A supervision post capacity portrait construction method and system based on a knowledge graph

Abstract

The application relates to the field of building engineering supervision information technology, and discloses a supervision post capacity portrait construction method and system based on a knowledge graph, wherein the construction method comprises the following steps: acquiring engineering basic data to generate a clause mapping relationship of a post capacity knowledge graph; receiving multi-source data reported by a mobile terminal to generate a component-level process event stream; outputting an active component inspection node set and a remaining inspection duration from an incremental process state machine; calculating a verification time delay budget to generate a hierarchical verification scheduling sequence; outputting a process completion degree value and a scene repetition identifier in a lightweight verification channel and a deep verification channel respectively; generating a pre-process procedure credibility abstract vector and outputting a time sequence dependent credibility; fusing content credibility to generate a re-inspection trigger item and an irreversible suspect record item; performing conflict resolution and merging, and outputting a post capacity portrait which is distinguished into three types of identification, i.e., a re-inspection closed type, a non-re-inspection type within a window, and an irreversible suspect type.

Description

Technical Field

[0001] This invention relates to the field of information technology in construction engineering supervision, and more specifically, to a method and system for constructing a competency profile of supervision positions based on knowledge graphs. Background Technology

[0002] With the large-scale application of prefabricated buildings, the quality of supervision of concealed works (such as sleeve grouting and pre-buried pipeline laying) has a significant impact on project safety. The industry is gradually introducing information technology such as image evidence, BIM technology and mobile terminal reporting to assist in the management of the supervision process, and is trying to build job competency evaluation indicators based on performance data.

[0003] However, concealed works are characterized by tight process connections, short coverage time limits, and unstable on-site network environments. On-site verification often faces technical bottlenecks such as large amounts of image data, weak or disconnected network connections at edge nodes, and untimely manual review due to approaching closure deadlines. The method of periodically counting the number of uploaded images is difficult to reflect the authenticity and temporal validity of the evidence of performance of duties, and it is also difficult to give verification conclusions in a timely manner before the components are covered.

[0004] Existing information technology solutions for supervision rely heavily on manual sampling of images or simple image quantity statistics to construct performance evaluation indicators for supervisors. They do not automatically verify the content of images and the completion status of procedures, nor do they establish constraints on the pre- and post-process dependencies between procedures. This makes it difficult to identify evidence of performance that is reused in a scene, misaligned in time sequence, or has questionable pre-process results.

[0005] The current supervision performance verification scheme does not take into account the causal relationship between the acceptance records of concealed works and the time dependence of the preceding and subsequent processes. This results in some performance evidence with misaligned timelines or questionable preceding processes being included in the competency evaluation, affecting the credibility of the job competency profile. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for constructing a competency profile of supervisory positions based on knowledge graphs, which solves the technical problems of inaccurate content credibility assessment and difficulty in quantifying and tracing the performance of supervisory positions due to the lack of process sequence context support in the verification of hidden works acceptance images.

[0007] This disclosure provides at least one embodiment of a method for constructing a competency profile for supervisory positions based on a knowledge graph. The method includes: generating a directed graph of process dependencies, a mapping relationship between closure conditions and clauses based on basic engineering data; generating a component-level process event flow from reported multi-source data, and outputting the remaining inspectable time through a process state machine; calculating a verification delay budget based on the remaining inspectable time, process importance, and computing power margin; assigning acceptance images to lightweight verification channels and deep verification channels according to the verification delay budget, and outputting process completion scores and scene duplication identifiers; calculating an attenuation coefficient along the directed graph of process dependencies using the credibility summary vector of closed components and applying a time-series conflict penalty, and outputting time-series dependency credibility; fusing the process completion score, the scene duplication identifier, and the time-series dependency credibility into content credibility; generating re-inspection trigger entries or irreversible questionable record entries according to the remaining inspectable time, and weighting and aggregating them according to the clause mapping relationship to output a competency profile for the supervisory position; wherein the verification delay budget is the minimum value between the time occupancy share and the computing power share.

[0008] For example, in the knowledge graph-based method for constructing a competency profile for supervisory positions provided in at least one embodiment of this disclosure, the basic engineering data includes a BIM component list, an inspection batch division table, inspection frequency parameters for the supervisory plan, and acceptance requirements for concealed works. The clause mapping relationship is organized into a set of triples, each triple containing a specification clause node identifier, a process node identifier, and a competency item node identifier. Each directed edge of the process-dependent directed graph is appended with a process importance score. With respect to the maximum exposure time parameter allowed by the process, For the first The importance of each process step The process number is used; the importance of the process is determined as follows: the acceptance level in the concealed works acceptance requirements is coded and converted into a numerical level value; the inspection frequency parameter of the supervision plan is normalized and mapped to the same value range as the level value; then the level value and the normalized inspection frequency parameter are combined according to a preset ratio to form a normalized value. interval .

[0009] For example, in the knowledge graph-based method for constructing a competency profile for supervisory positions provided in at least one embodiment of this disclosure, the state set of the process state machine includes four active states: pending inspection, inspected and awaiting inspection, on-site supervision in progress, and pending closed verification, as well as one terminated state: closed. The remaining inspectable time output by the process state machine includes: extracting the set of component numbers to which the current component-level process event flow belongs, and reading the current process state of the corresponding component from the active component subgraph of the edge node; determining the next process state based on the inspection event, on-site check-in event, image upload event, sensor data threshold event, and coverage process start event; calculating the margin according to the planned start time of the coverage process, the on-site pouring instruction for the coverage material, the maximum exposure time allowed by the process, and the first exposure time, and taking the minimum value among the three. Update components as the lower limit Remaining detectable time , For component identification, For components The remaining inspectable time; the process state is transferred to the closed component, which is removed from the active component subgraph and replaced by a boundary node. Components that do not appear in the component number set keep their process state and remaining inspectable time unchanged.

[0010] For example, in the knowledge graph-based method for constructing a competency profile for supervisory positions provided in at least one embodiment of this disclosure, the verification delay budget is calculated using the following formula:

[0011]

[0012] in, For the first Frame acceptance image verification delay budget, The frame number for the acceptance image. For the first The identifier of the component to which the frame image belongs. For components The remaining detectable time, For components The current number of image frames to be verified. The allowable occupancy factor for the remaining detectable time and its value range is: to , This is the sequence number of the current scheduling window. For the terminal in the first The total allocatable latency within each scheduling window, calculated from available computing power reserves. For the first The set of images to be verified within a scheduling window For set The frame number of the internal image is used to iterate through all images in the set. and The first Frame and the The importance of the process to which the frame image belongs. This indicates taking the minimum value among the values ​​within the parentheses. Not less than the baseline time for a single frame of depth verification channel The images are assigned to the depth verification channel. Single-frame latency benchmark located in the lightweight check channel and The images in between are assigned to the lightweight verification channel and are eligible to be transferred to the depth verification channel. Less than The images are assigned to lightweight verification channels and downsampled by a fixed short side length before processing. The hierarchical verification scheduling sequence is as follows: Sort in ascending order.

[0013] For example, in the knowledge graph-based method for constructing a competency profile for supervisory positions provided in at least one embodiment of this disclosure, the lightweight verification channel processes the acceptance image using a single-stage target detection network quantized with eight-bit integers. The classification branch of this network outputs the category identifier and construction status category of key components in the process, and the regression branch outputs the bounding box coordinates. The category identifier and bounding box coordinates are normalized to form a spatial layout identifier after being normalized to the location coordinates. The construction status category, together with the grouting pressure and temperature and humidity sensor data aligned to the process time axis, forms the construction status identifier. During the training phase, a weighted combination of classification loss and regression loss is used as the loss function, and the network parameters are updated by a stochastic gradient descent optimization algorithm. The quantization operation is performed after training. The initial judgment value of process completion and the distance to the discrimination boundary are calculated by the following formula:

[0014]

[0015] in, For the first Preliminary assessment of the completion rate of the process for each frame image. This refers to the number of completion status standard items listed in the process technical specifications for this process. The entry number is determined by... retrieve one by one , For the first The first frame of the image The matching value of the completion status criterion is taken and taken. or , For the first The weights of each standard are not less than the weights of all standards. Their sum is , To determine the boundary values ​​and the value range is , To determine the boundary distance.

