierarchical storage of task data, and tried to classify the task data by means of information modeling tools or data warehouses. At the same time, there are also applications that attempt to extract task semantics through feature compression or embedding methods for subsequent intelligent analysis modules such as progress prediction, resource scheduling, and risk identification.

[0004] However, the data standardization process of some existing construction management systems or engineering information platforms relies on artificial templates or fixed logics, with insufficient flexibility and poor adaptability to unstructured or new data formats. Secondly, most of the embedding models are statically trained, lacking the ability to adapt to time series, and are prone to "failure" after on-site changes. Moreover, the storage level division usually adopts simple thresholds or empirical rules, ignoring the information value distribution of the data itself, resulting in a high error rate in identifying hot and cold data. Therefore, the present invention provides a method for standardizing the processing and intelligent storage architecture of construction task data to solve the deficiencies existing in the prior art. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides a method for standardizing the processing and intelligent storage architecture of construction task data, which solves the problems that the existing data standardization process relies on artificial templates or fixed logics, with insufficient flexibility, poor adaptability to unstructured or new data formats, and the storage level division usually adopts simple thresholds or empirical rules, ignoring the information value distribution of the data itself, resulting in a high error rate in identifying hot and cold data.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for standardizing the processing and intelligent storage architecture of construction task data, including the following steps:

[0007] Collect multi-source construction task data from the construction site, and perform unified format encapsulation and structure initialization processing on the collected data;

[0008] Based on the minimum description length principle, multiple candidate structural models are established to reconstruct the multi-source construction task data, and the structural model with the minimum description length is selected to perform a standardized representation of the multi-source construction task data;

[0009] The standardized construction task data is embedded into a multi-scale information representation space, and the similarity relationship of the original data in the semantic structure is maintained during the embedding process;

[0010] Estimate the local information entropy and entropy gradient density of the embedded construction task data, and divide the data into different storage levels based on the magnitude of the information density;

[0011] According to the time variation of the construction task data, a self-supervised driven embedding update mechanism is constructed to complete the iterative optimization of the embedding function.

[0012] Preferably, the multi-source construction task data includes text data, image data, and sensor data, and the unified format encapsulation adopts a structured tuple form, and each piece of data is encapsulated into a triple structure including a field type, data content, and timestamp.

[0013] Preferably, the steps for performing a standardized representation of the multi-source construction task data include:

[0014] Calculate the corresponding structural coding length and data coding length for each candidate structural model, and add the two to obtain the total description length of the structural model;

[0015] Compare the total description lengths of all candidate structural models, and select the structural model with the minimum description length as the optimal structural model;

[0016] Based on the optimal structural model, perform field extraction, recombination, and mapping on the multi-source construction task data to obtain a unified structured data representation form.

[0017] Preferably, the steps for embedding the standardized construction task data into a multi-scale information representation space include:

[0018] By constructing a multi-level mapping function, map the standardized construction task data to embedding spaces of multiple scales to capture different-level features of the data;

[0019] During the embedding process, a similarity preservation method is adopted to ensure that the construction task data maintains semantic similarity and structural similarity between the original data after embedding;

[0020] The mapping function includes, but is not limited to, similarity metrics based on Euclidean distance, cosine similarity, or Gaussian kernel function, ensuring that the relative positions of the data in the embedding space reflect the semantic relationships of the original data.

[0021] Preferably, the step of partitioning data into different storage levels based on the information density size includes:

[0022] Estimate the local information entropy of each embedding vector, where the local information entropy represents the uncertainty of the probability distribution of the embedding vector in the current information space;

[0023] Based on the local information entropy of the embedding vector, calculate the entropy gradient density of each embedding vector, where the entropy gradient density represents the rate of change of entropy in the information space;

[0024] According to the calculated entropy gradient density, partition the embedded construction task data into different storage levels, including a hot layer, a warm layer, and a cold layer, where the partitioning of the storage levels is based on the size of the entropy gradient density.

[0025] Preferably, the formula for calculating the local information entropy is:

[0026] H(x i )=-log2P(x i );

[0027] Among them, H(x i ) represents the local information entropy of the data point x i ; x i represents the i-th construction task data point in the embedding vector space; P(x i ) is the probability distribution of the embedding vector x i within its local neighborhood.

