A building construction safety monitoring method based on big data collection

Abstract

The application discloses a kind of building construction safety monitoring methods based on big data acquisition, it is related to big data technical field, comprising: acquisition and pre-processing construction monitoring data, form trusted construction state;According to object, region, task and time, sample and associated record are constructed;State sample is input into improved DAGMM model, obtains reconfiguration state, initial risk energy and state migration code;Calculate reconfiguration residual, construct risk residual manifold;Extract state transition position and risk duration section, construct risk phase encoding body;Perform transfer entropy algorithm, obtain risk transfer record;Risk transfer record is written back to improved DAGMM model, reconstructs latent variable association, calculates associated risk energy and generates propagation path;Determine risk source, influence object, occurrence area and propagation time, generate safety warning result and monitoring record.The application realizes construction risk accurate identification and propagation traceability, improves safety warning efficiency.

Description

Technical Field

[0001] This invention relates to the field of big data technology, and in particular to a method for monitoring building construction safety based on big data collection. Background Technology

[0002] As construction scale expands and on-site sensing devices increase, data on personnel location, machinery operation, structural response, environmental conditions, on-site images, work tasks, and safety records are continuously generated. Existing construction safety monitoring methods typically perform time alignment, anomaly handling, and regional correlation on the collected data, and then use threshold judgment, statistical analysis, or anomaly detection models to identify personnel violations, machinery anomalies, structural deviations, and environmental risks, generating construction safety early warning information.

[0003] Existing methods often directly stitch together monitoring data from different construction objects and input them into the detection model, lacking a unified expression of the correlation between construction objects, construction areas, work tasks, and monitoring time. The risk impacts between personnel, machinery, structures, and environmental conditions are difficult to map into the model's latent variable space. Furthermore, when a single monitoring value fails to reach the alarm threshold, the cumulative risk resulting from continuous deviations of multiple objects is difficult to identify.

[0004] Construction risks are characterized by state migration, time lag, and cross-object transmission. Existing anomaly detection models focus on deviations from the current state and lack joint analysis of continuous changes in reconstructed residuals, risk phase switching, and risk propagation direction. This makes it difficult to identify the source, affected, and propagation time of risks, and easily leads to problems such as incomplete risk positioning, missing propagation paths, and delayed early warnings.

[0005] Therefore, how to provide a construction safety monitoring method based on big data collection is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a construction safety monitoring method based on big data collection. This invention combines an improved DAGMM model, risk residual manifold, risk phase encoding body, and transfer entropy algorithm to achieve construction risk identification, propagation analysis, and safety early warning. It has the advantages of accurate risk location, clear propagation path, and timely early warning.

[0007] A construction safety monitoring method based on big data collection according to an embodiment of the present invention includes: Collect construction monitoring data, perform preprocessing on the construction monitoring data, and generate reliable construction status data; Arrange reliable construction status data according to construction object number, construction area number, work task number and monitoring time to generate a construction status sample set and construction object association records; An improved DAGMM model is constructed by setting the risk mapping latent space structure, risk disturbance latent variables and state transition coding. The construction state sample set is input into the improved DAGMM model to generate the reconstruction state of the construction object, the initial risk energy and the state transition code. Based on reliable construction status data, reconstruction status of construction objects and associated records of construction objects, the reconstruction residual is calculated and a local residual neighborhood is constructed to generate a risk residual manifold. Based on the initial risk energy, state transition coding, and risk residual manifold, the state transition positions and risk persistence segments are extracted to generate the risk phase coding body; The risk transmission record is generated by performing a transfer entropy algorithm based on the risk phase encoding body and the risk residual manifold, through conditional transfer entropy variable filtering, adaptive lag order search and transfer entropy state transition encoding. The risk transmission records are written back to the improved DAGMM model. Based on the risk transmission records, the latent variable associations in the risk mapping latent space are reconstructed and the associated risk energy is recalculated. The risk propagation path is generated according to the object connection relationship and propagation sequence in the risk transmission records. Based on the associated risk energy and risk propagation path, the risk source object, risk impact object, risk occurrence area and risk propagation time are determined, and construction safety early warning results and construction safety monitoring records are generated.

[0008] Optionally, the construction monitoring data specifically includes construction personnel monitoring data, construction machinery operation data, construction structure monitoring data, construction environment monitoring data, construction site image data, construction operation task data, construction site layout data, and construction safety record data.

[0009] Optionally, the preprocessing of the construction monitoring data includes time alignment, anomaly handling, object binding, and region matching.

[0010] Optionally, generating the construction status sample set and the construction object association record includes: Read reliable construction status data and extract construction object identifier, location, status, monitoring time, on-site target status, work task number, work area, task status, construction area number and area location; The construction object is identified and registered as a construction object number, associated with the on-site target status, and the construction object area relationship and construction object task relationship are established based on the object location, work area and area location. Write the object status according to the order of personnel, machinery, structure, environment, image and task status fields, write non-corresponding fields into invalid field markers, and generate a fixed-dimensional construction status vector. Arrange fixed-dimensional construction status vectors according to construction object number and monitoring time, bind them with construction area number, work task number, trusted status marker and valid field marker to generate a construction status sample set, combine construction object number, construction area number, work task number, construction object area relationship, construction object task relationship and monitoring time to generate construction object association record.

[0011] Optionally, the generation of the construction object reconstruction state, initial risk energy, and state transition code includes: An improved DAGMM model is constructed, retaining the encoder, decoder, estimation network, Gaussian mixture estimation and sample energy calculation. A risk mapping latent space structure is set at the encoder output, risk perturbation latent variable processing is set at the estimation network input, and state transition coding processing is set within the estimation network. Based on the construction safety record data, historical reliable construction status data that has been removed by removing related safety events and risk handling records is used to train and improve the DAGMM model based on the remaining historical reliable construction status data. Input the construction status sample set into the encoder, read the fixed-dimensional construction status vector according to the valid field marking, perform compression encoding on the fixed-dimensional construction status vector, and generate initial latent variables. By reading the regional and task relationships of construction objects through the risk mapping latent space structure, a latent variable correlation matrix is ​​generated. Based on the latent variable correlation matrix, an association index is established between the initial latent variables. The adjacency order of the initial latent variables is adjusted according to the association index. The coordinates of latent variables with regional and task relationships of construction objects are arranged to continuous coordinate positions to generate risk mapping latent variables. Input the risk mapping latent variables into the decoder to perform state reconstruction, generate the reconstruction state of the construction object, calculate the field difference, reconstruction distance and directional similarity between the fixed-dimensional construction state vector and the reconstruction state of the construction object according to the effective field labels, and generate reconstruction features; The risk mapping latent variables and reconstructed features are input into the estimation network to generate Gaussian component membership degrees. The indexes of the principal Gaussian components and the direction of change of the principal Gaussian components corresponding to adjacent monitoring time windows are read to generate state transition codes. The state transition codes and Gaussian component membership degrees are combined according to the Gaussian component index to generate transition state membership degrees. Gaussian mixture parameters are calculated based on the transition state membership degrees, and the initial risk energy is calculated based on the risk mapping latent variables, reconstructed features, and Gaussian mixture parameters.

[0012] Optionally, the generation of the risk residual manifold includes: The reliable construction status data and the reconstruction status of the construction object are matched according to the construction object number and monitoring time. The difference of the effective fields is calculated based on the effective field markers to generate the reconstruction residual vector of the construction object. The residual vector of the reconstructed construction object is used as the residual space coordinate. The residual modulus, the change in residual modulus, and the direction of residual change are calculated. The continuous residual space coordinates are connected to generate the object residual trajectory. Based on the regional and task relationships of the construction objects, we screen the pairs of related construction objects, compare the changes in residual modulus and the direction of residual change within the same monitoring time window, and generate cross-object residual connections. Calculate the residual distance between adjacent residual spatial coordinates, establish a local residual neighborhood based on the object residual trajectory and cross-object residual connections, and combine the residual spatial coordinates, residual distance, object residual trajectory and local residual neighborhood to generate the risk residual manifold.

[0013] Optionally, generating the risk phase code includes: Read the initial risk energy, state transition code and risk residual manifold, calculate the change in risk energy and the change in residual modulus, set the stability threshold based on the median of the absolute value of the change corresponding to the historical normal construction state, and set the risk energy interval boundary based on the preset quantile value of the initial risk energy. According to the preset priority order, based on the initial risk energy crossing the risk energy range, the continuous growth of the risk energy change and the residual modulus change, the transition from the holding state to synchronous growth, synchronous decline and the remaining state, the risk level phase, risk expansion phase, risk rise phase, risk fall phase and state holding phase are marked in sequence. Connect the same risk phase markers to determine the state transition location and risk duration segment, and set the corresponding risk phase channel; The initial risk energy, state transition code, state transition location, risk duration segment, and residual trajectory index are written into the corresponding risk phase channel and arranged according to the construction object, monitoring time, and risk phase channel to generate a risk phase code body.

[0014] Optionally, generating the risk transfer record includes: Read the risk phase encoding body and risk residual manifold, register the construction object pairs with cross-object residual connections as candidate transfer object pairs, combine the risk phase channel encoding and state transition encoding, and generate the transfer entropy state transition sequence. Candidate condition objects that meet the conditions of common region relationship, common task relationship and cross-object residual connection are selected, and a set of condition variables is generated based on the change of propagation entropy and preset selection conditions. Calculate the positive time interval between the risk rise phase position of the source candidate object and the state transition position of the response candidate object to generate a set of candidate hysteresis orders; Based on the candidate lag order, the transfer entropy state transition sequence is moved, and the conditional transfer entropy is calculated in combination with the set of conditional variables. The source sequence is then subjected to cyclic displacement, and the preset quantile value of the cyclic displacement result is set as the effective transfer threshold. Candidate relationships with consistent directions within a continuous preset number of risk intervals and conditional transmission entropy not less than the effective transmission threshold are identified as effective risk transmission relationships. The target lag order, risk transmission direction, risk precursor object, and risk response object are determined, and the transmission benchmark monitoring time is bound to generate risk transmission records.

[0015] Optionally, the generation of risk propagation paths includes: Read risk transmission records, risk mapping latent variable time series records, reconstruct features and state transition codes to obtain risk precursor objects, risk response objects, target lag order and transmission baseline monitoring time; Based on the target lag order, the historical monitoring time is determined. The preceding historical latent variables and the response baseline latent variables are read according to the object number and monitoring time. The coordinate change is calculated, and the coordinates that meet the preset coordinate change conditions are identified as risks and written into the coordinate system. According to the risk transmission direction, the coordinate values ​​of the preceding historical latent variables at the risk writing coordinate are written into the response benchmark latent variables to generate associated risk disturbance latent variables; Input the latent variables of associated risk perturbation and the reconstructed features into the estimation network, recalculate the membership degree of the Gaussian component, read the Gaussian mixture parameters of the improved DAGMM model, and calculate the associated risk energy by combining state transition coding. The risk transmission records are arranged according to the monitoring time. Records that are consistent with the previous risk response object and the next risk precursor object are connected. Circular connections are deleted. The risk transmission object sequence, transmission time sequence and object association information are combined to generate the risk transmission path.