[0016] For example, in the knowledge graph-based method for constructing a competency profile for supervisory positions provided in at least one embodiment of this disclosure, the boundary distance is determined. Images below the resolution threshold are transferred to the depth verification channel. The depth verification channel processing includes: converting measured geometric quantities from bounding box scale combined with lens focal length and location coordinate calibration parameters; extracting and aggregating fixed-length construction environment feature vectors from the acceptance image pixel data using a convolutional feature extraction network; constraining the feature vector distance relationship between anchor point samples, positive samples from the same scene, and negative samples from different scenes using a triplet loss function, and updating network parameters using the Adam optimization algorithm; and calculating the construction environment feature vectors based on cosine similarity. and Environmental feature similarity between , , The first Frame and the The construction environment feature vector of the frame image, when the maximum environmental feature similarity is not less than the duplication determination threshold, is used to identify scene duplication. Pick Otherwise take , model When below the lower limit of the effective norm Not to be calculated Place Including areas with insufficient image quality; the process completion score is calculated using the following formula:

[0017]

[0018] in, For the first The completion rate of the process for each frame image is scored. The initial value proportion coefficient and the value range is: , This is the set of geometric measurement items for this process. For set The sequence number of the internal measurement item is used to iterate through all measurement items in the set. The number of measurement items, For the first Measured geometric quantities The value to be assigned to this item in the BIM component list. This is the allowable deviation for this item.

[0019] For example, in the knowledge graph-based method for constructing a competency profile for supervisory positions provided in at least one embodiment of this disclosure, the credibility summary vector is a fixed-length vector, and its components include closed components. The last check before closure corresponds to the first Frame image process completion score Scene repetition identifier Normalized closed time The process importance of the process corresponding to this component. Missing source , For component identification, For the corresponding image frame number, For the corresponding process sequence number, Take when no verification conclusion is obtained before the component is closed Otherwise take The confidence summary vector resides at the boundary nodes of the active component subgraph and is decoupled from the original component image and intermediate detection results; the attenuation coefficient and temporal dependency confidence are calculated using the following formula:

[0020]

[0021]

[0022] in, For components The direct successor component identifier, For components To the component Attenuation coefficient of propagation , , The attenuation weights for insufficient completion, repetitive scenarios, and missing sources are respectively, and none of them are less than 1. , The lower limit of the attenuation coefficient and its value range is , This indicates that the value will be truncated to... interval, For components The reliability of the temporal dependency of the corresponding concealed works records. For components The set of direct preceding components and Iterate through the collection. This indicates taking the product of all attenuation coefficients within the set. For the inflow component Furthermore, the number of directed edges whose image capture time or inspection time of the subsequent process is earlier than the closing time of the corresponding preceding component is also earlier. The intensity of the timing conflict penalty and its value range is [value range missing]. .

[0023] For example, in the knowledge graph-based method for constructing a competency profile for supervisory positions provided in at least one embodiment of this disclosure, the content credibility and the sorting key of the re-examination trigger item are calculated using the following formula:

[0024]

[0025] in, For the first The credibility of the content corresponding to the frame image. For its component The reliability of the temporal dependency of the corresponding concealed works records. , , The weights are to be merged and all are not less than 1. The sum of the three is , Sort key For the first The importance of the process to which the frame image belongs. For components The remaining detectable time, To prevent division by zero of positive numbers and to use the same method Same time unit, This indicates that the maximum value of the two values ​​within the parentheses is taken; images processed only by the Light Check Channel are... Substitution and Pick , Below the content credibility threshold and Generate a retest trigger entry for positive components and press Enter. Pushed to on-site mobile devices in descending order. For components that have been zeroed out, generate irreversible, questionable record entries and record the reason for zeroing out, taken from one of the three sources of the closing condition. The content credibility threshold is defined as follows, and the value range is as follows: .

[0026] For example, in the knowledge graph-based method for constructing a competency profile for supervisory positions provided in at least one embodiment of this disclosure, the method further includes: receiving re-examination reported images, re-performing hierarchical verification scheduling and content credibility calculation only on the corresponding components, concatenating the re-examination results with the original records according to component number data, and ensuring that the credibility of the re-examined content is not lower than that of the original records. Furthermore, if the re-examination image capture time falls within the range where the remaining detectable time is positive, the re-examination closure flag is taken. The reliability of the content obtained from the re-examination is still lower than that of the previous one. Furthermore, when the remaining detectable duration is zeroed upon completion of the re-inspection, the original re-inspection trigger entry is rewritten as an irreversible questionable record entry; upon network connectivity restoration, conflict resolution and merging are performed on records of the same component and clause according to the version vector and event timestamp. The conflict resolution and merging includes: comparing the version vectors of both sides component by component to determine the partial order relationship; the components of the version vector consist of the terminal identifier and the local sequence number of the event; if the version vector of one side is not less than the other side in all components, the record of that side is retained and the dominated side is discarded; if the version vectors of both sides do not dominate each other, irreversible questionable record entries take precedence over re-inspection trigger entries, and the re-inspection closure identifier is... The records cover questionable records of the same component and clause, and in other cases, the records are selected to be retained in order of the later event timestamp; the version vector of the retained record is set to the maximum value of each component of the version vectors on both sides, and the re-inspection response delay and re-inspection closure rate of the on-duty personnel are aggregated according to the clause mapping relationship, and the job capability profile is output to distinguish the three types of identifiers: re-inspection closure, no re-inspection within the window and irreversible questionable.

[0027] At least one embodiment of this disclosure also provides a knowledge graph-based system for constructing a competency profile for supervisory positions. The system includes a graph mapping unit, an event flow generation unit, a process status maintenance unit, a hierarchical scheduling unit, a lightweight verification unit, a deep verification unit, a summary propagation unit, a performance increment unit, a re-inspection closure unit, and a cloud merging unit. The graph mapping unit is configured to acquire basic engineering data and output a directed graph of process dependencies, closure conditions, and clause mapping relationships. The event flow generation unit is configured to receive inspection reporting events, on-site check-in events, acceptance image pixel data and EXIF ​​metadata, grouting pressure and temperature / humidity sensor data, and terminal positioning trajectories, and output a component-level process event flow. The process status maintenance unit is configured to output the process status, a set of mandatory inspection nodes, remaining inspectable time, and an active component subgraph from the component-level process event flow. The hierarchical scheduling unit is configured to calculate the verification delay budget based on the remaining inspectable time, process importance, and computing power margin, and output a hierarchical verification scheduling sequence; the lightweight verification unit is configured to output the initial judgment value of process completion and the distance to the discrimination boundary, and transfer images with a discrimination boundary distance lower than the resolution threshold to the deep verification unit; the deep verification unit is configured to output the process completion score and scene duplication identifier; the summary propagation unit is configured to generate a credibility summary vector and output the temporal dependency credibility along the process dependency directed graph; the job performance increment unit is configured to fuse content credibility and generate re-inspection trigger entries and irreversible doubtful record entries; the re-inspection closure unit is configured to receive re-inspection reported images and write the re-inspection closure identifier back to the job performance increment unit; the cloud merging unit is configured to perform conflict resolution merging and output a job capability profile when network connectivity is restored.

[0028] The beneficial effects of this invention are:

[0029] 1. By constructing a directed graph of dependencies between preceding and subsequent processes and closed conditions, and combining it with an incremental process state machine to maintain the active component subgraph, the remaining inspectable time and the set of mandatory inspection nodes are dynamically updated with the component-level process event flow. The computational cost of a single update is related to the number of events reported this time but not to the total number of components in the project, which is beneficial for edge nodes to continue processing under weak network or network outage conditions.

[0030] 2. By calculating the verification delay budget based on the remaining inspectable time, process importance, and terminal computing power margin, the acceptance images are assigned to the lightweight verification channel and the deep verification channel. This allows components with sufficient remaining time to go through the deep verification channel, while components that are close to being closed are given a preliminary judgment by the lightweight verification channel. This is beneficial for completing the content-level verification before the components are covered.

[0031] 3. By generating a credibility summary vector of the preceding process and propagating the attenuation coefficient along the process dependency directed graph to the subsequent components, and combining the temporal conflict penalty of the time violation dependency relationship, the credibility of the subsequent process record is jointly constrained by the verification results of the preceding process and the temporal compliance, which is conducive to identifying the performance records of the preceding process that are questionable or have misaligned timing.

[0032] 4. By integrating process completion scores, scene duplication identification, and temporal dependency credibility, content credibility is generated, and re-inspection triggered items and irreversible questionable record items are distinguished accordingly. This enables the job competency profile to separate and present three types of situations: re-inspection closure, non-re-inspection within the window, and irreversible questionable, thus avoiding the obscuring of performance evidence with different levels of credibility by the same image upload volume index.

[0033] 5. By using version vectors and event timestamps to resolve and merge conflicts, the performance evidence for the same component and the same clause is not counted repeatedly when the network is disconnected and reconnected or when multiple terminals report repeatedly, which helps to improve the consistency of the job competency profile aggregation results. Attached Figure Description

[0034] Figure 1 This is a flowchart of the method for constructing a competency profile for supervisory positions based on knowledge graphs, provided in an embodiment of the present invention.

