Engineering safety progress intelligent monitoring method based on multi-source data collaboration

Abstract

The invention discloses an engineering safety progress intelligent monitoring method based on multi-source data collaboration, which relates to the technical field of intelligent engineering monitoring, and comprises the following steps of: mapping multi-source engineering monitoring data into nodes and edges of a graph in real time by utilizing an incremental graph updating algorithm, generating a dynamic knowledge graph, and generating a dynamic mapping result; performing graph traversal on the dynamic knowledge graph through an association subgraph extraction algorithm, extracting a security event matrix and a project progress matrix, calculating an SPI index value by using a dynamic weighted fusion algorithm, synchronously performing multi-threshold interval grading on the SPI index value, generating an SPI early warning level, performing state coding on the SPI early warning level, and generating a comprehensive feature vector; according to the method, data of different types can be effectively integrated and a basis is provided for formulating a targeted engineering safety progress solution through an incremental graph updating algorithm and a Bayesian causal graph model, and node probability distribution characteristics are extracted by using a forward propagation layer. The method is advantaged in that the incremental graph updating algorithm and the Bayesian causal graph model are utilized to effectively integrate data of different types and provide a basis for formulating a targeted engineering safety progress solution.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent engineering monitoring, and in particular to an intelligent monitoring method for engineering safety progress based on multi-source data collaboration. Background Art

[0002] Intelligent monitoring of project safety and progress is an important part of modern construction project management. It aims to improve the scientific management and risk control capabilities of engineering projects through real-time perception, dynamic analysis and intelligent decision-making. In large and complex construction projects, it is crucial to ensure the safety of the construction process and the accuracy of the progress.

[0003] Traditional technologies typically use IoT-based structural health monitoring (SHM) models to collect and analyze structural response data in real time, identifying trends in project safety and predicting potential risk points. However, these traditional technologies still have several shortcomings. For one thing, existing SHM models struggle to effectively integrate data from disparate sources, leading to severe information silos. Furthermore, they lack the ability to conduct in-depth root-cause analysis of safety incidents or schedule deviations in complex engineering environments, often providing only superficial problem descriptions and failing to provide specific, targeted measures. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an intelligent monitoring method for engineering safety progress based on multi-source data collaboration to solve the problems of insufficient multi-source data fusion and lack of root cause analysis capabilities in the prior art.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides an intelligent monitoring method for engineering safety progress based on multi-source data collaboration, which includes using an incremental graph update algorithm to map multi-source engineering monitoring data into nodes and edges of a graph in real time to generate a dynamic knowledge graph;

[0008] The dynamic knowledge graph is traversed through the associated subgraph extraction algorithm to extract the security event matrix and the project progress matrix. The SPI index value is calculated using a dynamic weighted fusion algorithm. The SPI index value is simultaneously graded by multiple threshold intervals to generate the SPI warning level.

[0009] The SPI warning level is state-encoded to generate a comprehensive feature vector, which is then input into the Bayesian causal graph model. The forward propagation layer is used to extract the node probability distribution characteristics, and the counterfactual intervention layer is used to extract the blocking impact characteristics. The entropy weight method is then used to simultaneously perform weight distribution and generate an abnormal contribution matrix.

[0010] The hierarchical clustering algorithm is used to perform multi-level attribution aggregation on the anomaly contribution matrix to generate root cause analysis data. Based on the root cause analysis data, the node weights and edge strengths of the dynamic knowledge graph are incrementally updated to generate an engineering safety progress monitoring report.

[0011] As a preferred solution of the intelligent monitoring method for engineering safety progress based on multi-source data collaboration described in the present invention, the multi-source engineering monitoring data includes sensor monitoring data, engineering structure parameter data and construction process management data.

[0012] As a preferred solution of the engineering safety progress intelligent monitoring method based on multi-source data collaboration of the present invention, wherein: the generation of the dynamic knowledge graph specifically includes the following steps:

[0013] An incremental graph update algorithm is used to stream partition multi-source engineering monitoring data. Based on a time-series graph database, the topological structure is mapped through an adjacency table index mechanism to obtain the nodes and edges of the graph.

[0014] The weights of nodes and edges are adjusted through the Bayesian confidence dynamic update method to generate a dynamic knowledge graph.

[0015] As a preferred solution of the engineering safety progress intelligent monitoring method based on multi-source data collaboration described in the present invention, wherein: the SPI index value is calculated using a dynamic weighted fusion algorithm, specifically including the following steps:

[0016] The dynamic knowledge graph is subjected to semantic diffusion and subgraph extraction through the associated subgraph extraction algorithm, and the graph is traversed through depth-first search to generate an engineering safety progress topology network.

[0017] Graph embedding is used to map abnormal events to the project safety progress topology network to obtain a safety event matrix. The EVM algorithm is then used to quantify resource consumption and obtain a project progress matrix.

[0018] The safety event matrix and the project progress matrix are weighted and aggregated through a dynamic weighted fusion algorithm to generate the SPI index value.

