Smart construction site safety evaluation method and system based on data elements

Abstract

The present invention discloses a smart construction site safety evaluation method and system based on data elements, which relates to the technical field of computer platform load balancing, including collecting multi-source heterogeneous data of the construction site, pre-processing the multi-source heterogeneous data of the construction site; combining the long short-term memory network with the graph neural network to construct a safety evaluation model that integrates time series modeling and graph structure, performing feature selection and dimension compression operations on the processed multi-source heterogeneous data of the construction site to obtain input feature tensors, and inputting the input feature tensors into the safety evaluation model; calculating a comprehensive score according to the output results of the model, and generating classification results and data feedback through edge nodes. The method of the present invention The smart construction site safety evaluation method based on data elements provided by the present invention realizes joint modeling of time series dependency and spatial correlation by integrating the deep modeling structure of LSTM and GNN, effectively improving the risk identification accuracy under multi-source heterogeneous data.

Description

Technical Field

[0001] The present invention relates to the technical field of construction engineering informatization and intelligent safety management, and specifically to a smart construction site safety evaluation method and system based on data elements. Background Art

[0002] As the construction industry continues to improve its informatization, smart construction sites, as a crucial component of digital construction, have been widely adopted in large-scale infrastructure, urban renewal, and other projects. Leveraging technologies such as the Internet of Things, big data, and artificial intelligence, these smart construction sites enable real-time perception, dynamic scheduling, and data-driven management of construction sites. In particular, in terms of safety management, the deployment of various sensors, video surveillance equipment, and behavior recording terminals enables full-cycle monitoring of construction personnel, machinery, and the on-site environment.

[0003] The existing smart construction site safety evaluation system still has the following problems: most current safety evaluation methods rely on static indicators or manual inspection data, and fail to effectively integrate multi-source heterogeneous data, such as video images, text logs, sensor signals, etc., resulting in limited evaluation dimensions and a single data structure. Safety evaluation models mostly use rule-driven or single algorithms for scoring, lack the ability to model complex correlations and dynamic changes between data, and are difficult to adapt to the needs of integrating high-frequency, nonlinear, and unstructured data. Some systems only use fixed thresholds for risk assessment and lack adaptive adjustment mechanisms for different construction stages, scenario environments, and risk types, which can easily lead to false positives or missed reports. At the model deployment level, most of them are concentrated in cloud processing, and the response timeliness is insufficient, which cannot meet the requirements of localized rapid response and dynamic adjustment to major safety hazards.

[0004] Therefore, there is an urgent need for an intelligent safety assessment method and system for smart construction sites that can integrate multi-source heterogeneous data, has adaptive feature extraction and risk modeling capabilities, and supports edge deployment and multi-task output, so as to improve the safety management level and risk response efficiency of construction sites. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by the present invention is: the existing smart construction site safety evaluation methods have weak data fusion capabilities, low model generalization capabilities, and slow risk identification response, as well as how to construct a safety evaluation model with time series and structure joint modeling capabilities based on multi-source heterogeneous data, and realize efficient and explainable intelligent risk classification and feedback.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: a smart construction site safety evaluation method based on data elements, including collecting multi-source heterogeneous data on the construction site.

[0008] Preprocess multi-source heterogeneous data at the construction site.

[0009] By combining long short-term memory networks with graph neural networks, a security evaluation model that integrates time series modeling and graph structure is constructed.

[0010] Feature selection and dimension compression operations are performed on the processed multi-source heterogeneous data of the construction site to obtain the input feature tensor, which is then input into the safety assessment model.

[0011] The comprehensive score is calculated based on the model output results, and the classification results and data feedback are generated through the edge nodes.

[0012] Performing feature selection and dimensionality compression operations to obtain the input feature tensor includes extracting the semantic and structural features of text and images through BERT and ResNet respectively, and then concatenating them with the structured features after PCA dimensionality reduction to form a fused feature vector.

[0013] The Pearson correlation coefficient is used to filter redundant features, LASSO regression is used to achieve sparse selection, and random forest sorting is used to screen key variables to form a unified input feature tensor. This is then input into a security assessment model that integrates LSTM and GNN structures, and the channel attention mechanism is combined to achieve multi-task prediction output.