[0035] Figure 2 This is a schematic diagram of the component process state machine transition diagram provided in an embodiment of the present invention;

[0036] Figure 3 This is a schematic diagram comparing the importance of key component processes provided in the embodiments of the present invention;

[0037] Figure 4 This is a schematic diagram of the image frame verification delay budget and channel allocation provided in the embodiments of the present invention;

[0038] Figure 5 This is a schematic diagram of the initial judgment value of the lightweight verification channel completion and the distance to the judgment boundary provided in the embodiment of the present invention;

[0039] Figure 6 This is a schematic diagram of the geometric measurement deviation ratio of the depth verification channel provided in an embodiment of the present invention;

[0040] Figure 7 This is a schematic diagram of the content credibility fusion result and judgment distribution provided in the embodiments of the present invention;

[0041] Figure 8 This is a schematic diagram illustrating the contribution decomposition of personnel M-23 in the capability item CAP-grouting on-site control provided in an embodiment of the present invention. Detailed Implementation

[0042] like Figure 1 As shown, at least one embodiment of this disclosure provides a method for constructing a knowledge graph-based capability profile for supervisory positions, including the following steps S100~S1000. This method is applied to a supervisory business platform at a prefabricated building construction site. The platform consists of edge nodes deployed on-site and cloud nodes deployed in a computer room. Edge nodes independently complete image content verification and credibility propagation under weak network or network outage conditions, while cloud nodes complete record merging and capability item aggregation after network connectivity is restored.

[0043] Step S100: Obtain basic engineering data and generate the clause mapping relationship of the job competency knowledge graph.

[0044] In step S100, the BIM component list, inspection batch division table, supervision planning inspection frequency parameters and concealed works acceptance requirements are obtained. The directed graph of the dependencies between the preceding and subsequent processes of each component and the closure condition of each concealed process being covered by the next process are calculated. The data are then connected with the specification clause nodes and capability item nodes to generate the clause mapping relationship of the job capability knowledge graph.

[0045] For example, the component number, component type, elevation, and location coordinates in the BIM component list are combined with the inspection batch number and the set of components contained in the inspection batch division table to generate a sequence of process nodes for each component. Directed edges are then connected between process nodes according to the process sequence relationship, such as grouting of sleeves before sealing of composite slab joints and laying of pre-embedded pipelines before pouring of composite layers, forming a directed graph of pre- and post-process dependencies. The mandatory inspection nodes for each process are determined by the inspection frequency parameters of the supervision plan, including on-site supervision before inspection, on-site supervision during grouting, and actual measurement before closure. The closure conditions are determined by the requirements for acceptance of concealed works. The closure conditions are expressed as the first of the following three conditions to be met: the commencement event of the subsequent covering process, the instruction for pouring the covering material, or the longest exposure time allowed by the process.

[0046] In at least one example of the embodiments of this disclosure, the clause mapping relationship is organized into a set of triples, each triple containing a specification clause node identifier, a process node identifier, and a capability item node identifier, so that the performance evidence generated on the same component and the same process can be uniquely aggregated to the corresponding capability item along the triple, instead of relying on the image upload volume attribute written monthly or quarterly on the personnel node.

[0047] It should be noted that the importance of each process is assigned to each directed edge in the directed graph that determines the preceding and subsequent processes. And the maximum exposure duration parameter allowed by the process, where For the first The importance of each process step, indicated by the subscript. This refers to the process sequence number; the importance of the process is determined jointly by the inspection frequency parameter in the supervision plan and the acceptance level in the concealed works acceptance requirements. The acceptance level is categorical data, first coded and converted into a numerical level value. The inspection frequency parameter is normalized and mapped to the same value range as this level value, and then synthesized into a normalized value according to a preset ratio. Process importance in the interval To avoid frequency parameters and grade categories from directly participating in the synthesis calculation when their dimensions are inconsistent; process importance. The value of is consistently used in the delay budget calculation in step S400, the composition of the credibility summary vector of the preceding process in step S700, and the re-inspection sorting in step S800.

[0048] Step S200: Receive multi-source data reported by the mobile terminal and generate a component-level process event flow.

[0049] In step S200, the inspection report event, the on-site check-in event, the acceptance image pixel data and EXIF ​​metadata, the grouting pressure and temperature and humidity sensor data, and the terminal positioning trajectory reported by the mobile terminal are received, and connected according to the component number and the location coordinate data to generate a component-level process event flow.

[0050] For example, the inspection reporting event provides the component number, process identifier, inspection time, and on-duty personnel identifier. The inspection reporting event for the next process is recorded as a "covering process commencement event" from the perspective of its preceding process. The on-site pouring instruction for the covering material is reported along with the material status field carried in the inspection reporting event. The on-site check-in event is reported by the terminal when it completes on-site check-in at the pre-inspection, grouting process, or pre-closure measured inspection node, and includes an inspection node type identifier to distinguish these three types. The moment the acceptance image pixel data reaches the edge node is recorded as an image upload event. The moment the grouting pressure or temperature and humidity sensor data exceeds a preset threshold is recorded as a sensor data threshold event. XIF metadata provides the shooting time, lens focal length, and geographic coordinates, which are used to verify whether the image shooting time falls between the inspection reporting event and the time when the closure condition is met. Grouting pressure and temperature and humidity sensor data are aligned to the process timeline according to the sampling time and are used as circumstantial evidence for the construction status indicator in step S500. For example, if the duration of the grouting pressure plateau section is insufficient, the construction status indicator is set to the state of not reaching stable pressure. The terminal positioning trajectory and EXIF ​​geographic coordinates jointly participate in the connection of part coordinate data. When the deviation between the two exceeds the expansion tolerance of the component's outer contour, the image is incorporated into the candidate set of duplicate scene indicators in step S600. The aforementioned inspection reporting event, on-site check-in event, image upload event, sensor data threshold event, and coverage process start-up event together constitute the state transition trigger event of the incremental process state machine in step S300.

[0051] In at least one example of the embodiments of this disclosure, each event in the component-level process event stream is assigned an event timestamp and a terminal identifier, and carries a monotonically increasing local sequence number, which serves as a component source in the version vector merging in step S1000.

[0052] Step S300: The incremental process state machine outputs the set of inspection nodes and the remaining inspection time of the active components.

[0053] In step S300, the component-level process event flow is input into the incremental process state machine, the active component subgraph is loaded at the edge nodes, and the current process state of each component, the set of forced check nodes, and the remaining inspectable time before the closure condition is met are output.

[0054] The incremental process state machine's state set includes four active states: pending inspection, submitted for inspection and awaiting inspection, in progress under on-site supervision, and pending closed verification, as well as one terminated state: closed. Its processing includes:

[0055] a) Receive the current component-level process event flow and extract the set of component numbers to which the event belongs;

[0056] b) For each component in the component number set The current process status of the component is read from the active component subgraph of the edge node, where For component identification;

[0057] c) Based on the inspection report event, on-site check-in event, image upload event, sensor data threshold event, and coverage process commencement event, determine the next process status of the component according to the following conditions: Components in the pending inspection state will be transferred to the "inspected but awaiting inspection" state after receiving the inspection report event, and the on-site check-in event of the pre-inspection on-site inspection node associated with the inspection report event will be recorded as accompanying evidence; Components in the "inspected but awaiting inspection" state will be transferred to the "on-site inspection in progress" state after receiving the on-site check-in event of the grouting process on-site inspection node; Components in the "on-site inspection in progress" state will be transferred to the "awaiting closure verification" state after the sensor data threshold event indicates that the grouting pressure platform section has reached the pressure stabilization standard and an image upload event is received; Components in the "awaiting closure verification" state will be transferred to the "closed" state after meeting the closure conditions (i.e., receiving the coverage process commencement event, receiving the material status update corresponding to the on-site pouring instruction for the coverage material, or the expiration of the longest exposure time allowed by the process) and receiving the on-site check-in event of the pre-closure measured inspection node.

[0058] d) Calculate the remaining inspectable time based on the three sources of the closure conditions, and take the minimum value among the following: the margin calculated from the planned start time of the covering process, the margin calculated from the on-site pouring instruction of the covering material, and the margin calculated from the maximum allowable exposure time and the first exposure time (the time of the inspection event corresponding to the component). As the lower limit, update the remaining inspectable duration of the component, denoted as . ,in For components The remaining detectable time, i.e., the minimum value is less than hour Pick This enables subsequent steps to use The criterion for determining that the remaining detectable time has been exhausted by zeroing remains numerically valid;

[0059] e) When the process state of the component changes to closed, remove the component from the active component subgraph and replace its adjacent position in the subgraph with a boundary node.