[0016] Optionally, the generation of construction safety early warning results and construction safety monitoring records includes: The historical reliable construction status data is used to perform construction status sample construction, improved DAGMM analysis, risk residual manifold construction, risk phase encoding, transfer entropy analysis and risk transfer write-back to generate historical associated risk energy, and the associated risk judgment threshold is set with preset quantile values. Construction objects whose current associated risk energy is not less than the associated risk judgment threshold are identified as risk status objects and written into the risk status flag; Based on the risk propagation path, determine the risk source and the risk impact object, read the corresponding construction area number to generate the risk source area and the risk impact area, and read the propagation time series to generate the risk propagation time; By combining the risk source object, the risk impact object, the risk source area, the risk impact area, the work task number, the risk propagation time, the associated risk energy, and the risk propagation path, a construction safety early warning result is generated. This result is then associated with the construction object number, the construction area number, the work task number, and the monitoring time to generate a construction safety monitoring record.

[0017] The beneficial effects of this invention are: This invention collects monitoring data from construction personnel, construction machinery operation data, construction structure monitoring data, construction environment monitoring data, construction site image data, construction task data, construction site layout data, and construction safety record data. Through time alignment, anomaly handling, object binding, and region matching, it generates reliable construction status data. Furthermore, it constructs a construction status sample set and construction object association records according to the construction object number, construction area number, task number, and monitoring time. This enables a unified data representation of personnel, machinery, structure, environment, and task status, improves the integrity of the association of multi-source construction status, and reduces monitoring bias caused by scattered data sources and missing object correspondences.

[0018] This invention constructs an improved DAGMM model that includes a risk mapping latent space structure, risk disturbance latent variable processing, and state transition coding processing. It reconstructs the latent variable associations by utilizing the regional and task relationships of construction objects, and constructs a risk residual manifold and risk phase coding body by combining the reconstructed state of construction objects, initial risk energy, and state transition coding. This enables a unified expression of the state deviation, continuous change, phase switching, and cross-object association of construction risks, improves the identification ability of progressive risks, cumulative risks, and state transition risks, and reduces the risk missed detection when a single monitoring value does not meet the alarm conditions.

[0019] This invention is based on the risk phase encoding body and the risk residual manifold to execute the transfer entropy algorithm. It uses transfer entropy state transition encoding, conditional transfer entropy variable screening, and adaptive lag order search to generate risk transfer records. The risk transfer records are then written back to the improved DAGMM model to reconstruct the latent variable associations, generate associated risk energy and risk propagation paths, and enable construction safety monitoring results to determine the risk source object, risk impact object, risk area, and risk propagation time, thereby improving the accuracy of risk propagation direction identification, risk location tracing, and safety early warning response. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They explain the invention together with the embodiments of the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a construction safety monitoring method based on big data collection proposed in this invention; Figure 2 This is a schematic diagram of the improved DAGMM model structure of a construction safety monitoring method based on big data collection proposed in this invention; Figure 3 Here is a flowchart of the transfer entropy algorithm for a construction safety monitoring method based on big data collection proposed in this invention; Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figure 1 , Figure 2 and Figure 3 A construction safety monitoring method based on big data collection includes: Collect construction monitoring data, perform preprocessing on the construction monitoring data, and generate reliable construction status data; Arrange reliable construction status data according to construction object number, construction area number, work task number and monitoring time to generate a construction status sample set and construction object association records; An improved DAGMM model is constructed by setting the risk mapping latent space structure, risk disturbance latent variables and state transition coding. The construction state sample set is input into the improved DAGMM model to generate the reconstruction state of the construction object, the initial risk energy and the state transition code. Based on reliable construction status data, reconstruction status of construction objects and associated records of construction objects, the reconstruction residual is calculated and a local residual neighborhood is constructed to generate a risk residual manifold. Based on the initial risk energy, state transition coding, and risk residual manifold, the state transition positions and risk persistence segments are extracted to generate the risk phase coding body; The risk transmission record is generated by performing a transfer entropy algorithm based on the risk phase encoding body and the risk residual manifold, through conditional transfer entropy variable filtering, adaptive lag order search and transfer entropy state transition encoding. The risk transmission records are written back to the improved DAGMM model. Based on the risk transmission records, the latent variable associations in the risk mapping latent space are reconstructed and the associated risk energy is recalculated. The risk propagation path is generated according to the object connection relationship and propagation sequence in the risk transmission records. Based on the associated risk energy and risk propagation path, the risk source object, risk impact object, risk occurrence area and risk propagation time are determined, and construction safety early warning results and construction safety monitoring records are generated.

[0023] In this embodiment, the construction monitoring data specifically includes construction personnel monitoring data, construction machinery operation data, construction structure monitoring data, construction environment monitoring data, construction site image data, construction operation task data, construction site layout data, and construction safety record data.

[0024] In this embodiment, the preprocessing of construction monitoring data includes time alignment, anomaly handling, object binding, and region matching.

[0025] In this embodiment, generating the construction status sample set and the construction object association record includes: Read reliable construction status data and extract construction object identifier, location, status, monitoring time, on-site target status, work task number, work area, task status, construction area number and area location; The construction object is identified and registered as a construction object number, associated with the on-site target status, and the construction object area relationship and construction object task relationship are established based on the object location, work area and area location. Write the object status according to the order of personnel, machinery, structure, environment, image and task status fields, write non-corresponding fields into invalid field markers, and generate a fixed-dimensional construction status vector. Arrange fixed-dimensional construction status vectors according to construction object number and monitoring time, bind them with construction area number, work task number, trusted status marker and valid field marker to generate a construction status sample set, combine construction object number, construction area number, work task number, construction object area relationship, construction object task relationship and monitoring time to generate construction object association record.

