Contract graph-based project whole life cycle oa process automatic management method and system

Abstract

This application provides a method and system for automated management of project lifecycle OA processes based on contract graphs, relating to the technical field of project management. The method includes: collecting real-time point cloud sequences of the project performance area and extracting baseline data from the contract graph; then using a graph neural network to parse the contract graph to obtain first and second constraint rules; employing a point cloud segmentation network and segmenting the real-time point cloud sequence according to the first constraint rule to obtain entity point clouds, and registering them with the baseline data according to a search strategy to generate trajectory deviation data; next, constructing a hidden Markov model based on the second constraint rule to decode the trajectory deviation data and determine its state; and generating a task completion list when the trajectory deviation data meets the first acceptance condition and the state meets the second acceptance condition, otherwise generating task exception information; finally, generating the corresponding management process. This application achieves automated compliance verification and process triggering throughout the project lifecycle.

Description

Technical Field

[0001] This application relates to the technical field of project management, and in particular to a method and system for automated management of the entire project lifecycle OA process based on contract graphs. Background Technology

[0002] In the field of engineering project management, transforming contract terms, technical specifications, and performance requirements into executable control rules, and combining them with information systems to automate processes, has become an important means of improving project execution efficiency and reducing compliance risks. With the popularization of building information modeling and on-site sensing technologies, how to dynamically align the construction progress in the physical world with the abstract rules agreed in the contract is gradually becoming a core focus of the industry.

[0003] Currently, mainstream project lifecycle management systems mainly rely on documented process engines and preset node triggering mechanisms. By entering key contract time nodes into the system, the approval processes such as payment requests and acceptance are automatically initiated. Some systems have introduced basic structured data validation functions, which can perform format checks or threshold comparisons on the completed quantities reported on-site, thereby assisting managers in making decisions.

[0004] However, existing methods often treat contract rules as static checklists when analyzing project performance status, making it difficult to effectively utilize continuous dynamic data collected at the construction site. Furthermore, due to a lack of in-depth analysis capabilities regarding the complex relationships between spatial constraints, technical standards, and construction procedures in contract terms, the system is unable to accurately compare point cloud and other perception data with contract requirements. It is also difficult to assess the impact of deviations on the overall process in real time during the accumulation of deviations, resulting in a lag in the determination of performance status. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for automated management of the entire project lifecycle OA process based on contract graphs, so as to solve the problem in the prior art that it is difficult to independently verify the dynamic consistency between the physical performance status of the project and the abstract rules of the contract.

[0006] To address the aforementioned technical problems, firstly, this application provides a method for automated management of project lifecycle OA processes based on contract graphs, including: Collect real-time point cloud sequences of the project performance area and extract benchmark data corresponding to project performance from a pre-built contract map; The contract graph is analyzed using a graph neural network to obtain the first constraint rule and the second constraint rule; A point cloud segmentation network is used to segment the real-time point cloud sequence according to the first constraint rule to obtain the entity point cloud to be verified. The point cloud of the entity to be verified is iteratively registered with the reference data according to the search strategy defined in the first constraint rule to generate trajectory deviation data; Based on the second constraint rule, a hidden state space of the Hidden Markov Model is constructed, and the trajectory deviation data is synchronously decoded in the hidden state space to determine the corresponding state. When the trajectory deviation data of any generation meets the first acceptance condition defined in the first constraint rule and the state of the corresponding generation simultaneously meets the second acceptance condition defined in the second constraint rule, a task completion list is generated. If the trajectory deviation data continues to fail to meet the first acceptance condition, or the state of the corresponding generation fails to meet the second acceptance condition, task abnormality information is generated. Input the task completion list or the task exception information into the OA system to generate management processes corresponding to task completion and task exception respectively.

[0007] Optionally, the step of using a graph neural network to parse the contract graph to obtain the first constraint rule and the second constraint rule includes: Nodes and edges related to the current performance project are extracted from the contract graph to obtain a target subgraph, wherein the target subgraph includes a central node representing the entity to be verified, attribute nodes representing spatial location or technical standards, and event nodes representing the sequence of performance actions. Construct a two-channel graph neural network and input the target subgraph into the two-channel graph neural network; Through the first channel of the dual-channel graph neural network, based on the path distance between the attribute node and the center node in the shown contract graph, an attention convergence operation based on distance decay is performed to obtain the constraint feature vector; Through the second channel of the dual-channel graph neural network, based on the predefined causal sequence relationship between the event nodes, the weights of edges that conform to the causal order are enhanced, while edge connections that do not conform to the causal order are suppressed, to obtain the sequence feature vector; A rule generator decodes the constraint feature vector into a first constraint rule and the sequence feature vector into a second constraint rule.

[0008] Optionally, the step of decoding the constraint feature vector into a first constraint rule and the sequence feature vector into a second constraint rule using a rule generator includes: The constraint feature vector and the sequence feature vector are input into a rule generator, which consists of a feature fusion module based on an attention mechanism and a multi-output head. The feature fusion module concatenates the constraint feature vector with the sequence feature vector to form a joint feature vector, and performs self-attention calculation on the joint feature vector to obtain a fused representation vector; The fused representation vector is simultaneously input into two parallel sub-modules of the multi-output head; The first submodule maps the fused representation vector to an original parameter vector through a fully connected layer, and uses the parameter boundary obtained in real time from the contract graph to numerically constrain the original parameter vector through a boundary clipping function to obtain numerical parameters. Extract rule elements related to the current performance project from the contract graph, wherein the rule elements shall at least include spatial range, key geometric features and search strategy, and assemble the numerical parameters with the rule elements to obtain the first constraint rule; The second submodule inputs the fused representation vector into a sequence-to-sequence encoder-decoder structure to generate a state transition probability graph. Using the state transition constraints obtained in real time from the contract graph, the module performs a graph editing operation to adjust the structure of the state transition probability graph and obtain the second constraint rule.

[0009] Secondly, this application provides a project lifecycle OA process automation management system based on contract graphs, including: The data acquisition module is used to collect real-time point cloud sequences of the project performance area and extract the benchmark data corresponding to the project performance from the pre-built contract map. The parsing module is used to parse the contract graph using a graph neural network to obtain the first constraint rule and the second constraint rule; The segmentation module is used to segment the real-time point cloud sequence using a point cloud segmentation network according to the first constraint rule to obtain the entity point cloud to be verified. The registration module is used to iteratively register the point cloud of the entity to be verified with the reference data according to the search strategy defined in the first constraint rule, and generate trajectory deviation data. The decoding module is used to construct the hidden state space of the Hidden Markov Model based on the second constraint rule, and synchronously decode the trajectory deviation data in the hidden state space to determine the corresponding state. The first generation module is used to generate a task completion list when the trajectory deviation data of any generation meets the first acceptance condition defined in the first constraint rule and the state of the corresponding generation simultaneously meets the second acceptance condition defined in the second constraint rule. If the trajectory deviation data continues to fail to meet the first acceptance condition, or the state of the corresponding generation fails to meet the second acceptance condition, task abnormal information is generated. The second generation module is used to input the task completion list or the task exception information into the OA system and generate management processes corresponding to task completion and task exception respectively.

[0010] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor is used to execute the computer program to implement the steps of the project lifecycle OA process automation management method based on contract graphs as described in the first aspect above.

[0011] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the project lifecycle OA process automation management method based on contract graphs as described in the first aspect above.