[0060] f) For components that do not appear in the component number set, keep their process status and remaining inspectable time unchanged, and do not participate in this update, so that the calculation amount of a single update is proportional to the number of events reported this time and is independent of the total number of components in the project.

[0061] Step S400: Calculate the verification delay budget and generate a hierarchical verification scheduling sequence.

[0062] In step S400, based on the remaining testable time, process importance and the current available computing power of the terminal, the verification delay budget for each frame of acceptance image is calculated, and the images are assigned to the lightweight verification channel and the deep verification channel according to the budget value, generating a hierarchical verification scheduling sequence.

[0063] For example, the The frame acceptance image verification delay budget is calculated using the following formula:

[0064]

[0065] in, For the first Frame image verification delay budget, The frame number for the acceptance image. For the first The identifier of the component to which the frame image belongs. For components The remaining detectable time, For components The current number of image frames to be verified. The allowable occupancy factor for the remaining detectable time and its value range is: to , This is the sequence number of the current scheduling window. For the terminal in the first The total allocatable latency within each scheduling window, calculated from available computing power reserves. For the first The set of images to be verified within a scheduling window For set The frame number of the internal image is used to iterate through all images in the set. and The first Frame image and the first The process number of the process to which the frame image belongs. and The first Frame image and the first The importance of the process to which the frame image belongs. This indicates taking the minimum value among the values ​​within the parentheses.

[0066] In at least one example of the embodiments of this disclosure, the assignment rule is: when Not less than the depth channel single frame latency benchmark At that time, the image was directly assigned to the depth verification channel; when Located at the lightweight channel single frame latency benchmark and In between, the image is first assigned to the lightweight verification channel and retains its eligibility to be transferred to the depth verification channel; when Less than At that time, the image was assigned to a lightweight calibration channel and downsampled by a fixed short side length before processing. This serves as a benchmark for the time consumed in processing a single frame of image in the depth verification channel. This serves as a time benchmark for processing a single frame of image data in the lightweight verification channel. The hierarchical verification scheduling sequence is based on... Arrange them in ascending order so that components that are close to the enclosure are given priority in obtaining the initial judgment.

[0067] It should be noted that, due to Meanwhile, constrained by the remaining testable time and computing power margin, components with ample remaining time go through the deep verification channel, while components nearing closure are given an initial judgment value by the lightweight verification channel. The content-level verification conclusion is constrained to be generated before the component is covered.

[0068] Step S500: Output the initial judgment value of process completion and the judgment boundary distance in the lightweight verification channel.

[0069] In step S500, the acceptance image is processed using the parameter-quantified target detection model in the lightweight verification channel. The spatial layout identifiers and construction status identifiers of key components of the process are extracted and matched with the completion status standards in the process technical specifications. The initial judgment value of the process completion degree and the judgment boundary distance are output.

[0070] The quantized target detection model is a single-stage target detection network using 8-bit integer quantization. Its input layer receives pixel data from acceptance images, and its output layer includes a classification branch and a regression branch. The classification branch is a fully connected layer that outputs the category identifiers and construction status categories of key components in the process. The regression branch is also a fully connected layer that outputs the bounding box coordinates. The category identifiers and bounding box coordinates are normalized to form spatial layout identifiers after being normalized to location coordinates. The construction status category, along with the grouting pressure and temperature / humidity sensor data aligned in step S200, together form the construction status identifier. During the training phase, the acceptance image dataset labeled with component categories, bounding box coordinates, and construction status categories is used as training samples. A weighted combination of classification and regression losses is used as the loss function. The classification loss constrains the prediction results of component category identifiers and construction status categories, while the regression loss constrains the prediction results of bounding box coordinates. A stochastic gradient descent optimization algorithm is used to update the network parameters. Quantization is performed after training to keep the accuracy loss within an acceptable range.

[0071] For example, the initial judgment value of process completion and the distance to the judgment boundary are calculated using the following formula:

[0072]

[0073] in, For the first Preliminary assessment of the completion rate of the process for each frame image. The frame number for the acceptance image. This refers to the number of completion status standard items listed in the process technical specifications for this process. The entry sequence number for completing the status standard and by retrieve one by one , For the first The first frame of the image The matching value of the completion status criterion is as follows: or , For the first The weights of each standard are not less than the weights of all standards. Their sum is , To determine the boundary values, the value range is: , For the first Discrimination boundary distance of frame image express and The absolute value of the difference.

[0074] In one example, the completion criteria for the sleeve grouting process include three aspects: the grout outlet is properly sealed, the overflow height of the grout reaches the standard, and the pressure stabilization time reaches the standard. In another example, the completion criteria for the composite slab joint process include three aspects: the joint is clean, the additional reinforcing bars are in the correct position, and the support system has not been removed.

[0075] Step S600: Output the process completion score and scene duplication flag in the depth verification channel.

[0076] In step S600, the boundary distance will be determined. Below the resolution threshold (the resolution threshold range is...) The images are transferred to the depth verification channel to extract the geometric measurement features of the components and the features of the construction environment. The similarity of environmental features between different images of the same project is calculated, and the completion score of the process and the scene duplication mark are output.

[0077] For example, the geometric measurement features of components are converted into measured geometric quantities by combining the bounding box dimensions with lens focal length and location coordinate calibration parameters, and then compared item by item with the design values ​​in the BIM component list; construction environment features are extracted and aggregated from the pixel data of acceptance images by a convolutional feature extraction network. 3D construction environment feature vector, where The dimension of the construction environment feature vector is given. The input layer of the convolutional feature extraction network receives pixel data from the acceptance image, and the output layer is a fully connected layer with a fixed length. The dimension of the construction environment feature vector is 1.5. All images entering the depth verification channel within the same project have a construction environment feature vector of dimension 1. During the training phase, a triplet loss function is used as the loss function. The network parameters are constrained by the distance relationship between the feature vectors of anchor samples, positive samples from the same scene, and negative samples from different scenes, ensuring that the feature vector distance between image pairs from the same scene is less than that between image pairs from different scenes. The Adam optimization algorithm is used to update the network parameters. Environmental feature similarity is calculated using cosine similarity, with the input being the corresponding features of the two frames to be compared. 3D construction environment feature vector and ,in For the first Construction environment feature vector of frame image, For the first Construction environment feature vector of frame image, and Both are frame numbers of the acceptance images, and the output is the environmental feature similarity between the two. .

[0078] The process completion score is calculated using the following formula:

[0079]

[0080] in, For the first The completion rate of the process for each frame image is scored. For the first Preliminary assessment of the completion rate of the process for each frame image. The initial value proportion coefficient and the value range is: , This is the set of geometric measurement items for this process. For set The sequence number of the internal geometry measurement item is used to iterate through all measurement items in the set. For set The number of internal geometric measurements For the first Measured geometric quantities Its design value, For its allowable deviation, It represents the absolute value of the difference between the measured geometric quantity and the design value. This indicates taking the minimum value among the two values ​​within the parentheses.

[0081] In at least one example of the embodiments of this disclosure, when the first The maximum environmental feature similarity between a frame image and images with different component numbers or different inspection events within the same project is not less than the duplicate judgment threshold (the value range of the duplicate judgment threshold is...). When the scene is repeatedly identified, Pick Otherwise take ,in For the first Scene repetition markers in frame images.

[0082] It should be noted that the effective norm lower bound is set based on the norm distribution of the construction environment feature vectors of the normal construction environment image samples during the training phase; when When the image falls below the lower limit of the effective norm, meaning it contains no valid construction environment information due to underexposure, overexposure, or lens obstruction, Not included in the calculation. Place In addition, an image quality deficiency bit is added. The component to which this image belongs directly enters the determination branch of the re-examination trigger entry or the irreversible questionable record entry in step S800, to avoid the similarity losing its discriminative meaning when the vector degrades. For the first Construction environment feature vector of frame image The model.

[0083] Step S700: Generate a summary vector of the credibility of the preceding process and output the credibility of the temporal dependency.

[0084] In step S700, the completion score, scene repetition identifier, and closing time of the component that has achieved the closure condition before being removed from the active component subgraph are read, compressed into a previous process credibility summary vector, and resided at the boundary node of the active component subgraph; along the previous process dependency directed graph, the attenuation coefficient propagating to the subsequent component node is calculated based on the value of the previous process credibility summary vector, and a time conflict penalty is applied to the node whose time violates the dependency relationship, and the time dependency credibility of each hidden project record is output.