[0014] The channel attention mechanism involves assigning dynamic weights based on the importance of different modal features in the task objectives.

[0015] As a preferred solution of the data element-based smart construction site safety evaluation method described in the present invention, the collection of multi-source heterogeneous data on the construction site includes collecting personnel attendance data, video surveillance data, mechanical equipment operating status, environmental monitoring data, water and electricity monitoring data, and safety inspection records as data sources.

[0016] A unified data access interface is used for protocol conversion, and a high-precision clock synchronization mechanism is adopted to timestamp various data streams. The timestamp accuracy is controlled within seconds, and an edge buffer mechanism is used in high-frequency data collection.

[0017] As a preferred solution of the data element-based smart construction site safety evaluation method described in the present invention, the preprocessing of multi-source heterogeneous data on the construction site includes performing normalization, anomaly elimination and time alignment processing on the collected multi-source heterogeneous data on the construction site.

[0018] Normalization processing includes mapping the values ​​of each dimension to the interval [0,1] using the minimum-maximum scaling method;

[0019] Anomaly elimination includes processing based on the Z-score method, setting a deviation of three times the standard deviation as the threshold to mark and eliminate outliers.

[0020] Time alignment processing includes time window mapping of data from different sources based on the dynamic time warping algorithm, filling in missing time point data with a linear interpolation algorithm, and uniformly constructing time series structure data according to a preset sliding window structure.

[0021] As a preferred solution of the data element-based smart construction site safety evaluation method described in the present invention, the safety evaluation model that integrates time series modeling and graph structure is constructed, including using the long short-term memory network LSTM to process normalized time series data to capture the temporal dependency of construction behavior, equipment operation and environmental fluctuations.

[0022] A graph neural network (GNN) model is constructed based on construction site entities, with personnel, equipment, risk points, and environmental sources as nodes of the graph. The weights of the edges between nodes are set according to the historical co-occurrence probability and event interaction frequency.

[0023] The outputs of LSTM and GNN are spliced ​​through the fusion layer respectively, and the spliced ​​outputs of LSTM and GNN are used as the input of the model prediction layer. Joint modeling and multi-task output are performed through the input of the model prediction layer.

[0024] Splicing through fusion layers includes introducing a feature channel attention mechanism to assign dynamic weights to different modal information based on task-specific loss functions.

[0025] The multi-task output includes the predicted frequency of violation, abnormal state probability and risk coefficient vector.

[0026] The risk coefficient vector includes risk measurements in multiple dimensions, including construction personnel status, abnormal environmental parameters, and abnormal equipment operation.

[0027] As a preferred solution of the data element-based smart construction site safety evaluation method described in the present invention, the input feature tensor is obtained, including embedding the text features and image features in the pre-processed multi-source heterogeneous data using BERT encoding and ResNet convolutional network to extract semantic vectors and structural graph features.

[0028] BERT encoding involves using a pre-trained language model based on the Transformer architecture to perform semantic vectorization on text data, and using BERT encoding to perform high-dimensional semantic modeling on the natural language text of construction logs and safety inspection records.

[0029] The ResNet convolutional network includes feature extraction of image data such as video surveillance images, and alleviates the gradient disappearance problem in deep network training through the residual connection structure.

[0030] A principal component analysis dimensionality reduction operation is performed on the feature subsets of text features and image features to compress the embedding dimension to within 64 dimensions.

[0031] Principal component analysis involves extracting principal components by constructing a feature covariance matrix, retaining the most important direction of change between features, reducing dimensions and retaining information.

[0032] The structured class features, compressed semantic features and image features are uniformly spliced ​​to construct a fused feature vector, and feature screening operations are performed based on the fused feature vector.

[0033] The feature selection operation includes calculating the Pearson correlation coefficient between each feature, identifying redundant feature pairs with an absolute correlation value greater than 0.8, and retaining one of them.

[0034] On the feature set after redundant filtering, the LASSO regression algorithm is introduced to construct a sparse feature selector with L1 regularization term to screen out features that do not contribute significantly to the model response variable.

