A method and system for installation risk based on video monitoring partition

Abstract

The application relates to the technical field of building construction safety monitoring, and particularly provides a method and system for installing a partition wall based on video monitoring, the method comprising the following steps: acquiring a monitoring video stream of a partition wall installation operation area in real time; constructing a current knowledge graph representing a construction site state based on the video stream; simultaneously, acquiring image interference factor data of the current environment and forming an image interference factor vector; according to the image interference factor data, dynamically selecting a corresponding weight set from a preset weight library and loading the weight set into a lightweight graph neural network; inputting the current knowledge graph into the network loaded with the dynamic weight, and outputting a safety risk prediction value. Through a dynamic weight adaptation mechanism, the model can be self-adapted to environmental changes such as light and height, online training is not needed, the lightness requirement of edge computing and the accuracy of risk identification are considered, and efficient and real-time safety monitoring of the partition wall installation process is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of construction safety monitoring, in particular to a partition board installation risk method and system based on video monitoring. BACKGROUND

[0002] In the field of building decoration, the installation of partition boards is a common but risky operation. Construction workers may fall from a high place, be hit by objects, or cause other safety accidents due to improper operation, incorrect tool use, or unsafe positions. Currently, safety monitoring at construction sites relies on safety inspectors or traditional video monitoring systems. The former is labor-intensive and prone to omissions, while the latter relies mainly on manual review of playback, lacking real-time and intelligent features.

[0003] With the development of computer vision technology, some target detection and behavior recognition methods based on deep learning have emerged, attempting to automatically identify safety hazards. However, these methods face serious challenges in real construction sites: first, the construction site environment is complex and variable, and the image quality captured by the camera is affected by factors such as installation height, lighting conditions, weather, angle, etc., causing the model recognition accuracy to fluctuate dramatically. Second, to achieve high accuracy, advanced models such as graph neural networks usually have large parameter quantities and complex calculations, making it difficult to deploy on edge devices with limited computing resources for real-time response. If a lightweight model is used for real-time performance, it will sacrifice the ability to understand complex scenes and relationships between entities, leading to false positives and false negatives.

[0004] Therefore, there is a fundamental contradiction in the existing technology: complex models cannot meet the real-time requirements of edge devices, while simple models cannot maintain stable accuracy in a variable environment. There is an urgent need for an innovative technical solution that can adapt to environmental changes in a lightweight computing framework and achieve accurate and real-time perception of construction risks. SUMMARY

[0005] To solve the above technical problems, the technical solution adopted by the present application is as follows:

[0006] According to the first aspect of the present application, a partition board installation risk method based on video monitoring is provided, which comprises the following steps:

[0007] Real-time acquisition of monitoring video stream of partition board installation operation area.

[0008] Based on the monitoring video stream, a current knowledge graph representing the state of the construction site is constructed.

[0009] Obtain image interference factor data of the current environment and form an image interference factor vector.

[0010] According to the image interference factor vector, a corresponding weight set is dynamically selected from a preset weight library and loaded into a lightweight graph neural network.

[0011] The current knowledge graph is input into the lightweight graph neural network loaded with the weight set, and a safety risk prediction value of the construction site is output.

[0012] According to the second aspect of the present application, a system for monitoring the installation risk of a partition wall based on a video is provided, and the system comprises:

[0013] A monitoring device is arranged in the partition wall installation operation area to obtain a monitoring video stream in real time.

[0014] A data processing module is configured to construct a current knowledge graph representing the state of the construction site based on the monitoring video stream.

[0015] A sensor module is used to collect image interference factor data of the current environment.

[0016] A data preprocessing module is coupled to the sensor module and is configured to convert the image interference factor data collected by the sensor module into an image interference factor vector.

[0017] A weight management module is coupled to the data preprocessing module and pre-stores a weight library; the weight management module is configured to dynamically select a corresponding weight set from the weight library according to the image interference factor vector and load it into a lightweight graph neural network inference module.

[0018] A lightweight graph neural network inference module is coupled to the data processing module and the weight management module and is configured to input the current knowledge graph as an input and output a safety risk prediction value of the construction site.

[0019] According to the third aspect of the present application, a system for monitoring the installation risk of a partition wall based on a video is provided, and the system comprises:

[0020] A plurality of monitoring devices are arranged at different positions in the cooperative partition wall installation operation area to obtain monitoring video streams from different perspectives in real time.

[0021] A plurality of edge processing units are coupled to one of the monitoring devices and are configured to perform the following operations:

[0022] Construct a local knowledge graph based on the video stream obtained by the monitoring device connected to the edge processing unit.

[0023] Obtain image interference factor data from the perspective of the monitoring device connected to the edge processing unit and form an image interference factor vector.

[0024] According to the image interference factor vector, a weight set is dynamically selected from a pre-stored weight library and loaded into a lightweight graph neural network.

[0025] The local risk prediction value is output by processing the local knowledge graph by using the lightweight graph neural network.

[0026] A central decision module is coupled to all the edge processing units and is configured to fuse all the local risk prediction values to generate a global security risk prediction value.

[0027] The present application introduces a core mechanism of dynamically selecting pre-trained weights based on image interference factors, so that the lightweight graph neural network can adapt to changes in environment such as illumination and viewing angle, effectively solving the inherent contradiction between limited computing resources at the edge and high risk identification accuracy without online training, and ultimately realizing real-time, accurate and robust perception of safety risks in complex construction sites.

[0028] It should be understood that the contents described in this part are not intended to identify key or important features of the embodiments of the present application, nor are they used to limit the scope of the present application. Other features of the present application will become apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0030] Figure 1 A flowchart of a method for installing a partition wall based on video monitoring according to an embodiment of the present application is provided.

[0031] Figure 2 A structural block diagram of a system for installing a partition wall based on video monitoring according to an embodiment of the present application is provided. DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0033] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description herein is for describing particular embodiments only and is not intended to be limiting of the application. As used herein, the term "and / or" includes any and all combinations of one or more of the associated listed items.