[0085] The process of generating the credibility summary vector of the preceding process and calculating it by propagating it along the directed graph of dependencies between preceding and preceding processes includes:

[0086] a) For components that have achieved closure conditions. Read the process completion score, scenario duplication identifier, and closure time obtained from the last verification before closure, and generate a fixed-length five-dimensional credibility summary vector of the preceding process. ,in For components The credibility summary vector of the preceding process. The component is identified and the closure condition has been met. For components The sequence number of the acceptance image frame corresponding to the last verification before closure. For the first The completion rate of the process for each frame image is scored. For the first Scene repetition markers in frame images For the normalized closed moment, For components The corresponding process number, The importance of this process is determined by its process order. When the source is missing and no verification results were obtained before the component was closed, take Otherwise take Furthermore, the credibility summary vector of the preceding process is retained at the boundary nodes of the active component subgraph, decoupling it from the original component image and intermediate detection results;

[0087] b) Read the components along the directed graph of dependencies between preceding and subsequent processes. direct successor components ,in For components The direct successor component identifier;

[0088] c) Based on the credibility summary vector of the preceding process The value of is calculated by the following formula from the component. Towards successor components Attenuation coefficient of propagation:

[0089]

[0090] in, For components Towards successor components Attenuation coefficient of propagation For components The last check before closure corresponds to the first The completion rate of the process for each frame image is scored. This serves as a scene repetition marker for the current frame of the image. For components The source is missing. , , The attenuation weights for insufficient completion, repetitive scenarios, and missing sources are respectively, and none of them are less than 1. , The lower limit of the attenuation coefficient and its value range is , This indicates that the value will be truncated to... Interval; Pre-process credibility summary vector In The subsequent step d) is used to determine whether the annotation time violates the dependency relationship. Along with the other four components, they reside at the boundary node and are transferred to step S1000 as a whole with the performance record. This allows the cloud node to confirm the importance level of the corresponding process when aggregating capability items according to the clause mapping relationship. They do not participate in the numerical calculation of the attenuation coefficient and the time-series dependency reliability in this step. The attenuation coefficient and the time-series dependency reliability only reflect the reliability of the verification result of the preceding component itself.

[0091] d) Statistical inflow components And the number of directed edges that violate dependency rules in the annotation time. That is, the time of image capture or inspection of the subsequent process is earlier than the time of closure of the corresponding preceding component (and) The number of edge chains (values ​​before normalization at the same time), where For the inflow component Furthermore, the number of directed edges that violate dependency rules during the annotation period;

[0092] e) Calculate the components according to the following formula Reliability of temporal dependencies corresponding to concealed works records:

[0093]

[0094] in, For components The reliability of the temporal dependency of the corresponding concealed works records. For components The set of direct preceding components on the directed graph that the preceding and following processes depend on. For set The direct predecessor of the set is traversed, and all components in the set are traversed. Represents a set Internal components Corresponding attenuation coefficient Take the product of all products. For the inflow component Furthermore, the number of directed edges that violate dependency rules in the annotation time section is also mentioned. The intensity of the timing conflict penalty and its value range is [value range missing]. .

[0095] It should be noted that if a preceding component fails to obtain a verification result before closure, that component will not have a corresponding frame number. Its pre-process credibility summary vector is based on , Take the specified default lower limit, Pick Participate in dissemination in the form of [unclear], and The computational link remains intact, so the decay constraint of the preceding questionable component on the subsequent component does not fail when the edge node only loads the active component subgraph and is offline.

[0096] Step S800: Integrate content credibility and generate re-examination trigger entries and irreversible questionable record entries.

[0097] In step S800, the process completion score, scene duplication identifier, and temporal dependency credibility are weighted and fused into a content credibility score. For content credibility scores below the content credibility threshold... Furthermore, for components with a positive remaining inspectable time, re-inspection trigger entries are generated and sorted according to the remaining inspectable time and process importance. For components with a content credibility lower than [a certain threshold], [further details are needed]. Furthermore, for components with remaining inspectable duration reduced to zero, irreversible questionable record entries are generated and written into the edge node performance increment table. The content credibility threshold is defined as follows, and the value range is as follows: .

[0098] For example, content credibility is calculated using the following formula:

[0099]

[0100] in, For the first The credibility of the content corresponding to the frame image. The frame number for the acceptance image. For the first The completion rate of the process for each frame image is scored. For the first Scene repetition markers in frame images For the first The identifier of the component to which the frame image belongs. For components The reliability of the temporal dependency of the corresponding concealed works records. , , The weights are to be merged and all are not less than 1. The sum of the three is For images that have only passed through the light verification channel and have not been transferred to the depth verification channel, by Substitute and Pick .

[0101] In at least one example of the embodiments of this disclosure, the sorting key of the re-detection trigger entry is calculated using the following formula:

[0102]

[0103] in, For the first The sorting key for the re-detection trigger entry corresponding to the frame image. For the first The process number of the process to which the frame image belongs. The importance of this process is determined by its process order. For the first Components to which the frame image belongs The remaining detectable time, To prevent division by zero of positive numbers and to use AND... Same time unit, This indicates taking the maximum value of the two values ​​within the parentheses; the re-detection trigger entry is... Sort in descending order and push to the on-site mobile device; when Below the content credibility threshold and When the value has been reduced to zero, the component generates an irreversible questionable record entry, which records the reason for the zeroing from one of the three sources of the closing condition.

[0104] Step S900: Receive the re-inspection reported image and output the re-inspection closure identifier.

[0105] In step S900, the re-inspection reported image is received. Only for that component, the hierarchical verification scheduling and content reliability calculation from steps S400 to S800 are re-executed. The re-inspection result is concatenated with the original record according to the component number, and a re-inspection closure flag is output. When the content reliability obtained from the re-inspection is not lower than... Furthermore, if the re-examination image capture time still falls within the range where the remaining detectable time is positive, the re-examination closure marker is taken. When the credibility of the content obtained from the re-examination is still lower than ; Furthermore, when the remaining detectable time reaches zero upon completion of the re-inspection, the original re-inspection trigger entry is rewritten as an irreversible questionable record entry. The re-inspection response delay is the difference between the capture time of the re-inspection reported image and the generation time of the corresponding re-inspection trigger entry.

[0106] Step S1000: Perform conflict resolution and merging and output job competency profiles.

[0107] In step S1000, when network connectivity is restored, the re-inspection trigger entries, irreversible questionable record entries, content credibility values ​​and previous process credibility summary vectors in the job performance increment table are merged according to version vectors, event timestamps and records of the same component and clause in the job competency knowledge graph to perform conflict resolution. The contribution of the job performance record to the competency item is weighted by the content credibility. The re-inspection response delay and re-inspection closure rate within the window of the on-duty personnel are aggregated according to the clause mapping relationship. The job competency profile is output, which distinguishes between three types of identifiers: re-inspection closure, no re-inspection within the window and irreversible questionable.

[0108] The conflict resolution and merging process includes: a) taking the version vectors of the records on both sides to be merged, comparing them component by component to determine the partial order relationship. The components of the version vector are composed of the terminal identifier and the local sequence number in step S200; b) when the version vector of one side is not less than that of the other side in all components, retaining the record of that side and discarding the dominated side; c) when the version vectors of both sides are not dominated by each other, prioritizing irreversible questionable record entries over re-examination trigger entries, and using the re-examination closure identifier as... The records cover questionable records of the same component and clause; in other cases, the records are selected to be retained in order of the later timestamp of the event; d) the version vector of the retained record is set to the maximum value of each component of the version vectors on both sides, so that the performance evidence of the same component and clause is not counted repeatedly due to network disconnection and reconnection or repeated reporting by multiple terminals.

[0109] For example, personnel In the ability category The weighted contribution and in-window re-detection closure rate are calculated using the following formula:

[0110]

[0111] in, For personnel In the ability category The weighted contribution value on, For identification of on-duty personnel, For capability item identification, Aggregated to capability items through clause mapping relationships And the staff on duty are A collection of job performance records, For set The sequence number of the internal performance record is used to iterate through all records in the set. For the first The credibility of the content corresponding to each performance record is taken, along with the credibility of the content of the acceptance image frame on which the record is based. For personnel Re-examination closure mark under name The collection of records, For personnel A collection of records under your name that have generated re-check trigger entries. and Sets With sets The number of internal records, For personnel The in-window re-detection closure rate. When Time, i.e., personnel No re-check trigger entries have been generated under this name. This person will not be included in the calculation of the re-examination closure rate within this round of windowing, and this indicator will not be included in this round of profile update.

[0112] In at least one example of the embodiments of this disclosure, three types of identifiers are written side by side on each competency item of the job competency profile: re-inspection closure, no re-inspection within the window, and irreversible doubt, so that the values ​​of doubtful items that have been re-shot and closed and doubtful items that cannot be verified permanently are separated on the competency profile and are no longer covered by the same image upload volume index.