[0012] The project lifecycle OA process automation management method based on contract graph provided in this application has the following beneficial effects: First, this application collects real-time point cloud sequences of the project performance area and extracts benchmark data from the contract graph, thus building a foundation for the association between physical space and contract space for automated verification; then, it uses graph neural networks to perform deep analysis of the contract graph, decoupling complex contract terms into spatial constraint rules and process logic rules, which can realize the structured expression of abstract contract semantics; subsequently, it uses a point cloud segmentation network to accurately extract entities to be verified from the field data according to spatial constraint rules, and iteratively registers them with benchmark data to generate trajectory deviation data, and then quantifies the construction quality of the entities into calculable dynamic offset information; Simultaneously, a hidden Markov model is constructed based on the process logic rules to perform time-series decoding of deviation data, thereby realizing the status determination of the current performance stage; finally, through the joint verification of the dual conditions of spatial constraints and process logic, the task completion list or abnormal information is input into the OA system to automatically trigger the corresponding management process, which can realize the full-chain automated closed loop from on-site perception, status determination to process driving, and thus realize the autonomous verification of the dynamic consistency between the physical performance status of the project and the abstract rules of the contract.

[0013] Furthermore, this application extracts a target subgraph containing a central node, attribute nodes, and event nodes from the contract graph, and employs a dual-channel graph neural network to perform distance-based attention convergence and causal-based edge weight enhancement, respectively. This enables the separate extraction and collaborative learning of spatial constraint features and process sequence features, thereby effectively separating and strengthening two types of core rules from the complex contract association structure. This provides semantically clear and logically complete rule inputs for subsequent spatial registration and state decoding, and improves the accuracy of understanding contract terms and the robustness of rule generation. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A flowchart illustrating an automated management method for the entire project lifecycle OA process based on a contract graph, provided in this application embodiment; Figure 2 This application provides a schematic diagram illustrating a specific implementation of an automated management method for the entire project lifecycle OA process based on contract graphs. Figure 3 A schematic diagram of the structure of a project lifecycle OA process automation management system based on contract graphs is provided in this application embodiment; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0016] In existing project lifecycle management methods, contract terms and technical specifications are usually transformed into static checklists or pre-set approval nodes. This results in a disconnect between rules and dynamic on-site data. On the one hand, the system lacks the ability to deeply analyze the complex relationships between spatial constraints, technical standards, and process logic in contract terms, making it difficult to accurately compare point cloud and other perception data collected on the construction site with abstract rules. On the other hand, the determination of performance status relies heavily on the results of discrete nodes, making it impossible to continuously track and identify the progress of processes based on real-time trajectory deviations. This results in a lag in verification results and makes it difficult to trigger effective control processes in the early stages of deviation accumulation.

[0017] To address this, this application proposes an automated management method for the entire project lifecycle OA process based on contract graphs. The core of this method lies in: firstly, using a graph neural network to deeply analyze the contract graph, decoupling complex clauses into spatial constraint rules and process logic rules; then, employing a point cloud segmentation network to extract entities to be verified from on-site data based on spatial constraints, and generating quantifiable trajectory deviation data through iterative registration; simultaneously, constructing a Hidden Markov Model based on process logic rules to perform temporal decoding of the deviation data, accurately determining the current performance stage; and finally, through dual-condition joint verification, automatically inputting task completion or anomaly information into the OA system to trigger the corresponding management process. This method achieves an automated closed loop from on-site perception and dynamic verification to process-driven processes, effectively overcoming the problems of rule-data fragmentation, delayed status determination, and disconnection from control in existing technologies.

[0018] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] The core of this application is to provide an automated management method for the entire project lifecycle OA process based on contract graphs. A flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes: S101. Collect the real-time point cloud sequence of the project performance area and extract the benchmark data corresponding to the project performance from the pre-built contract map.

[0020] Among them, the contract graph refers to a knowledge base that organizes contract terms and related information in a graph structure. Nodes represent entity elements in the contract, such as key equipment models, material specifications, installation technical specifications, and acceptance standards, while edges represent the relationships between elements, such as the binding relationship between equipment and installation location, and the sequence of procedures. Benchmark data refers to standard information extracted from contract drawings for subsequent comparison and verification, including reference values ​​such as the design three-dimensional dimensions, installation coordinate positions, allowable deviation ranges, and acceptance index thresholds of key equipment; real-time point cloud sequence refers to a high-density set of three-dimensional coordinate points that change continuously over time, collected by means of laser scanning or photogrammetry at the arrival site, installation work surface, or acceptance area of ​​key equipment, used to reflect the actual spatial state and installation accuracy of equipment and materials.

[0021] In step S101, firstly, a handheld laser scanner or a stand-alone 3D laser scanning device is used to continuously collect data on the storage area, installation work surface, or acceptance part of the key equipment to obtain a high-density real-time point cloud sequence covering the site. At the same time, benchmark data related to the current batch of equipment and materials is queried and extracted from the pre-built contract map to provide accurate reference for the on-site verification, installation process monitoring, and subsequent acceptance comparison of the batch of materials.

[0022] S102. Using a graph neural network, the contract graph is analyzed to obtain the first constraint rule and the second constraint rule.

[0023] In one specific implementation, such as Figure 2 As shown, step S102 includes: Step 1021: Extract the nodes and edges related to the current performance project from the contract graph to obtain the target subgraph, wherein the target subgraph includes a central node representing the entity to be verified, attribute nodes representing spatial location or technical standards, and event nodes representing the sequence of performance actions.

[0024] The central node refers to the node in the subgraph that represents the entity to be verified, such as key equipment like the "main transformer" or "steel structure beam". Attribute nodes are nodes that characterize spatial location, such as installation coordinates, or technical standards, such as dimensional tolerances and material requirements, and are used to describe the constraints of the central node. Event nodes are nodes that represent a sequence of performance actions, such as process nodes like "equipment arrival", "hoisting and positioning", and "precision acceptance". The edges between nodes represent the order of the processes.

[0025] In step 1021, based on the identifier of the current performance project, such as the equipment batch number or component number, all nodes and edges related to the equipment are retrieved and extracted from the pre-built contract graph to form a target subgraph. This subgraph contains three types of nodes: the central node represents the key equipment to be verified, the attribute node represents the technical parameters of the equipment such as installation coordinates, design dimensions, and allowable deviations, and the event node represents each process of the equipment from arrival, installation to acceptance. The edges between nodes are used to represent the relationship, including the "having attribute" edge between the attribute node and the central node, and the "sequence" edge between the event nodes.

[0026] Step 1022: Construct a two-channel graph neural network and input the target subgraph into the two-channel graph neural network.

[0027] The key parameters in the network include: the number of graph convolutional layers, which is usually set to 2 to 3 layers to cover the path distances between the center node, attribute nodes, and event nodes; the attention decay coefficient β, which controls the degree of influence of path distance on attention weights. The larger β is, the faster the weights of nodes farther away decay. This parameter can be set as a learnable parameter and optimized along with the network, or it can be preset based on experience, for example, β=0.5; the feature vector dimension, which is the dimension of the constraint feature vector, sequence feature vector, and fused feature vector, such as 128-dimensional or 256-dimensional, which is determined according to the model complexity and the amount of data. This application does not make a specific limitation; and the number of self-attention heads, such as 4 or 8.

[0028] Each training sample of a dual-channel graph neural network consists of two parts: the input is the target subgraph; the output is the label data corresponding to the target subgraph, that is, the true values ​​of the first and second constraint rules to be generated. This label data can be labeled by experts based on historical project data, such as labeling the correct spatial search range, deviation threshold, and correct process state transition path.