[0026] In this embodiment, generating the reconstruction state of the construction object, the initial risk energy, and the state transition code includes: An improved DAGMM model is constructed, retaining the encoder, decoder, estimation network, Gaussian mixture estimation, and sample energy calculation. A risk mapping latent space structure is added at the encoder output, risk perturbation latent variable processing is added at the estimation network input, and state transition encoding is added within the estimation network. Specifically, the improved DAGMM model is constructed as follows: The number of fields read from the fixed-dimensional construction state vector is denoted as D. The encoder is set up with a D-dimensional input layer, a 128-dimensional first encoding layer, a 64-dimensional second encoding layer, and a 16-dimensional latent variable output layer. The decoder is set up with a 16-dimensional input layer, a 64-dimensional first decoding layer, a 128-dimensional second decoding layer, and a D-dimensional state reconstruction layer. The reconstruction distance and direction deviation are calculated based on the effective fields and combined to form a 2-dimensional reconstruction feature. The 16-dimensional latent variable is concatenated with the 2-dimensional reconstruction feature to form an 18-dimensional estimated input vector. The estimation network is set up with an 18-dimensional input layer, a 32-dimensional hidden layer, and a 4-dimensional Gaussian component output layer. The original DAGMM directly inputs the 16-dimensional initial latent variable generated by the encoder into the solution. The coder inputs 16 initial latent variables and 2-dimensional reconstructed features into the estimation network. The improved DAGMM inputs 16 initial latent variables into the risk mapping latent space structure to generate 16-dimensional risk mapping latent variables. The risk mapping latent variables are then input into the decoder and input into the estimation network after being concatenated with the 2-dimensional reconstructed features. The estimation network outputs 4-dimensional Gaussian component membership degrees. The state transition coding process converts the 4-dimensional Gaussian component membership degrees into 4-dimensional transition state membership degrees. The Gaussian mixture estimation calculates the mixing coefficients, 18-dimensional mean vector, and 18-row 18-column covariance matrix corresponding to the 4-dimensional transition state membership degrees, and calculates the risk energy. The 16-dimensional initial latent variables corresponding to the N construction object state samples are arranged into an N-row, 16-column initial latent variable matrix. An N-row, N-column latent variable association matrix is ​​constructed based on the regional and task relationships of the construction objects. Elements corresponding to the same object, and objects with the same region and task, are set to 1; elements corresponding to objects with the same region but different tasks are set to 0.6; elements corresponding to objects with different regions but the same task are set to 0.4; and elements corresponding to objects with no regional or task relationship are set to 0. Row normalization is performed on the latent variable association matrix, and the row normalization result is multiplied by the initial latent variable matrix. An N-row, 16-column association latent variable matrix is ​​generated. The association strength of the 16 latent variable coordinates in all training data is calculated. A fixed coordinate rearrangement index is generated according to the association strength from largest to smallest. The association latent variable matrix is ​​rearranged columnwise using the fixed coordinate rearrangement index to generate an N-row, 16-column risk mapping latent variable matrix. The risk mapping latent variables replace the initial latent variables in the original DAGMM as the decoder input and the estimation network input in the initial risk analysis stage. The regional relationship weight of 0.6 and the task relationship weight of 0.4 are determined by the normalization and minimization of the reconstruction error and risk state distinction error in the validation data by the candidate weight combination. In the initial risk analysis phase, the input of risk transmission records is turned off, and the writing of latent variable coordinates is not performed. The 16-dimensional risk mapping latent variables are connected with the 2-dimensional reconstruction features and then input into the estimation network. In the risk transmission write-back phase, the risk precursor object, risk response object, target lag order, and transmission benchmark monitoring time are read to obtain the 16-dimensional precursor historical latent variables and the 16-dimensional response benchmark latent variables. A 16-dimensional risk writing marker vector is generated. The coordinates marked as 1 are set as the corresponding coordinate values ​​of the precursor historical latent variables, and the coordinates marked as 0 are set as the corresponding coordinate values ​​of the response benchmark latent variables. A 16-dimensional associated risk disturbance latent variable is generated. The associated risk disturbance latent variable replaces the risk mapping latent variable and is connected with the corresponding 2-dimensional reconstruction features to form an 18-dimensional associated estimation input vector. The vector is then re-input into the estimation network to calculate the associated risk energy. Read the membership degrees of the 4-dimensional Gaussian components output by the estimation network, extract the index of the principal Gaussian component corresponding to the adjacent monitoring time window, generate the growth direction code 1, the maintenance direction code 0, and the fall direction code -1 according to the risk sequence change of the Gaussian components, combine the previous principal Gaussian component index, the current principal Gaussian component index, and the change direction to form the state transition code, read the corresponding rows of the 4-row 4-column Gaussian state transition probability matrix according to the previous principal Gaussian component index, generate the 4-dimensional historical transition probability, multiply the 4-dimensional current Gaussian component membership degree by 0.8, multiply the 4-dimensional historical transition probability by 0.2, add the calculation results of the corresponding Gaussian component positions, and generate the 4-dimensional migration state membership degree. The migration state membership degree replaces the Gaussian component membership degree in the original DAGMM and is input into the Gaussian mixture estimation. 0.8 and 0.2 are determined by the normalization and minimization of the risk energy fluctuation value and state misclassification rate of the candidate combination in the validation data. This invention retains the encoder, decoder, estimation network, Gaussian mixture estimation, and risk energy calculation structure of the traditional DAGMM model. It sets up a risk mapping latent space structure, writes the regional and task relationships of construction objects into the latent variable expression, sets up risk disturbance latent variable processing, writes the state of risk precursor objects back to the latent variables of risk response objects, sets up state transition coding processing, and writes the Gaussian state changes of adjacent monitoring time windows into the Gaussian mixture estimation. This enables the improved DAGMM model to jointly analyze the association of construction objects, state evolution, and risk propagation, and improves the ability to identify progressive risks, state transition risks, and cross-object propagation risks. Based on historical reliable construction status data, which excludes associated safety events and risk management records from construction safety record data, an improved DAGMM model is trained using the remaining historical reliable construction status data. Specifically, training the improved DAGMM model based on the remaining historical reliable construction status data involves: Read the safety incident object number, safety incident occurrence time, risk handling object number, risk handling start time and risk handling end time, set the 30 minutes before and after the safety incident occurrence time as the safety incident associated time interval, set the 30 minutes before the risk handling start time to the 30 minutes after the risk handling end time as the risk handling associated time interval, delete the historical reliable construction status data with matching object numbers and monitoring times within the corresponding time intervals, arrange the remaining data according to construction object number and monitoring time, and construct normal construction training samples using a 60-second monitoring time window; According to the monitoring time, the normal construction training samples were divided into 70% training data, 15% validation data, and 15% test data. Each batch of 128 samples was formed into a training batch. The compression encoding, latent variable association mapping, state reconstruction, reconstruction feature calculation, Gaussian component membership calculation, state transition encoding, and Gaussian mixture estimation were performed in sequence. Risk propagation and write-back were turned off during the training phase. The reconstruction error, energy error, covariance stability error, and transition consistency error were combined into training objectives of 0.5, 0.25, 0.15, and 0.1, respectively. The Adam optimizer was used to update the encoder, decoder, and estimated network parameters. The initial learning rate was set to 0.001, and the maximum number of training rounds was set to 300. Training was stopped when the rate of change of the validation objective was less than 0.0005 for 8 consecutive rounds. The network parameters, Gaussian mixture parameters, and Gaussian state transition probability matrix corresponding to the minimum number of rounds of validation objective were saved. The construction state sample set is input into the encoder. A fixed-dimensional construction state vector is read according to the valid field markers. Compression encoding is performed on the fixed-dimensional construction state vector to generate initial latent variables, where: The construction status sample set is input into the encoder, and a fixed-dimensional construction status vector is read according to the valid field markers, specifically: Read the construction status sample set in ascending order of construction object number and monitoring time. Write each fixed-dimensional construction status vector as 1 row into the input matrix. Keep the field values ​​marked as 1 in the field validity marker vector and set the field values ​​marked as 0 to 0. Multiply each field value in the input matrix with the corresponding valid field marker to generate a valid status input matrix. Save the construction object number, construction area number, work task number and monitoring time as the row index of the input matrix. The fixed-dimensional construction state vector is compressed and encoded to generate initial latent variables, specifically: The valid state input matrix is ​​input into the first coding layer of the encoder. Matrix multiplication is performed on each row of input data with the parameter matrix of the first coding layer, and a bias vector is added. Results less than 0 are set to 0, generating a 128-dimensional first coding feature. The first coding feature is input into the second coding layer and the same calculation is performed, generating a 64-dimensional second coding feature. The second coding feature is input into the latent variable output layer. Matrix multiplication is performed on the second coding feature with the parameter matrix of the latent variable output layer, and a bias vector is added, generating a 16-dimensional initial latent variable. All initial latent variables are arranged according to the row index of the input matrix to generate the initial latent variable matrix. By reading the regional and task relationships of construction objects through the risk mapping latent space structure, a latent variable correlation matrix is ​​generated. Based on this matrix, initial latent variable correlation indices are established, and the adjacency order of these indices is adjusted. Finally, the coordinates of latent variables with regional and task relationships are arranged in continuous coordinate positions to generate risk-mapped latent variables. Generate the latent variable correlation matrix, specifically as follows: According to the row index of the initial latent variable matrix, read the construction area number and task number corresponding to every two construction objects. Set the matrix element corresponding to the same construction object to 1. Set the matrix element corresponding to the construction object with the same construction area number and task number to 1. Set the matrix element corresponding to the construction object with the same construction area number but different task number to 0.6. Set the matrix element corresponding to the construction object with different construction area number but the same task number to 0.4. Set the matrix element corresponding to the construction object with different construction area number and different task number to 0. Generate a latent variable association matrix. Test the combinations of regional relationship weight and task relationship weight of 0.5 and 0.5, 0.6 and 0.4, and 0.7 and 0.3 respectively. After normalizing the verification data reconstruction error and risk status distinction error, add them together. Set the smallest sum of 0.6 and 0.4 as the regional relationship weight and task relationship weight. The regional relationship weight is set to 0.6 to enhance the immediate risk association of construction objects in the same area. The task relationship weight is set to 0.4 to retain the cross-regional task association. The association index between the initial latent variables is established based on the latent variable association matrix, specifically as follows: Calculate the sum of all elements in each row of the latent variable association matrix, divide each matrix element by the sum of the elements in its row to generate a row-normalized latent variable association matrix. When the sum of the elements in a row is 0, set the corresponding diagonal element to 1. Perform matrix multiplication between the row-normalized latent variable association matrix and the initial latent variable matrix to generate an association latent variable matrix. Extract the associated construction object numbers according to the column numbers corresponding to the non-zero elements in the latent variable association matrix. Arrange the associated construction object numbers in descending order of matrix elements to generate an association index sequence corresponding to each construction object. Generate risk mapping latent variables, specifically: Calculate the matrix product of the transpose of the initial latent variable matrix and the associated latent variable matrix. Use the sum of the absolute values ​​of each row of the matrix product as the association strength of the corresponding latent variable coordinates. Arrange the 16 latent variable coordinate numbers in descending order of association strength to generate a coordinate rearrangement index. Rearrange the 16 coordinates of the associated latent variable matrix according to the coordinate rearrangement index. Arrange the latent variable coordinates with higher association strength in the lower consecutive number positions and the latent variable coordinates with lower association strength in the higher consecutive number positions to generate risk mapping latent variables. Bind the risk mapping latent variables to the construction object number and monitoring time. Arrange them according to the construction object number and monitoring time to generate a time series record of risk mapping latent variables. The risk mapping latent variables are input into the decoder to perform state reconstruction, generating the reconstructed state of the construction object. Based on the effective field labels, the field differences, reconstruction distance, and directional similarity between the fixed-dimensional construction state vector and the reconstructed state of the construction object are calculated to generate reconstruction features, where: Generate the reconstruction status of the construction object, specifically as follows: Each 16-dimensional risk mapping latent variable is input into the first decoding layer. Matrix multiplication is performed between the risk mapping latent variable and the parameter matrix of the first decoding layer, and a bias vector is added. Results less than 0 are set to 0, generating a 64-dimensional first decoding feature. The first decoding feature is input into the second decoding layer and the same calculation is performed, generating a 128-dimensional second decoding feature. The second decoding feature is input into the state reconstruction layer. Matrix multiplication is performed between the second decoding feature and the parameter matrix of the state reconstruction layer, and a bias vector is added. The output values ​​are restricted to 0 to 1 by a logical function. The output values ​​are arranged according to the field order of the fixed-dimensional construction state vector to generate the construction object reconstruction state. The field differences between the fixed-dimensional construction state vector and the reconstruction state of the construction object are calculated based on the valid field labels, specifically: Match the fixed-dimensional construction status vector with the reconstruction status of the construction object according to the same construction object number and monitoring time. Read the field validity marker vector, subtract the corresponding reconstruction field value from the original field value marked as 1 for the valid field to generate the valid field difference, set the field difference marked as 0 to 0, and arrange all field differences in the order of fixed-dimensional fields to generate the reconstruction residual vector. The reconstruction distance and orientation similarity are calculated as follows: Calculate the sum of squares of the differences between each effective field in the reconstructed residual vector, divide the sum of squares by the number of effective fields and take the square root to generate the reconstruction distance. Multiply the fixed-dimensional construction state vector with the corresponding effective field values ​​in the reconstruction state of the construction object item by item and sum them. Calculate the sum of squares of the two sets of effective field values ​​and take the square root of each. Divide the sum of the product of the two sums of squares by the sum of the square roots of the two sums of squares and 0.000001 to generate the directional similarity. Subtract the directional similarity from 1 to generate the directional deviation. Arrange the reconstruction distance and directional deviation in the order of reconstruction distance first and directional deviation second to generate a 2D reconstruction feature. The risk mapping latent variables and reconstructed features are input into the estimation network to generate Gaussian component membership degrees. The indices and directions of change of the principal Gaussian components corresponding to adjacent monitoring time windows are read to generate state transition codes. These state transition codes are combined with the Gaussian component membership degrees according to the Gaussian component indices to generate transition state membership degrees. Gaussian mixture parameters are calculated based on the transition state membership degrees. Finally, the initial risk energy is calculated based on the risk mapping latent variables, reconstructed features, and Gaussian mixture parameters. Generate Gaussian component membership, specifically: Each 16-dimensional risk mapping latent variable is concatenated with its corresponding 2-dimensional reconstructed feature in the order of risk mapping latent variable first and reconstructed feature last to generate an 18-dimensional estimated input vector. The 18-dimensional estimated input vector is then input into a 32-neuron estimated hidden layer. Matrix multiplication is performed on the estimated input vector and the estimated hidden layer parameter matrix, and a bias vector is added. Results less than 0 are set to 0. The estimated hidden layer output during the training phase is randomly set to 0 according to a deactivation ratio of 0.1. The estimated hidden layer output is then input into a 4-neuron Gaussian component output layer. The four output values ​​are converted into Gaussian component membership degrees with a sum of 1 using a normalized exponential function. The state transition codes are generated as follows: The Gaussian component number with the largest value among the membership degrees of each Gaussian component is selected as the main Gaussian component index. The main Gaussian component index of adjacent monitoring time windows is matched according to the construction object number. The average reconstruction distance of each Gaussian component in the training data is calculated. The risk sequence number is configured for the four Gaussian components in ascending order of average reconstruction distance. The risk sequence number of the current monitoring time window is subtracted from the risk sequence number of the previous monitoring time window. When the difference is greater than 0, the growth direction code 1 is written. When the difference is equal to 0, the maintenance direction code 0 is written. When the difference is less than 0, the fallback direction code -1 is written. The previous main Gaussian component index, the current main Gaussian component index and the change direction of the main Gaussian component are combined in a fixed order to generate the state transition code. The membership degree of the transition state is generated as follows: The number of times the previous principal Gaussian component enters the current principal Gaussian component in the statistical training data is counted to construct a 4x4 Gaussian state transition counting matrix. Each matrix element is incremented by 1 for smoothing. Each element is divided by the sum of all elements in its row to generate a Gaussian state transition probability matrix. The corresponding row of the Gaussian state transition probability matrix is ​​read according to the index of the previous principal Gaussian component in the state transition code to generate a 4-dimensional historical transition probability. The membership degree of the current Gaussian component is multiplied by 0.8, and the historical transition probability is multiplied by 0.2. The two results at the corresponding Gaussian component index position are added to generate the transition state membership degree. On the validation data, the combination ratio of the membership degree of the current Gaussian component and the historical transition probability is tested to be 0.6 and 0.4, 0.7 and 0.3, 0.8 and 0.2, and 0.9 and 0.1. The risk energy fluctuation value and state misclassification rate of the validation data are normalized and added together. The smallest sum, 0.8 and 0.2, is set as the combination ratio. 0.8 is used to retain the current construction state information, and 0.2 is used to write the state continuity information of adjacent monitoring time windows. The Gaussian mixture parameters are calculated based on the membership degree of the migration state, specifically as follows: For each Gaussian component within the same training batch, sum the membership degrees of the transition states. Divide the sum by the number of samples in the training batch to generate the Gaussian component mixing coefficients. Multiply each 18-dimensional estimated input vector by the corresponding Gaussian component transition state membership degree and sum them. Divide the sum by the total sum of the corresponding Gaussian component transition state membership degrees to generate the Gaussian component mean vector. Calculate the cross product of the difference between each 18-dimensional estimated input vector and the Gaussian component mean vector. Multiply the cross product by the corresponding transition state membership degree and sum them. Divide the sum by the total sum of the corresponding Gaussian component transition state membership degrees to generate the Gaussian covariance matrix. Add 0.000001 to each diagonal element of the Gaussian covariance matrix to generate the Gaussian covariance matrix used for matrix inversion and determinant calculation. The initial risk energy is calculated based on the risk mapping latent variables, reconstructed features, and Gaussian mixture parameters, specifically as follows: An 18-dimensional energy calculation vector is generated in the order of risk mapping latent variables first and reconstructed features last. The Gaussian probability density of the energy calculation vector under the mean vector and covariance matrix of the four Gaussian components is calculated respectively. Each Gaussian probability density is multiplied by the mixing coefficient of the corresponding Gaussian component and summed. The natural logarithm of the summation result is taken and the sign is changed to generate the initial risk energy of the current construction state sample. The initial risk energy is bound to the construction object number, construction area number, work task number and monitoring time to generate the initial risk energy record of the construction object.