[0034] It is to be understood that some of the example embodiments are described in terms of a process or method depicted as a flowchart. Although a flowchart can describe operations as a sequential process, many of the operations can be performed in parallel, concurrently or simultaneously. In addition, the order of the operations can be re-arranged. A process is terminated when its operations are completed. A process can correspond in part to a procedure, function, routine, subroutine, or the like. This description uses "an embodiment," or "one embodiment,” to convey the understanding that a description using these terms indicates that a particular feature, structure, or characteristic is included in at least one embodiment. The appearance of the phrases "in one embodiment” or "in an embodiment” in various places in the specification are not necessarily all referring to the same embodiment.

[0035] (E1)

[0036] The embodiment provides a method for installing a video monitoring partition wall, aiming to solve the problem that the prior art cannot identify construction risks in real time and accurately at the edge end. As shown in Figure 1 the method comprises the following steps:

[0037] S100, real-time acquisition of a monitoring video stream of a partition wall installation work area.

[0038] In a preferred embodiment of the present application, the current monitoring video stream of the partition wall installation area is acquired in real time by a single monitoring device deployed at the construction site.

[0039] The partition wall installation work area is defined as a physical space range covering all processes involved in the partition wall from material stacking, carrying, positioning to final fixing in the construction site. The determination of the area can be realized by any of the following technical means:

[0040] Virtual fence demarcation method: one or more static or dynamic virtual geographic fences are preset in the video stream picture of the monitoring device, and the area surrounded by the virtual fence is directly defined as the installation work area. This method is simple to implement and can effectively focus on the analysis range and exclude the interference of irrelevant areas.

[0041] Dynamic visual recognition method: based on a computer vision model, key entities (including erected partition walls, partition walls to be installed, and supporting installation tools) in the scene are identified in real time, and the dynamic space range jointly occupied by these entities and in which work activities are carried out is determined as the installation work area. This method is more adaptable and can dynamically adjust the monitoring focus following the construction progress.

[0042] In this embodiment, the monitoring device is a single device and is fixedly deployed at a position that can completely cover the partition panel installation work area. In an illustrative embodiment of the present application, the monitoring device is a high-definition network camera fixedly installed at a key position of the construction site, which continuously collects raw video data of the partition panel installation work area at a predetermined frame rate and resolution. The monitoring device is in communication connection with an edge computing node, and the raw video data is transmitted to the edge computing node in real time to form a current monitoring video stream for subsequent processing. This deployment mode ensures the comprehensiveness of video source coverage, the real-time nature of data acquisition, and provides a stable data basis for subsequent lightweight edge analysis.

[0043] In this paper, the real-time acquisition of the monitoring video stream is defined as: the end-to-end processing delay from image acquisition to the generation of safety risk prediction value is controlled within a predetermined time threshold to ensure that the system can make timely warning for the sudden safety risk of the construction site. Specifically, this real-time is achieved through the following technical features:

[0044] Edge processing architecture: the acquisition and subsequent analysis of the video stream are completed on the edge computing node at the construction site, avoiding the network transmission delay caused by uploading data to the cloud server.

[0045] Lightweight model design: the computer vision model and graph neural network used are designed to be lightweight, with small model size and fast inference speed, which can meet the processing needs of high-frequency frame rate under limited edge computing resources.

[0046] Clear performance indicators: the predetermined time threshold can be set according to actual safety needs, usually not more than 2 seconds, and in the preferred embodiment, it can be controlled within 500 milliseconds. This time scale is much faster than the response speed of manual inspection, which can provide an effective time window for risk intervention.

[0047] S200, based on the monitoring video stream, a current knowledge graph representing the state of the construction site is constructed.

[0048] In an embodiment of the present application, the computing task of constructing a dynamic knowledge graph is completed by a collaborative processing system deployed on the edge computing node at the construction site. The system processing system is composed of a visual perception subsystem and a relationship reasoning subsystem:

[0049] 1. Visual perception subsystem (i.e. computer vision model system): responsible for extracting structured data from video pixels, including the following core components:

[0050] Lightweight object detection and tracking model (such as YOLO series): responsible for real-time identification and continuous tracking of entity targets such as construction personnel, partition panels, installation tools, etc. from the video stream, providing entity nodes for the knowledge graph.

[0051] Lightweight attribute recognition model (such as MobileNet series): responsible for analyzing the detected entity target image area to determine its state attribute (such as whether the construction personnel wears a safety helmet, whether the partition board is in the lifted state).

[0052] Visual geometry analysis module: based on camera calibration parameters and target detection results, calculate the spatial position attribute of the entity (such as three-dimensional coordinates, moving speed and direction), provide data basis for constructing the spatial relationship between entities.

[0053] 2. Relationship reasoning and graph generation subsystem: receives the structured data output by the aforementioned perception subsystem, and based on the preset construction behavior logic rule library, performs high-level semantic reasoning. Its core tasks are:

[0054] Determine spatial relationship: based on the spatial position attribute of the entity, determine the relationship such as "under", "contact", "close" through geometric calculation.

[0055] Infer behavior relationship: integrate spatial relationship and state attribute, and infer semantic behavior relationship such as "installing", "operating", "violating" according to the preset logic.

[0056] Generate knowledge graph: take the entity as the node and the determined relationship as the edge, and finally assemble into a structured current knowledge graph for subsequent graph neural network risk prediction. All the above models and modules are pruned and quantized before deployment to strictly adapt to the computing resource constraints of edge devices. More importantly, their running cooperates with the "dynamic weight adaptation" mechanism: image interference factors (such as light, camera height) not only affect the recognition confidence of the above models, but also directly participate in the weight selection of the lightweight graph neural network, so as to jointly guarantee the accuracy, robustness and real-time of risk identification from the perception and reasoning two levels.