[0113] The knowledge graph-based supervisory post competency profile construction system provided in at least one embodiment of this disclosure includes at least one edge node, a graph mapping unit, an event flow generation unit, a process status maintenance unit, a hierarchical scheduling unit, a lightweight verification unit, a deep verification unit, a summary propagation unit, a performance increment unit, a re-inspection closure unit, and a cloud merging unit. The graph mapping unit acquires the BIM component list, inspection batch division table, supervision planning inspection frequency parameters, and concealed works acceptance requirements, and outputs a directed graph of pre- and post-process dependencies, closure conditions, and clause mapping relationships; the event flow generation unit receives inspection events, acceptance image pixel data and EXIF ​​metadata, grouting pressure and temperature and humidity sensor data, and terminal positioning trajectory, and outputs the component-level process event flow to the process status maintenance unit; the process status maintenance unit outputs the remaining inspectable time, the set of mandatory inspection nodes, and the active component subgraph to the hierarchical scheduling unit, and outputs closure events to the summary propagation unit; the hierarchical scheduling unit outputs the hierarchical verification scheduling sequence to the lightweight verification unit and the deep verification unit; the lightweight verification unit transfers images with a boundary distance lower than the resolution threshold to the deep verification unit; the deep verification unit and the summary propagation unit respectively output the process completion score, scene duplication identifier, and temporal dependency credibility to the performance increment unit; the performance increment unit outputs re-inspection trigger entries to the re-inspection closure unit, and the re-inspection closure unit writes the re-inspection closure identifier back to the performance increment unit; the cloud merging unit reads the performance increment table and outputs the job capability profile when network connectivity is restored.

[0114] For a detailed explanation of the graph mapping unit, please refer to the relevant description of step S100 in the embodiments of the above method; for a detailed explanation of the event stream generation unit, please refer to the relevant description of step S200; for a detailed explanation of the process status maintenance unit, please refer to the relevant description of step S300; for a detailed explanation of the hierarchical scheduling unit, please refer to the relevant description of step S400; for a detailed explanation of the lightweight verification unit, please refer to the relevant description of step S500; for a detailed explanation of the deep verification unit, please refer to the relevant description of step S600; for a detailed explanation of the digest propagation unit, please refer to the relevant description of step S700; for a detailed explanation of the duty increment unit, please refer to the relevant description of step S800; for a detailed explanation of the re-check closure unit, please refer to the relevant description of step S900; and for a detailed explanation of the cloud merging unit, please refer to the relevant description of step S1000. These details will not be repeated here.

[0115] At least one embodiment of this disclosure also provides an electronic device and a non-transitory computer-readable storage medium. The electronic device includes a processor and a memory. The memory stores computer-executable instructions, which, when executed by the processor, implement steps S100 to S1000 of the above-described method. The storage medium stores corresponding computer-executable instructions. The technical effects of this system, electronic device, and storage medium are the same as those of the above-described method for constructing a competency profile of supervisory positions based on knowledge graphs, and will not be repeated here.

[0116] A specific application example is shown below, such as Figure 2-8 The corresponding chart is shown below;

[0117] Step S100: Obtain basic engineering data and generate clause mapping relationships for the job competency knowledge graph.

[0118] Based on the BIM component list, inspection batch division table, supervision plan inspection frequency parameters, and concealed works acceptance requirements, the dependent directed graph and closure conditions of the preceding and subsequent processes are calculated, and connected with the data of specification clause nodes and capability item nodes to generate clause mapping relationships; at the same time, the process importance is synthesized according to the acceptance level code value and the normalized value of inspection frequency.

[0119] This example scenario is a prefabricated residential building project (codename "Project C"). The supervision business platform consists of an edge node (number SI-04) in the on-site computer room and a cloud node in the headquarters computer room. The project is currently under construction up to the third floor. Inspection batch JYP-07 includes precast column PC-101, composite slab PB-205, and adjacent components PC-108, PC-095, and PB-210. The grouting process of PC-101 precedes the joint sealing process of the composite slab PB-205. The process sequence is expressed as a directed edge from PC-101 to PB-205. In the acceptance level, the grouting process is marked as a critical process with a code value of 0.90, and the joint sealing process is marked as a general critical process with a code value of 0.70. In the supervision plan, the grouting inspection frequency is 3 times per week, normalized to 0.75, and the joint sealing inspection frequency is 2 times per week, normalized to 0.50. The composite proportion coefficient is 0.60. According to the formula... (in For the first The importance of each process step This is the normalized value for the acceptance level code. To check the frequency normalization value, Substituting the proportion coefficient (for synthesis) into the equation, the grouting process yields... The joint sealing process must be completed. .

[0120] Table 1. Mapping Relationship between Basic Engineering Data and Clauses

[0121]

[0122] Step S200: Receive multi-source data reported by the mobile terminal and generate a component-level process event flow.

[0123] The inspection events, on-site check-in events, image pixel data and EXIF ​​metadata, grouting pressure and temperature and humidity sensor data, and terminal positioning trajectory reported by the mobile terminal are connected according to the component number and location coordinate data to generate a component-level process event flow.

[0124] For PC-101, the on-site supervisor (personnel identification M-23) uses terminal T-08 to complete the inspection before grouting. Subsequently, four types of events are triggered in sequence: on-site check-in, sensor threshold, and image upload. Each event carries a monotonically increasing local sequence number for use in subsequent version vector merging.

[0125] Table 2 PC-101 Component-Level Process Event Flow

[0126]

[0127] Step S300: The incremental process state machine outputs the set of inspection nodes for active components and the remaining inspectable time.

[0128] Input the component-level process event flow into the incremental process state machine, load the active component subgraph, and output the current process state, the set of mandatory check nodes, and the remaining checkable time. .

[0129] According to the event sequence in Table 2, PC-101 changed from "in progress" to "pending closure verification"; PB-205 is in "reported for inspection and awaiting testing"; and PC-108 is in "in progress". The remaining inspection time for the three is determined by the minimum of the following three values: the allowance for start-up of the coverage process, the allowance for pouring instructions, and the allowance for the longest exposure time.

[0130] Table 3. Set of active component inspection nodes and remaining inspection time

[0131]

[0132] Step S400: Calculate the verification delay budget and generate a hierarchical verification scheduling sequence.

[0133] Based on the remaining inspectable time, process importance, and available computing power at the terminal, according to the formula... Calculate the frame check delay budget and allocate channels.

[0134] The scheduling window contains three frames of video; the occupancy coefficient is taken. Total latency that can be allocated by the terminal The sum of the importance of the processes within the window (from PC-101, PB-205, and PC-108) Adding them together gives 0.84 + 0.62 + 0.84 = 2.30), the depth channel latency benchmark. Lightweight channel latency benchmark .Will , , Substituting, we get Not less than Directly assigned to the deep verification channel; , , Substituting, we get ,lie in and In between, it is assigned to the lightweight verification channel while retaining its eligibility to be transferred to the deep channel; , , Substituting, we get less than It is assigned to a lightweight verification channel and downsampled according to a fixed short side length.

[0135] Table 4 Verification Delay Budget and Channel Allocation

[0136]

[0137] Step S500: Output the initial judgment value of process completion and the judgment boundary distance in the lightweight verification channel.

[0138] Using the parameter-quantified target detection model, the spatial layout and construction status indicators of key components in the process are extracted, and then processed according to the formula. , Output the initial judgment value of process completion and the distance to the judgment boundary, and the value of the judgment boundary. The resolution threshold is set to 0.15.

[0139] Frame F3 (PC-108) corresponds to three completion status standards for the sleeve grouting process, with weights of 0.40, 0.35, and 0.25 respectively. The sealing of the grout outlet and the achievement of the grout overflow height standard are both matched as 1, while the pressure stabilization time standard is matched as 0 (corresponding to the insufficient pressure stabilization duration shown in Table 2's sensor data threshold events). Substituting these values, we get... , If the resolution is below the threshold, it switches to the depth verification channel. All three criteria for frame F1 (PC-101) match to 1, resulting in... , Frame F2 (PB-205) corresponds to three criteria: clean joints, correct placement of additional reinforcing bars, and no removal of the support system. The weights are 0.30, 0.40, and 0.30 respectively. All three criteria match to 1, resulting in... , It will not be transferred to the deep verification channel.

[0140] Table 5 Preliminary results of lightweight verification channel

[0141]

[0142] Step S600: Output the process completion score and scene duplication flag in the depth verification channel.