[0027] In this embodiment, generating the risk residual manifold includes: Based on the construction object number and monitoring time, reliable construction status data is matched with the reconstructed status of the construction object. The difference between valid fields is calculated based on the valid field markers to generate a construction object reconstruction residual vector, where: The reliable construction status data is matched with the reconstructed status of the construction object according to the construction object number and monitoring time, specifically as follows: Read the construction object number and monitoring time from the trusted construction status data, read the construction object number and monitoring time corresponding to the reconstruction status of the construction object, establish a status matching relationship between two records with the same construction object number and the same monitoring time. In this embodiment, the monitoring time window length is set to 60 seconds. Trusted construction status data that are located in the same 60-second monitoring time window but have different collection times are arranged in ascending order according to the collection time. Calculate the average value of each effective field in the monitoring time window, and use the average value as the trusted construction status value corresponding to the monitoring time window. Write the reconstruction status of the construction object and the trusted construction status value of the corresponding monitoring time window into the same residual calculation record. Generate the residual vector of the construction object reconstruction, specifically as follows: A 60-second monitoring time window is used to perform time alignment on the reliable construction status data. The fixed-dimensional construction status vector is matched with the reconstruction status of the construction object according to the construction object number and the monitoring time. The reconstruction value is subtracted from the status value of the valid field, and the difference of the invalid field is set to 0. All field differences are arranged in the order of fixed-dimensional fields to generate the reconstruction residual vector of the construction object, and it is bound to the construction object number, construction area number, work task number and monitoring time. Using the reconstructed residual vector of the construction object as the residual space coordinates, the residual modulus, the change in residual modulus, and the direction of residual change are calculated. Connecting the continuous residual space coordinates generates the object residual trajectory, where: The residual vector of the reconstructed construction object is used as the coordinates in the residual space, specifically: Read all field differences in the reconstructed residual vector of each construction object, set the position of each field as a residual space coordinate axis, set the field difference as the coordinate value on the corresponding residual space coordinate axis, set the total number of fields of the fixed-dimensional construction status vector as the residual space dimension, write the reconstructed residual vector of the construction object into the residual space coordinate table according to the construction object number and monitoring time, and connect the construction object number and monitoring time to form the residual space coordinate index; The residual modulus and the change in residual modulus are calculated as follows: Read the residual vector and corresponding field validity marker vector of the construction object reconstruction, calculate the sum of squares of the differences of valid fields, take the square root of the sum of squares to generate the residual modulus, match the residual modulus of the current monitoring time window with the previous monitoring time window according to the construction object number, subtract the residual modulus of the previous monitoring time window from the residual modulus of the current monitoring time window to generate the residual modulus change, when the construction object lacks the record of the previous monitoring time window, set the residual modulus change to 0, and write the residual modulus and residual modulus change to the corresponding residual space coordinate record; The direction of residual change is calculated as follows: Reconstructing residual vectors for the construction objects corresponding to the current and previous monitoring time windows according to the construction object number; subtracting the corresponding field difference from the previous construction object's reconstruction residual vector from the difference of each field in the current construction object's reconstruction residual vector to generate a residual change vector; calculating the sum of squares of each field value in the residual change vector and taking the square root to generate the residual change vector's magnitude; dividing each field value in the residual change vector by the sum of the residual change vector's magnitude and 0.000001 to generate a unit residual change vector; using the unit residual change vector as the residual change direction; and setting all directional components to 0 when the residual change vector's magnitude is 0. Based on the regional and task relationships of the construction objects, pairs of related construction objects are selected. The changes in residual modulus and direction of residual change within the same monitoring time window are compared to generate cross-object residual connections, where: Based on the regional and task relationships of the construction objects, related construction object pairs are selected, specifically as follows: Read the construction object number, construction area number, work task number, construction object area relationship and construction object task relationship from the construction object association record. Read any two different construction objects according to the same monitoring time window. When the construction area numbers corresponding to the two construction objects are the same, write the corresponding construction object pair into the area association object set. When the work task numbers corresponding to the two construction objects are the same, write the corresponding construction object pair into the task association object set. Merge the area association object set and the task association object set, delete the records with duplicate construction object number combinations, and generate a set of associated construction object pairs. By comparing the changes in residual modulus and the direction of residual change within the same monitoring time window, cross-object residual connections are generated, specifically as follows: Different construction objects with the same construction area number or the same work task number are registered as associated construction object pairs. The inner product of the residual change direction of the two objects within the same monitoring time window is calculated and the residual direction consistency is generated. The 50th percentile of the absolute value of the residual modulus change in the historical normal construction state is set as the effective change threshold. When the absolute value of the residual modulus change of the two objects is not less than the effective change threshold, the change is consistent, and the residual direction consistency is not less than 0.7, a cross-object residual connection is established between the corresponding residual space coordinates. Calculate the residual distance between adjacent residual spatial coordinates, establish a local residual neighborhood based on the object residual trajectory and cross-object residual connections, and combine the residual spatial coordinates, residual distance, object residual trajectory, and local residual neighborhood to generate the risk residual manifold, where: Calculate the residual distance between adjacent residual spatial coordinates, specifically as follows: Read the residual spatial coordinate pairs with time connection and residual spatial coordinate pairs with cross-object residual connection in the residual trajectory of the object. Read the coordinate values ​​in the two residual spatial coordinates in the order of fixed dimension fields, subtract the corresponding coordinate values, calculate the sum of squares of all coordinate differences, and take the square root of the sum of squares to generate the residual distance. When the two residual spatial coordinates correspond to different construction object categories, first calculate the residual modulus, residual modulus change and residual direction consistency of the residual spatial coordinates respectively. After normalizing the three values, form a residual description vector. Then calculate the sum of squares of the differences of corresponding components between the two residual description vectors and take the square root to generate the cross-category residual distance. Local residual neighborhoods are established based on object residual trajectories and cross-object residual connections, specifically as follows: Calculate the square root of the sum of squares of the coordinate differences between two residual spatial coordinates that have time connection or cross-object residual connection, generate the residual distance, set the 90th percentile of the residual distance of the historical normal construction state as the neighborhood distance threshold, write the adjacent coordinates in the object residual trajectory and the cross-object coordinates whose residual distance is not greater than the neighborhood distance threshold into the local residual neighborhood, and when the number of cross-object coordinates exceeds 6, retain the first 6 coordinates in ascending order of residual distance; Generate the risk residual manifold, specifically: Establish a residual spatial coordinate table, a time connection table, a cross-object residual connection table, a residual distance table, and a local residual neighborhood index table. Write the construction object number, construction area number, job task number, and monitoring time into the residual spatial coordinate table. Write the start index, end index, and time sequence corresponding to the object residual trajectory into the time connection table. Write the associated construction object number, residual direction consistency, and connection time into the cross-object residual connection table. Write the residual distance corresponding to each connection into the residual distance table. Write the time neighborhood index and cross-object neighborhood index corresponding to each residual spatial coordinate into the local residual neighborhood index table. Link the five data tables according to the residual spatial coordinate index to generate a risk residual manifold.