[0019] As a preferred solution of the engineering safety progress intelligent monitoring method based on multi-source data collaboration of the present invention, wherein: the generating of the SPI warning level specifically includes the following steps:

[0020] Sliding window normalization is used to smooth and unify the SPI index values to obtain a standardized SPI sequence. The standardized SPI sequence is then divided into intervals according to a three-level threshold to generate an SPI warning level.

[0021] As a preferred solution of the engineering safety progress intelligent monitoring method based on multi-source data collaboration of the present invention, wherein: the generation of the abnormal contribution matrix specifically includes the following steps:

[0022] A hierarchical clustering algorithm is used to cluster the SPI warning levels based on similarity, and singular value decomposition is used to encode the state and generate a comprehensive feature vector.

[0023] The forward propagation layer and the counterfactual intervention layer are parameterized and stacked through the automatic differentiation mechanism to construct a Bayesian causal graph model;

[0024] The comprehensive feature vector is input into the Bayesian causal graph model, and the forward propagation layer uses causal convolution to aggregate the time dimension and spatial density to obtain the node probability distribution characteristics;

[0025] The counterfactual intervention layer blocks the causal path through Gumbel-max reparameterization to obtain the blocking impact characteristics;

[0026] The node probability distribution features and blocking impact features are weighted by the entropy weight method, and the attention channel is used to splice them to generate the abnormal contribution matrix.

[0027] As a preferred solution of the engineering safety progress intelligent monitoring method based on multi-source data collaboration of the present invention, wherein: generating root cause analysis data specifically includes the following steps:

[0028] The hierarchical clustering algorithm is used to perform full-connection clustering on the anomaly contribution matrix to obtain multi-level anomaly cluster nodes. The causal reasoning method is then used to perform multi-level attribution aggregation on the multi-level anomaly cluster nodes to generate root cause analysis data.

[0029] As a preferred solution of the engineering safety progress intelligent monitoring method based on multi-source data collaboration of the present invention, wherein: the generating of the engineering safety progress monitoring report specifically includes the following steps:

[0030] The Flink engine is used to input root cause analysis data into a dynamic knowledge graph. The node weights are updated using a dynamic decay algorithm. The Dijkstra algorithm is used to optimize the edge strength path and obtain the causal link vector.

[0031] The transfer matrix decomposition method is used to convert the causal link vector into structured safety assessment parameters and real-time progress parameters, which are then integrated into a project safety progress monitoring report.

[0032] In a second aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the intelligent monitoring method for engineering safety progress based on multi-source data collaboration as described in the first aspect of the present invention is implemented.

[0033] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the method for intelligent monitoring of engineering safety progress based on multi-source data collaboration as described in the first aspect of the present invention.

[0034] The present invention has the following beneficial effects: Utilizing an incremental graph update algorithm, multi-source engineering monitoring data is mapped into a dynamic knowledge graph in real time, effectively integrating different types of data and ensuring its timeliness and accuracy. Through an associated subgraph extraction algorithm and a Bayesian causal graph model, the specific causes of problems and their impact paths are identified, and potential causal relationships are deeply explored, providing a basis for developing targeted solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0036] Figure 1 This is a flow chart of the intelligent monitoring method for engineering safety progress based on multi-source data collaboration.

[0037] Figure 2 Flowchart generated for a dynamic knowledge graph.

[0038] Figure 3 This is a flow chart of the SPI index calculation and early warning process.

[0039] Figure 4 Flowchart generated for root cause analysis data. DETAILED DESCRIPTION

[0040] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0041] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0042] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0043] Reference Figures 1 to 4 , is an embodiment of the present invention, which provides an intelligent monitoring method for engineering safety progress based on multi-source data collaboration, including the following steps:

[0044] S1. Use the incremental graph update algorithm to map multi-source engineering monitoring data into graph nodes and edges in real time to generate a dynamic knowledge graph.

[0045] The specific steps include:

[0046] S1.1. Collect multi-source engineering monitoring data, including sensor monitoring data, engineering structure parameter data, and construction process management data;

[0047] Among them, sensor monitoring data includes structural deformation data, temperature and humidity data, and vibration displacement data. Structural deformation data is collected by strain gauge sensors; temperature and humidity data is collected by temperature and humidity sensors; vibration displacement data is collected by accelerometers. Structural deformation data, temperature and humidity data, and vibration displacement data are integrated through a centralized data collector to form sensor monitoring data.

[0048] Engineering structure parameter data includes geometric dimension data, material property data, and position coordinate data. Geometric dimension data is collected by 3D laser scanners; material property data is collected by non-destructive testing equipment; and position coordinate data is collected by total stations or GPS positioning devices. The geometric dimension data, material property data, and position coordinate data are integrated through a centralized data processing platform to form engineering structure parameter data.