[0057] Further, S200 specifically includes the following steps:

[0058] S201, identify the entity of interest from the monitoring video stream, the entity of interest includes construction personnel, partition board, installation tool and support structure.

[0059] Based on target detection and tracking algorithm (such as YOLO, DeepSORT), automatically identify and continuously track specific entity objects from the image frames of the monitoring video stream, which constitute the nodes of the knowledge graph. Among them, the construction personnel refers to all workers in the work area; the partition board includes the installed board and the to-be-installed board; the installation tool is, for example, a drill, a crowbar, a cutting machine, etc.; the support structure is, for example, a temporary support, a keel, etc.

[0060] S202, identify the state attribute and the spatial position attribute of each entity.

[0061] For each identified entity, further identify two types of key attributes:

[0062] State attribute: refers to semantic information describing the instantaneous state of the entity itself, such as whether the construction worker wears a safety helmet, whether the partition board is in the lifted state, etc. The state attribute is automatically extracted by a lightweight attribute recognition model deployed on the edge node. The lightweight attribute recognition model adopts the optimal lightweight network architecture for different attribute types (such as static appearance and dynamic behavior). For example, for static appearance and dynamic behavior attributes, MobileNet classification model or time sequence pose analysis model is used for recognition, such as: for the static appearance attribute of "whether the construction worker wears a safety helmet", a binary classification model based on Mobile Net architecture is used for recognition; for the dynamic behavior attribute of "whether the partition board is in the lifted state", the motion trajectory and the posture key points of the entity in the continuous frames are analyzed for comprehensive judgment. These state information is the key basis for judging whether the behavior relationship violates the safety rules.

[0063] Spatial position attribute: refers to the quantitative physical quantity obtained by camera calibration and visual geometry calculation for describing the entity in the real world, mainly including the three-dimensional spatial coordinates, movement speed vector and direction of the entity.

[0064] The acquisition of spatial position attribute is responsible by the visual geometry analysis module deployed on the edge node. The workflow of the module is as follows: first, based on the inherent parameters and installation position of the monitoring device, the mapping relationship between the image coordinate system and the world coordinate system is established through camera calibration technology; then, the approximate three-dimensional position, movement speed vector and direction of the construction worker, partition board and other entities in the real world are calculated by using the geometric information (such as center point, bottom edge) of the target detection frame, or combining with the depth information estimated by the stereo vision / deep learning model.

[0065] The spatial position attribute obtained in this way is a set of quantitative physical quantities describing the spatial state of the entity in the real world. They are input into the dynamic knowledge graph as the core data basis, used to build accurate spatial relationships (such as "located below", "approaching") between entities, and provide support for subsequent risk reasoning based on behavior logic.

[0066] S203, determine the relationship between each entity based on the spatial position attribute and the preset construction behavior logic, which is used as the edge of the knowledge graph, and finally form a structured current knowledge graph.

[0067] Among them, the relationship mainly includes:

[0068] Spatial relationship: refers to the judgment result of the static or dynamic geometric correlation between entities based on the quantified spatial position attribute. It describes the relative pose of entities in the physical world, such as "person A is below partition B" and "tool C contacts partition D". Spatial relationship is the bottom layer geometric basis for high-level behavior reasoning.

[0069] Behavior relationship: refers to the inference of the semantic interaction or action between entities based on spatial relationship and the combination of state attributes and their changes. It describes "who does what to what", such as "person A is installing partition B" (the relationship inference depends on the spatial relationship of "person A is close to partition B", and the state attributes of "partition B is in the state of being lifted" and "tool is running"). Behavior relationship is the core element of knowledge graph that is directly used to judge whether safety rules are violated.

[0070] The preset construction behavior logic is a knowledge base of formal safety rules and normal operation processes stored in the system. It encapsulates the definition of safety and risk behavior in partition installation operation by domain experts (such as safety engineers), and its essence is a series of IF-THEN conditional judgment rules. These rules map low-level spatial position and state attributes to high-level semantic relationships and risk labels. The preset construction behavior logic includes at least one of the following rules:

[0071] If the spatial relationship between a construction personnel entity and an unfixed partition entity is "under", it is determined that there is a "dangerous area" behavior relationship;

[0072] If the state attribute of a construction personnel entity is "not wearing a safety helmet" and its spatial position is in the "installation operation area", it is determined that there is a "violation of safety regulations" behavior relationship;

[0073] If the state attribute of an installation tool entity is "running with power on", and its spatial relationship with a construction personnel entity is "abnormal contact", it is determined that there is a "tool operation improper" behavior relationship.

[0074] The safety risk prediction value is generated based on whether the behavior relationship violates the preset safety rules. For example, when there is a relationship of "person is under unfixed partition" in the graph, the "high-altitude falling risk" rule is triggered, resulting in an increase in the final risk prediction value.

[0075] In the embodiment of the application, the relationship between each entity is determined based on the spatial position attribute and the preset construction behavior logic, which is realized by a relationship reasoning and graph generation subsystem, specifically by the following steps:

[0076] S2031, spatial relationship calculation: based on the entity spatial position attributes provided by the visual geometry analysis module, through geometric operation and threshold comparison, the static and dynamic spatial correlation between entities is quantitatively calculated. Specifically, it includes:

[0077] Based on the three-dimensional coordinates between entities, the relative distance and azimuth are calculated to determine the "close" and "far away" relationships.

[0078] Based on the coordinate difference of entities in the vertical direction and the preset safety threshold, the "above" or "below" relationship is determined. For example, calculate the distance D between the bottom bounding box center point of construction personnel A and the bottom bounding box of the partition plate B in the vertical direction. According to the preset safety threshold, if the distance D is less than the threshold and the vertical coordinate of the construction personnel A is less than the vertical coordinate of the partition plate B, it is determined that the spatial relationship is "below".

[0079] Based on the overlap or nearest point distance of the entity bounding box, the "contact" and "overlap" spatial relationships are determined.