[0028] Preferably, the formula for calculating the entropy gradient density is:

[0029]

[0030] Among them, x i represents the i-th construction task data point in the embedding space; H(x i ) represents the local information entropy of the data point x i ; is the gradient vector of the information entropy at x i ; I(x i ) is the entropy gradient density index; is the modulus of the information entropy gradient, and ∈ is a smoothing factor.

[0031] Preferably, the steps of constructing a self-supervised driven embedding update mechanism include:

[0032] According to the time change of the construction task data, continuously monitor the difference between the embedding vector of the data and the historical embedding vector, calculate the difference metric function, and identify the change pattern of the data through this difference metric function;

[0033] Based on the data change pattern, dynamically adjust the mapping parameters of the embedding space, and perform self-supervised update on the embedding function through an optimization algorithm;

[0034] By comparing the updated embedding vectors with the historical embedding vectors, evaluate the change trend of the embedding model, and optimize the embedding update strategy according to the evaluation results.

[0035] Preferably, the embedding function is expressed as:

[0036] Φ s (s i ) = W·s i +b;

[0037] where s i is the standardized construction task data; W is the mapping weight matrix of the embedding space; b is the bias term; Φ s (s i ) is the vector representation after embedding.

[0038] There is also provided a standardized processing and intelligent storage architecture system for construction task data, including:

[0039] A data acquisition module for collecting and encapsulating multi-source construction task data;

[0040] A structure modeling module for selecting a structure model and performing standardized reconstruction on the data based on the minimum description length principle;

[0041] An embedding compression module for embedding the standardized data into a multi-scale information representation space;

[0042] An information density analysis module for estimating the information entropy density of the embedded data and determining the storage level;

[0043] A storage stratification module for automatically allocating the data to the hot layer, warm layer or cold layer according to the density determination;

[0044] An embedding update module for dynamically updating the embedding function based on a self-supervised mechanism to adapt to the data distribution under time changes.

[0045] The present invention provides a method for standardized processing and intelligent storage architecture of construction task data. It has the following beneficial effects:

[0046] 1. The present invention adopts a structure modeling method based on the minimum description length principle to perform unified model expression on multi-source heterogeneous construction data, achieving the technical effects of reducing redundant information and enhancing the clarity of the data structure. Compared with the existing solutions that often rely on template matching or static field mapping, it solves the problems of poor adaptability to dynamic construction scenarios and weak semantic abstraction ability.

[0047] 2. The present invention constructs a density analysis mechanism that combines information entropy and entropy gradient, which can accurately identify the distribution characteristics of data in the embedding space and achieve the technical goal of intelligently dividing storage levels. Compared with traditional static rules or manual classification storage strategies, it breaks through the limitations of fuzzy cold and hot data identification and low storage efficiency.

[0048] 3. By introducing a self-supervised embedding update mechanism, the embedding function of the present invention can be automatically adjusted over time to maintain semantic consistency with the changes in new data. In traditional methods, once the embedding expression is trained, it is fixed and prone to failure, unable to cope with real-time changes at the construction site. This solution effectively solves the problem of model rigidity of this kind.

[0049] 4. By constructing a multi-scale embedding compression mechanism, the present invention takes into account both global semantics and local details, enabling data to retain more useful information without being bloated during expression. Different from the existing solutions that use a single scale or high-dimensional redundant embeddings, this approach reduces the storage burden and improves the usability of subsequent analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flowchart of the method steps of the present invention;

[0051] Figure 2 is a system architecture diagram of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0053] Please refer to the attached Figure 1 , the embodiments of the present invention provide a method for standardizing processing and intelligent storage architecture of construction task data, including the following steps:

[0054] S1. Collect multi-source construction task data from the construction site, and perform unified format encapsulation and structure initialization processing on the collected data;

[0055] S2. Establish multiple candidate structure models based on the minimum description length principle, perform structure reconstruction on the multi-source construction task data, and select the structure model with the minimum description length to perform standardized expression on the multi-source construction task data;

[0056] S3. Embed the standardized construction task data into a multi-scale information representation space, and maintain the similarity relationship of the original data in the semantic structure during the embedding process;

[0057] S4. Estimate the local information entropy and entropy gradient density of the embedded construction task data, and divide the data into different storage levels based on the magnitude of the information density;

[0058] S5. According to the time variation of the construction task data, construct a self-supervised driven embedding update mechanism to complete the iterative optimization of the embedding function.