[0035] The random forest algorithm is applied to the retained features to calculate the feature importance scores, and the top 20 percent of the variables are selected to form the final input feature tensor.

[0036] The feature set is used as the fused input feature tensor and input into the safety assessment model for joint modeling and classification prediction.

[0037] The input feature tensor includes a multimodal feature set that has undergone preprocessing, embedded coding, dimensionality reduction compression, and screening, and a high-dimensional structured input matrix formed after dimensional unification and structural rearrangement.

[0038] Structured features include subsets of multi-source data represented in the form of numerical values ​​and time series.

[0039] As a preferred solution of the data element-based smart construction site safety evaluation method described in the present invention, the calculation of the comprehensive score includes weighted combination of the results of multiple task outputs, setting a comprehensive score calculation formula, and setting an interval level mapping according to the score results. The interval level mapping is divided into three levels: A, B, and C.

[0040] Grade A indicates excellent condition, Grade B indicates controllable risks, and Grade C indicates obvious safety hazards that need to be rectified.

[0041] As a preferred solution of the data element-based smart construction site safety evaluation method described in the present invention, the generation of classification results and data feedback includes compressing the trained safety evaluation model into a lightweight model and then deploying it to the construction site edge server. The edge server receives sensor data streams and data batches, and performs real-time reasoning operations by calling the deployed model. The reasoning cycle is no more than once per hour.

[0042] When any output classification level in the inference result is A, the score and summary information are written to the local log as long-term learning data for the model.

[0043] When any output classification level in the inference result is B, it is recorded. When B level appears for three consecutive cycles, the system determines it as a risk trend state and performs feedback operation.

[0044] When any output classification level in the inference result is C, the data feedback mechanism is triggered, and the current scoring level, risk dimension distribution and data segment identifier are sent to the remote smart construction site platform through the configured HTTP interface, and recorded in the edge log database for subsequent audit and model correction sampling.

[0045] Another object of the present invention is to provide a smart construction site safety evaluation system based on data elements, which can solve the problems of current smart construction site safety management technology in low efficiency of heterogeneous data processing, high model response delay and lack of risk trend prediction ability through real-time reasoning and multi-dimensional feedback mechanism of lightweight models deployed at the edge.

[0046] As a preferred solution of the data element-based smart construction site safety evaluation system described in the present invention, it includes: a data acquisition module, a data preprocessing module, a model building module, a safety evaluation module, and a scoring and data feedback module.

[0047] The data acquisition module is used to collect multi-source heterogeneous data on the construction site.

[0048] The data preprocessing module is used to preprocess multi-source heterogeneous data on the construction site.

[0049] The model building module is used to build a security evaluation model that integrates time series modeling and graph structure by combining long short-term memory networks with graph neural networks.

[0050] The safety assessment module is used to perform feature selection and dimension compression operations on the processed multi-source heterogeneous data of the construction site to obtain an input feature tensor, and input the input feature tensor into the safety assessment model.

[0051] The scoring and data feedback generation module is used to calculate the comprehensive score based on the model output results, and generate classification results and data feedback through edge nodes.

[0052] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a step of a smart construction site safety evaluation method based on data elements.

[0053] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a smart construction site safety evaluation method based on data elements.

[0054] Beneficial effects of the present invention: The data element-based smart construction site safety evaluation method provided by the present invention realizes the joint modeling of temporal dependency and spatial correlation by integrating the deep modeling structure of LSTM and GNN, effectively improving the risk identification accuracy under multi-source heterogeneous data; by introducing BERT and ResNet to perform deep semantic extraction of text and image information, and combining the channel attention mechanism to optimize the feature weight distribution, the model's adaptability to complex site environments is improved; by deploying lightweight inference models and hierarchical feedback mechanisms on the edge, the system's performance in response speed and local handling capabilities is improved. The present invention achieves better results in data fusion efficiency, model prediction accuracy and safety response timeliness. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0056] Figure 1 An overall flow chart of the data element-based smart construction site safety assessment method provided for the first embodiment of the present invention. DETAILED DESCRIPTION

[0057] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0058] Example 1, with reference to Figure 1 , which is an embodiment of the present invention, provides a smart construction site safety assessment method based on data elements, including:

[0059] S1: Collect multi-source heterogeneous data on the construction site.