[0080] S2032, behavior relationship reasoning: combine the calculated spatial relationship with the state attribute provided by the lightweight attribute recognition model, input into the preset construction behavior logic rule library for matching and reasoning, and infer the high-level semantic behavior relationship. Specifically: the reasoning engine detects whether the conditions of the preset construction behavior logic rule library are met, and if so, triggers the rule to generate the corresponding behavior relationship. For example: IF [Entity: Personnel A] [Spatial relationship: below] [Entity: Partition plate B] AND [Partition plate B state: not fixed] THEN [Behavior relationship: Personnel A violates the dangerous area].

[0081] S2033, knowledge graph assembly and update: the identified entity is used as a graph node, and the various relationships determined in steps S2031 and S2032 are used as edges connecting the nodes, and time stamp information is added, to finally generate or update a structured, dynamic knowledge graph that can represent the current construction site state.

[0082] Suppose the following information is identified at a certain time point:

[0083] Entity: personnel (ID: P1), partition plate (ID: W1), impact drill (ID: T1).

[0084] State attribute: personnel P1 does not wear a safety helmet; partition plate W1 is in a suspended and unfixed state; impact drill T1 is in a powered running state.

[0085] Spatial relationship: personnel P1 is below partition plate W1; impact drill T1 is held by personnel P1.

[0086] Based on the above data, the system will generate or update a knowledge graph fragment composed of a series of (subject-predicate-object) triples as follows:

[0087] (Person P1, has status, not wearing safety helmet);

[0088] (Person P1, spatial relationship, under the partition wall W1);

[0089] (Person P1, behavior relationship, operating impact drill T1);

[0090] (Wall W1, has status, suspended and not fixed);

[0091] (Impact drill T1, has status, powered on).

[0092] Metadata:

[0093] Graph context: update timestamp: X year X month X day XX:XX:XX.XXX (for example: 2024-11-07 10:05:30.050).

[0094] This knowledge graph fragment clearly represents the high-risk scenario of "an unsafe worker operating an impact drill under a suspended and unfixed partition wall." The spatial relationship "under" combined with the status attribute "suspended and unfixed" directly triggers the derived behavior relationship "in a dangerous area" through the behavior logic rule base. This structured semantic scenario provides direct and explicit input for subsequent lightweight graph neural networks to accurately predict "object impact" and "high fall" risks.

[0095] S300, acquire image interference factor data of the current environment, and form an image interference factor vector.

[0096] In the embodiments of the present application, the current environment specifically refers to a set of dynamic changing physical conditions and device inherent parameters that affect the imaging quality of the monitoring device. These factors will directly cause the attenuation or distortion of the features extracted by the computer vision model, thereby affecting the accuracy of the subsequent knowledge graph and the reliability of the risk prediction.

[0097] The selection basis of image interference factor data is that it must have a significant and quantifiable impact on image quality, and it can be directly or indirectly obtained through sensors or system configuration. Specifically, the image interference factor data includes at least one of the following:

[0098] The installation height of the monitoring device: this data is a static inherent parameter, measured during device installation and stored in the system configuration file. The installation height determines the width of the shooting angle and the scale of the target in the image. The higher the height, the larger the coverage of the scene, but the lower the pixel resolution of the specific target, and the more blurred the detailed features.

[0099] The ambient light intensity of the shooting area: this data is a dynamic parameter, obtained in real time through the ambient light sensor integrated by the monitoring device or the average pixel brightness of the video stream analysis. Insufficient light will increase image noise and reduce contrast; excessive exposure will cause feature details to be lost, which is one of the most important dynamic factors affecting model recognition accuracy.

[0100] The shooting pitch angle of the monitoring device: this data is a static or semi-static parameter, determined and recorded during device installation and adjustment. The pitch angle determines the perspective relationship of the picture, directly affecting the accuracy of spatial position calculation. Under different angles, the same entity relationship in the image will be completely different.

[0101] Current weather conditions: this data is a dynamic macro parameter, including but not limited to the following types that can cause image quality degradation:

[0102] Precipitation type: such as rain, snow, which can form attachments on the lens surface or moving stripes in the air, blocking the view and introducing noise.

[0103] Suspended matter type: such as fog, haze, dust, which can produce scattering effects in the air, causing significant reduction in image contrast and clarity, and color distortion.

[0104] Light abnormality type: such as strong light, glare, which can cause local overexposure, resulting in loss of feature details.

[0105] The weather condition data can be obtained by one or a combination of the following methods:

[0106] Method one: access the authorized meteorological data service interface to obtain real-time weather information of the construction site area.

[0107] Method two: measure locally through special meteorological sensors (such as raindrop sensors, humidity sensors) integrated by the monitoring device.

[0108] Method three: based on the monitoring video stream itself, analyze and identify the current picture in real time through a specially trained weather classification model.

[0109] Height information of installation operation surface: vertical distance (height difference) between installation operation surface and a preset reference plane. The parameter is a time-varying parameter, which accurately represents the spatial position state of construction progress. The installation operation surface refers to the operation active plane formed by the construction personnel and the to-be-installed plate during the installation of the partition plate, which moves vertically upward with the construction progress. The plane is the dynamic working reference plane of the entire installation operation.

[0110] In an illustrative embodiment, the preset reference plane is the ground level. At this time, the height information of the installation operation surface is the absolute height, which is stable and uniform, and can reliably reflect the stage of the macro construction progress.

[0111] In another preferred embodiment, the preset reference plane is the installation plane of the monitoring device. At this time, the height information of the installation operation surface is the relative height, which most directly determines the scale and clarity of the imaging target, because it is the primary geometric factor affecting the image perspective relationship and target pixel ratio.

[0112] In the embodiment of the application, the height information of the installation operation surface can be obtained by visual recognition method. Specifically:

[0113] System preset and calibration: On the building structure column or core tube, a visual scale ruler or coded marker system that has been accurately calibrated is preset. The zero point or specific reference scale of the ruler has been aligned with a preset reference plane (such as the ground level or the installation plane of the monitoring device) during installation.