[0059] For step S1, in this embodiment, high-quality acquisition of multi-source heterogeneous data at the construction site is achieved, and preliminary organization in a unified format is completed to make it parsable for subsequent embedding and modeling processing. The data sources may involve construction voice records, text descriptions, working condition images, equipment logs, time series sensor data, construction video frame sequences, etc. Different types of data have characteristics such as inconsistent formats, asynchronous time dimensions, and non-uniform semantic granularity, which directly affect the structural unity for intelligent processing.

[0060] Generally, in this embodiment, data access is completed through edge acquisition nodes deployed at the construction site. The nodes include, but are not limited to, industrial camera modules, audio acquisition modules, RFID readers, laser scanners, environmental sensor arrays and other devices. All data are marked with a unified timestamp to construct a unified time axis index structure.

[0061] Specifically, in a possible implementation manner, the system organizes the original data in the form of structured triples. Each piece of data is encapsulated in the following form:

[0062] D i =<T i ,C i ,τ i >;

[0063] Among them, D i represents the i-th piece of encapsulated construction task data; T i is the field type identifier, represented by a fixed-length code indicating the data source type (such as text, image, sensor, etc.); C i is the original content or compressed representation of this piece of data, which can be an original byte stream, an image tensor, a voice spectrum vector, etc.; τ i represents the timestamp information, which is the standard UTC time when the data is collected, in milliseconds.

[0064] In some embodiments, the system will also perform basic normalization processing on the C i part, including value range standardization, image gray level adjustment, frame rate unification, etc., to improve the processing consistency in the subsequent feature extraction link.

[0065] As an option, the field type T iIt can be defined through a static mapping table. This mapping table includes contents such as field ID, data pattern, expected processing module, etc. For example:

[0066] Image data, adapted to the image convolution network;

[0067] Voice data, connected to the spectrogram generator;

[0068] Text description, used for the embedding encoding module;

[0069] Temperature and humidity sensor data, suitable for time window aggregation processing.

[0070] To ensure the feasibility of structure initialization in the subsequent modeling process, a structure tuple extension mechanism is also introduced in this embodiment. Under this mechanism, the original triple data D i will be further encapsulated into an enhanced structure tuple D′ i , and its format is:

[0071] D′ i =<T i ,C i ,τ i ,S i ,M i >;

[0072] Among them, T i is the field type identifier, represented by a fixed-length encoding to indicate the data source type (such as text, image, sensor, etc.); C i is the original content or compressed representation of this piece of data, which can be the original byte stream, image tensor, voice spectrum vector, etc.; τ i represents the timestamp information, which is the standard UTC time when this data is collected, and the unit is milliseconds; S i is the structure position identifier of the data, such as the component number it belongs to, construction stage, operation partition number, etc., serving as the initial grouping condition for structure reconstruction modeling; M i is the metadata description, including information such as data collection accuracy, device ID, sampling frequency, signal-to-noise ratio, geographical coordinates, etc.

[0073] In some embodiments, to improve the processing efficiency and data transfer speed, the structured data after encapsulation will be temporarily cached in the edge node buffer area and regularly uploaded to the central server or cloud processing system in batches. The upload strategy can be adaptively controlled based on data priority, bandwidth pressure, or scenario events.

[0074] As an extended method, the present invention also supports the lightweight preprocessing ability based on the edge computing unit. Under this implementation, the edge side can first complete part of the feature extraction, anomaly detection, and redundancy filtering, and then upload the encapsulated results.

[0075] In addition, in this embodiment, to avoid data ambiguity and structural conflicts during the encapsulation process, the system introduces a data consistency verification mechanism before encapsulation. This mechanism performs integrity analysis on the original collected data based on the field type specification and temporal continuity. If abnormal collection or field conflicts are found, the fault-tolerant retransmission mechanism is activated.