[0029] Next, loss functions are designed for the two output channels separately, and then the first channel loss and the second channel loss are jointly optimized. The first channel loss is the error between the numerical parameters in the first constraint rule and the actual numerical parameters, typically using mean squared error loss. For example, if the rule includes a search radius r and a deviation threshold, the loss is... ,in, This represents the loss value of the first channel in a two-channel graph neural network. This represents the model prediction value of the point cloud search radius in the first constraint rule. This represents the expert-annotated or contractually agreed-upon true value of the point cloud search radius in the first constraint rule. This represents the model prediction value of the registration deviation threshold in the first constraint rule. This represents the expert-annotated or contractually agreed-upon true value of the registration deviation threshold in the first constraint rule; The second channel loss is the error between the state transition probability map in the second constraint rule and the actual state transition path. It is usually the cross-entropy loss, which compares the predicted probability with the actual path for each state transition pair i to j. The total loss function is the weighted sum of the two channel losses.

[0030] The training process is as follows: the backpropagation algorithm and stochastic gradient descent optimizer are used to iteratively update the network parameters. During training, the target subgraph is fed into the dual-channel network, and the predicted constraint rules and state transition graph are obtained through forward propagation. The total loss is then calculated, and the weight parameters of each layer in the network are updated through backpropagation. The training continues for multiple rounds until the loss converges or the preset number of iterations is reached. Finally, a network model that can accurately parse the two types of rules from the contract subgraph is obtained.

[0031] In step 1022, a dual-channel graph neural network with two parallel graph convolution processing branches is constructed. This network is pre-trained to extract accurate constraint features and sequence features as the target for optimization. After the target subgraph is input into the network, it is simultaneously fed into the two channels for parallel processing, preparing for subsequent feature extraction.

[0032] Step 1023: Using the first channel of the dual-channel graph neural network, perform a distance-attention convergence operation based on the path distance between the attribute node and the center node in the shown contract graph to obtain the constraint feature vector.

[0033] Among them, the distance decay-based attention convergence operation refers to a mechanism that allocates attention weights according to the path distance between nodes. The closer the node is, the greater its contribution, and the farther the node is, the greater its contribution decays, thereby focusing on the attribute information closely related to the central node.

[0034] In step 1023, within the first channel, for the central node, the path distance between it and each attribute node is first calculated, where the path distance is measured by the shortest path hop count. Then, based on these distance values, a distance-decay-based attention convergence operation is performed. That is, for each attribute node, its attention weight is proportional to the exponential decay function of the path distance; the closer the distance, the greater the weight. Assuming the path distance between the central node c and attribute node i is... Then attention weights are passed through Calculate, where, Let A be the attention decay coefficient, and A be the set of all attribute nodes. The attention weights are for each attribute node. This represents the path distance between the center node c of the target subgraph of the contract graph and the i-th attribute node. The path distance between the center node c of the target subgraph of the contract graph and the j-th attribute node is represented; then, the features of all attribute nodes are weighted and summed according to the calculated weights to obtain the constraint feature vector.

[0035] Step 1024: Through the second channel of the dual-channel graph neural network, based on the predefined causal sequence relationship between the event nodes, enhance the weight of edges that conform to the causal order and suppress edge connections that do not conform to the causal order to obtain the sequence feature vector.

[0036] Among them, the causal sequence is a predefined logic that must be followed between processes. For example, "equipment arrival" must precede "hoisting and positioning". This relationship comes from contract terms or construction specifications. The conformity of the causal sequence is a fact, which means whether the actual connection between the event nodes being processed conforms to the above rules. If there is an edge in the diagram that points from "equipment arrival" to "hoisting and positioning", then this edge conforms to the causal sequence; conversely, if there is an edge that points from "hoisting and positioning" to "equipment arrival", then it does not conform to the causal sequence.

[0037] In step 1024, in the second channel, the event nodes are processed according to the predefined causal sequence relationship, which refers to the inherent logical order between processes. Based on these relationships, the weights of the edges between event nodes in the target subgraph are adjusted: for edges that conform to the causal order, their weights are multiplied by an enhancement factor greater than 1; for edges that do not conform to the order, their weights are multiplied by an inhibition factor less than 1 or set to zero. Then, graph convolution operation is performed on the adjusted target subgraph to extract the sequence feature vector that reflects the logical dependency relationship between processes.

[0038] Step 1025: Using a rule generator, decode the constraint feature vector into a first constraint rule and decode the sequence feature vector into a second constraint rule.

[0039] Step 1025 may specifically include the following steps: Step a1: Input the constraint feature vector and the sequence feature vector into a rule generator, which consists of a feature fusion module based on an attention mechanism and a multi-output head.

[0040] In step a1, the constraint feature vector and the sequence feature vector are input into the rule generator. The generator contains two core components: one is a feature fusion module based on the attention mechanism, which is used to fully interact and fuse the two types of features; the other is a multi-output head, which contains two parallel sub-modules for generating the first constraint rule and the second constraint rule, respectively.

[0041] Step a2: The feature fusion module concatenates the constraint feature vector with the sequence feature vector to form a joint feature vector, and performs self-attention calculation on the joint feature vector to obtain the fused representation vector.

[0042] In step a2, the feature fusion module first combines the constrained feature vectors With sequence feature vector Perform a concatenation operation to form a joint feature vector containing information from both types of data. Then, self-attention calculation is performed on the joint feature vector. Specifically, the query matrix Q, key matrix K, and value matrix V are obtained through linear transformation. Then, attention weights are calculated and the values ​​are weighted and summed to obtain the fused representation vector. The self-attention calculation formula is as follows: ,in, Let be the dimension of the key vector, and be a scaling factor. This represents the transpose of the key matrix K. Represents the normalized exponential function, This represents the output of the self-attention calculation. This step enables the spatial constraint features and process sequence features to fully interact, thereby achieving deep information fusion.

[0043] Step a3: Input the fused representation vector into the two parallel sub-modules of the multi-output head simultaneously.

[0044] In step a3, the fused representation vector is simultaneously sent to two parallel sub-modules of the multi-output head. The first sub-module is responsible for generating rules related to spatial constraints, and the second sub-module is responsible for generating rules related to process logic. The two sub-modules work in parallel based on the same fused feature and generate different types of rules respectively. For example, the first sub-module includes a fully connected layer consisting of a 2-layer fully connected network and a ReLU activation function, and the mapping dimension can be set to 128 or 256 dimensions. The second sub-module includes an encoding layer of a sequence-to-sequence encoder-decoder structure, and its dimension can be set to 128 or 256 dimensions.

[0045] Step a4: The first submodule maps the fused representation vector into an original parameter vector through a fully connected layer, and uses the parameter boundary obtained in real time from the contract graph to numerically constrain the original parameter vector through a boundary clipping function to obtain numerical parameters.

[0046] Among them, the parameter boundary refers to the allowable range of each numerical parameter in the first constraint rule obtained from the contract graph, such as the minimum and maximum values ​​of the search radius and the minimum and maximum values ​​of the deviation threshold. The main function of the parameter boundary is to ensure that the numerical parameters generated from the neural network conform to the specifications of the contract terms and play a role in compliance constraints and security guarantees. A boundary clipping function is a function that restricts the original parameter vector within the parameter boundaries to ensure that the generated numerical parameters conform to the contract specifications. This embodiment does not limit the expression of this function.