[0049] Construction process management data includes construction plans, quality inspection reports, and field work logs. Construction plans and resource allocation records are collected through project management software (such as Primavera P6); quality inspection reports are collected through mobile applications (such as Procore); and field work logs are collected through the field manager's log entry software (such as Fieldwire). Construction plans, resource allocation records, quality inspection reports, and field work logs are integrated into construction process management data through the ERP unit within the project management.

[0050] S1.2. Preprocess the collected multi-source engineering monitoring data. In the specific operation, first, the multi-source engineering monitoring data is decomposed into multiple layers through wavelet decomposition filtering to obtain high-frequency wavelet coefficients, and the high-frequency wavelet coefficients are set to zero using soft thresholding to eliminate high-frequency noise; the KNN interpolation algorithm is selected to perform nearest neighbor weighted interpolation on the multi-source engineering monitoring data, and STL is used to perform decomposition residual fitting to retain trend characteristics; then, the multi-source engineering monitoring data is uniformly converted into JSON-LD format, and the time base deviations of different devices are aligned using the dynamic time warping (DTW) algorithm to ensure the synchronization of cross-source data on the time axis; then, the Min-Max normalization is used to perform global parameter linear mapping on the multi-source engineering monitoring data to eliminate dimensional differences, and finally the preprocessed multi-source engineering monitoring data is generated to provide high-quality input for subsequent knowledge graph construction and dynamic analysis;

[0051] It should be noted that the soft threshold is defined based on the threshold function of the high-frequency wavelet coefficients, and its value range is [0.1σ, 0.5σ].

[0052] S1.3. Utilize an incremental graph update algorithm to stream partition multi-source engineering monitoring data and map the topological structure through an adjacency table index mechanism to obtain graph nodes and edges. Specifically, within the Kafka stream processing framework, the incremental graph update algorithm extracts metadata such as timestamps and sensor IDs from the multi-source engineering monitoring data. It then performs dual partitioning by time window and project section to obtain spatiotemporal correlation data shards, ensuring timeliness and spatial relevance.

[0053] It should be noted that the project section refers to the spatial division unit of the project, which is defined based on the pile number range or three-dimensional coordinate interval of the construction drawings;

[0054] Next, we use an adjacency list indexing mechanism based on a time-series graph database (such as Nebula Graph or Neo4j) and call a dual index structure of hash index and B+ tree index through a graph query language (such as Cypher or nGQL) to map the spatiotemporal correlation data shards into nodes and edges of the knowledge graph: for sensor monitoring data, we convert it into "monitoring point-entity" nodes and establish "affiliation relationship" edges to connect to the corresponding engineering structure nodes; for engineering structure parameter data, we convert it into "structure attribute" nodes and connect them to the monitoring indicator nodes through "parameter association" edges; for construction process management data, we convert it into "construction process" nodes and use "timing constraint" edges to connect to the engineering progress nodes.

[0055] S1.4. Adjust the weights of nodes and edges through the Bayesian confidence dynamic update method to generate a dynamic knowledge graph. In specific operations, the Bayesian confidence update algorithm is used to adjust the weight values in real time according to the time decay factor and sensor accuracy level: the time decay factor dynamically reduces the contribution of historical multi-source engineering monitoring data through an exponential decay function to ensure that attention is paid to recent multi-source engineering monitoring data. For example, the weights of nodes and edges of data in the last hour are retained at 90%, and the weights of nodes and edges of data from 24 hours ago are decayed to 30%; the sensor accuracy level is adjusted according to the accuracy level. For example, the weights of nodes and edges of high-precision sensors are increased by 20%, while the weights of nodes and edges of low-precision sensors are reduced by 15%;

[0056] It should be noted that the time decay factor is defined based on the timeliness requirements of multi-source engineering monitoring data and has a value range of [0.01, 0.1]. The sensor accuracy level is defined based on the factory indicators of the sensor equipment.

[0057] Based on the dynamically adjusted weights of edges and nodes, the nodes and edges are spatiotemporally aggregated and weighted embedded using the weighted averaging method to obtain a graph vector representation with weight constraints. The graph vector representation is incrementally stored and visualized using the real-time update interface of the graph database (such as Nebula Graph's upsert) to generate a dynamic knowledge graph. The dynamic knowledge graph can not only reflect the evolution of the project safety status in real time, but also improve the accuracy of risk identification and the intelligence level of decision support.

[0058] S2. Use the associated subgraph extraction algorithm to traverse the dynamic knowledge graph, extract the security event matrix and the project progress matrix, and use the dynamic weighted fusion algorithm to calculate the SPI index value. Simultaneously, perform multi-threshold interval classification on the SPI index value to generate the SPI warning level.

[0059] The specific steps include:

[0060] S2.1. Use the associated subgraph extraction algorithm to perform semantic diffusion and subgraph extraction on the dynamic knowledge graph, and then perform graph traversal through depth-first search to generate an engineering safety progress network. Specifically, first, use the associated subgraph extraction algorithm to expand the key monitoring nodes in the dynamic knowledge graph outward to form a three-layer neighborhood node.