[0060] Collect personnel attendance data, video surveillance data, mechanical equipment operating status, environmental monitoring data, water and electricity monitoring data, and safety inspection records as data sources.

[0061] A unified data access interface is used for protocol conversion, and a high-precision clock synchronization mechanism is adopted to timestamp various data streams. The timestamp accuracy is controlled within seconds, and an edge buffer mechanism is used in high-frequency data collection.

[0062] S2: Preprocess multi-source heterogeneous data on the construction site.

[0063] Normalization, anomaly elimination and time alignment are performed on the multi-source heterogeneous data collected at the construction site.

[0064] Normalization involves mapping the values ​​of each dimension to the interval [0, 1] using the minimum-maximum scaling method.

[0065] A preferred solution for normalization is:

[0066] ;

[0067] in, Represents the normalized eigenvalues, Indicates the data values, Indicates the minimum value of the feature dimension in the sample set, Indicates the maximum value of the feature dimension in the sample set.

[0068] Anomaly elimination includes processing based on the Z-score method, setting a deviation of three times the standard deviation as the threshold to mark and eliminate outliers.

[0069] A preferred solution based on the Z-score method is:

[0070] ;

[0071] in, Indicates the The Z score of the sample points, represents the sample mean, represents the sample standard deviation.

[0072] Set The absolute value of the Z score of each sample point cannot exceed 3. If it exceeds 3, it is considered as abnormal data.

[0073] Time alignment processing includes time window mapping of data from different sources based on the dynamic time warping algorithm, filling in missing time point data with a linear interpolation algorithm, and uniformly constructing time series structure data according to a preset sliding window structure.

[0074] S3: By combining long short-term memory networks with graph neural networks, a security assessment model that integrates time series modeling and graph structure is constructed.

[0075] The long short-term memory (LSTM) network is used to process normalized time series data to capture the temporal dependencies among construction behavior, equipment operation, and environmental fluctuations.

[0076] A graph neural network (GNN) model is constructed based on construction site entities, with personnel, equipment, risk points, and environmental sources as nodes of the graph. The weights of the edges between nodes are set according to the historical co-occurrence probability and event interaction frequency.

[0077] The outputs of LSTM and GNN are spliced ​​through the fusion layer respectively, and the spliced ​​outputs of LSTM and GNN are used as the input of the model prediction layer. Joint modeling and multi-task output are performed through the input of the model prediction layer.

[0078] Splicing through fusion layers includes introducing a feature channel attention mechanism to assign dynamic weights to different modal information based on task-specific loss functions.

[0079] ;

[0080] in, Represents the weighted The fusion characteristics of the channel, Indicates the original The eigenvector of the channel, represents the channel weight coefficient, Indicates a task The corresponding loss function value is, Indicates the total number of tasks or modal channels, Indicates the channel index.

[0081] It should be noted that the number of channels is equal to the number of tasks.

[0082] The multi-task output includes the predicted frequency of violation, abnormal state probability and risk coefficient vector.

[0083] The risk coefficient vector includes risk measurements in multiple dimensions, including construction personnel status, abnormal environmental parameters, and abnormal equipment operation.

[0084] S4: Perform feature selection and dimension compression operations on the processed multi-source heterogeneous data of the construction site to obtain the input feature tensor, and input the input feature tensor into the safety assessment model.

[0085] The text and image features in the preprocessed multi-source heterogeneous data are embedded using BERT encoding and ResNet convolutional network to extract semantic vectors and structural graph features.

[0086] BERT encoding involves using a pre-trained language model based on the Transformer architecture to perform semantic vectorization on text data, and using BERT encoding to perform high-dimensional semantic modeling on the natural language text of construction logs and safety inspection records.

[0087] The ResNet convolutional network includes feature extraction of image data such as video surveillance images, and alleviates the gradient disappearance problem in deep network training through the residual connection structure.

[0088] A principal component analysis dimensionality reduction operation is performed on the feature subsets of text features and image features to compress the embedding dimension to within 64 dimensions.