[0047] In step a4, the first submodule first inputs the fused representation vector into the fully connected layer and maps it to an original parameter vector through a linear transformation using the ReLU function. This vector contains the values ​​required by the first constraint rule, such as the radius range of the point cloud search and the registration deviation threshold. At the same time, the parameter boundaries corresponding to these parameters are obtained in real time from the contract graph. For example, the search radius ranges from [Lr, Ur] and the deviation threshold ranges from [Lt, Ut]. Then, the original parameter vector praw is numerically constrained using a boundary clipping function, that is, values ​​exceeding the boundary are forcibly restricted to within the boundary, resulting in the final numerical parameter pclipped. Specifically, for the i-th parameter, the clipping function is calculated as follows: ,in, and These are the lower and upper bounds of the corresponding parameter boundaries, respectively. This represents the final compliance value of the i-th numerical parameter in the first constraint rule after boundary clipping. This represents the original predicted value of the i-th numerical parameter in the first constraint rule.

[0048] Step a5: Extract rule elements related to the current performance project from the contract graph, wherein the rule elements shall at least include spatial range, key geometric features and search strategy, and assemble the numerical parameters with the rule elements to obtain the first constraint rule.

[0049] In step a5, rule elements related to the current performance project are extracted from the contract graph. These elements include at least the spatial range, such as the designated area for equipment installation; key geometric features, such as the external dimensions and key feature points of the equipment; and search strategies, such as nearest neighbor search or normal consistency search. Then, numerical parameters are assembled and integrated with these rule elements to form a complete first constraint rule. For example, the rule can be expressed as: within the spatial range R, the nearest neighbor search strategy is used to search for point clouds that match the geometric features of the benchmark data, and the matching deviation shall not exceed the threshold δ.

[0050] Step a6: The second submodule inputs the fused representation vector into a sequence-to-sequence encoder-decoder structure to generate a state transition probability graph. Using the state transition constraints obtained in real time from the contract graph, the structure of the state transition probability graph is adjusted through a graph editing operation to obtain the second constraint rule.

[0051] Among them, the state transition probability graph refers to a graph structure that represents the transition probability between process states, where nodes represent states and edges represent transition probabilities; State transition constraints refer to the preconditions or logical restrictions that must be met for the transition of certain states in a process flow. In other words, they are rigid logical rules that must be followed between processes as stipulated in a contract or specification. Graph editing operations refer to the process of adjusting the state transition probability graph according to the state transition constraints in the contract graph, such as deleting transition edges that do not meet the preconditions.

[0052] In step a6, the second submodule inputs the fused representation vector into a sequence-to-sequence encoder-decoder structure, which generates a state transition probability map in matrix form by learning the dependency patterns between processes. It means that, among them, N R represents the total number of process states, and R represents the real number field, meaning that all transition probabilities in the matrix are real numbers between 0 and 1. At the same time, state transition constraints are obtained in real time from the contract graph. For example, some state transitions must meet specific preconditions, such as the "equipment arrival" state must be experienced before the "lifting in place" state. Then, the generated P matrix is ​​adjusted through graph editing operations, that is, the transition probabilities of those that violate the queried state transition constraints are set to zero, and each row of the matrix is ​​re-normalized so that the sum of the probabilities of each row is 1. The adjusted state transition graph is the final second constraint rule.

[0053] This application decouples complex contract terms into spatial constraint rules and process logic rules, enabling subsequent point cloud verification and state recognition to be performed based on clear numerical conditions and temporal relationships, thereby improving the accuracy and interpretability of rule parsing and laying the rule foundation for automated performance verification.

[0054] S103. Using a point cloud segmentation network, the real-time point cloud sequence is segmented according to the first constraint rule to obtain the entity point cloud to be verified.

[0055] In one specific implementation, step S103 includes: Step 1031: Parse the spatial range and key geometric features of the target entity from the first constraint rule, and generate an attention weight map based on the spatial range and the key geometric features.

[0056] The spatial range refers to the area where the target entity may exist, such as the specified coordinate range for equipment installation; the key geometric features refer to the technical parameters stipulated in the contract, such as the external dimensions and key structural features of the target entity.

[0057] In step 1031, the spatial range and key geometric features of the target entity to be verified are first parsed from the first constraint rule, and an attention weight map corresponding to the point cloud space to be processed is generated based on this information. This map is a three-dimensional weight matrix whose size corresponds completely to the spatial range covered by the point cloud. Higher initial weights are given to the regions within the spatial range that have a high degree of matching with the key geometric features, and lower initial weights are given to the regions outside the spatial range or regions where the features do not match.

[0058] Step 1032: Input the real-time point cloud sequence into a pre-trained point cloud segmentation network, which includes a feature extraction encoder, a feature interaction fusion unit, and a semantic segmentation decoder.

[0059] The feature extraction encoder of the point cloud segmentation network adopts a multi-layer convolutional structure, and each layer contains random sampling and local feature aggregation modules. Its input layer receives coordinate and color information of real-time point cloud sequences, with a total of 6 channels. The encoder has 4 layers, and the number of points decreases and the feature dimension increases after downsampling in each layer. The feature interaction fusion unit is designed as a cross-scale feature fusion structure, which receives multi-scale features from each layer of the encoder and fuses them through a spatial attention mechanism. The semantic segmentation decoder adopts an upsampling and skip connection structure, and gradually restores the point cloud density and outputs the class probability of each point.

[0060] The loss function is designed with weighted cross-entropy loss to penalize classification errors of different categories, and a higher weight is set for the target entity category to enhance the network's learning ability for that category. The training process uses a stochastic gradient descent optimizer with an initial learning rate of 0.01, a batch size of 8, and 50 training iterations. During training, the network parameters are continuously updated through backpropagation to make the prediction results gradually approach the true labels.

[0061] In terms of parameter design, the input point cloud has a fixed number of 4096 points, and the output feature dimensions of each layer of the encoder are 64, 128, 256, and 512, respectively; the number of neighborhood points in the local feature aggregation module is set to 16; the random sampling ratio is set to retain 25% of the points in each layer; the activation function is the ReLU function, and the Dropout probability is set to 0.5 to prevent overfitting.

[0062] In step 1032, the acquired real-time point cloud sequence is input into a pre-trained point cloud segmentation network. This network adopts an encoder-decoder architecture and consists of three parts: the first part is a feature extraction encoder, which is responsible for extracting features of the point cloud layer by layer; the second part is a feature interaction fusion unit, which is responsible for cross-scale feature fusion and introducing attention guidance; and the third part is a semantic segmentation decoder, which is responsible for generating point-by-point classification results.

[0063] Step 1033: Using the feature extraction encoder, perform layer-by-layer convolution and downsampling operations on the real-time point cloud sequence to obtain multi-layer point cloud features.

[0064] Among them, layer-by-layer convolution and downsampling refer to the process of multi-level processing of point cloud data. The convolution operation is used to extract the local features of each point and its neighborhood, and the downsampling operation is used to reduce the number of points, thereby expanding the receptive field and extracting higher-level features.

[0065] In step 1033, after the real-time point cloud sequence is input into the feature extraction encoder, the encoder begins to process layer by layer. In the first layer, for each point, the local feature aggregation module collects the points in its neighborhood and uses a multilayer perceptron to extract features from these neighborhood points to obtain a new feature vector for that point. Then, a random downsampling operation is performed, and some points are randomly discarded according to a certain proportion, retaining representative points as input for the next layer. The second layer repeats the above process on the sampled points to further expand the receptive field and extract higher-level features. The third and fourth layers continue this process, and the number of points decreases after each downsampling, but the feature dimension represented by each point increases, and the receptive field also expands accordingly.