[0076] In terms of time alignment, the system uses a synchronization protocol (such as PTP) to uniformly calibrate the time of all edge nodes, ensuring that the data timestamps have a millisecond-level alignment accuracy. This mechanism plays an important supporting role in subsequent data modeling and temporal graph construction.

[0077] For step S2, in this embodiment, an optimal structure model that adapts to the current data distribution characteristics is constructed and selected. First, the system predefines a set of candidate structure model collections, and each model has different structural expression forms, field hierarchy combinations, or nesting relationships. The model can be a field-level tree, a directed acyclic graph, a nested JSON structure template, or other data representation structures suitable for the construction scenario.

[0078] Generally, the system will sequentially perform coding evaluations on each candidate structure model. Specifically, for any candidate structure model M k , its total description length L(M k , D) is calculated, including the coding length of the structure itself and the reconstruction length of the data under this model. The formula is as follows:

[0079] L(M k , D) = L s (M k ) + L d (D | M k );

[0080] Among them, L(M k , D) is the total description length; L s (M k ) represents the structural coding length of model M k , reflecting its expression complexity, which is usually related to the number of fields, hierarchical depth, reference times, etc. of the model; L d (D | M k ) represents the coding length required to map data D to model M k , reflecting the compactness degree of the model's fitting to the data.

[0081] As an option, the L s (M k ) can be evaluated for compression using methods such as Huffman coding or entropy coding to measure the structural complexity. And L d (D | M k ) can be quantified through indicators such as the compression ratio after field filling and the mapping matching loss.

[0082] In a possible implementation, the system introduces a penalty term during actual operation to control the overfitting risk, and rewrites the objective function as:

[0083] L * (M k , D) = L s (M k ) + L d (D|M k ) + λ · C(M k );

[0084] Among them, L s (M k ) represents the structural encoding length of the model M k , reflecting its expression complexity, which is usually related to the number of fields, hierarchical depth, reference times, etc. of the model; L d (D|M k ) represents the encoding length required to map the data D to the model M k , reflecting the compactness degree of the model's fitting to the data; C(M k ) represents the complexity penalty term of the structural model, which can be defined according to the number of redundant fields or the number of nested layers of the model; λ is the regularization coefficient, and its value range is generally between [0.01, 0.2], which is used to balance the model compression rate and generalization ability. After comparing the L * (M k , D) values of all candidate models, the system selects the model M * that minimizes this value as the optimal structural expression of the current data. This model will be used for subsequent field reorganization and semantic normalization mapping.

[0085] In this embodiment, after the structural modeling is completed, the system, based on the selected optimal structural model M * , performs field extraction, field sorting, and field remapping operations on all encapsulated entries D i in the dataset. The original data is re-encoded into a standardized representation of a unified structure, and the form can be defined as:

[0086]

[0087] Among them, s i is the i-th standardized construction task data; represents the normalized representation of the j-th field of the i-th data under the structural model M * , and the data type can be scalar, vector, or tensor.

[0088] In some embodiments, the system also normalizes the field vectors. For example, real number fields are standardized to the interval [0, 1], and categorical fields are one-hot encoded or index encoded to improve the convergence speed and expression consistency of subsequent embedding functions.

[0089] As a supplementary mechanism, after the structural model is selected, the system also performs a consistency check process to ensure compatibility between the selected model and historical structural versions, and to avoid subsequent module interruptions or data interpretation failures. This consistency check includes processes such as field alignment checks, missing field filling strategy confirmations, and model switching logging.

[0090] For step S3, in this embodiment, each construction task data has been mapped to a field vector s with a unified structure i . The system uses this as input to construct an embedding function Φ s , and converts it into a vector representation x i in the embedding space. The formal expression of this mapping process is as follows:

[0091] x i = Φ s (s i ) = W · s i + b;

[0092] Where, represents the i-th standardized data vector after structural modeling, with a dimension of n; is the mapping weight matrix in the embedding space, which maps the input vector from the original dimension n to the embedding dimension m; is the bias term, used to adjust the displacement deviation after linear transformation; is the final embedding vector, representing the semantic expression result in the embedding space.