[0089] An optimal solution for principal component analysis dimensionality reduction is:

[0090] ;

[0091] in, represents the original high-dimensional feature matrix, Represents the principal component transformation matrix calculated by PCA, Represents the feature representation after dimensionality reduction.

[0092] The feature representation after dimensionality reduction is obtained by eigendecomposition of the covariance matrix.

[0093] Principal component analysis involves extracting principal components by constructing a feature covariance matrix, retaining the most important direction of change between features, reducing dimensions and retaining information.

[0094] The structured class features, compressed semantic features and image features are uniformly spliced ​​to construct a fused feature vector, and feature screening operations are performed based on the fused feature vector.

[0095] The feature selection operation includes calculating the Pearson correlation coefficient between each feature, identifying redundant feature pairs with an absolute correlation value greater than 0.8, and retaining one of them.

[0096] A preferred method for calculating the Pearson correlation coefficient between the features is:

[0097] ;

[0098] in, Representation characteristics and features The correlation coefficient of Representation characteristics Middle The corresponding eigenvalues ​​of the samples after normalization, 、 represents the mean of each feature, the numerator is the covariance term, and the denominator is the product of standard deviations.

[0099] On the feature set after redundant filtering, the LASSO regression algorithm is introduced to construct a sparse feature selector with L1 regularization term to screen out features that do not contribute significantly to the model response variable.

[0100] An optimal solution for constructing a sparse feature selector with an L1 regularization term is:

[0101] ;

[0102] in, Indicates the The input feature vector of samples, Indicates the The regression coefficient of the feature, represents the L1 regularization hyperparameter, represents the number of samples, Indicates the number of feature dimensions.

[0103] The random forest algorithm is applied to the retained features to calculate the feature importance scores, and the top 20 percent of the variables are selected to form the final input feature tensor.

[0104] A preferred solution for calculating feature importance scores using the random forest algorithm is:

[0105] ;

[0106] in, represents the complete feature set, Indicates the Features, Indicates the The ranking of features sorted by importance, represents the total number of features, Represents the top 20% key feature set retained.

[0107] The feature set is used as the fused input feature tensor and input into the safety assessment model for joint modeling and classification prediction.

[0108] The input feature tensor includes a multimodal feature set that has undergone preprocessing, embedded coding, dimensionality reduction compression, and screening, and a high-dimensional structured input matrix formed after dimensional unification and structural rearrangement.

[0109] Structured features include subsets of multi-source data represented in the form of numerical values ​​and time series.

[0110] S5: Calculate the comprehensive score based on the model output results, and generate classification results and data feedback through edge nodes.

[0111] The results of multi-task output are weighted and combined, and a comprehensive scoring calculation formula is set. The interval level mapping is set according to the scoring results. The interval level mapping is divided into three levels: A, B, and C.

[0112] A preferred method for calculating the comprehensive score is:

[0113] ;

[0114] in, Indicates the final comprehensive score, represents the personnel behavior sub-score, Indicates the equipment operation sub-score, represents the environmental status sub-score, , , Indicates the weight coefficient corresponding to each scoring item, , , The sum is equal to 1.

[0115] A preferred solution for setting interval level mapping based on the scoring results is:

[0116] ;

[0117] Among them, Grade A indicates excellent condition, Grade B indicates controllable risks, and Grade C indicates obvious safety hazards that need to be rectified.

[0118] The trained safety assessment model is compressed and converted into a lightweight model and then deployed to the edge server on the construction site. The edge server receives sensor data streams and data batches, and performs real-time inference operations by calling the deployed model. The inference cycle is no more than once per hour.

[0119] When any output classification level in the inference result is A, the score and summary information are written to the local log as long-term learning data for the model.

[0120] When any output classification level in the inference result is B, it is recorded. When B level appears for three consecutive cycles, the system determines it as a risk trend state and performs feedback operation.

[0121] When any output classification level in the inference result is C, the data feedback mechanism is triggered, and the current scoring level, risk dimension distribution and data segment identifier are sent to the remote smart construction site platform through the configured HTTP interface, and recorded in the edge log database for subsequent audit and model correction sampling.