[0066] After four layers of processing, multi-layer point cloud features are obtained, including shallow detail feature maps output by the first layer, mid-level structural feature maps output by the second layer, and high-level semantic feature maps output by the third and fourth layers. These feature maps have different spatial resolutions and feature dimensions. Shallow features retain rich spatial detail information, while deep features contain stronger semantic information.

[0067] Step 1034: Input the multi-layer point cloud features into the feature interaction fusion unit and introduce the attention weight map. The feature interaction fusion unit performs cross-scale fusion of the multi-layer point cloud features through a spatial attention mechanism to obtain fused features.

[0068] In step 1034, the feature interaction fusion unit receives multi-layer point cloud features from the encoder and introduces an attention weight map. The specific implementation process of its cross-scale fusion is as follows: First, the shallow feature map and the deep feature map are spatially aligned. Since the shallow feature map has a higher spatial resolution and the deep feature map has a lower resolution, it is necessary to use an interpolation method to make the resolution of the deep feature map consistent with that of the shallow feature map, and use bilinear interpolation to upsample the deep feature map to the size of the shallow feature map.

[0069] Then, the aligned multi-layer feature maps are concatenated along the channel dimension to form a joint feature tensor containing multi-scale information. A spatial attention mechanism is then introduced for weighted processing. This spatial attention mechanism is designed based on the content of the attention weight map and the feature map itself. The specific calculation process is as follows: the joint feature tensor is average-pooled along the channel dimension to obtain a single-channel spatial feature map, which reflects the overall response intensity at each spatial location. This spatial feature map is then multiplied element-wise with the attention weight map to obtain the fused spatial attention weights. Finally, these fused spatial attention weights are multiplied element-wise with the original joint feature tensor to obtain the weighted fused features.

[0070] Step 1035: The fused features are upsampled and classified point by point using the semantic segmentation decoder to obtain a segmentation mask. A probability threshold-based binarization operation is applied to the segmentation mask, and three-dimensional connected component analysis is performed to obtain multiple point cloud clusters.

[0071] In step 1035, the semantic segmentation decoder receives the fused features and gradually recovers the point cloud density through upsampling operations. The upsampling adopts the nearest neighbor interpolation method to gradually map the low-resolution feature map back to the original point cloud resolution. After upsampling at each layer, it is spliced ​​with the feature map of the corresponding layer in the encoder through skip connections to supplement the detailed information that may be lost during the downsampling process of the encoder, so as to obtain a feature map with the same number of points as the original point cloud.

[0072] Subsequently, a point-by-point classification operation is performed on this feature map after resolution restoration. That is, a fully connected layer is used to map the feature vector of each point into a class probability vector, and the length of this probability vector is equal to the preset total number of classes. Then, the class corresponding to the maximum probability is taken as the predicted class of that point to obtain the initial segmentation mask.

[0073] To achieve this, a probability threshold-based binarization operation is applied to the segmentation mask. This involves setting a confidence threshold, such as 0.7, and then relabeling points that are predicted to be the target entity category but have a probability value lower than this threshold as background. Next, a 3D connected component analysis is performed on the binarized mask. This involves determining whether the spatial distance between points is less than a preset connectivity threshold to aggregate spatially adjacent points with the same target label into the same point cloud cluster, ultimately resulting in multiple independent point cloud clusters.

[0074] Step 1036: Verify the matching degree between the geometric features of each point cloud cluster and the key geometric features corresponding to the first constraint rule to obtain the entity point cloud to be verified.

[0075] Among them, geometric features refer to quantitative indicators that describe the shape of point cloud clusters, such as length, width, height, volume, surface area, principal axis direction, etc.

[0076] In step 1036, for each point cloud cluster obtained, its geometric features are calculated. The specific calculation method includes calculating the three principal axis directions of the point cloud cluster and their corresponding eigenvalues ​​through principal component analysis, and determining the extension length of the point cloud cluster in the three principal axis directions based on the magnitude of the eigenvalues, thereby obtaining the approximate circumscribed cuboid size; at the same time, the total volume of the point cloud cluster is calculated, and the space occupied by the entity is estimated through the point cloud density, and then the coordinates of the center point of the point cloud cluster are calculated to determine the actual installation position of the entity.

[0077] Subsequently, these calculated geometric features are compared item by item with the key geometric features specified in the first constraint rule. For example, the rule specifies that the target equipment has an external dimension of 5 meters in length, 2 meters in width, and 3 meters in height, with an allowable deviation of ±0.1 meters. Then, for each point cloud cluster, its calculated length, width, and height dimensions are compared with the standard values. If all three dimensions are within the allowable deviation range, the geometric features of the point cloud cluster are considered to meet the contract requirements; otherwise, the point cloud cluster fails the verification. Next, for all point cloud clusters that pass the matching degree verification, their set is output as the entity point cloud to be verified.

[0078] This application realizes the mapping from contract rules to physical entities, which can provide an accurate data foundation for subsequent deviation analysis and status determination.

[0079] S104. The point cloud of the entity to be verified is iteratively registered with the reference data according to the search strategy defined in the first constraint rule to generate trajectory deviation data.

[0080] In one specific implementation, step S104 includes: Step 1041: Construct a search function based on the search strategy defined in the first constraint rule.

[0081] The search function refers to a rule execution tool used to find the corresponding point of the point to be verified in the benchmark data. This function transforms the search strategy defined in the first constraint rule into specific mathematical search logic, and determines how to find the most suitable matching point for each point to be verified. This embodiment does not limit the expression of the function.

[0082] In step 1041, the search strategy defined in the first constraint rule is parsed out. The search strategy specifies how to perform point cloud matching, such as using the nearest neighbor principle for searching, or requiring the consistency of the normal vector direction of the points to be considered at the same time. Then, a search function is constructed based on this strategy. If the search strategy is specified as nearest neighbor search, the constructed search function will use a spatial index structure to organize the reference data points in order to quickly find the nearest reference point corresponding to each point to be verified.

[0083] Step 1042: Using the search function, perform a corresponding point search in the benchmark data for the points in the point cloud of the entity to be verified, forming corresponding point pairs.

[0084] In step 1042, each point in the entity point cloud to be verified is used as the search object, and the search function is called to search for the corresponding matching point in the reference data. The search function finds the nearest point in the reference data for each point to be verified according to a preset strategy, or a point that simultaneously satisfies the normal consistency condition. Each time a match is found, a corresponding point pair is formed.

[0085] Step 1043: Calculate a spatial transformation parameter based on the corresponding point pair, and calculate the registration error of this iteration based on the spatial transformation parameter and the corresponding point pair.

[0086] Registration error refers to the average distance between all corresponding point pairs after spatial transformation, and is used to measure the accuracy of the current registration.

[0087] In step 1043, based on the corresponding point pairs, the singular value decomposition method in point cloud registration is used to calculate the optimal spatial transformation parameters. That is, firstly, the center points of all corresponding point pairs are calculated, then the centering process is performed, the covariance matrix is ​​constructed and matrix decomposition is performed, and finally the spatial transformation parameters containing the rotation matrix and translation vector are obtained. These parameters are then applied to the current point cloud to be verified to calculate the distance between all corresponding point pairs, and the average of these distances is calculated to obtain the registration error of this iteration.