[0143] For images where the boundary distance is below the resolution threshold, extract the geometric measurement features of the components and the construction environment features, and then apply the formula... Calculate the process completion score and categorize it according to environmental feature similarity. Determine if a scene is duplicated.

[0144] The measured exposed height of the grouting sleeve for frame F3 (PC-108) is 18.00 mm. The marked value is 20.00. Permissible deviation: 3.00 The actual thickness of the slurry outlet plug was 5.00 mm. The labeled value is 5.00. Permissible deviation 1.00 The measured peak pressure of the regulated pressure is 0.28. The labeled value is 0.30. Allowable deviation 0.05 Take the initial judgment value proportion coefficient. Substituting into Based on environmental feature similarity comparison, the image similarity between frame F3 and historical component PC-095 reached 0.91, which is not less than the duplication threshold of 0.85, thus indicating scene duplication. .

[0145] Table 6. Geometric Measurement and Scoring Results of Depth Verification Channel

[0146]

[0147] Frame F3 finally got , .

[0148] Step S700: Generate a confidence summary vector of the preceding process and output the time-dependent confidence.

[0149] For components that have met the closure conditions, read the process completion score, scenario duplication flag, and closure time from the last verification before closure, and compress them into a five-dimensional pre-process credibility summary vector. And remain at the boundary nodes, along the directed graph of dependencies between preceding and subsequent processes according to the formula Calculate the attenuation coefficient, then use the formula Calculate the reliability of time-series dependencies.

[0150] Before PC-101 is closed, the last check of the corresponding frame F1 is performed. ;Pick , , , Substituting into The prerequisite component for PC-108 is the grouting preparation process node, and its abstract vector is taken as... Substituting into Let the intensity of the timing conflict penalty be set. The number of edges that violate dependency rules when flowing into the two links All are 0, substituting them gives , .

[0151] Table 7. Credibility Summary Vector of Pre-process and Credibility of Temporal Dependency

[0152]

[0153] Step S800: Integrate content credibility and generate re-examination trigger entries and irreversible questionable record entries.

[0154] According to the formula Calculate content credibility and take the fusion weight. , , Content credibility threshold For frame F3 (PC-108), substitute... , , ,have to lower than And the remaining detectable time Minutes are positive, according to the formula (Pick Substituting (minutes) into the sort key Generate a re-detection trigger entry. For frame F2 (PB-205), substitute... , , ,have to This is a normal record. For historical component PC-095, its remaining testable time has been reduced to zero. Substitute... , , have to lower than and Reset to zero, generating an irreversible questionable record entry. The reason for resetting to zero is "the longest exposure time has expired".

[0155] Table 8 Results of Content Credibility Fusion and Re-examination Trigger / Questionable Items

[0156]

[0157] Step S900: Receive the re-inspection reported image and output the re-inspection closure identifier.

[0158] For the component that generates the re-inspection trigger entry, re-execute hierarchical verification scheduling and content credibility calculation, connect the re-inspection result with the original record according to the component number, and output the re-inspection closure identifier; the re-inspection response delay is the difference between the re-inspection reported image capture time and the re-inspection trigger entry generation time.

[0159] Field terminal receives sorting key After the corresponding re-inspection trigger entry is pushed, personnel M-23 takes a supplementary image of PC-108. The image capture time is 45 seconds later than the entry generation time, meaning the re-inspection response delay is 45 seconds. The re-inspection image is then processed by re-executing S400 to S800, resulting in... , (The reshoots avoided scenes similar to those in PC-095). Substitute have to No less than If the re-examination image capture time still falls within the range where the remaining detectable time is positive (0.50 minutes remaining), the re-examination closure flag is set to 1.

[0160] Step S1000: Perform conflict resolution and merging, and output job competency profiles.

[0161] After network connectivity is restored, the cloud node reads the edge node's duty performance increment table, performs conflict resolution and merging on records of the same component and clause according to version vector and event timestamp, weights the contribution of duty performance records to capability items according to content credibility, and aggregates the re-inspection response latency and re-inspection closure rate of on-duty personnel according to clause mapping relationship.

[0162] The cloud node verifies the version vectors and event timestamps of the PC-101 initial judgment record, PC-108 re-inspection record, and PC-095 questionable record reported by the edge node SI-04. It confirms that the three records belong to different components and there are no duplicate reports of the same clause within the same component. No overwrite operation in the conflict resolution and merging process is triggered, and all three records directly enter the weighted aggregation stage. For personnel M-23 in the capability item CAP-grouting on-site control, the set of performance records aggregated to this capability item through the clause mapping relationship includes PC-101 (…). ), re-inspection of PC-108 after closure ( ) and the PC-095 (which is of questionable irreversibility) According to the formula Substituting, we get The set of records under this person's name that have generated re-detection trigger entries only includes PC-108 (PC-095 was not included in this set because its remaining detectable duration was already zero at the beginning of its generation and it directly entered the irreversible questionable record entry branch). Records with a re-detection closure flag of 1 are also PC-108, according to the formula. Substitution The re-inspection response delay for the only re-inspection record (PC-108) in window M-23 is 45 seconds. After aggregation, the aggregated re-inspection response delay for this personnel within the mapped window of this clause is 45 seconds. The job competency profile includes three types of identifiers in parallel on the competency item CAP-Grouting On-site Control: Re-inspection Closure Count 1 (PC-108), Non-Re-inspection Count within the Window 0, and Irreversible Question Count 1 (PC-095), ensuring that the re-inspected closure and permanently unverifiable results are separated in the personnel's competency values. The aggregated delay value is 45 seconds. The job competency profile includes three types of identifiers in parallel on the competency item CAP-Grouting On-site Control: Re-inspection Closure Count 1 (PC-108), Non-Re-inspection Count within the Window 0, and Irreversible Question Count 1 (PC-095), ensuring that the re-inspected closure and permanently unverifiable results are separated in the personnel's competency values.

[0163] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for constructing a competency profile for supervisory positions based on knowledge graphs, characterized in that, include: Generate a directed graph of process dependencies and mapping relationships between closure conditions and clauses based on the basic engineering data; The component-level process event flow is generated from the reported multi-source data, and the remaining inspectable time is output through the process state machine. Calculate the verification delay budget based on the remaining inspectable time, process importance and computing power margin, and assign the acceptance images to the lightweight and deep verification channels according to the budget, and output the process completion score and scene duplication mark. The attenuation coefficient is calculated along the directed graph using the confidence summary vector of the closed component, and a timing conflict penalty is applied to output the timing dependency confidence. The scoring, identification, and credibility are integrated into content credibility. Re-inspection trigger entries or irreversible questionable record entries are generated according to the remaining detectable time. The job competency profile is output by weighted aggregation according to the mapping relationship of the terms. The latency budget is the minimum value of the time occupancy share and the computing power share.

2. The method according to claim 1, characterized in that, The basic engineering data includes a BIM component list, an inspection batch division table, supervision planning inspection frequency parameters, and concealed works acceptance requirements. The clause mapping relationships are organized into sets of triples, each triple containing a specification clause node identifier, a process node identifier, and a capability item node identifier. Each directed edge of the process-dependent directed graph is appended with a process importance score. With respect to the maximum exposure time parameter allowed by the process, For the first The importance of each process step The process number is used as the process importance number, and the value of the process importance is determined as follows: The acceptance levels in the concealed works acceptance requirements are converted into numerical level values ​​through coding. The frequency parameter of the supervision plan inspection is normalized and mapped to the same value range as the grade value. The grade value and the normalized inspection frequency parameter are combined according to a preset ratio to form a normalized value. interval .

3. The method according to claim 1, characterized in that, The state set of the process state machine includes four active states: pending inspection, reported and pending inspection, in progress under on-site supervision, and pending closed verification, as well as one terminated state: closed. The remaining inspectable time output by the process state machine includes: Extract the set of component numbers to which the current component-level process event flow belongs, and read the current process status of the corresponding component from the active component subgraph of the edge node; The status of the next process is determined based on the inspection report event, the on-site check-in event, the image upload event, the sensor data threshold event, and the start-up event of the coverage process. The margin is calculated based on the planned start time of the covering process, the delivery and pouring instructions of the covering material, the maximum allowable exposure time, and the initial exposure time, respectively. The minimum value among the three is then taken. Update components as the lower limit Remaining detectable time , For component identification, For components The remaining detectable time; When a process state transitions to a closed component, the component is removed from the active component subgraph and its adjacent position is replaced by a boundary node. Components that do not appear in the component number set retain their process state and remaining inspectable time.