[0061] Use traversal query statements (such as Cypher) to traverse neighboring nodes and generate a preliminary related subgraph. Specifically, starting from the key monitoring node, perform the first-level traversal to obtain directly connected nodes and edges. Then, using the node traversed from the first level as the new starting point, perform the second-level traversal. Finally, repeat the third-level traversal and use the query result merging function of the graph database (such as Neo4j's UNION ALL) to remove duplicates and structure the results to form a preliminary related subgraph.

[0062] It should be noted that key monitoring nodes include displacement exceeding limit points, stress abnormal points and frequency mutation points, which are extracted from multi-source engineering monitoring data through threshold comparison algorithm;

[0063] Using a preset similarity threshold, the semantic relevance between neighborhood nodes of the preliminary association subgraph is evaluated in real time. For example, the neighborhood nodes are vectorized and quantified using Word2Vec word embedding to generate a semantic similarity score. The similarity threshold is set to 0.7. When the semantic similarity score of a neighborhood node is higher than 0.7, it is judged to have high semantic relevance. When the semantic similarity score of a neighborhood node is lower than 0.7, it is judged to have low semantic relevance.

[0064] The similarity threshold is defined based on the hierarchical distance of neighboring nodes and its value range is [0.6, 0.8];

[0065] At the same time, the strong connection edges of "affiliation" and "parameter association" in the preliminary association subgraph are retained, and weak connection edges with confidence below a strength threshold (such as 0.6) are filtered out. The domain nodes with high semantic relevance and strong connection edges are then eliminated through subgraph pruning to eliminate redundant edges, ultimately generating an optimized association subgraph.

[0066] The intensity threshold is defined based on the false alarm rate associated with different confidence intervals;

[0067] After semantic diffusion is complete, subgraph extraction is performed. First, based on the dependencies of the project structure, the key paths in the associated subgraphs are identified and optimized to obtain the core topology structure. Node attributes are supplemented to the core topology structure using graph database index queries (such as Neo4j's node attribute index). A weighted filtering algorithm is used to remove low-weight edges. A time alignment algorithm is used to calibrate the time window, generating a spatiotemporally aligned project subgraph.

[0068] It should be noted that the dependency relationship of engineering structures refers to the physical connection relationship between components, such as beam-column joints, and the contact between support anchors and surrounding rock, which is called through the API interface of BIM software;

[0069] A depth-first search algorithm is used to traverse the spatiotemporally aligned project subgraphs, establishing a topological sequence along the logical path of "project location → monitoring indicator → construction process → environmental factors." During the traversal, topological sequences within the same section are dynamically merged, and construction phases (such as support and pouring) are marked. The D3.js visualization engine is used to reconstruct and visually map the marked topological sequences into a tree structure, generating a project safety and progress topological network. The root node is the abnormal event node, and branch nodes represent changes in project location distribution, monitoring indicator status, construction process progress, and environmental parameters. This allows engineering personnel to intuitively understand the complete impact scope and progress-related factors of the abnormal event node.

[0070] S2.2. Use graph embedding to map abnormal event associations within the project safety progress topology network to obtain a safety event matrix. The EVM algorithm is then used to quantify resource consumption and obtain a project progress matrix. Specifically, Node2Vec graph embedding is used to convert the project safety progress topology network into a vectorized representation. A biased random walk is then used to capture the topological relationships between "project location, monitoring indicator, and process" in the vectorized representation, generating a multidimensional node embedding vector.

[0071] The abnormal event association mapping is performed on the multi-dimensional node embedding vector. In the specific operation, first, the abnormal event node is extracted from the engineering safety progress topology network, and the abnormal vector is converted through the GNN encoder. The cosine similarity value between the abnormal vector and the multi-dimensional node embedding vector is calculated using the approximate nearest neighbor (ANN) search algorithm. The specific mathematical formula is as follows:

[0072] ;

[0073] in, Represents the cosine similarity value; Unified dimensionality representing anomaly vectors and multi-dimensional node embedding vectors; The index representing the dimension of the abnormal vector and the multi-dimensional node embedding vector; represents the normalized outlier vector Dimensional component; The first node embedding vector representing the multi-dimensional Dimensional component;

[0074] When the cosine similarity value exceeds the preset dynamic threshold, an abnormal event association map is established through the graph attention mechanism, and a security event association matrix is constructed through the sparse matrix compression storage method. In the specific operation, a four-head attention mechanism is enabled, in which each attention head independently performs query-key value matching. Finally, the outputs of multiple attention heads are spliced to generate a node representation containing global topology information. Subsequently, the node representation of the global topology information is weighted and quantified through the sparsification processing engine to construct a sparse association matrix. The non-zero elements are stored in the compressed sparse row (CSR) format. The maximum value is automatically merged for the repeated elements in the sparse association matrix, and finally the security event association matrix is generated. The security event association matrix not only accurately depicts the propagation path between abnormal event nodes, but also quantifies the propagation intensity through weight values, allowing engineers to quickly locate the propagation hub nodes of abnormal events.