[0028] In this embodiment, generating the risk phase code includes: Read the initial risk energy, state transition code, and risk residual manifold; calculate the change in risk energy and the change in residual modulus; set a stability threshold based on the median of the absolute values ​​of the changes corresponding to historical normal construction states; and set the risk energy interval boundaries based on the preset quantile values ​​of the initial risk energy, where: The changes in risk energy and residual modulus are calculated as follows: The risk phase calculation records are grouped according to the construction object number. The records corresponding to the same construction object are arranged in ascending order of monitoring time. The risk energy change is generated by subtracting the initial risk energy of the previous monitoring time window from the initial risk energy of the current monitoring time window. The residual modulus change is generated by subtracting the residual modulus of the previous monitoring time window from the residual modulus of the current monitoring time window. When the construction object lacks the record of the previous monitoring time window, the risk energy change and the residual modulus change are written to 0, and the corresponding monitoring time window is marked as the initial time window. The stability threshold is set based on the median of the absolute values ​​of changes corresponding to historical normal construction conditions, specifically: Read the normal construction training data retained after removing safety event-related samples and risk disposal-related samples during the model training phase, generate historical initial risk energy and historical residual modulus, calculate the historical risk energy change and historical residual modulus change according to construction object category and monitoring time, convert the two types of changes into absolute values ​​and arrange them in ascending order, set the absolute value of the historical risk energy change that is in the middle of the data quantity in the sorting position as the risk energy stability threshold, set the absolute value of the historical residual modulus change that is in the middle of the data quantity in the sorting position as the residual modulus stability threshold, when the number of effective changes corresponding to the construction object category is less than 30, read all the historical changes corresponding to the construction object categories to calculate the two stability thresholds; The risk energy interval boundary is set based on the preset quantile value of the initial risk energy, specifically as follows: Read the initial risk energy corresponding to the normal construction training data, group the initial risk energy according to the construction object category, arrange the initial risk energy of each group in ascending order of value, set the initial risk energy that reaches 50% of the corresponding data in the sorted position as the first risk energy boundary, set the initial risk energy that reaches 75% of the corresponding data in the sorted position as the second risk energy boundary, set the initial risk energy that reaches 90% of the corresponding data in the sorted position as the third risk energy boundary, register the range that is not greater than the first risk energy boundary as risk energy interval 0, register the range that is greater than the first risk energy boundary but not greater than the second risk energy boundary as risk energy interval 1, register the range that is greater than the second risk energy boundary but not greater than the third risk energy boundary as risk energy interval 2, and register the range that is greater than the third risk energy boundary as risk energy interval 3. According to a preset priority order, based on the initial risk energy crossing the risk energy range, the continuous increase in risk energy change and residual modulus change, the transition from a holding state to synchronous growth, synchronous decline, and the remaining state, the risk level-crossing phase, risk expansion phase, risk rise phase, risk decline phase, and state holding phase are marked sequentially, where: According to the preset priority order, specifically: Set the risk leap phase as priority 1, the risk expansion phase as priority 2, the risk rise phase as priority 3, the risk fall phase as priority 4, and the status maintenance phase as priority 5. Perform phase condition judgments in order of priority value from small to large. When multiple phase conditions are met in the same monitoring time window, retain the risk phase mark with the smallest priority value. Only write 1 risk phase mark for each monitoring time window. The risk level transition phase is marked based on the initial risk energy crossing the risk energy interval, specifically as follows: Read the risk energy interval number corresponding to the previous monitoring time window and the current monitoring time window, subtract the previous risk energy interval number from the current risk energy interval number, read the change direction of the principal Gaussian component in the state transition code, when the interval number difference is greater than 0 and the change direction of the principal Gaussian component is the growth direction code 1, mark the current monitoring time window as the risk level crossing phase, write the interval number difference into the risk level crossing amplitude field, and proceed to the next priority judgment when the interval number difference is not greater than 0. The risk expansion phase is marked by the continuous increase of changes in risk energy and residual modulus, specifically: Read the risk energy change and residual modulus change corresponding to the current monitoring time window and the previous monitoring time window. When the risk energy change of two consecutive monitoring time windows is greater than the risk energy stability threshold, and the residual modulus change of two consecutive monitoring time windows is greater than the residual modulus stability threshold, the change direction of the principal Gaussian component in the current state transition code is not the negative 1 of the fall direction code, and the current monitoring time window is not marked as the risk expansion phase; The risk rise phase is marked by the transition from a state of maintenance to synchronous growth. Specifically: Read the risk phase marker of the previous monitoring time window, the current risk energy change, the current residual modulus change, and the current direction of change of the principal Gaussian component. Mark the previous monitoring time window as a state-maintaining phase. The current risk energy change is greater than the risk energy stability threshold. The current residual modulus change is greater than the residual modulus stability threshold. The current direction of change of the principal Gaussian component is the growth direction code 1. If the current monitoring time window does not mark the risk level transition phase or the risk expansion phase, mark the current monitoring time window as the risk rise phase. Based on the synchronous pullback, the risk pullback phase is marked as follows: Read the current risk energy change, the current residual modulus change, and the current direction of change of the principal Gaussian component. When the risk energy change is less than the negative of the risk energy stability threshold, the residual modulus change is less than the negative of the residual modulus stability threshold, and the direction of change of the principal Gaussian component is a falling direction encoded as negative 1, and the current monitoring time window has not been written with the risk cross-level phase, risk expansion phase, or risk rise phase, mark the current monitoring time window as the risk fall phase. The marked state maintains the phase, specifically: After completing the judgment of risk cross-level phase, risk expansion phase, risk rise phase and risk fall phase, read the monitoring time window that has not been written with risk phase mark, write the corresponding monitoring time window into the status maintenance phase mark, and bind the status maintenance phase mark with the construction object number and monitoring time. Connect the same risk phase markers to determine the state transition location and risk duration segment, and set the corresponding risk phase channel, wherein: Connecting the same risk phase markers and determining the state transition positions are specifically as follows: Risk phase markers are grouped according to the construction object number. Risk phase markers corresponding to the same construction object are arranged in ascending order of monitoring time. When the risk phase markers of adjacent monitoring time windows are the same and the monitoring time interval is one 60-second monitoring time window, a phase continuity connection is established. When the risk phase markers of adjacent monitoring time windows are different or the monitoring time interval exceeds one 60-second monitoring time window, the previous phase continuity connection is terminated. The index of the current monitoring time window is written into the state transition position field and the state transition position marker is set to 1. The state transition position marker corresponding to the monitoring time window that has not undergone phase switching is set to 0. The risk duration segment is identified and a risk phase channel is established, specifically as follows: Read the start and end monitoring times in each group of phases in continuous connection, write the start monitoring time into the risk duration segment start field, write the end monitoring time into the risk duration segment end field, count the number of continuous monitoring time windows contained from the start monitoring time to the end monitoring time, generate the risk duration, set the state holding phase as risk phase channel 0, set the risk escalation phase as risk phase channel 1, set the risk expansion phase as risk phase channel 2, set the risk level crossing phase as risk phase channel 3, and set the risk fallback phase as risk phase channel 4; The initial risk energy, state transition code, state transition location, risk duration segment, and residual trajectory index are written into the corresponding risk phase channel. These are then arranged according to the construction object, monitoring time, and risk phase channel to generate a risk phase code body, in which: The initial risk energy, state transition code, state transition location, risk duration segment, and residual trajectory index are written into the corresponding risk phase channel, specifically: Read the risk phase marker corresponding to each construction object in each monitoring time window, locate the risk phase channel according to the risk phase marker, write the initial risk energy into the energy field, write the previous principal Gaussian component index, the current principal Gaussian component index and the direction of change of the principal Gaussian component in the state transition code into the transition field in sequence, write the state transition position marker into the transition field, write the risk duration start point, risk duration end point and risk duration duration into the segment field in sequence, write the residual trajectory index in the risk residual manifold into the trajectory field, and write all fields corresponding to the risk phase channel that is not selected by the current risk phase marker to 0; The risk phase code body is generated as follows: The number of construction objects is counted and recorded as the object dimension length. The number of monitoring time windows is counted and recorded as the time dimension length. The number of risk phase channels is set to 5. An object dimension index is established in ascending order of construction object number, and a time dimension index is established in ascending order of monitoring time. A phase dimension index is established according to risk phase channels 0 to 4. At the corresponding positions of each object dimension index, time dimension index, and phase dimension index, the initial risk energy, the previous principal Gaussian component index, the current principal Gaussian component index, the direction of change of principal Gaussian component, the state transition position marker, the start point of the risk duration segment, the end point of the risk duration segment, the risk duration duration, and the residual trajectory index are stored. A risk phase encoding body composed of object dimension, time dimension, and risk phase channel dimension is generated.