[0114] Real-time identification: The current positions of the construction personnel and the partition plate on the calibrated ruler are identified in real time by a computer vision model.

[0115] Analysis and calculation: The corresponding scale value is analyzed. The scale value itself is the height reading relative to the preset reference plane, and the accurate height information of the current operation surface can be directly obtained by reading the value.

[0116] This method converts the complex three-dimensional space measurement problem into a relatively simple task of reading the ruler reading in the two-dimensional image by deploying a physical ruler aligned with the reference plane in the scene, greatly reducing the computational complexity and implementation cost of the system, while ensuring the accuracy and reliability of the measurement.

[0117] Including the installation operation surface height information into the image interference factor system is based on the following deep technical insights:

[0118] (1) Solve the scale attenuation effect

[0119] The increase of the operation surface height leads to the systematic reduction of the scale of the imaging target;

[0120] The pixel proportion of the same entity in the image decreases significantly with the increase of height;

[0121] The feature extraction model thus faces the technical challenge of gradual loss of detailed information.

[0122] (2) Deal with the influence of perspective distortion

[0123] The change in height causes the shooting angle to transition from a horizontal view to a downward view;

[0124] Causing geometric perspective distortion of the image, directly affecting the accuracy of the calculation of the spatial position attribute;

[0125] The fixed visual model is difficult to adapt to such systematic and predictable changes in viewing angle.

[0126] (3) Achieve consistency throughout the cycle

[0127] The traditional fixed weight model performs well during low-level construction, but its performance significantly deteriorates during high-level construction;

[0128] By taking height as the core control variable, the system can dynamically select the optimal expert model for different construction stages;

[0129] Thus ensuring that the risk identification accuracy remains stable and at a high level throughout the entire construction cycle from the bottom to the top.

[0130] In summary, by real-time perception of height changes, the evolution trend of image quality can be foreseen in advance, and accordingly switched to the most suitable visual analysis model for the current construction stage. This mechanism is the core technical innovation to achieve high-precision monitoring throughout the cycle.

[0131] Further, S300 specifically includes:

[0132] S301, collecting multi-source and heterogeneous original interference factor data.

[0133] In the embodiment of the application, the image interference factor data is collected from multiple sources periodically or according to a preset threshold. The data is multi-source and heterogeneous, including but not limited to installation height in meters, ambient light intensity in lux, weather conditions as category data, and installation height information in meters, etc.

[0134] S302, by a data preprocessing module, normalizing and vectorizing the original interference factor data to obtain processed data, and splicing the processed data into a unified one-dimensional feature vector, i.e. the interference factor vector.

[0135] S302 specifically includes:

[0136] Normalization: Numerical data (e.g. height, illumination) is converted to a uniform numerical range (e.g. [0, 1]) through linear scaling or other normalization methods.

[0137] Vectorization: Categorical data (e.g. weather) is converted to one-hot encoding or other forms.

[0138] Concatenation: All processed data is concatenated into a one-dimensional feature vector in a predetermined order, i.e. the interference factor vector is obtained.

[0139] This interference factor vector, as a quantitative and structured description of the entire environmental state, is input into the weight management module to dynamically select the most matching weight set from the weight library.

[0140] S400, according to the image interference factor vector, dynamically select the corresponding weight set from the preset weight library, and load it into a lightweight graph neural network.

[0141] In the embodiments of the present application, the preset weight library stores a plurality of weight sets, each of which is associated with one or more typical interference factor vectors as indexes.

[0142] In the embodiments of the present application, the weight set is a complete set of model state snapshots of the lightweight graph neural network, which contains all trainable parameters in each layer of the network. These parameters collectively define the entire behavior pattern of the network in extracting features from input data and finally performing risk inference. Specifically, a weight set includes but is not limited to: weight matrices and bias vectors in graph convolution layers or fully connected layers. In graph attention mechanisms, trainable weight vectors or matrices used to calculate attention scores. In the embodiments of the present application, the preset weight library is generated offline in the following way:

[0143] (1) Collect historical video data of construction sites under different combinations of image interference factors to cover various typical construction environments.

[0144] (2) For each representative combination of image interference factors, train a graph neural network model respectively to form a series of "expert models" focusing on specific environments.

[0145] (3) Save the weights of all trained graph neural network models to form the preset weight library. Essentially, this weight library is a mapping table, with typical values or value ranges of interference factor vectors as indexes, and corresponding pre-trained weight sets as contents.

[0146] A typical value of the interference factor vector refers to a specific interference factor vector adopted in the offline training phase of constructing the weight library, which can represent a typical or discrete environmental state of the construction site. It is a "snapshot" or "anchor point" of a common environmental condition. For example: [equipment installation height: 6 meters] + [working surface height: 5 meters] + [light: 500 lux] + [pitch angle: 10 degrees] + [weather: sunny] can constitute a typical value vector V_low_day_sunny representing "low-level operation, flat viewing angle, daytime, sunny". [equipment installation height: 6 meters] + [working surface height: 25 meters] + [light: 50 lux] + [pitch angle: 30 degrees] + [weather: sunny] can constitute a typical value vector V_mid_night_clear representing "mid-level operation, downward viewing angle, night, sunny". [equipment installation height: 6 meters] + [working surface height: 60 meters] + [light: 200 lux] + [pitch angle: 45 degrees] + [weather: rain] can constitute a typical value vector V_high_rain_foggy representing "high-level operation, large pitch angle, rainy day, low visibility". In real-time operation, the Euclidean distance between the current vector and all such "anchor points" is calculated, and the closest one is selected to achieve matching.