[0075] The EVM (Earned Value Management) algorithm is used to quantify resource consumption in the engineering safety progress topology network. First, the EVM algorithm is used to analyze node attributes and extract spatiotemporal correlation features in the engineering safety progress topology network to obtain a process node value dataset. A graph database time series query interface (such as Cypher) is used to extract the planned value, actual cost, and completion percentage of the construction process from the process node value dataset. The planned value of the construction process is multiplied by the completion percentage to obtain the earned value.

[0076] Subtract the earned value from the planned value of the construction process to generate the progress deviation, and subtract the earned value from the actual cost to obtain the cost deviation; use the CP decomposition method to reorganize the progress deviation and cost deviation in the time and space dimensions to obtain the project progress matrix. In the specific operation, first construct a three-dimensional tensor structure for the progress deviation and cost deviation according to the three dimensions of section, process, and time; use the CP decomposition algorithm to iteratively solve the three-dimensional tensor structure using the alternating least squares method, and use L1 regularization to prevent overfitting during the solution process. At the same time, a GPU-accelerated parallel computing framework is used to process large-scale tensor operations to obtain the core factor matrix; the core factor matrix is reconstructed through tensor to generate a low-rank approximation of the project progress matrix; the project progress matrix not only reflects the comprehensive deviation status at different time points, but also provides data support for subsequent end-cloud collaborative project management, enhancing the foresight and scientific nature of decision-making.

[0077] S2.3. The security event matrix and the project progress matrix are weighted and aggregated using a dynamic weighted fusion algorithm to generate the SPI index value. Specifically, the time-varying weight coefficients of the security event matrix and the project progress matrix are calculated based on the entropy weight method. The specific mathematical formula is as follows:

[0078] ;

[0079] in, represents the time-varying weight coefficient, represents the risk warning coefficient, represents the information entropy of the security event matrix, represents the construction phase coefficient, Represents the information entropy of the project progress matrix;

[0080] It should be noted that the risk warning coefficient is defined based on the abnormal node ratio in the security event matrix, and its value range is [0.3, 0.5, 0.7]. The construction phase coefficient is defined based on the critical path process ratio in the construction process, and its value range is [0.4, 0.6].

[0081] Based on the time-varying weight coefficients, the safety event matrix and the project progress matrix are aligned in two dimensions using a tensor contraction operation. An exponentially weighted moving average is performed along the time dimension to smooth short-term fluctuations. Finally, cross-dimensional aggregation is performed through PCA to generate the SPI index value. The Sigmoid function is used to nonlinearly suppress extreme deviation values of the SPI index value to ensure its stability.

[0082] S2.4. SPI index values are graded using multiple threshold intervals to generate SPI warning levels. Specifically, sliding window normalization is used to divide the SPI index values into time windows. The SPI index values within each time window are normalized using Z-scores, and missing data are filled using cubic spline interpolation. The scale is then unified using to generate a standardized SPI sequence.

[0083] The standardized SPI sequence is divided into intervals according to a three-level threshold to generate an SPI warning level. Specifically, the three-level threshold is set based on the statistical distribution of historical standardized SPI sequences. For example, the range of the first threshold is set to [0.8, 1.0]. When the SPI ∈ [0.8, 1.0], it is the green safety zone. At this time, the monitoring dashboard needs to be automatically updated and no warning is required. The range of the second threshold is set to [0.6, 0.8). When the SPI ∈ [0.6, 0.8), it is the yellow warning zone. At this time, an alert notification needs to be sent to the responsible person's mobile terminal to trigger manual review. The range of the third threshold is set to [0, 0.6). When the SPI ∈ [0, 0.6), it is the red danger zone. At this time, the sound and light alarm device is automatically activated and the emergency plan is initiated. The interval classification process uses a hysteresis mechanism to prevent frequent threshold jumps, ensuring the stability of the response and the reliability of the warning results in the end-cloud collaborative environment.

[0084] S3. Encode the SPI warning level to generate a comprehensive feature vector, which is then input into the Bayesian causal graph model. The forward propagation layer is used to extract the node probability distribution characteristics, and the counterfactual intervention layer is used to extract the blocking impact characteristics. The entropy weight method is used to simultaneously perform weight distribution and generate an abnormal contribution matrix.