[0029] In this embodiment, generating the risk transmission record includes: Read the risk phase encoding body and risk residual manifold, register construction object pairs with cross-object residual connections as candidate transit object pairs, combine risk phase channel encoding and state transition encoding to generate a transit entropy state transition sequence, where: Construction object pairs with cross-object residual connections are registered as candidate transfer object pairs, specifically: Read the connection start construction object number, connection end construction object number, connection monitoring time, construction area number, job task number, and residual direction consistency from the cross-object residual connection table. Form an object pair index by combining the connection start construction object number and the connection end construction object number in ascending order. Merge cross-object residual connection records with the same object pair index. Count the number of cross-object residual connections for each object pair. Register construction object pairs with a number of cross-object residual connections of not less than 3 and whose cross-object residual connections are distributed in not less than 2 monitoring time windows as candidate transfer object pairs. Write the object pair index, cross-object residual connection index, construction area number, and job task number into the candidate transfer object pair record. By combining risk phase channel encoding and state transition encoding, a transfer entropy state transition sequence is generated, specifically as follows: Read the risk phase channel number corresponding to each construction object in each monitoring time window, set the status maintenance phase channel number to 0, the risk escalation phase channel number to 1, the risk expansion phase channel number to 2, the risk level transition phase channel number to 3, and the risk decline phase channel number to 4. Read the previous principal Gaussian component index and the current principal Gaussian component index in the state transition code, subtract 1 from each index and convert them to index values ​​from 0 to 3. Convert the negative 1 of the principal Gaussian component change direction to 0, and convert the 0 of the principal Gaussian component change direction to 1. The process involves converting the change direction of the principal Gaussian component from 1 to 2, multiplying the risk phase channel number by 48, multiplying the previous principal Gaussian component index conversion value by 12, multiplying the current principal Gaussian component index conversion value by 3, and adding the four calculation results to generate a state transition symbol with a value range of 0 to 239. When the construction object lacks the previous monitoring time window record, the previous principal Gaussian component index is set as the current principal Gaussian component index and the change direction of the principal Gaussian component is set to 0. The state transition symbols are grouped according to the construction object number and arranged in ascending order according to the monitoring time to generate a transfer entropy state transition sequence. Candidate condition objects that meet the conditions of common region relationship, common task relationship, and cross-object residual connection are selected. A set of condition variables is generated based on the change of propagation entropy and preset selection conditions, where: Candidate object filtering is performed based on changes in propagation entropy and filtering conditions, specifically as follows: The first candidate object in the candidate transmission object pair is used as the source candidate object and the response candidate object in turn. The second candidate object is set to the opposite role. Based on the state transition sequence of the source candidate object, the state transition sequence of the response candidate object, and the state transition symbol of the previous monitoring time window of the response candidate object, the basic transmission entropy of the unincluded condition object is calculated. The state transition symbol of the previous monitoring time window of one candidate condition object is written into the condition state. The condition transmission entropy is recalculated. The transmission entropy reduction is generated by subtracting the condition transmission entropy from the basic transmission entropy. The transmission entropy reduction is divided by the sum of the basic transmission entropy and 0.000001 to generate the transmission entropy reduction ratio. Candidate condition objects with a transmission entropy reduction greater than 0 and a transmission entropy reduction ratio not less than 0.1 are retained. Set the threshold for filtering candidate condition objects based on the change in transit entropy, specifically as follows: Keeping the time order of the source candidate object state transition sequence and the response candidate object state transition sequence unchanged, the candidate condition object state transition sequence is subjected to 100 cyclic shifts. The shift length of the k-th cyclic shift is set to 1 plus k minus 1 divided by the sequence length minus 1. The reduction in propagation entropy corresponding to each of the 100 cyclic shifts is calculated. All the reductions in propagation entropy are sorted in ascending order. The reduction in propagation entropy that reaches 95% of the data in the sorted position is set as the condition variable screening threshold. Candidate condition objects whose actual reduction in propagation entropy is not less than the condition variable screening threshold are retained. Calculate the positive time interval between the risk rise phase position of the source candidate object and the state transition position of the response candidate object, and generate a set of candidate hysteresis orders, where: The positive time interval between the risk rise phase position of the source candidate object and the state transition position of the response candidate object is calculated as follows: For each candidate transmission object pair, establish direction records for the first candidate object pointing to the second candidate object and the second candidate object pointing to the first candidate object respectively. Read the monitoring time of the risk rise phase channel number 1 and the state transition position mark 1 in the risk phase encoding body of the source candidate object, and register the corresponding monitoring time as the risk rise phase position. Read the monitoring time of the response candidate object with the state transition position mark 1, subtract the monitoring time corresponding to the risk rise phase position of the source candidate object from the monitoring time of the response candidate object, divide the time difference by 60 seconds and take the integer to generate the positive time interval, and only retain the positive time interval with a value greater than 0. Generate a set of candidate lag orders, specifically as follows: Retain positive time intervals with values ​​from 1 to 15, register each positive time interval as the lag order of the corresponding direction, count the occurrence of each lag order, retain lag orders with an occurrence frequency of not less than 2, sort the retained lag orders in descending order of occurrence frequency, when the number of retained lag orders is greater than 5, select the top 5 lag orders in the sorting position, when the occurrence frequencies are the same, arrange the lag orders with smaller values ​​first, when there are no lag orders with an occurrence frequency of not less than 2, sort all positive time intervals in ascending order of value, read the positive time interval in the middle of the sorting position and round it to an integer, set the corresponding integer as the candidate lag order, when there are no positive time intervals in the range of 1 to 15, delete the corresponding direction record; Based on the candidate lag order, the transfer entropy state transition sequence is moved, and the conditional transfer entropy is calculated using the set of conditional variables. A cyclic shift is performed on the source sequence, and the preset quantile value of the cyclic shift result is set as the effective transfer threshold, where: The state transition sequence is based on the candidate lag order and the entropy transfer sequence, specifically: Read the state transition sequences of the source candidate objects and the state transition sequences of the response candidate objects. Establish a set of time alignment records for each candidate lag order. Write the state transition symbols of the source candidate objects corresponding to the current monitoring time minus the candidate lag order and 60-second monitoring time windows into the source historical state field. Write the state transition symbols of the response candidate objects corresponding to the current monitoring time into the response current state field. Write the state transition symbols of the response candidate objects corresponding to the current monitoring time minus one 60-second monitoring time window into the response historical state field. Connect the state transition symbols of each construction object in the condition variable set corresponding to the current monitoring time minus one 60-second monitoring time window in ascending order of construction object number to generate a condition state combination. Delete any missing time alignment records in any state field. Stop calculating the corresponding candidate lag order when the number of retained time alignment records is less than 30. The conditional transitive entropy is calculated by combining the set of condition variables, specifically as follows: Read the source historical state symbol, response current state symbol, response historical state symbol and conditional state combination from each time-aligned record. Count the number of times the response current state symbol, response historical state symbol, source historical state symbol and conditional state combination appear simultaneously. Count the number of times the response historical state symbol, source historical state symbol and conditional state combination appear simultaneously. Count the number of times the response current state symbol, response historical state symbol and conditional state combination appear simultaneously. Count the number of times the response historical state symbol and conditional state combination appear simultaneously. Divide each count by the total number of time-aligned records to generate the corresponding joint probability. Use the joint probability to calculate the information difference of the response current state before and after adding the source historical state when the response historical state and conditional state are known. Multiply the information differences of all state combinations by the corresponding joint probability and sum them to generate the conditional transit entropy. The information difference is calculated using the natural logarithm. When the conditional variable set is empty, only the response historical state is used to calculate the transit entropy. Perform a cyclic shift on the source sequence, specifically: Keeping the order of the current response state sequence, the historical response state sequence, and the condition state combination unchanged, perform 100 cyclic shifts on the source historical state sequence. Set the shift length of the k-th cyclic shift to 1 plus the remainder obtained by dividing k minus 1 by the length of the source historical state sequence minus 1. Write the state transition symbols that are shifted out from the end of the source historical state sequence into the starting position of the sequence in the original order. Recalculate the conditional propagation entropy based on the source historical state sequence after each cyclic shift to generate 100 cyclic shift conditional propagation entropies. Arrange the 100 cyclic displacement condition transfer entropies in ascending order of value, set the sorting position number to 95, set the cyclic displacement condition transfer entropy at the 95th sorting position as the effective transfer threshold for the corresponding candidate transfer direction and candidate lag order, subtract the effective transfer threshold from the actual condition transfer entropy to generate the effective transfer difference, and write the candidate transfer direction, candidate lag order, actual condition transfer entropy, effective transfer threshold, and effective transfer difference into the transfer test record. Candidate relationships with consistent direction within a continuously preset risk duration and conditional propagation entropy not less than the effective propagation threshold are identified as effective risk propagation relationships. The target lag order, risk propagation direction, risk precursor object, and risk response object are determined, and the propagation baseline monitoring time is bound to generate risk propagation records. Among these: Candidate relationships whose conditional propagation entropy is not less than the effective propagation threshold and whose direction is consistent within a continuously preset risk duration segment are identified as effective risk propagation relationships. Specifically: Read the transmission test records, retain candidate transmission directions where the actual condition transmission entropy is not less than the effective transmission threshold and the effective transmission difference is greater than 0, group the retained records according to the candidate transmission object pairs and candidate transmission directions, compare the effective transmission difference of different candidate lag orders corresponding to the same candidate transmission direction, register the candidate lag order with the largest effective transmission difference as the lag order to be confirmed, and when multiple candidate lag orders correspond to the same effective transmission difference, select the candidate lag order with the smallest value to generate the risk transmission relationship to be confirmed; Read the corresponding risk duration segments of candidate transmission objects in ascending order of the starting monitoring time of the risk duration segment. Set two adjacent risk duration segments as continuous risk duration segments. Read the risk rise phase position of the source candidate object in each risk duration segment. Add the 60-second monitoring time window of the lag order to be confirmed to the risk rise phase position to generate the response matching time. Read the state transition position of the response candidate object within a 60-second monitoring time window before and after the response matching time. When there are state transition positions of response candidate objects in both of the two continuous risk duration segments and there is a cross-object residual connection between the risk rise phase position and the state transition position of the response candidate object, register the two risk duration segments as segments with the same direction. When either risk duration segment does not meet the corresponding condition, cancel the corresponding risk transmission relationship to be confirmed. The target lag order and risk transmission direction are determined as follows: The lag order to be confirmed is set as the target lag order if the directions of two consecutive risk duration segments are consistent. The effective transmission difference between the first candidate object and the second candidate object is read respectively. When only one direction passes the effective transmission judgment, the corresponding direction is set as the risk transmission direction. When both directions pass the effective transmission judgment, the direction with the larger effective transmission difference is set as the risk transmission direction. When the absolute value of the difference between the effective transmission differences of the two directions is not greater than 0.000001, the direction judgment of the corresponding candidate transmission object pair is cancelled. The risk precursors and risk response targets are identified as follows: Read the determined risk transmission direction, register the construction object corresponding to the starting point of the risk transmission direction as the risk precursor object, register the construction object corresponding to the ending point of the risk transmission direction as the risk response object, and write the risk precursor object number, risk response object number, target lag order, condition variable set index and cross-object residual connection index into the risk transmission object record. Generate a risk transmission record, specifically as follows: Read the state transition position corresponding to the risk response object within the consistent direction segment, set the monitoring time corresponding to the state transition position as the transmission benchmark monitoring time, subtract the target lag order and a 60-second monitoring time window from the transmission benchmark monitoring time to generate the risk precursor monitoring time, and write the risk precursor object number, risk response object number, risk transmission direction, target lag order, risk precursor monitoring time, transmission benchmark monitoring time, actual condition transmission entropy, effective transmission threshold, effective transmission difference, condition variable set index, construction area number, work task number, risk persistence segment index, and cross-object residual connection index into the same record. Sort all records in ascending order according to the transmission benchmark monitoring time to generate a risk transmission record.