[0147] The value range of the interference factor vector refers to a continuous region defined around a typical value or a center point in the interference factor vector space. It defines the boundary of the environmental conditions applicable to the same set of weights. For example, a range can be set for V_low_day_sunny: equipment installation height: [5, 7] meters; installation working surface height: [3, 15] meters (covering low-level operation); environmental light intensity: [300, 700] lux (covering moderate light); device shooting pitch angle: [5, 20] degrees (covering flat view and small angle downward view); weather condition: ["sunny", "cloudy"]. As long as the current environmental vector falls within this range, the weight set is used. In real-time operation, matching is achieved by determining which pre-defined region the current vector falls into.

[0148] In another embodiment of the present application, the preset weight library is generated by an expert method, which includes:

[0149] According to historical experience, domain experts (such as safety engineers and algorithm experts) define several typical construction site environmental modes, and each construction site environmental mode is defined by a specific combination of image interference factors. Examples of defined typical environmental modes are as follows:

[0150] Mode 1: low-level operation - daytime - strong light

[0151] Interference factor combination: [installation working surface height: 3 meters] + [light: 800 lux] + [weather: sunny];

[0152] Mode description: This mode corresponds to the initial stage of construction, the working surface is low, the target imaging scale is large, and the details are rich, but it may face strong light irradiation.

[0153] Mode two: middle layer operation-night-lack of lighting

[0154] Interference factor combination: [installation working surface height: 25 meters] + [lighting: 50 lux] + [weather: sunny];

[0155] Mode description: This mode corresponds to the middle stage of construction, the working surface is lifted, the target scale is reduced, and the ambient light is insufficient, mainly relying on artificial lighting.

[0156] Mode three: high-level operation-rainy day-low visibility

[0157] Interference factor combination: [installation working surface height: 60 meters] + [lighting: 300 lux] + [weather: rain];

[0158] Mode description: This mode corresponds to the later stage of construction, the working surface is high, the target scale is small, and it is affected by rain and fog weather, and the image contrast and clarity are significantly reduced.

[0159] For each of the construction site environment modes described, one or more lightweight graph neural network weight sets are assigned that are pre-trained and perform well in the corresponding environment;

[0160] The mapping relationship between the construction environment mode and the weight set is stored to form the preset weight library.

[0161] In the embodiment of the application, the weight management module is specifically responsible for performing the dynamic selection and loading of the weight set. The module receives the interference factor vector from the data preprocessing module as input, and internally pre-stores the weight library. The weight management module selects the most matched weight set from the preset weight library based on the input vector through a matching algorithm, and loads the selected weight set to each layer of the lightweight graph neural network. When the weight management module dynamically selects a weight set and loads it, it is equivalent to instantaneously switching the entire network to an expert mode optimized for a specific environmental condition (such as a certain combination of lighting and height).

[0162] Further, the dynamic selection is completed by matching the current image interference factor vector with the index in the weight library, and the specific way includes but is not limited to:

[0163] Exact / nearest neighbor matching: find the weight set corresponding to the index with the smallest Euclidean distance from the current image interference factor vector from the preset weight library. This method has fast response speed and is suitable for scenes with relatively discrete interference factor states.

[0164] Interpolation matching: if the current image interference factor vector is between multiple indexes in the preset weight library, a new weight set most matched with the current image interference factor vector is generated based on interpolation calculation (such as linear interpolation) of the current image interference factor vector and adjacent indexes. This method can realize smooth transition of weights between different environments, is suitable for scenes with continuous changes of interference factors, and can provide finer adaptation capability.

[0165] In the embodiment of the present application, the lightweight graph neural network is a risk reasoning core specially designed for an edge computing environment, which follows the principle of "deep light and narrow width" to realize real-time and reliable risk prediction under limited computing resources.

[0166] Specifically, the lightweight characteristics are guaranteed through the following collaborative design:

[0167] Structural simplification: the number of layers of the network is strictly limited to three or less. This shallow architecture effectively avoids the huge computing overhead and memory occupation brought by deep networks, and significantly reduces the inference delay, ensuring that the processing speed on the edge node can meet the requirements of real-time monitoring.

[0168] Parameter compression: the number of neurons in each layer of the network is strictly compressed and optimized. This is achieved through advanced model compression techniques such as channel pruning and weight sparsification after model training, aiming to remove redundant neuron connections and significantly reduce the total parameter quantity of the model, thereby reducing storage space requirements and computational complexity.

[0169] Hardware adaptation optimization: before deployment, the weights of the network have been converted to low-precision data types (such as INT8). This quantization operation not only further compresses the model size, but also fully utilizes the integer operation units of edge computing hardware (such as GPU / NPU), achieving several times of inference acceleration.

[0170] Task-oriented customized structure: the lightweight graph neural network selects lightweight variants such as graph convolutional network (GCN) or graph attention network (GAT) as the basic skeleton, and its structure is customized for processing graph structure data composed of "entity-relation" from the knowledge graph, avoiding a large number of parameters required for processing general image data.

[0171] The customized structure of the lightweight graph neural network mainly reflects that its input features, graph structure definition and message passing mechanism are designed for processing construction safety knowledge graphs, wherein the initial feature vector of the node of the graph neural network is composed of the state attribute and spatial position attribute of the corresponding entity, and different propagation weights are used in the message passing process of the graph neural network according to different edge relation types. The customized structure specifically includes:

[0172] (1) Node feature customization

[0173] Feature fusion: The initial feature vector of each entity node (such as personnel, partition plate) is composed of the splicing and fusion of multiple state attributes and spatial position attributes of the corresponding entity. For example, the features of a "construction personnel" node may be composed of attributes such as [whether to wear a safety helmet (0 / 1), whether to be in a moving state (0 / 1), three-dimensional spatial coordinates (x, y, z)].

[0174] Semantic coding: Unlike general model processing pixel features, the input features of the lightweight graph neural network are structured data rich in semantic information extracted from the knowledge graph.

[0175] (2) Edge relationship customization

[0176] Relationship typing: The edges in the graph are explicitly assigned types, such as "under", "operating", "contacting", etc. This is fundamentally different from the general GCN / GAT, which treats all edges as the same type.