[0085] The specific steps include:

[0086] S3.1. Encode the SPI warning level and generate a comprehensive feature vector. In the specific operation, first, align the time series of the SPI warning level based on the hierarchical clustering algorithm to obtain the similarity distance matrix; use the DTW constraint window to dynamically fit the similarity distance matrix to time warp, and perform layer-by-layer aggregation through Ward minimum variance to construct a hierarchical clustering tree; search for the best cutting point of the hierarchical clustering tree through the silhouette coefficient to obtain the optimal clustering value; stop merging when the inter-class distance of the hierarchical clustering tree exceeds the optimal clustering value, and form a warning status cluster;

[0087] It should be noted that the silhouette coefficient is defined based on the ratio of the intra-class compactness to the inter-class separation of the hierarchical clustering tree, and its value range is [-1, 1];

[0088] Truncated singular value decomposition (Truncated SVD) is used to extract the first three principal components of the SPI warning level in each warning status cluster to obtain the principal component characteristics. The feature importance is weighted through principal component load analysis to obtain the pattern feature vector. Based on the pattern feature vector, the state is encoded through a One-Hot encoder to generate a discrete state identifier. The weighted average method is used to perform cross-dimensional fusion of the principal component characteristics and discrete state identifiers to generate a comprehensive feature vector with unified dimensions.

[0089] S3.2. Construct and train a Bayesian causal graph model. Specifically, in the PyTorch framework, call the forward propagation layer through the nn.Module parameter and initialize it through the xavier_uniform function: set the hidden layer dimension to 128 and the activation function to GeLU; simultaneously call the counterfactual intervention layer using the Do-calculus parameter and initialize it with a zero-mean Gaussian, setting the intervention intensity to 0.5 and the noise scale to 0.1; the automatic differentiation mechanism couples the initialized forward propagation layer and the counterfactual intervention layer with bidirectional gradients, and parameterizes them through the Bayesian parameterization stacker. Simultaneously, use the Hamiltonian Monte Carlo (HMC) algorithm to achieve efficient exploration of the parameter space and construct a Bayesian causal graph model.

[0090] Next, the Bayesian causal graph model is trained. Specifically, the comprehensive feature vector is first divided into a sample set, a training set, and a validation set. On the sample set, the comprehensive feature vector is mapped into a latent space using a variational autoencoder to obtain a low-dimensional embedding representation. Markov chain Monte Carlo sampling is used to estimate the posterior distribution parameters of the low-dimensional embedding representation to generate the initial Bayesian causal graph model parameters. On the training set, the initial Bayesian causal graph model parameters are iterated using the Adam optimizer, and backpropagation gradient updates are performed using the evidence lower bound (ELBO) loss function to obtain the optimized Bayesian causal graph model parameters. On the validation set, when the optimized Bayesian causal graph model parameters reach the convergence threshold for ten consecutive epochs, training is terminated, and the trained Bayesian causal graph model is output synchronously using the TorchScript script.

[0091] It should be noted that the convergence threshold is defined based on the relative change rate of the ELBO loss value of the validation set, and the value range is [0.05%, 0.5%];

[0092] The trained Bayesian causal graph model can not only achieve efficient causal analysis and prediction of complex engineering events through end-cloud collaboration, but also improve the stability and generalization ability of the Bayesian causal graph model in practical applications, ensuring accurate risk assessment and decision support in dynamic environments.

[0093] S3.3. Generate anomaly contribution matrix using Bayesian causal graph model. Specifically, the comprehensive feature vector is first input into the forward propagation layer of the Bayesian causal graph model. One-dimensional dilated causal convolution is used to process the temporal dependency of the comprehensive feature vector. Zero padding is used to keep the sequence length unchanged. Gated activation units (GLUs) are synchronously superimposed to control the information flow and generate an intermediate feature tensor. Spatial correlation weighting is used to perform spatial density aggregation on the intermediate feature tensor, and a spatiotemporal fusion feature matrix is output through residual connection. A variational inference layer is used to map the spatiotemporal fusion feature matrix to the mean and standard deviation of the Gaussian distribution. Reparameterized sampling is used to perform differentiable random sampling on the mean and standard deviation of the Gaussian distribution to obtain node probability distribution characteristics.

[0094] The counterfactual intervention layer uses Gumbel-max reparameterization to block causal paths and obtain blocking impact features. Specifically, the comprehensive feature vector is firstly sampled differently based on the Gumbel-max reparameterization technique, and causal contribution ranking and path significance screening are performed through Top-k sparsification to obtain the causal path. The causal path is then directionally blocked using a mask matrix, and the intervention effect of the blocked causal path is calculated through a two-layer graph convolution. Skip connections are simultaneously used to preserve the original information flow and generate counterfactual feature representations. Finally, a feature distribution calibrator is used to perform distribution alignment optimization on the counterfactual feature representation to output the blocking impact features.

[0095] The node probability distribution features and blocking influence features are weighted by the entropy weight method, and the attention channel is used to splice them to generate the abnormal contribution matrix. In the specific operation, the information entropy of the node probability distribution features and the blocking influence features are first calculated respectively, and the feature weight value is obtained according to the inverse of the entropy value and normalization; according to the feature weight value, the node probability distribution features and the blocking influence features are projected into the query, key, and value space through the multi-head attention channel, and the feature strength interaction is performed by scaling the dot product attention to obtain the original attention score matrix; the original attention score matrix is fused with multi-head attention using Softmax normalization to generate the abnormal contribution matrix;

[0096] The anomaly contribution matrix can not only achieve refined quantitative analysis of the impact of various factors in engineering safety incidents through end-cloud collaboration, but also improve anomaly attribution efficiency and decision-making support capabilities in complex multi-source data environments.