[0093] Generally, the value of the embedding dimension m is set according to the task complexity and downstream model requirements, usually between 64 and 512. The embedding function Φ s can use linear transformation, or can form a multi-layer perceptron architecture by combining non-linear activation (such as ReLU, Tanh) in actual deployment to enhance the expression ability. Specifically, in order to ensure that the embedding process preserves the semantic similarity between the original data, this embodiment introduces a similarity-preserving constraint mechanism. This mechanism pulls similar data pairs closer and pushes dissimilar data pairs farther away during the training process of the embedding function Φ s , so as to maintain the semantic topological structure of the data in the embedding space.

[0094] In a possible implementation, the system constructs a triplet loss function for embedding training:

[0095]

[0096] Among them, represents the triple loss value; s a represents the anchor sample; s p represents the positive sample semantically similar to the anchor; s n represents the negative sample semantically irrelevant to the anchor; α is the margin hyperparameter, generally taken between 0.1 and 0.5; Φ s (·) represents the vector representation function after mapping the sample s to the embedding space; the symbol ∥·∥ represents the Euclidean distance metric. This loss function encourages positive samples to be closer to the anchor and negative samples to be farther from the anchor in the embedding space.

[0097] As an option, in some embodiments, the system may also use cosine similarity as the similarity metric in the embedding space, and its expression form is as follows:

[0098]

[0099] Among them, are two embedding vectors, representing the dot product of vectors; ∥·∥ represents the L2 norm of the vector. This similarity metric is used to measure the relative directions of two data samples in the semantic space.

[0100] In some embodiments, to enhance the expression ability of data at different abstraction levels, the present invention also constructs a multi-scale embedding space. The multi-scale design means mapping each standardized data to multiple embedding subspaces respectively, and each subspace focuses on different levels of feature granularity. For example:

[0101] The first-layer embedding captures the original field feature relationships;

[0102] The second-layer embedding models the context dependencies between fields;

[0103] The third-layer embedding introduces the task-level structural organization. Finally, a joint embedding representation is formed through a concatenation or fusion strategy, and the form is as follows:

[0104]

[0105] Among them, x i represents the joint embedding vector after multi-layer fusion; represents the k-th layer embedding function; Concat represents the vector concatenation operation, which is used to retain the feature distributions of each scale embedding space.

[0106] To improve the robustness of the embedding process, during the training phase, the system also introduces a noise perturbation mechanism. By injecting a small amount of random perturbation into the input vector s i the adaptability of the model to data variation is improved.

[0107] For step S4, in this embodiment, for each embedded construction task data vector x i the local information entropy is estimated. The local information entropy is used to measure the distribution uncertainty of the embedded vector in its neighborhood and is expressed as:

[0108]

[0109] where, represents the i-th embedded vector; x j is an embedded vector in the neighborhood ; represents the neighborhood set of x i in the embedding space, usually determined by the K-nearest neighbor or radius neighborhood method; P(x j |x i ) is the conditional probability of x j relative to x i , indicating the likelihood of x i appearing in the local area centered on x j ; log2 is the logarithmic function with base 2, used to calculate the bit unit of information entropy.

[0110] Generally, to avoid interference from outliers in the local entropy estimation, the system normalizes the neighborhood distribution. As an option, P(x j |x i ) can be estimated by the Gaussian kernel function, expressed as follows:

[0111]

[0112] where, represents the i-th embedded vector; x j is an embedded vector in the neighborhood ; represents the neighborhood set of x i in the embedding space, usually determined by the K-nearest neighbor or radius neighborhood method; P(x j |x i ) is the conditional probability of x j relative to x i , indicating the likelihood of x i appearing in the local area centered on x j ; ||x i - x j || represents the Euclidean distance; σ is the bandwidth parameter of the kernel function, controlling the neighborhood influence range, usually taking values between 0.1 and 1.0.

[0113] After estimating the local information entropy, this embodiment further calculates the entropy gradient density of each embedding vector to evaluate the severity of information change. The calculation formula of the entropy gradient density is as follows:

[0114]

[0115] where x i represents the i-th construction task data point in the embedding space; H(x i ) represents the local information entropy of the data point x i ; is the gradient vector of the information entropy at x i ; I(x i ) is the entropy gradient density index; is the norm of the information entropy gradient, and ∈ is the smoothing factor.