[0088] Step 1044: Apply the spatial transformation parameters to the point cloud of the entity to be verified to obtain an updated point cloud.

[0089] In step 1044, spatial transformation parameters are applied to the current entity point cloud to be verified. For each point in the point cloud, its new position is calculated by rotation matrix and translation vector. After all points are transformed, an updated point cloud is formed.

[0090] Step 1045: Repeat the corresponding point search and update steps until the preset iteration stop condition is met, so as to obtain the final spatial transformation parameters and the final registration error of the current frame.

[0091] In step 1045, steps 1042 to 1044 are repeated as a complete iterative loop. At the beginning of each iteration, the updated point cloud obtained in the previous round is used as input to re-execute the corresponding point search to form new corresponding point pairs, calculate new spatial transformation parameters and registration errors, and update the point cloud position again. The iterative loop continues, and each iteration further aligns the point cloud to be verified with the reference data.

[0092] After each iteration, it is determined whether the preset iteration stopping conditions are met. The iteration stopping conditions include three cases: First, the current registration error is less than the preset accuracy threshold, indicating that the point cloud has been aligned to a sufficiently accurate level; second, the current iteration number has reached the preset maximum iteration number limit; third, the change between the spatial transformation parameters calculated in this iteration and the transformation parameters in the previous iteration is less than the preset change threshold. As long as any one of these conditions is met, the iteration process stops, and the spatial transformation parameters calculated in the last iteration are used as the final spatial transformation parameters, and the registration error of the last iteration is used as the final registration error for output.

[0093] Step 1046: Combine the final spatial transformation parameters and the final registration error of each frame according to the time sequence to generate trajectory deviation data.

[0094] In step 1046, the final spatial transformation parameters and final registration errors of all frames are arranged and combined in chronological order of acquisition time to form a time series data. Each time point in this series corresponds to a set of spatial transformation parameters and registration error values, which together constitute the trajectory deviation data.

[0095] This application generates trajectory deviation data that reflects the changes in the spatial position and accuracy of the entity over time by iteratively registering the point cloud of the entity to be verified with the contract reference data frame by frame, thereby realizing the quantitative tracking of the displacement and deformation of the entity during construction.

[0096] S105. Based on the second constraint rule, construct the hidden state space of the Hidden Markov Model, and synchronously decode the trajectory deviation data in the hidden state space to determine the corresponding state.

[0097] In one specific implementation, step S105 includes: Step 1051: Input the state sequence defined by the second constraint rule into a rule parser. The rule parser outputs a state transition diagram according to the order of the states and maps the structure of the state transition diagram to the state transition matrix and initial state probability distribution of a hidden Markov model.

[0098] In step 1051, the state sequence defined in the second constraint rule is first input into the rule parser. The rule parser constructs a state transition diagram based on the predefined order between these states. The nodes in the diagram represent the state of each process, and the directed edges represent the allowed process transition directions.

[0099] Then, the structure of this state transition graph is mapped to the state transition matrix of a hidden Markov model. Each element in the matrix represents the probability of transitioning from one state to another. For directed edges that exist in the state transition graph, an initial transition probability is assigned. For edges that do not exist in the graph, the corresponding transition probability is set to zero. At the same time, the initial state probability distribution is determined according to contractual provisions or engineering experience, that is, which process state the equipment is most likely to be in when construction begins.

[0100] Step 1052: Perform multi-scale time window segmentation on the trajectory deviation data, extract the deviation statistical features within each time window, input the deviation statistical features into a temporal convolutional neural network for pattern recognition processing, and obtain the observation symbol corresponding to each time window.

[0101] In step 1052, the trajectory deviation data is first segmented into multiple time windows, that is, multiple time windows of different lengths are used. For example, short windows are used to capture rapid changes, and long windows are used to capture the overall trend. The continuous deviation data is divided into multiple time segments. Then, the trajectory deviation data in each time window is extracted to extract the deviation statistical features, including the average value of the deviation in the window, the fluctuation range, the trend of change, etc., thereby forming a feature vector.

[0102] Subsequently, these feature vectors are input into a pre-trained temporal convolutional neural network for pattern recognition processing. The network extracts local patterns of the time series through convolutional layers, maps the feature vector of each time window to class probabilities through fully connected layers, and outputs the observation symbol most likely to correspond to the time window. The observation symbol represents a typical pattern of trajectory deviation data, such as stable, slowly increasing, rapidly increasing, or drastic fluctuation.

[0103] Step 1053: Based on the candidate state at the current moment, query the corresponding deviation tolerance range from the contract graph, and calculate the likelihood probability of the current observation symbol based on the deviation tolerance range.

[0104] In step 1053, for each possible candidate state at the current moment, such as the equipment being in the hoisting and positioning state, the allowable deviation range corresponding to that state is queried from the contract graph. For example, in the hoisting and positioning state, the allowable deviation range for the horizontal position is ±5 mm. Then, the observation symbol of the current time window is compared with this allowable deviation range. If the deviation pattern reflected by the observation symbol is completely within the allowable range, a high likelihood probability is assigned; if the deviation reflected by the observation symbol exceeds the allowable range, a low likelihood probability is assigned.

[0105] Step 1054: If the transition path from the current candidate state to the next state conflicts with the preconditions obtained from the contract graph, then the probability of the corresponding transition path is set to zero.

[0106] In step 1054, during the state decoding process, for each transition path from the current candidate state to the next state, the preconditions that the transition path must meet are queried from the contract graph. For example, the precondition for transitioning from the hoisting and positioning state to the precision acceptance state may be that the equipment has been fixed. If the actual construction situation at the current moment does not meet these preconditions, the transition probability of this transition path is set to zero, indicating that the transition is prohibited at the current moment.

[0107] Step 1055: Based on the state transition matrix, the initial state probability distribution, the likelihood probability, and the transition paths not set to zero, the Viterbi decoding algorithm is used to perform a global optimal path search on the observed symbols to obtain a hidden state path, and path filtering is performed on the hidden state path to determine the corresponding state.

[0108] In step 1055, based on the state transition matrix and the initial state probability distribution, the calculated likelihood probability, and the transition paths that are not set to zero after correction, the Viterbi decoding algorithm is used to perform a global optimal path search for all observed symbols. That is, starting from the initial time, the Viterbi algorithm is used to recursively calculate the maximum probability path to each state, record the previous state in the path, and after traversing all time windows, backtrack to obtain the hidden state path with the highest probability, that is, the most likely process state corresponding to each time.

[0109] Then, path filtering is performed on the obtained hidden state path. The path filtering adopts a sliding window mechanism and sets a window with a length of three to check whether the state at the middle moment of the window is continuous with the state at the two moments before and after. If the state at the middle moment is not continuous with the states before and after, for example, if the states before and after are both in the hoisting state and the acceptance state suddenly appears in the middle, it is determined that the jump is caused by noise, and the state at the middle moment is corrected to be consistent with the states before and after. Finally, the state corresponding to each moment after filtering is output, that is, the actual process stage of the equipment in each time window.

[0110] This application combines trajectory deviation data with contract process logic and uses a hidden Markov model to perform time-series decoding of the construction process in order to accurately identify the real-time process status of the equipment, thereby providing a reliable status basis for subsequent acceptance condition determination.

[0111] S106. When the trajectory deviation data of any generation meets the first acceptance condition defined in the first constraint rule and the state of the corresponding generation simultaneously meets the second acceptance condition defined in the second constraint rule, a task completion list is generated. If the trajectory deviation data continues to fail to meet the first acceptance condition, or the state of the corresponding generation fails to meet the second acceptance condition, task abnormality information is generated.