[0097] S4. Perform multi-level attribution aggregation on the abnormal contribution matrix through a hierarchical clustering algorithm to generate root cause analysis data. Based on the root cause analysis data, incrementally update the node weights and edge strengths of the dynamic knowledge graph to generate an engineering safety progress monitoring report.

[0098] The specific steps include:

[0099] S4.1. Multi-level attribution aggregation is performed on the anomaly contribution matrix through a hierarchical clustering algorithm to generate root cause analysis data. In the specific operation, the anomaly contribution matrix is first divided into multiple samples, and the distance between samples is calculated using Ward minimum variance. Based on the distance between samples, full-connection clustering is performed through hierarchical clustering to obtain multi-level anomaly cluster nodes. The multi-level anomaly cluster nodes are traversed hierarchically in a bidirectional manner to construct a clustering tree with multi-resolution characteristics. Based on the clustering tree, multi-level attribution aggregation is performed using the causal propagation inference method: at each cluster level of the clustering tree, the top k largest causal flow paths are selected based on Gumbel-Softmax reparameterization, and the causal flow paths are gradient reversed through the do-calculus operator to obtain hierarchical reasoning results. Bayesian average fusion is used to aggregate the multi-level reasoning results to finally generate root cause analysis data.

[0100] It should be noted that the do-calculus operator is defined based on the intensity of the intervention effect and its value range is [0,1].

[0101] S4.2. Based on the root cause analysis data, the node weights and edge strengths of the dynamic knowledge graph are incrementally updated to generate a causal link vector. In the specific operation, first, the node connection number statistics method is used to assign node weights to the dynamic knowledge graph to obtain the node weights, and the co-occurrence frequency calculation is simultaneously applied to quantify the edge strength to obtain the edge strength; in the Flink engine, the root cause analysis data is stream-aligned and timestamp-corrected through the event time window mechanism and input into the dynamic knowledge graph; next, the dynamic decay algorithm is used to perform temporal decay and incremental updates on the node weights to obtain the node state vector; Gaussian kernel density estimation is used to extract local features of the node state vector, and probability density smoothing is performed through a sliding window to form a spatiotemporal feature distribution matrix; the attention weighting method is used to recalibrate the spatiotemporal feature distribution matrix to generate a weighted feature tensor, and based on the weighted feature tensor, gradient backpropagation is used to update the contribution to obtain an enhanced node embedding representation;

[0102] The incremental Dijkstra algorithm is used to search for the shortest causal path on edge strength to obtain the original causal link set, and the original causal link set is filtered for path reliability to obtain a high-confidence causal subgraph. Spectral clustering is used to divide the high-confidence causal subgraph into communities to generate a multi-cluster causal structure. Based on the multi-cluster causal structure, the gradient backpropagation optimizer is used to perform extended path optimization to obtain the optimized causal path set.

[0103] The enhanced node embedding representation and the optimized causal path set are cross-modally fused to generate a causal link vector.

[0104] S4.3. Generate a project safety progress monitoring report based on the causal link vector. In the specific operation, first use the transfer matrix decomposition method to perform tensor conversion on the causal link vector, construct a three-dimensional transfer matrix, and extract the latent factors of the three-dimensional transfer matrix through Tucker decomposition; use the Sigmoid function to perform nonlinear probability conversion on the latent factors to obtain safety assessment parameters; use Min-Max normalization to calibrate the dynamic range of the latent factors to generate real-time progress parameters; then use the dynamic weighted fusion algorithm to integrate the parameters: linearly weight the safety assessment parameters and real-time progress parameters; at the same time, use the thermal-polyline hybrid rendering engine for visual rendering to generate a project safety progress monitoring report containing a four-dimensional risk heat map and a progress deviation trend map.

[0105] This embodiment also provides a computer device, which is suitable for the case of an intelligent monitoring method for engineering safety progress based on multi-source data collaboration, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the intelligent monitoring method for engineering safety progress based on multi-source data collaboration as proposed in the above embodiment.

[0106] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0107] This embodiment also provides a storage medium having a computer program stored thereon. When the program is executed by a processor, the program implements the method for intelligent monitoring of engineering safety progress based on multi-source data collaboration proposed in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0108] In summary, this invention effectively integrates different types of data by mapping multi-source engineering monitoring data into a dynamic knowledge graph in real time, ensuring its timeliness and accuracy. Using a correlation subgraph extraction algorithm and a Bayesian causal graph model, it identifies the specific causes of problems and their impact paths, deeply exploring potential causal relationships and providing a basis for developing targeted solutions.