[0030] In this embodiment, generating the risk propagation path includes: Read risk transmission records, risk mapping latent variable time series records, reconstruct features and state transition codes to obtain risk precursor objects, risk response objects, target lag order and transmission baseline monitoring time; Historical monitoring time is determined based on the target lag order. Precursor historical latent variables and response baseline latent variables are read according to the object number and monitoring time. Coordinate changes are calculated, and coordinates that meet preset coordinate change conditions are identified as risks and written into the coordinate system. Specifically: The calculation of coordinate changes is as follows: Read the precursor historical latent variables and precursor reference latent variables sequentially according to the latent variable coordinate numbers 0 to 15. Subtract the corresponding coordinate value of the precursor reference latent variable from the coordinate value of the precursor historical latent variable to generate the precursor coordinate change. Read the response baseline latent variable and response reference latent variable. Subtract the corresponding coordinate value of the response reference latent variable from the coordinate value of the response baseline latent variable to generate the response coordinate change. Subtract the response coordinate change from the precursor coordinate change and take the absolute value to generate the propagation coordinate difference. Arrange the 16 precursor coordinate changes, response coordinate changes and propagation coordinate differences according to the coordinate numbers. Read the time series records of risk mapping latent variables corresponding to the normal construction training data during the model training phase. Calculate the change in coordinates of each latent variable within an adjacent 60-second monitoring time window according to the construction object number and monitoring time. Convert the change in each coordinate to an absolute value and sort them in ascending order. Set the values ​​that reach 90% of the corresponding data position as the coordinate change threshold for the corresponding latent variable coordinate. When the number of valid changes corresponding to a single latent variable coordinate is less than 30, read the absolute values ​​of the changes in all latent variable coordinates and set the values ​​that reach 90% of the total data position as the corresponding coordinate change threshold. Coordinates that meet the coordinate change conditions are identified as risks and written into the coordinate system, specifically as follows: Read the preceding coordinate change, response coordinate change, propagation coordinate difference, and coordinate change threshold corresponding to each latent variable coordinate. The absolute value of the preceding coordinate change is not less than the coordinate change threshold, the propagation coordinate difference is not less than the coordinate change threshold, the product of the preceding coordinate change and the response coordinate change is not less than 0, and when the absolute value of the preceding coordinate change is greater than the absolute value of the response coordinate change, register the corresponding latent variable coordinate as a risk writing coordinate. Arrange all risk writing coordinates in descending order of propagation coordinate difference. When the number of risk writing coordinates is greater than 8, retain the top 8 risk writing coordinates in the sorting position. Limit the number of risk writing coordinates to 8 to retain no less than 8 original coordinates in the response baseline latent variable. According to the risk transmission direction, the coordinate values ​​of the preceding historical latent variables at the risk writing coordinate are written into the response benchmark latent variables to generate associated risk disturbance latent variables; The latent variables of associated risk perturbation and the reconstructed features are input into the estimation network, the membership degrees of the Gaussian components are recalculated, the Gaussian mixture parameters of the improved DAGMM model are read, and the associated risk energy is calculated by combining state transition coding. Specifically, the associated risk energy is calculated by combining state transition coding as follows: Calculate the difference between the 18-dimensional correlation estimation input vector and the corresponding element of the mean vector of each Gaussian component. Calculate the Gaussian distance of the difference vector using the inverse matrix of the corresponding Gaussian covariance matrix. Read the determinant of the corresponding Gaussian covariance matrix and calculate the Gaussian probability density. Multiply the four Gaussian probability densities by the corresponding correlation component coefficients and sum them. Take the natural logarithm of the summation result and the sum of 0.000001 and change the sign to generate the correlation risk energy. Read the initial risk energy corresponding to the risk response object at the transmission benchmark monitoring time. Subtract the initial risk energy from the correlation risk energy to generate the correlation risk energy increment. Write the correlation risk energy, correlation risk energy increment, risk into the coordinate number, and transmission benchmark monitoring time into the correlation risk energy record. The risk transmission records are arranged according to the monitoring time. Records that are consistent with the previous risk response object and the next risk precursor object are connected. Circular connections are deleted. The risk transmission object sequence, transmission time sequence and object association information are combined to generate the risk transmission path.

[0031] In this embodiment, generating construction safety early warning results and construction safety monitoring records includes: The historical reliable construction status data is used to perform construction status sample construction, improved DAGMM analysis, risk residual manifold construction, risk phase encoding, transfer entropy analysis, and risk transfer write-back to generate historical associated risk energy. A threshold for determining associated risks is set using preset quantile values, where, specifically: Historical associated risk energy records are grouped according to the first two digits of the construction object number. Personnel object category code 01, machinery object category code 02, structural monitoring object category code 03, and environmental monitoring object category code 04 are each established as a historical associated risk energy set. The historical associated risk energy in each set is arranged in ascending order of value. The number of historical associated risk energies in each set is counted and recorded as N. 0.95 is multiplied by N and rounded up to generate a threshold sorting position. The historical associated risk energy corresponding to the threshold sorting position is read and set as the associated risk judgment threshold for the corresponding construction object category. The 95th percentile is used to retain 95% of the associated risk energy under normal historical construction conditions and identifies associated risk energy exceeding the upper range of the normal distribution as a risk state. When the number of historical associated risk energies corresponding to a single construction object category is less than 30, the historical associated risk energies corresponding to the four construction object categories are merged and the 95th percentile is recalculated. When the number of historical associated risk energies after merging is still less than 30, the historical data reading range is increased by 30 days and the historical associated risk energy calculation is re-executed. The historical data reading range is extended to a maximum of 365 days. Construction objects with current associated risk energy not less than the associated risk judgment threshold are identified as risk status objects and marked with a risk status tag. Specifically, this involves: Read the risk response object number, associated risk energy, risk propagation path number and transmission benchmark monitoring time from the current associated risk energy record. Read the first two digits of the risk response object number as the object category code. Read the corresponding associated risk judgment threshold according to the object category code. When the same risk response object corresponds to multiple risk propagation paths in the same 60-second monitoring time window, sort all associated risk energies from largest to smallest. Set the associated risk energy at the first position of the sort as the associated risk energy of the current object. Set the corresponding risk propagation path number as the main risk propagation path number. Write the remaining risk propagation path numbers into the associated path index. Compare the associated risk energy of the current object with the associated risk judgment threshold of the corresponding construction object category one by one. When the associated risk energy of the current object is not less than the associated risk judgment threshold, register the corresponding construction object number as a risk status object and write the risk status mark to 1. When the associated risk energy of the current object is less than the associated risk judgment threshold, write the risk status mark to 0. Subtract the associated risk judgment threshold from the associated risk energy of the current object to generate the risk energy excess. Write the risk status object number, risk status mark, associated risk energy of the current object, associated risk judgment threshold, risk energy excess, main risk propagation path number and monitoring time into the risk status record. Based on the risk propagation path, determine the risk source and the risk impact object, read the corresponding construction area number to generate the risk source area and the risk impact area, and read the propagation time series to generate the risk propagation time; By combining the risk source object, the risk impact object, the risk source area, the risk impact area, the work task number, the risk propagation time, the associated risk energy, and the risk propagation path, a construction safety early warning result is generated. This result is then associated with the construction object number, the construction area number, the work task number, and the monitoring time to generate a construction safety monitoring record.

[0032] Example 1: In a construction project with multiple work areas operating in parallel, the site was divided into a structural work area, a material transfer area, and a temporary storage area. Personnel, machinery, structural, environmental, and image acquisition devices were deployed. Theoretically, 186,420, 42,860, 31,680, and 19,200 monitoring records were generated for construction personnel, construction machinery, construction structures, and the construction environment, respectively. 8,640 frames of construction site images, 1,260 construction task records, and 438 construction safety records were also generated. After filtering out communication packet loss, duplicate records, and invalid data, 185,937, 42,718, 31,496, 19,083, 8,574 frames, 1,237, and 438 records were retained, respectively. The sampling intervals for personnel, machinery, structure, and environmental data were 5 seconds, 2 seconds, 10 seconds, and 10 seconds, respectively. However, issues arose on-site, including inconsistent sampling times, dispersed object relationships, and some individual statuses not exceeding limits.

[0033] The system employs a 60-second monitoring time window for time alignment. It calculates the average of valid fields for the same object within the same time window, writes 0 for missing fields and marks them as invalid, resulting in 25,740 reliable construction status records. Personnel, machinery, structure, environment, image, and task status each have 8 field slots, forming a 48-dimensional construction status vector. Model training reads data from multiple historical monitoring periods, removing associated samples within 30 minutes before and after a safety incident, and from 30 minutes before to 30 minutes after the start of risk management, resulting in 186,000 normal construction samples. These are divided into 130,200 training samples, 27,900 validation samples, and 27,900 test samples. Sample 010125 corresponds to personnel, with normalized values ​​of 0.31, 0.44, and 0.52 for speed, dwell time, and personnel density, respectively. Sample 020048 corresponds to machinery, with normalized values ​​of 0.63, 0.28, and 0.74 for load rate, rotation speed, and task load, respectively. Each training batch consists of 128 samples. The Adam optimizer is used with an initial learning rate of 0.001 and four error weights of 0.5, 0.25, 0.15 and 0.1. The target change rate is verified to be less than 0.0005 for eight consecutive rounds in the 164th round, and the parameters are saved.

[0034] During the simulation, the crane lifted components, and 12 personnel performed installation within the structural work area. Traditional methods set alarm thresholds of 85%, 0.5 revolutions per minute, 0.4 meters, 0.8 people per square meter, 90 seconds, 6 millimeters, and 0.6 meters per square second for the mechanical load rate, rotation speed, hook swing amplitude, personnel density, personnel dwell time, structural lateral displacement, and root mean square acceleration. Within the 8th monitoring window, the mechanical load rate increased from 62% to 78%, the rotation speed increased from 0.24 revolutions per minute to 0.36 revolutions per minute, and the hook swing amplitude increased from 0.18 meters to 0.33 meters. Within the 9th monitoring window, the personnel density increased from 0.48 people per square meter to 0.66 people per square meter, the dwell time increased from 41 seconds to 59 seconds, and the image occlusion rate increased from 12% to 27%. During the 10th monitoring window, the lateral displacement of the supporting component increased from 2.9 mm to 4.7 mm, and the root mean square vibration acceleration increased from 0.25 m / s² to 0.43 m / s². All values ​​were below the alarm threshold, and traditional methods failed to issue a warning.

[0035] After inputting construction status samples into the improved DAGMM model, 16-dimensional initial latent variables were generated. The initial risk energy of the 7th monitoring time window was 1.68, and the residual modulus was 0.23, marked as the state-maintaining phase. The historical data corresponded to a risk energy stability threshold of 0.42, a residual modulus stability threshold of 0.11, and risk energy interval boundaries of 1.4, 3.2, and 4.5. The initial risk energies of the 8th to 10th monitoring time windows were 2.14, 2.83, and 3.71, and the residual modulus was 0.36, 0.51, and 0.74, respectively. The 8th window was marked as the risk escalation phase, the 9th window as the risk expansion phase, and the 10th window as the risk level transition phase due to the risk energy crossing 3.2. These three phase segments continuously formed a risk evolution process.

[0036] In the risk residual manifold, the residual direction consistency between the lifting machinery and the personnel group is 0.81, while that between the personnel group and the supporting components is 0.76. The transfer entropy algorithm filters out environmental monitoring objects that share a common region and task with both the lifting machinery and the personnel group, and that have cross-object residual connections. The conditional transfer entropy from the lifting machinery to the personnel group is 0.186 nat, with an effective transfer threshold of 0.112 nat; the conditional transfer entropy from the personnel group to the supporting components is 0.163 nat, with an effective transfer threshold of 0.105 nat. The target lag order is 1, generating a risk propagation path consisting of the lifting machinery, the personnel group, and the supporting components.

[0037] After writing back the risk transmission records to improve the DAGMM model, five risk coordinates were identified. The risk energy associated with personnel groups increased from 3.08 to 4.26, and the risk energy associated with supporting components increased from 3.71 to 5.02, exceeding the judgment thresholds of 3.84 and 4.21, respectively. The system identified the lifting machinery as the risk source, personnel groups and supporting components as the risk impact objects, and the material transfer area and structural operation area as the risk source area and risk impact area, respectively. The total risk propagation time was 120 seconds, generating early warning results and monitoring records.