[0122] Example 2 is an embodiment of the present invention, which provides a smart construction site safety evaluation system based on data elements, including a data acquisition module, a data preprocessing module, a model building module, a safety evaluation module, and a scoring and data feedback module.

[0123] The data acquisition module is used to collect multi-source heterogeneous data at the construction site.

[0124] The data preprocessing module is used to preprocess multi-source heterogeneous data at the construction site.

[0125] The model building module is used to build a security evaluation model that integrates time series modeling and graph structure by combining long short-term memory networks with graph neural networks.

[0126] The safety evaluation module is used to perform feature selection and dimension compression operations on the processed multi-source heterogeneous data of the construction site to obtain the input feature tensor, and input the input feature tensor into the safety evaluation model.

[0127] The scoring and data feedback generation module is used to calculate the comprehensive score based on the model output results, and generate classification results and data feedback through edge nodes.

Claims

1. A smart construction site safety assessment method based on data elements, characterized by: include: Collect multi-source heterogeneous data on the construction site; Preprocess multi-source heterogeneous data on construction sites; By combining long short-term memory networks with graph neural networks, a security assessment model integrating time series modeling and graph structure is constructed; Perform feature selection and dimension compression on the processed multi-source heterogeneous data of the construction site to obtain the input feature tensor, and input the input feature tensor into the safety assessment model; Calculate the comprehensive score based on the model output results, and generate classification results and data feedback through edge nodes; Perform feature selection and dimensionality compression operations to obtain the input feature tensor, including extracting the semantic and structural features of text and image through BERT and ResNet respectively, and then concatenating them with the structural features after PCA dimensionality reduction to form a fused feature vector; The Pearson correlation coefficient is used to filter redundant features, LASSO regression is used for sparse selection, and random forest sorting is used to screen key variables. This results in a unified input feature tensor, which is then fed into a safety assessment model that integrates LSTM and GNN structures. The channel attention mechanism is then used to achieve multi-task prediction output. The channel attention mechanism includes assigning dynamic weights according to the importance of different modal features under the task objectives; The construction of a security evaluation model integrating time series modeling and graph structure includes: Use the long short-term memory (LSTM) network to process normalized time series data and capture the temporal dependencies between construction behavior, equipment operation, and environmental fluctuations. A graph neural network (GNN) model is constructed based on construction site entities, with personnel, equipment, risk points, and environmental sources as nodes of the graph. The weights of the edges between nodes are set based on the historical co-occurrence probability and the frequency of event interaction. The outputs of LSTM and GNN are spliced ​​through the fusion layer respectively, and the spliced ​​LSTM and GNN outputs are used as the input of the model prediction layer. Joint modeling and multi-task output are performed through the input of the model prediction layer; The concatenation through the fusion layer includes introducing a feature channel attention mechanism to assign dynamic weights to different modal information based on a task-specific loss function; The multi-task output includes the predicted frequency of violation behavior, abnormal state probability and risk coefficient vector; The risk coefficient vector includes risk measurements in multiple dimensions, including construction personnel status, abnormal environmental parameters, and abnormal equipment operation.

2. The data element-based smart construction site safety assessment method according to claim 1, characterized in that: The collection of multi-source heterogeneous data on the construction site includes: Collect personnel attendance data, video surveillance data, mechanical equipment operating status, environmental monitoring data, water and electricity monitoring data, and safety inspection records as data sources; A unified data access interface is used for protocol conversion, and a high-precision clock synchronization mechanism is adopted to timestamp various data streams. The timestamp accuracy is controlled within seconds, and an edge buffer mechanism is used in high-frequency data collection.

3. The data element-based smart construction site safety assessment method according to claim 2, characterized in that: The preprocessing of multi-source heterogeneous data on the construction site includes: Perform normalization, anomaly elimination, and time alignment on the multi-source heterogeneous data collected at the construction site; Normalization processing includes mapping the values ​​of each dimension to the interval [0,1] using the minimum-maximum scaling method; Abnormal elimination includes processing based on the Z-score method, setting a deviation of three times the standard deviation as the threshold to mark and eliminate outliers; Time alignment processing includes time window mapping of data from different sources based on the dynamic time warping algorithm, filling in missing time point data with a linear interpolation algorithm, and uniformly constructing time series structure data according to a preset sliding window structure.