[0109] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An intelligent monitoring method for engineering safety progress based on multi-source data collaboration, characterized by: include, An incremental graph update algorithm is used to map multi-source engineering monitoring data into graph nodes and edges in real time, generating a dynamic knowledge graph. The dynamic knowledge graph is traversed through the associated subgraph extraction algorithm to extract the security event matrix and the project progress matrix. The SPI index value is calculated using a dynamic weighted fusion algorithm. The SPI index value is simultaneously graded by multiple threshold intervals to generate the SPI warning level. The SPI warning level is state-encoded to generate a comprehensive feature vector, which is then input into the Bayesian causal graph model. The forward propagation layer is used to extract the node probability distribution characteristics, and the counterfactual intervention layer is used to extract the blocking impact characteristics. The entropy weight method is then used to simultaneously perform weight distribution and generate an abnormal contribution matrix. The hierarchical clustering algorithm is used to perform multi-level attribution aggregation on the anomaly contribution matrix to generate root cause analysis data. Based on the root cause analysis data, the node weights and edge strengths of the dynamic knowledge graph are incrementally updated to generate an engineering safety progress monitoring report.

2. The method for intelligent monitoring of engineering safety progress based on multi-source data collaboration according to claim 1, characterized in that: The multi-source engineering monitoring data includes sensor monitoring data, engineering structure parameter data and construction process management data.

3. The method for intelligent monitoring of engineering safety progress based on multi-source data collaboration according to claim 1, characterized in that: The generation of a dynamic knowledge graph specifically includes the following steps: An incremental graph update algorithm is used to stream partition multi-source engineering monitoring data. Based on a time-series graph database, the topological structure is mapped through an adjacency table index mechanism to obtain the nodes and edges of the graph. The weights of nodes and edges are adjusted through the Bayesian confidence dynamic update method to generate a dynamic knowledge graph.

4. The method for intelligent monitoring of engineering safety progress based on multi-source data collaboration according to claim 1, characterized in that: The method of calculating the SPI index value using a dynamic weighted fusion algorithm specifically includes the following steps: The dynamic knowledge graph is subjected to semantic diffusion and subgraph extraction through the associated subgraph extraction algorithm, and the graph is traversed through depth-first search to generate an engineering safety progress topology network. Graph embedding is used to map abnormal events to the project safety progress topology network to obtain a safety event matrix. The EVM algorithm is then used to quantify resource consumption and obtain a project progress matrix. The safety event matrix and the project progress matrix are weighted and aggregated through a dynamic weighted fusion algorithm to generate the SPI index value.

5. The method for intelligent monitoring of engineering safety progress based on multi-source data collaboration according to claim 4 is characterized in that: The generation of the SPI warning level specifically includes the following steps: Sliding window normalization is used to smooth and unify the SPI index values to obtain a standardized SPI sequence; The standardized SPI sequence is divided into intervals according to the three-level threshold and the SPI warning level is generated.

6. The method for intelligent monitoring of engineering safety progress based on multi-source data collaboration according to claim 1, characterized in that: The generation of the abnormal contribution matrix specifically includes the following steps: A hierarchical clustering algorithm is used to cluster the SPI warning levels based on similarity, and singular value decomposition is used to encode the state and generate a comprehensive feature vector. The forward propagation layer and the counterfactual intervention layer are parameterized and stacked through the automatic differentiation mechanism to construct a Bayesian causal graph model; The comprehensive feature vector is input into the Bayesian causal graph model, and the forward propagation layer uses causal convolution to aggregate the time dimension and spatial density to obtain the node probability distribution characteristics; The counterfactual intervention layer blocks the causal path through Gumbel-max reparameterization to obtain the blocking impact characteristics; The node probability distribution features and blocking impact features are weighted by the entropy weight method, and the attention channel is used to splice them to generate the abnormal contribution matrix.

7. The method for intelligent monitoring of engineering safety progress based on multi-source data collaboration according to claim 6, characterized in that: Generating root cause analysis data specifically includes the following steps: The hierarchical clustering algorithm is used to perform full-connection clustering on the anomaly contribution matrix to obtain multi-level anomaly cluster nodes. The causal reasoning method is then used to perform multi-level attribution aggregation on the multi-level anomaly cluster nodes to generate root cause analysis data.

8. The method for intelligent monitoring of engineering safety progress based on multi-source data collaboration according to claim 1, characterized in that: The generation of the project safety progress monitoring report specifically includes the following steps: The Flink engine is used to input root cause analysis data into a dynamic knowledge graph. The node weights are updated using a dynamic decay algorithm. The Dijkstra algorithm is used to optimize the edge strength path and obtain the causal link vector. The transfer matrix decomposition method is used to convert the causal link vector into structured safety assessment parameters and real-time progress parameters, which are then integrated into a project safety progress monitoring report.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the engineering safety progress intelligent monitoring method based on multi-source data collaboration described in any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the engineering safety progress intelligent monitoring method based on multi-source data collaboration described in any one of claims 1 to 8 are implemented.

Images