[0112] In step S106, for each frame of point cloud data, the trajectory deviation data and the corresponding state obtained by synchronous decoding are judged under two conditions. The first acceptance condition may be: whether the registration error is less than the deviation threshold specified in the contract. The second acceptance condition may be: whether the current acceptance process state has been reached.

[0113] If the trajectory deviation data of a certain frame simultaneously meets the first and second acceptance conditions, that is, the installation accuracy of the frame meets the contract requirements and has reached the correct acceptance process, the task completion process is triggered, and a task completion list is generated. The list records in detail the entity information, acceptance time, deviation value and acceptance conclusion of this acceptance.

[0114] If the trajectory deviation data of all frames from the first frame to the last frame fails to meet the first acceptance condition, or if at a certain moment the deviation data meets the requirements but the corresponding state fails to meet the process specified by the second acceptance condition, then it is determined that there is an anomaly in the performance process. At this time, the anomaly handling process is triggered to generate task anomaly information. The information records whether the anomaly type is a continuous deviation exceeding the limit or a process state mismatch, as well as the earliest time point when the anomaly occurred and the deviation value at that time.

[0115] This application enables joint verification of both spatial accuracy and process status during construction, and ensures that a task completion list is generated only when the physical installation meets the contract requirements and the process is executed correctly, thus avoiding acceptance omissions due to unqualified accuracy or disordered processes.

[0116] S107. Input the task completion list or the task exception information into the OA system to generate management processes corresponding to task completion and task exception respectively.

[0117] In one specific implementation, step S107 includes: Step 1071: Based on the task completion list or the task exception information, query the contract graph in real time to obtain a set of process definition parameters.

[0118] Among them, process definition parameters refer to the configuration data required to create specific management processes in the OA system, which includes information such as process type, approval nodes, processing roles, and time limit requirements.

[0119] In step 1071, the task completion list or task exception information is used as a query condition to access the contract graph in real time. That is, based on the task type, equipment information, exception category and other elements contained in the list or information, the corresponding process definition parameters are retrieved from the contract graph. For example, if a task completion list is generated, the approval nodes and personnel configuration required for the acceptance confirmation process are retrieved; if task exception information is generated, the processing department and time limit requirements required for the rectification process are retrieved.

[0120] Step 1072: Based on the process definition parameters, call an instantiation engine to generate a structured process instance.

[0121] In step 1072, the process definition parameters are input into the instantiation engine. Based on these parameters, the instantiation engine selects the corresponding template from the process template library and fills the parameters into the corresponding positions of the template, thereby generating a complete and executable structured process instance. This process instance contains all necessary information such as process name, initiator, current node, subsequent nodes, and approval rules.

[0122] Step 1073: Convert the process instance into a set of API call instructions that can be recognized by the OA system through an interface adapter.

[0123] In step 1073, the process instance is passed to the interface adapter. The interface adapter parses the data in the process instance and reorganizes the data into a format that meets the interface requirements according to the API interface specifications provided by the target OA system. After the conversion is completed, a set of API call instructions is output. These instructions contain all the parameters and calling methods required to create the process.

[0124] Step 1074: Execute the API call instruction to create and start a management process in the OA system that corresponds to the task completion list or the task exception information.

[0125] In step 1074, the API call instruction is sent to the interface address of the OA system. After receiving the instruction, the OA system parses the parameters and creates a new process instance in the system, and starts the process according to the instruction content. Finally, a management process corresponding to the original task completion list or task exception information is generated in the OA system, such as an acceptance approval process or a rectification process.

[0126] This application achieves a complete closed loop from on-site perception and status determination to management process triggering by directly connecting the results of automated verification to the OA system, thereby ensuring that the completion or abnormality of each performance node can be promptly transformed into traceable and executable management tasks.

[0127] Figure 3 This application provides a schematic diagram of a specific implementation of a project lifecycle OA workflow automation management system based on contract graphs, as illustrated in the embodiments of this application. Figure 3 The system may include: The acquisition module 31 is used to acquire real-time point cloud sequences of the project performance area and extract the benchmark data corresponding to the project performance from the pre-built contract map.

[0128] The parsing module 32 is used to parse the contract graph using a graph neural network to obtain the first constraint rule and the second constraint rule.

[0129] The segmentation module 33 is used to segment the real-time point cloud sequence using a point cloud segmentation network according to the first constraint rule to obtain the entity point cloud to be verified.

[0130] The registration module 34 is used to iteratively register the point cloud of the entity to be verified with the reference data according to the search strategy defined in the first constraint rule to generate trajectory deviation data.

[0131] The decoding module 35 is used to construct the hidden state space of the hidden Markov model based on the second constraint rule, and synchronously decode the trajectory deviation data in the hidden state space to determine the corresponding state.

[0132] The first generation module 36 is used to generate a task completion list when the trajectory deviation data of any generation meets the first acceptance condition defined in the first constraint rule and the state of the corresponding generation simultaneously meets the second acceptance condition defined in the second constraint rule. If the trajectory deviation data continues to fail to meet the first acceptance condition, or the state of the corresponding generation fails to meet the second acceptance condition, task abnormal information is generated.

[0133] The second generation module 37 is used to input the task completion list or the task exception information into the OA system and generate management processes corresponding to task completion and task exception respectively.

[0134] The project lifecycle OA process automation management system based on contract graphs in this application is used to implement the aforementioned project lifecycle OA process automation management method based on contract graphs. Therefore, the specific implementation of the project lifecycle OA process automation management system based on contract graphs can be found in the embodiment section of the project lifecycle OA process automation management method based on contract graphs above. The specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0135] like Figure 4 As shown, this application also provides an electronic device, including: a memory 41 for storing a computer program; and a processor 42 for executing the computer program to implement the steps of any of the above-described project lifecycle OA process automation management methods based on contract graphs.

[0136] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described project lifecycle OA process automation management methods based on contract graphs.

[0137] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0138] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the project lifecycle OA process automation management method based on contract graphs.

[0139] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0140] The foregoing has provided a detailed description of the project lifecycle OA process automation management method and system based on contract graphs provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for automated management of the entire project lifecycle OA process based on contract graphs, characterized in that, include: Collect real-time point cloud sequences of the project performance area and extract benchmark data corresponding to project performance from a pre-built contract map; The contract graph is analyzed using a graph neural network to obtain the first constraint rule and the second constraint rule; A point cloud segmentation network is used to segment the real-time point cloud sequence according to the first constraint rule to obtain the entity point cloud to be verified. The point cloud of the entity to be verified is iteratively registered with the reference data according to the search strategy defined in the first constraint rule to generate trajectory deviation data; Based on the second constraint rule, a hidden state space of the Hidden Markov Model is constructed, and the trajectory deviation data is synchronously decoded in the hidden state space to determine the corresponding state. When the trajectory deviation data of any generation meets the first acceptance condition defined in the first constraint rule and the state of the corresponding generation simultaneously meets the second acceptance condition defined in the second constraint rule, a task completion list is generated. If the trajectory deviation data continues to fail to meet the first acceptance condition, or the state of the corresponding generation fails to meet the second acceptance condition, task abnormality information is generated. Input the task completion list or the task exception information into the OA system to generate management processes corresponding to task completion and task exception respectively.