4. The data element-based smart construction site safety assessment method according to claim 3, characterized in that: The obtaining of the input feature tensor includes: The text and image features in the pre-processed multi-source heterogeneous data are embedded using BERT encoding and ResNet convolutional network to extract semantic vectors and structural graph features; BERT encoding involves using a pre-trained language model based on the Transformer architecture to perform semantic vectorization on text data. BERT encoding is used to perform high-dimensional semantic modeling on natural language text in construction logs and safety inspection records. The ResNet convolutional network includes feature extraction for image data such as video surveillance images, and alleviates the gradient vanishing problem in deep network training through the residual connection structure; Perform principal component analysis dimensionality reduction on the feature subsets of text features and image features to compress the embedding dimension to within 64 dimensions; Principal component analysis involves extracting principal components by constructing a feature covariance matrix, retaining the most important direction of change between features, reducing dimensions and retaining information; The structured class features, compressed semantic features and image features are uniformly spliced ​​to construct a fused feature vector, and feature screening operations are performed based on the fused feature vector; The feature screening operation includes calculating the Pearson correlation coefficient between each feature, identifying redundant feature pairs with an absolute correlation value greater than 0.8, and retaining one of them; On the feature set after redundant filtering, the LASSO regression algorithm is introduced to construct a sparse feature selector with an L1 regularization term to filter out features that do not contribute significantly to the model response variable; Apply the random forest algorithm to the retained features to calculate the feature importance score, and select the top 20% of the variables to form the final input feature tensor; The feature set is used as the fused input feature tensor and input into the safety assessment model for joint modeling and classification prediction; Input feature tensor Including, after preprocessing, embedding coding, dimensionality reduction compression and screening of multimodal feature sets, and after dimension unification and structural rearrangement, a high-dimensional structured input matrix is ​​formed; Structured features include subsets of multi-source data represented in the form of numerical values ​​and time series.

5. The data element-based smart construction site safety assessment method according to claim 4, characterized in that: The calculation of the comprehensive score includes, Perform weighted combination of the results of multi-task output, set the comprehensive scoring calculation formula, and set the interval level mapping according to the scoring results. The interval level mapping is divided into three levels: A, B, and C; Grade A indicates excellent condition, Grade B indicates controllable risks, and Grade C indicates obvious safety hazards that need to be rectified.

6. The data element-based smart construction site safety assessment method according to claim 5, characterized in that: Generating classification results and data feedback includes: The trained safety assessment model is compressed and converted into a lightweight model and then deployed to the edge server on the construction site. The edge server receives sensor data streams and data batches, and performs real-time inference operations by calling the deployed model. The inference cycle is no more than once per hour. When any output classification level in the inference result is A, the score and summary information are written to the local log as long-term learning data for the model; When any output classification level in the inference result is B, it is recorded. When B level appears for three consecutive cycles, the system determines it as a risk trend state and performs feedback operation; When any output classification level in the inference result is C, the data feedback mechanism is triggered, and the current scoring level, risk dimension distribution and data segment identifier are sent to the remote smart construction site platform through the configured HTTP interface, and recorded in the edge log database for subsequent audit and model correction sampling.

7. A system using the data element-based smart construction site safety assessment method according to any one of claims 1 to 6, characterized in that: It includes data acquisition module, data preprocessing module, model building module, safety evaluation module, scoring generation and data feedback module; The data acquisition module is used to collect multi-source heterogeneous data on the construction site; The data preprocessing module is used to preprocess multi-source heterogeneous data on the construction site; The model building module is used to build a security assessment model that integrates time series modeling and graph structure by combining long short-term memory networks with graph neural networks; The safety assessment module is used to perform feature selection and dimension compression operations on the processed multi-source heterogeneous data of the construction site to obtain an input feature tensor, and input the input feature tensor into the safety assessment model; The scoring and data feedback generation module is used to calculate the comprehensive score based on the model output results, and generate classification results and data feedback through edge nodes.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the smart construction site safety evaluation method based on data elements described in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the smart construction site safety evaluation method based on data elements described in any one of claims 1 to 6 are implemented.

Images