[0116] Specifically, the larger the value of the entropy gradient, the more severe the change in the information density of the data region, and it may be at the task critical node or the dynamic boundary region; when the gradient approaches zero, it indicates that the information change at this point is gentle and may belong to the stable or redundant region.

[0117] In a possible implementation, the system constructs a storage partitioning strategy with the information entropy H(x i ) and the entropy gradient density I(x i ) as indicators. The partitioning criteria can adopt the following multi-threshold determination:

[0118] If I(x i ) ≥ θ1, it is determined as "hot data", which needs to be accessed frequently and stored in the fast response layer;

[0119] If θ2 ≤ I(x i ) < θ1, it is classified as "warm data" and is suitable for the periodic call scenario;

[0120] If I(x i ) < θ2, it is classified as "cold data" and can be transferred to the archive or compressed storage layer.

[0121] where θ1 and θ2 are empirical threshold parameters, which can be dynamically adjusted according to the construction task characteristics and system load. Generally, θ1 = 0.8 and θ2 = 0.3 are acceptable initial settings.

[0122] In some embodiments, the system also combines the above information density index with the timestamp factor to construct an "information density - time decay" joint scoring model to further improve the timeliness sensitivity of data level partitioning. This method can be implemented through a weighted scoring function, which is expressed as follows:

[0123]

[0124] Among them, S(x i ) is the final hierarchical score; I(x i ) is the entropy gradient density index; t now represents the current system time; e refers to the base of the natural logarithm; τ i is the timestamp of the data; α ∈ [0, 1] controls the information density and time weight; λ is the time decay coefficient, which controls the downweighting speed of historical data.

[0125] For step S5, in this embodiment, by taking the time change of the embedding vector as the monitoring basis, a difference metric function is constructed to determine whether the current data has a significant semantic shift. Specifically, the system uses the embedding vector at the current moment t and the embedding vector at the historical moment t - Δt for comparison. The difference metric function is defined as follows:

[0126]

[0127] Among them, is the difference metric value of the i-th sample at time t; represents the embedding representation of the i-th data at the current moment; t - Δt is the historical moment; t is the current moment; Δt represents the time window length, and the unit can be minutes, hours or days; ∥·∥ represents the L2 norm, which measures the Euclidean distance between the two embedding vectors; is the embedding change amplitude of the i-th data.

[0128] Generally, when exceeds the set threshold θ d , the system considers that the semantics of the current data have undergone a substantial shift, thereby triggering the update process of the embedding function. As an option, the value range of θ d can be set from 0.1 to 1.0 and adjusted according to the embedding space dimension and task stability.

[0129] Specifically, the update of the embedding function adopts a self-supervised training strategy and does not rely on manually labeled information. The system constructs pseudo-labels from the existing historical data for training a new embedding mapping function In a possible implementation, the training objective function is the embedding stability contrast loss, which is expressed as follows:

[0130]

[0131] Among them, is the training loss for the update of the embedding function; is the structured input of the i-th data at the current moment; is the embedding function to be updated; is the embedding result of the previous round. This loss function encourages the new embedding to maintain local consistency with the old embedding and prevents representation oscillation.

[0132] In some embodiments, to further improve the robustness and generalization ability of the update strategy, the system introduces a dynamic weight regulation mechanism to differentially allocate the training weights of different samples. The formula is as follows:

[0133]

[0134] where is the difference metric value of the i-th sample at time t; is the weight coefficient of the i-th sample in the current update round; is the sum of the difference metric values of all samples at time t; the weight value is proportional to the degree of embedding change, ensuring that significant data is focused on.

[0135] After the update is completed, the system verifies the vector results generated by the new embedding function. The verification methods include metrics such as distribution comparison with historical vectors, consistency analysis of embedding clustering structures, and semantic retention evaluation. When performance degradation or abnormal distribution occurs, the system can execute a parameter rollback mechanism to restore to the previous version of the embedding function to ensure the stability of the system.