4. The method according to claim 3, characterized in that, The verification delay budget is calculated using the following formula: ; in, For the first Frame acceptance image verification delay budget, The frame number for the acceptance image. For the first The identifier of the component to which the frame image belongs. For components The remaining detectable time, For components The current number of image frames to be verified. The allowable occupancy factor for the remaining detectable time and its value range is: to , This is the sequence number of the current scheduling window. For the terminal in the first The total allocatable latency within each scheduling window, calculated from available computing power reserves. For the first The set of images to be verified within a scheduling window For set The frame number of the internal image is used to iterate through all images in the set. and The first Frame and the The importance of the process to which the frame image belongs. This indicates taking the minimum value among the values ​​within the parentheses; in, Not less than the baseline time for a single frame of depth verification channel The images are assigned to the depth verification channel. Single-frame latency benchmark located in the lightweight check channel and The images in between are assigned to the lightweight verification channel and are eligible to be transferred to the depth verification channel. Less than The images are assigned to lightweight verification channels and downsampled by a fixed short side length before processing. The hierarchical verification scheduling sequence is as follows: Sort in ascending order.

5. The method according to claim 4, characterized in that, The lightweight verification channel processes the acceptance image using a single-stage target detection network quantized with eight bits. The network's classification branch outputs the category identifier and construction status category of key components in the process, and the regression branch outputs the bounding box coordinates. The category identifier and bounding box coordinates are normalized to form a spatial layout identifier after being normalized to the location coordinates. The construction status category, together with the grouting pressure and temperature and humidity sensor data aligned to the process time axis, forms the construction status identifier. During the training phase, a weighted combination of classification loss and regression loss is used as the loss function, and the network parameters are updated by a stochastic gradient descent optimization algorithm. The quantization operation is performed after training. The initial judgment value of process completion and the distance to the judgment boundary are calculated using the following formula: ; in, For the first Preliminary assessment of the completion rate of the process for each frame image. This refers to the number of completion status standard items listed in the process technical specifications for this process. The entry number is determined by... retrieve one by one , For the first The first frame of the image The matching value of the completion status criterion is taken and taken. or , For the first The weights of each standard are not less than the weights of all standards. Their sum is , To determine the boundary values ​​and the value range is , To determine the boundary distance.

6. The method according to claim 5, characterized in that, Determine boundary distance Images below the resolution threshold are transferred to the depth verification channel, and the processing of the depth verification channel includes: The measured geometric quantities are calculated by combining the bounding frame dimensions with the lens focal length and the location coordinate calibration parameters. The convolutional feature extraction network extracts and aggregates the pixel data of the acceptance image into a fixed-length construction environment feature vector. The network uses a triplet loss function to constrain the distance relationship between the feature vectors of anchor point samples, positive samples in the same scene and negative samples across scenes, and updates the network parameters using the Adam optimization algorithm. Calculate the construction environment feature vector based on cosine similarity. and Environmental feature similarity between , , The first Frame and the The construction environment feature vector of the frame image, when the maximum environmental feature similarity is not less than the duplication determination threshold, is used to identify scene duplication. Pick Otherwise take , model When below the lower limit of the effective norm Not to be calculated Place And add images with insufficient quality; Calculate the process completion score using the following formula: ; in, For the first The completion rate of the process for each frame image is scored. The initial value proportion coefficient and the value range is: , This is the set of geometric measurement items for this process. For set The sequence number of the internal measurement item is used to iterate through all measurement items in the set. The number of measurement items, For the first Measured geometric quantities The value to be assigned to this item in the BIM component list. This is the allowable deviation for this item.

7. The method according to claim 1, characterized in that, The credibility summary vector is a fixed-length vector, and its components include closed components. The last check before closure corresponds to the first Frame image process completion score Scene repetition identifier Normalized closed time The process importance of the process corresponding to this component. Missing source , For component identification, For the corresponding image frame number, For the corresponding process sequence number, Take when no verification conclusion is obtained before the component is closed Otherwise take The confidence summary vector resides at the boundary nodes of the active component subgraph and is decoupled from the original component image and intermediate detection results; the attenuation coefficient and temporal dependency confidence are calculated using the following formula: ; ; in, For components The direct successor component identifier, For components To the component Attenuation coefficient of propagation , , The attenuation weights for insufficient completion, repetitive scenarios, and missing sources are respectively, and none of them are less than 1. , The lower limit of the attenuation coefficient and its value range is , This indicates that the value will be truncated to... interval, For components The reliability of the temporal dependency of the corresponding concealed works records. For components The set of direct preceding components and Iterate through the collection. This indicates taking the product of all attenuation coefficients within the set. For the inflow component Furthermore, the number of directed edges whose image capture time or inspection time of the subsequent process is earlier than the closing time of the corresponding preceding component is also earlier. The intensity of the timing conflict penalty and its value range is [value range missing]. .

8. The method according to claim 6, characterized in that, The credibility of the content and the sorting key of the re-examination trigger entry are calculated using the following formula: ; ; in, For the first The credibility of the content corresponding to the frame image. For its component The reliability of the temporal dependency of the corresponding concealed works records. , , The weights are to be merged and all are not less than 1. The sum of the three is , Sort key For the first The importance of the process to which the frame image belongs. For components The remaining detectable time, To prevent division by zero of positive numbers and to use the same method Same time unit, This indicates that the maximum value of the two values ​​within the parentheses is taken; images processed only by the Light Check Channel are... Substitution and Pick , Below the content credibility threshold and Generate a retest trigger entry for positive components and press Enter. Pushed to on-site mobile devices in descending order. For components that have been zeroed out, generate irreversible, questionable record entries and record the reason for zeroing out, taken from one of the three sources of the closing condition. The content credibility threshold is defined as follows, and the value range is as follows: .

9. The method according to claim 8, characterized in that, Also includes: Upon receiving the re-inspection reported images, only the corresponding components are re-executed with hierarchical verification scheduling and content reliability calculation. The re-inspection results are then concatenated with the original records according to component number. The content reliability obtained from the re-inspection is no less than [a certain value]. Furthermore, if the re-examination image capture time falls within the range where the remaining detectable time is positive, the re-examination closure flag is taken. The reliability of the content obtained from the re-examination is still lower than that of the previous one. Furthermore, when the remaining detectable duration reaches zero upon completion of the re-detection, the original re-detection trigger entry is rewritten as an irreversible questionable record entry; and When network connectivity is restored, conflict resolution and merging are performed on records of the same component and the same clause according to the version vector and event timestamp, wherein the conflict resolution and merging includes: The partial order relationship is determined by comparing the version vectors on both sides component by component. The components of the version vector consist of the terminal identifier and the local sequence number of the event. If the version vector of one side is not less than that of the other side in all components, the record of that side is retained and the dominated side is discarded. When the version vectors on both sides do not dominate each other, irreversible questionable record entries take precedence over re-examination trigger entries, and the re-examination closure flag is set as follows. The records shall cover any questionable records of the same component and clause; in other cases, the records shall be selected and retained in order of the later timestamp of the event. Set the version vector of the retained record to the maximum value of each component of the version vectors on both sides, and aggregate the re-examination response delay and the re-examination closure rate within the window of the on-duty personnel according to the mapping relationship of the aforementioned clauses, and output a job competency profile that distinguishes between three categories of identifiers: re-examination closure, no re-examination within the window, and irreversible doubt.

10. A knowledge graph-based system for constructing a competency profile for supervisory positions, used to execute the method described in any one of claims 1 to 9, characterized in that, include: The graph mapping unit is configured to acquire basic engineering data and output a directed graph of process dependencies, closure conditions and clause mapping relationships; The event stream generation unit is configured to receive inspection reporting events, on-site check-in events, acceptance image pixel data and EXIF ​​metadata, grouting pressure and temperature and humidity sensor data, terminal positioning trajectory, and output component-level process event streams. The process status maintenance unit is configured to output the process status, the set of forced inspection nodes, the remaining inspectable time, and the active component subgraph from the component-level process event flow. The hierarchical scheduling unit is configured to calculate the verification delay budget based on the remaining inspectable time, process importance, and computing power margin, and output the hierarchical verification scheduling sequence. The lightweight verification unit is configured to output the initial judgment value of process completion and the distance to the discrimination boundary, and to transfer images whose discrimination boundary distance is lower than the resolution threshold to the depth verification unit. The depth verification unit is configured to output a process completion score and a scenario duplication flag. The summary propagation unit is configured to generate a credibility summary vector and output the temporal dependency credibility along the process dependency directed graph; The duty performance increment unit is configured to integrate content credibility and generate re-examination trigger entries and irreversible questionable record entries. The re-inspection closure unit is configured to receive the re-inspection reported image and write the re-inspection closure identifier back to the duty increment unit; as well as The cloud-based merging unit is configured to perform conflict resolution merging and output job competency profiles when network connectivity is restored.

Images