[0038] The comparative verification used the same 186,000 historical samples to train both the traditional method and the method of this invention. The traditional method uses the original DAGMM to identify abnormal states and infers the propagation relationship based on the abnormal energy exceeding the limit time and the correlation coefficient of the object state. The test set contains 4,200 samples: 3,360 normal samples, 420 asymptotic risk samples, 280 cross-object propagation risk samples, and 140 state transition risk samples. The traditional method had 396 false positives and 335 false negatives, while the method of this invention had 175 false positives and 102 false negatives. The accuracy rate of risk status identification improved from 82.6% to 93.4%, the accuracy rate of progressive risk identification improved from 68.8% to 89.5%, the accuracy rate of cross-object propagation risk identification improved from 54.3% to 88.2%, the accuracy rate of risk source object location improved from 61.7% to 91.6%, and the accuracy rate of risk propagation direction identification improved from 57.9% to 89.7%. The false alarm rate decreased from 11.8% to 5.2%, and the false negative rate decreased from 39.9% to 12.1%. The lead time, calculated by subtracting the warning generation time from the time of occurrence of the risk event after manual verification, is 48 seconds on average with traditional methods, while this invention provides an average lead time of 132 seconds.

[0039] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A construction safety monitoring method based on big data collection, characterized in that, include: Collect construction monitoring data, perform preprocessing on the construction monitoring data, and generate reliable construction status data; Arrange reliable construction status data according to construction object number, construction area number, work task number and monitoring time to generate a construction status sample set and construction object association records; An improved DAGMM model is constructed by setting the risk mapping latent space structure, risk disturbance latent variables and state transition coding. The construction state sample set is input into the improved DAGMM model to generate the reconstruction state of the construction object, the initial risk energy and the state transition code. Based on reliable construction status data, reconstruction status of construction objects and associated records of construction objects, the reconstruction residual is calculated and a local residual neighborhood is constructed to generate a risk residual manifold. Based on the initial risk energy, state transition coding, and risk residual manifold, the state transition positions and risk persistence segments are extracted to generate the risk phase coding body; The risk transmission record is generated by performing a transfer entropy algorithm based on the risk phase encoding body and the risk residual manifold, through conditional transfer entropy variable filtering, adaptive lag order search and transfer entropy state transition encoding. The risk transmission records are written back to the improved DAGMM model. Based on the risk transmission records, the latent variable associations in the risk mapping latent space are reconstructed and the associated risk energy is recalculated. The risk propagation path is generated according to the object connection relationship and propagation sequence in the risk transmission records. Based on the associated risk energy and risk propagation path, the risk source object, risk impact object, risk occurrence area and risk propagation time are determined, and construction safety early warning results and construction safety monitoring records are generated.

2. The construction safety monitoring method based on big data collection according to claim 1, characterized in that, The construction monitoring data specifically includes construction personnel monitoring data, construction machinery operation data, construction structure monitoring data, construction environment monitoring data, construction site image data, construction operation task data, construction site layout data, and construction safety record data.

3. The construction safety monitoring method based on big data collection according to claim 1, characterized in that, The preprocessing of construction monitoring data includes time alignment, anomaly handling, object binding, and region matching.

4. The construction safety monitoring method based on big data collection according to claim 1, characterized in that, The generation of the construction status sample set and the construction object association record includes: Read reliable construction status data and extract construction object identifier, location, status, monitoring time, on-site target status, work task number, work area, task status, construction area number and area location; The construction object is identified and registered as a construction object number, associated with the on-site target status, and the construction object area relationship and construction object task relationship are established based on the object location, work area and area location. Write the object status according to the order of personnel, machinery, structure, environment, image and task status fields, write non-corresponding fields into invalid field markers, and generate a fixed-dimensional construction status vector. Arrange fixed-dimensional construction status vectors according to construction object number and monitoring time, bind them with construction area number, work task number, trusted status marker and valid field marker to generate a construction status sample set, combine construction object number, construction area number, work task number, construction object area relationship, construction object task relationship and monitoring time to generate construction object association record.

5. The construction safety monitoring method based on big data collection according to claim 1, characterized in that, The generation of the construction object reconstruction state, initial risk energy, and state transition code includes: An improved DAGMM model is constructed, retaining the encoder, decoder, estimation network, Gaussian mixture estimation and sample energy calculation. A risk mapping latent space structure is set at the encoder output, risk perturbation latent variable processing is set at the estimation network input, and state transition coding processing is set within the estimation network. Based on the construction safety record data, historical reliable construction status data that has been removed by removing related safety events and risk handling records is used to train and improve the DAGMM model based on the remaining historical reliable construction status data. Input the construction status sample set into the encoder, read the fixed-dimensional construction status vector according to the valid field marking, perform compression encoding on the fixed-dimensional construction status vector, and generate initial latent variables. By reading the regional and task relationships of construction objects through the risk mapping latent space structure, a latent variable correlation matrix is ​​generated. Based on the latent variable correlation matrix, an association index is established between the initial latent variables. The adjacency order of the initial latent variables is adjusted according to the association index. The coordinates of latent variables with regional and task relationships of construction objects are arranged to continuous coordinate positions to generate risk mapping latent variables. Input the risk mapping latent variables into the decoder to perform state reconstruction, generate the reconstruction state of the construction object, calculate the field difference, reconstruction distance and directional similarity between the fixed-dimensional construction state vector and the reconstruction state of the construction object according to the effective field labels, and generate reconstruction features; The risk mapping latent variables and reconstructed features are input into the estimation network to generate Gaussian component membership degrees. The indexes of the principal Gaussian components and the direction of change of the principal Gaussian components corresponding to adjacent monitoring time windows are read to generate state transition codes. The state transition codes and Gaussian component membership degrees are combined according to the Gaussian component index to generate transition state membership degrees. Gaussian mixture parameters are calculated based on the transition state membership degrees, and the initial risk energy is calculated based on the risk mapping latent variables, reconstructed features, and Gaussian mixture parameters.

6. The construction safety monitoring method based on big data collection according to claim 1, characterized in that, The generation of the risk residual manifold includes: The reliable construction status data and the reconstruction status of the construction object are matched according to the construction object number and monitoring time. The difference of the effective fields is calculated based on the effective field markers to generate the reconstruction residual vector of the construction object. The residual vector of the reconstructed construction object is used as the residual space coordinate. The residual modulus, the change in residual modulus, and the direction of residual change are calculated. The continuous residual space coordinates are connected to generate the object residual trajectory. Based on the regional and task relationships of the construction objects, we screen the pairs of related construction objects, compare the changes in residual modulus and the direction of residual change within the same monitoring time window, and generate cross-object residual connections. Calculate the residual distance between adjacent residual spatial coordinates, establish a local residual neighborhood based on the object residual trajectory and cross-object residual connections, and combine the residual spatial coordinates, residual distance, object residual trajectory and local residual neighborhood to generate the risk residual manifold.

7. The construction safety monitoring method based on big data collection according to claim 1, characterized in that, The generation of the risk phase code includes: Read the initial risk energy, state transition code and risk residual manifold, calculate the change in risk energy and the change in residual modulus, set the stability threshold based on the median of the absolute value of the change corresponding to the historical normal construction state, and set the risk energy interval boundary based on the preset quantile value of the initial risk energy. According to the preset priority order, based on the initial risk energy crossing the risk energy range, the continuous growth of the risk energy change and the residual modulus change, the transition from the holding state to synchronous growth, synchronous decline and the remaining state, the risk level phase, risk expansion phase, risk rise phase, risk fall phase and state holding phase are marked in sequence. Connect the same risk phase markers to determine the state transition location and risk duration segment, and set the corresponding risk phase channel; The initial risk energy, state transition code, state transition location, risk duration segment, and residual trajectory index are written into the corresponding risk phase channel and arranged according to the construction object, monitoring time, and risk phase channel to generate a risk phase code body.

8. The construction safety monitoring method based on big data collection according to claim 1, characterized in that, The generation of risk transmission records includes: Read the risk phase encoding body and risk residual manifold, register the construction object pairs with cross-object residual connections as candidate transfer object pairs, combine the risk phase channel encoding and state transition encoding, and generate the transfer entropy state transition sequence. Candidate condition objects that meet the conditions of common region relationship, common task relationship and cross-object residual connection are selected, and a set of condition variables is generated based on the change of propagation entropy and preset selection conditions. Calculate the positive time interval between the risk rise phase position of the source candidate object and the state transition position of the response candidate object to generate a set of candidate hysteresis orders; Based on the candidate lag order, the transfer entropy state transition sequence is moved, and the conditional transfer entropy is calculated in combination with the set of conditional variables. The source sequence is then subjected to cyclic displacement, and the preset quantile value of the cyclic displacement result is set as the effective transfer threshold. Candidate relationships with consistent directions within a continuous preset number of risk intervals and conditional transmission entropy not less than the effective transmission threshold are identified as effective risk transmission relationships. The target lag order, risk transmission direction, risk precursor object, and risk response object are determined, and the transmission benchmark monitoring time is bound to generate risk transmission records.

9. A construction safety monitoring method based on big data collection according to claim 1, characterized in that, The generation of risk propagation paths includes: Read risk transmission records, risk mapping latent variable time series records, reconstruct features and state transition codes to obtain risk precursor objects, risk response objects, target lag order and transmission baseline monitoring time; Based on the target lag order, the historical monitoring time is determined. The preceding historical latent variables and the response baseline latent variables are read according to the object number and monitoring time. The coordinate change is calculated, and the coordinates that meet the preset coordinate change conditions are identified as risks and written into the coordinate system. According to the risk transmission direction, the coordinate values ​​of the preceding historical latent variables at the risk writing coordinate are written into the response benchmark latent variables to generate associated risk disturbance latent variables; Input the latent variables of associated risk perturbation and the reconstructed features into the estimation network, recalculate the membership degree of the Gaussian component, read the Gaussian mixture parameters of the improved DAGMM model, and calculate the associated risk energy by combining state transition coding. The risk transmission records are arranged according to the monitoring time. Records that are consistent with the previous risk response object and the next risk precursor object are connected. Circular connections are deleted. The risk transmission object sequence, transmission time sequence and object association information are combined to generate the risk transmission path.

10. A construction safety monitoring method based on big data collection according to claim 1, characterized in that, The generation of construction safety early warning results and construction safety monitoring records includes: The historical reliable construction status data is used to perform construction status sample construction, improved DAGMM analysis, risk residual manifold construction, risk phase encoding, transfer entropy analysis and risk transfer write-back to generate historical associated risk energy, and the associated risk judgment threshold is set with preset quantile values. Construction objects whose current associated risk energy is not less than the associated risk judgment threshold are identified as risk status objects and written into the risk status flag; Based on the risk propagation path, determine the risk source and the risk impact object, read the corresponding construction area number to generate the risk source area and the risk impact area, and read the propagation time series to generate the risk propagation time; By combining the risk source object, the risk impact object, the risk source area, the risk impact area, the work task number, the risk propagation time, the associated risk energy, and the risk propagation path, a construction safety early warning result is generated. This result is then associated with the construction object number, the construction area number, the work task number, and the monitoring time to generate a construction safety monitoring record.

Images