[0136] As a supplementary mechanism, a time window sliding strategy is also designed in this embodiment to make the embedding update coherent and adaptable. The sliding of the time window can be triggered based on a fixed period (such as once a day) or event-driven (such as information density mutation, sharp increase in data scale) to achieve real-time response to changes at the construction site.

[0137] A system for standardizing processing and intelligent storage architecture of construction task data described below can be correspondingly referred to with a method for standardizing processing and intelligent storage architecture of construction task data described above.

[0138] Please refer to the appendix Figure 2 , the present invention also provides a system for standardizing processing and intelligent storage architecture of construction task data, including:

[0139] A data acquisition module for collecting and encapsulating multi-source construction task data; the data types can include structured progress plans, unstructured text records, image or video information, sensor logs, etc. Generally, the module performs standardized encapsulation through a unified data interface to support data synchronization and docking between the edge side, field devices, and cloud platforms.

[0140] In some embodiments, the data acquisition module has a built-in data preprocessing mechanism, including outlier removal, timestamp alignment, and field standardization operations to improve downstream processing efficiency.

[0141] A structure modeling module for selecting a structure model and performing standardized reconstruction on data based on the principle of minimum description length. Specifically, the module uses a greedy search or compression rate optimization strategy to generate a data structure tree, and performs semantic mapping and simplification modeling on the relationships between nodes.

[0142] As an option, this module supports methods such as interpretable causal graph modeling and template matching reconstruction to adapt to the data expression requirements in different construction business scenarios.

[0143] An embedding compression module for embedding the standardized data into a multi-scale information representation space. In a possible implementation, this module uses a graph neural network, multi-scale or self-attention mechanism to generate task embedding vectors.

[0144] The output result of the module can be used as a basis for semantically unified representation for subsequent tasks such as density analysis and semantic retrieval.

[0145] An information density analysis module for estimating the information entropy density of the embedded data and determining the storage level. Generally, this module combines factors such as timestamps and task contexts, and judges the importance and usage frequency of the data through information density formulas and trend analysis.

[0146] In some embodiments, the module also introduces an attention weighting mechanism to enhance the local sensitivity of density estimation.

[0147] A storage stratification module for automatically allocating data to the hot layer, warm layer or cold layer according to the density determination. The module can be docked with a distributed object storage system to achieve dynamic scheduling and data migration based on tags and density metrics.

[0148] As an option, the module has historical access statistics and migration backtracking functions to enhance the time-effect adaptation ability of the data stratification strategy.

[0149] An embedding update module for dynamically updating the embedding function based on a self-supervised mechanism to adapt to the data distribution under time changes. Specifically, the module generates pseudo-labels and optimizes the embedding network according to the embedding difference metric or semantic drift signal to make its embedding expression consistent with the time evolution.

[0150] In a possible implementation, the module supports incremental learning, elastic rollback and model cold start mechanisms to ensure the stability and continuity of embedding updates.

[0151] The system of this embodiment can be used to execute the method embodiments described above. The principles and technical effects are similar and will not be elaborated here.

[0152] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for standardizing the processing and intelligent storage architecture of construction task data, characterized in that, Including the following steps: Collect multi-source construction task data from the construction site, and perform unified format encapsulation and structure initialization processing on the collected data; Based on the principle of minimum description length, establish multiple candidate structure models, perform structure reconstruction on the multi-source construction task data, and select the structure model with the minimum description length to perform standardized expression on the multi-source construction task data; Embed the standardized construction task data into a multi-scale information representation space, and keep the similarity relationship of the original data in the semantic structure during the embedding process; Estimate the local information entropy and entropy gradient density of the embedded construction task data, and divide the data into different storage levels based on the magnitude of the information density; According to the time change of the construction task data, construct a self-supervised driven embedding update mechanism to complete the iterative optimization of the embedding function.

2. The standardized processing and intelligent storage architecture method for construction task data according to claim 1, wherein, The multi-source construction task data includes text data, image data and sensor data. The unified format encapsulation adopts a structured tuple form, and each piece of data is encapsulated into a triple structure including field type, data content and timestamp.