[0177] Differential propagation: In the message passing process, the lightweight graph neural network will propagate information according to different edge relationship types using different trainable parameters specific to the edge relationship type. The trainable parameters specific to the type are weight matrices specific to each relationship type.

[0178] In the embodiments of the present application, the lightweight graph neural network realizes differential propagation through relationship-sensitive message passing rules. Specifically, it assigns different sets of trainable parameters to different relationship types (such as "operating", "under") in the knowledge graph. These parameters can be weight matrices specific to each relationship (such as R-GCN), or attention calculation modules specific to each relationship (such as GAT). This allows the network to weight different types of information streams differently when aggregating neighbor node information, allowing it to more accurately capture interaction patterns strongly related to construction risks. Further, in the embodiments of the present application, the lightweight graph neural network realizes differential propagation based on relationship types through one or a combination of the following two core mechanisms:

[0179] 1) Relationship graph convolution network (R-GCN) mechanism

[0180] Under this mechanism, the lightweight graph neural network assigns a dedicated, lightweight weight matrix W r to each relationship type r. The message passing process is as follows: h i (L+1) = σ(∑ r∈R ∑ j∈N(i，r) [(1 / c ir )·Wr ·h j (L) + W0·h i (L), where h i (L) and h i (L+1) are the feature vectors of node i at the L-th and L+1-th layers of the lightweight graph neural network, respectively, i takes values from 1 to F(L), F(L) is the number of nodes at the L-th layer, L takes values from 1 to N, N is the total number of layers of the lightweight graph neural network. R is the set of all predefined relation types (e.g., “underneath”, “operating”), N(i, r) is the set of neighbor nodes adjacent to node i under relation type r, W r is the trainable weight matrix defined specifically for relation type r, c ir is a normalization constant, such as the number of elements in N(i, r), |N(i, r)|, used to balance the different number of neighbor nodes and stabilize the training process. W0is the weight matrix of the node’s own loop, used to preserve the original information of the node itself. σ is a nonlinear activation function such as ReLU. h j (L) represents the feature vector of neighbor node j of node i at the L-th layer, which is derived from the state and spatial attributes of entity j in the knowledge graph.

[0181] Through the above mechanism, the network can enable different feature transformation modes for different semantic relations. For example, when processing the “operating” relation, the network uses a trainable weight matrix specifically for the “operating” relation type, which has learned how to effectively fuse the features of tool state and personnel posture; when processing the “nearby” relation, it switches to a trainable weight matrix specifically for the “nearby” relation type, which pays more attention to spatial proximity rather than operational details. This precise, relation-based parameter sharing mechanism is the key to the lightweight graph neural network of the present application to deeply understand the semantics of the construction scene.

[0182] 2) Relation-aware graph attention network (GAT) mechanism

[0183] Under this mechanism, the lightweight graph neural network dynamically calculates the importance of different neighbor node information through a trainable attention mechanism customized for each relation type. The original attention coefficient e(ij, r) and the normalized attention weight a(ij, r) between the center node i and its neighbor node j under relation type r are calculated as follows:

[0184] e(ij, r) = LeakyReLU(a r T · [W·h i ||W·h j ] ) ;

[0185] a(ij,r) = softmax j (e(ij,r)) = exp(e(ij,r)) / [∑ k∈N(i，r) exp(e(ij,k))].

[0186] where h i denotes the current layer feature vector of the center node i, h j denotes the current layer feature vector of the neighbor node j of the center node i. e(ij,k) denotes the original attention coefficient between the center node i and its neighbor node k under the relationship type r, k ∈ N(i,r). W is a shared, trainable weight matrix for uniform linear transformation of the features of all nodes. [·||·] represents the concatenation operation of two transformed feature vectors. r is a trainable attention vector specially defined for the relationship type r, which is the core parameter to realize the differentiated perception of relationships. LeakyReLU is a nonlinear activation function to introduce nonlinearity in the attention score calculation. softmax j is a normalization exponential function to normalize the original attention scores of all neighbor nodes j ∈ N(i,r) of the node i for the relationship type r, ensuring that the sum of all weights is 1.

[0187] Through the above mechanism, the network can autonomously learn the priority of information transmission under different relationship types. For example:

[0188] For the "operating" relationship, the corresponding attention vector will learn the strong semantic association between the tool and the operator, thereby assigning a high attention weight to this type of connection, making it dominant in information aggregation. For the "located nearby" relationship, the corresponding attention vector will learn to judge that the "located nearby" relationship is weakly related background spatial information, thereby assigning a relatively low attention weight.

[0189] The final feature update of the node is the result of weighted aggregation of the features of all its neighbor nodes by the attention weight.

[0190] (3) Network structure customization

[0191] Shallow aggregation design: Given that the relationship paths in a knowledge graph are usually short (for example, risks often originate from the interaction between directly connected entities), there is no need for a deep network to capture long-range dependencies. Therefore, a network structure of 2 to 3 layers is sufficient to effectively aggregate the key information of neighboring nodes, which directly supports the lightweight of the network.

[0192] Risk-oriented reading mechanism: The last layer of the network (reading layer) is not to perform standard node classification or link prediction, but to be customized to globally pool the entire graph and output a single prediction value representing the global risk state of the entire construction site. This customization makes the network goal consistent with the macro goal of safety monitoring.

[0193] The core of this customized structure is that it directly embeds the prior semantics (entity type, relationship type) in the knowledge graph into the message passing rules and architecture design of the graph neural network, so that the network does not need to learn these rules from scratch, thereby completing the risk reasoning task more efficiently and accurately with fewer parameters and computational complexity.

[0194] In the embodiment of the application, the lightweight graph neural network successfully solves the core contradiction of deploying complex graph models on resource-constrained edge devices through the cooperation of shallow structure, parameter compression, low-precision quantization, and task-oriented customization, providing key technical support for real-time and accurate judgment of construction site risks.