2. The method according to claim 1, characterized in that, The step of using a graph neural network to parse the contract graph to obtain the first constraint rule and the second constraint rule includes: Nodes and edges related to the current performance project are extracted from the contract graph to obtain a target subgraph, wherein the target subgraph includes a central node representing the entity to be verified, attribute nodes representing spatial location or technical standards, and event nodes representing the sequence of performance actions. Construct a two-channel graph neural network and input the target subgraph into the two-channel graph neural network; Through the first channel of the dual-channel graph neural network, based on the path distance between the attribute node and the center node in the shown contract graph, an attention convergence operation based on distance decay is performed to obtain the constraint feature vector; Through the second channel of the dual-channel graph neural network, based on the predefined causal sequence relationship between the event nodes, the weights of edges that conform to the causal order are enhanced, while edge connections that do not conform to the causal order are suppressed, to obtain the sequence feature vector; A rule generator decodes the constraint feature vector into a first constraint rule and the sequence feature vector into a second constraint rule.

3. The method according to claim 2, characterized in that, The step of decoding the constraint feature vector into a first constraint rule and the sequence feature vector into a second constraint rule using a rule generator includes: The constraint feature vector and the sequence feature vector are input into a rule generator, which consists of a feature fusion module based on an attention mechanism and a multi-output head. The feature fusion module concatenates the constraint feature vector with the sequence feature vector to form a joint feature vector, and performs self-attention calculation on the joint feature vector to obtain a fused representation vector; The fused representation vector is simultaneously input into two parallel sub-modules of the multi-output head; The first submodule maps the fused representation vector to an original parameter vector through a fully connected layer, and uses the parameter boundary obtained in real time from the contract graph to numerically constrain the original parameter vector through a boundary clipping function to obtain numerical parameters. Extract rule elements related to the current performance project from the contract graph, wherein the rule elements shall at least include spatial range, key geometric features and search strategy, and assemble the numerical parameters with the rule elements to obtain the first constraint rule; The second submodule inputs the fused representation vector into a sequence-to-sequence encoder-decoder structure to generate a state transition probability graph. Using the state transition constraints obtained in real time from the contract graph, the module performs a graph editing operation to adjust the structure of the state transition probability graph and obtain the second constraint rule.

4. The method according to claim 1, characterized in that, The process of constructing a hidden state space for a hidden Markov model based on the second constraint rule, and synchronously decoding the trajectory deviation data in the hidden state space to determine the corresponding state, includes: The state sequence defined by the second constraint rule is input into a rule parser. The rule parser outputs a state transition diagram according to the order between states and maps the structure of the state transition diagram to the state transition matrix and initial state probability distribution of a hidden Markov model. The trajectory deviation data is segmented into multi-scale time windows, and the deviation statistical features within each time window are extracted. The deviation statistical features are then input into a temporal convolutional neural network for pattern recognition processing to obtain the observation symbols corresponding to each time window. Based on the candidate state at the current moment, the corresponding deviation tolerance range is queried from the contract graph, and the likelihood probability of the current observation symbol is calculated based on the deviation tolerance range; If the transition path from the current candidate state to the next state conflicts with the preconditions obtained from the contract graph, the probability of the corresponding transition path is set to zero. Based on the state transition matrix, the initial state probability distribution, the likelihood probability, and the transition paths not set to zero, the Viterbi decoding algorithm is used to perform a global optimal path search on the observed symbols to obtain a hidden state path, and path filtering is performed on the hidden state path to determine the corresponding state.

5. The method according to claim 1, characterized in that, The step of employing a point cloud segmentation network to segment the real-time point cloud sequence according to the first constraint rule to obtain the entity point cloud to be verified includes: The spatial extent and key geometric features of the target entity are parsed from the first constraint rule, and an attention weight map is generated based on the spatial extent and the key geometric features. The real-time point cloud sequence is input into a pre-trained point cloud segmentation network, which includes a feature extraction encoder, a feature interaction fusion unit, and a semantic segmentation decoder. The feature extraction encoder performs layer-by-layer convolution and downsampling operations on the real-time point cloud sequence to obtain multi-layer point cloud features. The multi-layer point cloud features are input into the feature interaction fusion unit, and the attention weight map is introduced. The feature interaction fusion unit performs cross-scale fusion of the multi-layer point cloud features through a spatial attention mechanism to obtain fused features. The semantic segmentation decoder performs upsampling and point-by-point classification operations on the fused features to obtain a segmentation mask. A probability threshold-based binarization operation is applied to the segmentation mask, and three-dimensional connected component analysis is performed to obtain multiple point cloud clusters. The matching degree of the geometric features of each point cloud cluster with the key geometric features corresponding to the first constraint rule is verified to obtain the entity point cloud to be verified.

6. The method according to claim 1, characterized in that, The step of iteratively registering the point cloud of the entity to be verified with the reference data according to the search strategy defined in the first constraint rule to generate trajectory deviation data includes: Based on the search strategy defined in the first constraint rule, construct a search function; Using the search function, a corresponding point search is performed in the benchmark data for the points in the point cloud of the entity to be verified, forming corresponding point pairs; Calculate a spatial transformation parameter based on the corresponding point pair, and calculate the registration error of this iteration based on the spatial transformation parameter and the corresponding point pair; The spatial transformation parameters are applied to the point cloud of the entity to be verified to obtain an updated point cloud; The corresponding point search and update steps are executed repeatedly until the preset iteration stop condition is met, so as to obtain the final spatial transformation parameters and the final registration error of the current frame. The final spatial transformation parameters of each frame are combined with the final registration error in a temporal sequence to generate trajectory deviation data.

7. The method according to claim 1, characterized in that, The step of inputting the task completion list or the task exception information into the OA system to generate management processes corresponding to task completion and task exception, respectively, includes: Based on the task completion list or the task exception information, query the contract graph in real time to obtain a set of process definition parameters; Based on the process definition parameters, an instantiation engine is invoked to generate a structured process instance; The process instance is converted into a set of API call instructions that can be recognized by the OA system through an interface adapter; Execute the API call instruction to create and start a management process in the OA system that corresponds to the task completion list or the task exception information.

8. A project lifecycle OA process automation management system based on contract graphs, characterized in that, include: The data acquisition module is used to collect real-time point cloud sequences of the project performance area and extract the benchmark data corresponding to the project performance from the pre-built contract map. The parsing module is used to parse the contract graph using a graph neural network to obtain the first constraint rule and the second constraint rule; The segmentation module is used to segment the real-time point cloud sequence using a point cloud segmentation network according to the first constraint rule to obtain the entity point cloud to be verified. The registration module is used to iteratively register the point cloud of the entity to be verified with the reference data according to the search strategy defined in the first constraint rule, and generate trajectory deviation data. The decoding module is used to construct the hidden state space of the Hidden Markov Model based on the second constraint rule, and synchronously decode the trajectory deviation data in the hidden state space to determine the corresponding state. The first generation module is used to generate a task completion list when the trajectory deviation data of any generation meets the first acceptance condition defined in the first constraint rule and the state of the corresponding generation simultaneously meets the second acceptance condition defined in the second constraint rule. If the trajectory deviation data continues to fail to meet the first acceptance condition, or the state of the corresponding generation fails to meet the second acceptance condition, task abnormal information is generated. The second generation module is used to input the task completion list or the task exception information into the OA system and generate management processes corresponding to task completion and task exception respectively.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the project lifecycle OA process automation management method based on contract graphs as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the automated management method for the entire project lifecycle OA process based on contract graphs as described in any one of claims 1 to 7.

Images