3. The standardized processing and intelligent storage architecture method for construction task data according to claim 1, characterized in that, The steps of performing standardized expression on the multi-source construction task data include: Calculate the corresponding structure coding length and data coding length for each candidate structure model, and add the two to obtain the total description length of the structure model; Compare the total description lengths of all candidate structure models, and select the structure model with the minimum description length as the optimal structure model; Based on the optimal structure model, perform field extraction, recombination and mapping on the multi-source construction task data to obtain a unified structured data representation form.

4. A method for standardizing processing and intelligent storage architecture of construction task data according to claim 1, characterized in that The steps of embedding the standardized construction task data into a multi-scale information representation space include: By constructing a multi-level mapping function, map the standardized construction task data to multiple scales of embedding spaces to capture different levels of features of the data; During the embedding process, adopt a similarity preservation method to ensure that the construction task data maintains semantic similarity and structural similarity between the original data after embedding; The mapping function includes, but is not limited to, similarity metrics based on Euclidean distance, cosine similarity or Gaussian kernel function, ensuring that the relative positions of the data in the embedding space reflect the semantic relationships of the original data.

5. A method for standardizing processing and intelligent storage architecture of construction task data according to claim 1, characterized in that, The steps of dividing the data into different storage levels based on the magnitude of the information density include: Estimate the local information entropy of each embedding vector, and the local information entropy is expressed as the uncertainty of the probability distribution of the embedding vector in the current information space; Based on the local information entropy of the embedding vector, calculate the entropy gradient density of each embedding vector, and the entropy gradient density is expressed as the change rate of entropy in the information space; According to the calculated entropy gradient density, divide the embedded construction task data into different storage levels, including hot layer, warm layer and cold layer. The division of the storage levels is based on the magnitude of the entropy gradient density.

6. A method for standardizing processing and intelligent storage architecture of construction task data according to claim 5, characterized in that, The formula for calculating the local information entropy is: H(x i ) = -log2P(x i ); Among them, H(x i ) represents the local information entropy of the data point x i ; x i represents the i-th construction task data point in the embedded vector space; P(x i ) is the probability distribution of the embedded vector x i within its local neighborhood.

7. A method for standardizing the processing and intelligent storage architecture of construction task data according to claim 5, characterized in that, The formula for calculating the entropy gradient density is: where x i represents the i-th construction task data point in the embedding space; H(x i ) represents the local information entropy of the data point x i ; is the gradient vector of the information entropy at x i ; I(x i ) is the entropy gradient density index; is the magnitude of the information entropy gradient, and ∈ is the smoothing factor.

8. A method for standardizing processing and intelligent storage architecture of construction task data according to claim 1, characterized in that The steps of constructing a self-supervised driven embedding update mechanism include: According to the time variation of the construction task data, monitor the difference between the embedding vector of the real-time monitoring data and the historical embedding vector, calculate the difference metric function, and identify the data change pattern through this difference metric function; Based on the data change pattern, dynamically adjust the mapping parameters of the embedding space, and perform self-supervised update on the embedding function through an optimization algorithm; By comparing the updated embedding vector with the historical embedding vector, evaluate the change trend of the embedding model, and optimize the embedding update strategy according to the evaluation result.

9. The standardized processing and intelligent storage architecture method for construction task data according to claim 1, characterized in that, The embedding function is expressed as: Φ s (s i ) = W·s i + b; Among them, s i is the standardized construction task data; W is the mapping weight matrix of the embedding space; b is the bias term; Φ s (s i ) is the vector representation after embedding.

10. A standardized processing and intelligent storage architecture system for construction task data, which is applied to the method for standardized processing and intelligent storage architecture of construction task data according to any one of claims 1-9, and is characterized in that, Including: A data acquisition module for acquiring and encapsulating multi-source construction task data; A structure modeling module for selecting a structure model and performing standardized reconstruction on the data based on the principle of minimum description length; An embedding compression module for embedding the standardized data into a multi-scale information representation space; An information density analysis module for estimating the information entropy density of the embedded data and judging the storage level; A storage stratification module for automatically allocating the data to the hot layer, warm layer or cold layer according to the density determination; An embedding update module for dynamically updating the embedding function based on a self-supervised mechanism to adapt to the data distribution under time variation.

Images