[0195] S500, input the current knowledge graph into the lightweight graph neural network loaded with the weight set, and output a safety risk prediction value of the construction site.

[0196] In the embodiment of the application, after loading the weight set matched with the environment, the lightweight graph neural network performs forward propagation and information aggregation on the input knowledge graph. Its reasoning process fully considers the attributes of all entity nodes in the graph, their spatial relationships and behavior relationships, and finally extracts global risk features from complex scene information. The output layer of the network generates one or more safety risk prediction values. The prediction value can be:

[0197] A continuous scalar value (for example, a probability value between 0.0 and 1.0), and the higher the value, the higher the risk level.

[0198] A discrete classification label (for example: "no risk", "low risk", "high risk").

[0199] A multi-dimensional vector, where each dimension corresponds to the probability of a specific type of risk (such as "object impact risk", "high-altitude falling risk", "electricity safety risk").

[0200] It should be noted that the forward propagation and information aggregation steps performed by the lightweight graph neural network are algorithms that belong to the prior art. The innovation of the present application lies in providing a structured input (current knowledge graph) for this reasoning step and creating a dynamically optimized reasoning condition (weight set matched with the environment), so that it can efficiently and accurately solve the specific technical problem of construction safety risk identification in the edge computing environment.

[0201] Further, the method further comprises:

[0202] S600, acquiring an audio stream of the construction site.

[0203] In the embodiment of the present application, the environmental audio stream can be acquired in real time through the microphone array deployed at the construction site.

[0204] S610, identifying whether there is an abnormal audio event related to a safety risk from the audio stream.

[0205] An abnormal audio event related to a safety risk is identified from the audio stream using a lightweight audio analysis model. The abnormal audio event includes but is not limited to object breaking sound, abnormal impact sound of metal tools, heavy object falling sound, and people's exclamation or shouting sound.

[0206] S620, the generation of the safety risk prediction value, which integrates the output result of the lightweight graph neural network and the identification result of the abnormal audio event, and is specifically implemented by one of the following methods:

[0207] Confidence weighted fusion method: a confidence score is assigned to the identification result of the abnormal audio event, and the confidence score and the original safety risk prediction value output by the lightweight graph neural network are weighted and averaged to generate a final comprehensive risk value = γ × visual risk prediction value + β × audio confidence, wherein γ and β are preset weight coefficients, and γ + β = 1. The weight coefficients can be statically configured according to prior experience, for example, in a regular construction site, visual information is usually more stable and reliable, and γ = 0.7 and β = 0.3 can be set, or dynamically adjusted according to the current image interference factor to balance the contribution of different perception modalities in the final decision, for example, when the environment is extremely poor (the image interference factor vector shows that the light value is extremely low), the visual reliability decreases, and the system can automatically adjust γ (such as 0.5) and correspondingly increase β (such as 0.5) to realize adaptive fusion.

[0208] Rule triggered promotion method: the generation of the safety risk prediction value is realized by logical decision of the preset cross-modal correlation rule. The input of this method is the original risk prediction value output by the lightweight graph neural network, and the output is the final safety risk prediction value corrected by the audio information. The specific implementation process is as follows:

[0209] Basic risk acquisition: the lightweight graph neural network outputs a basic, quantitative original risk prediction value (for example, a value between 0 and 1, or a low, medium, and high level) based on the current knowledge graph.

[0210] Rule matching and decision making: Real-time judgment of whether the preset cross-modal correlation rule is met. That is, whether a specific risk type and its associated abnormal audio event are identified at the same time. The preset cross-modal correlation rule is a pre-stored and extensible logical rule library. Each rule clearly specifies the decision action to be taken when a specific visual risk event and a specific audio event occur at the same time. The formal definition of the rule is: IF [condition: (visual risk type = A) and (abnormal audio event = B) are true at the same time]. THEN [action: execute risk escalation operation C]. The logical rule library includes the following core rules (exemplary and not exhaustive):

[0211] Rule one (object impact risk confirmation)

[0212] Condition: (visual risk type is "wallboard breakage") and (abnormal audio event is "object breaking sound");

[0213] Action: Increase the confidence level of the risk to the highest level and immediately trigger the object impact risk alarm.

[0214] Rule two (violent operation risk confirmation)

[0215] Condition: (visual risk type is "improper tool use") and (abnormal audio event is "abnormal impact sound of metal tool");

[0216] Action: Significantly increase the risk level and mark it as "high-risk operation behavior".

[0217] Rule three (personnel injury risk confirmation)

[0218] Condition: (visual risk type is "person falling down" or "high-altitude falling sign") and (abnormal audio event is "person's exclamation sound");

[0219] Action: Regardless of the original confidence level of the visual risk, directly trigger the highest level of personnel safety accident alarm.

[0220] Rule four (electrical safety risk confirmation)

[0221] Condition: (visual risk type is "cable breakage" or "violation of live-line operation") and (abnormal audio event is "electric shock burst sound");

[0222] Action: Increase the risk level and generate a disposal suggestion of "suspected electric shock, suggest emergency power-off".

[0223] The preset cross-modal correlation rule can be summarized and induced by safety experts and algorithm engineers based on a large number of construction site accident cases and historical data, and can be continuously optimized and expanded during operation.

[0224] Risk value correction: If the rule is triggered, the original risk prediction value will be promoted according to the preset logic to generate the final safety risk prediction value. The promotion methods include but are not limited to:

[0225] Value superposition: Add a fixed penalty value to the original value. Final value = original value + fixed penalty value;

[0226] Coefficient amplification: Multiply the original value by an amplification coefficient greater than 1. Final value = original value x amplification coefficient;

[0227] Grade transition: directly promote the risk grade to a preset highest or higher level (such as directly from medium to high).

[0228] For example, it is identified that the visual risk partition plate is broken, and the original risk prediction value is 0.7 (the high risk threshold is set to 0.8). At the same time, the audio model identifies the sound of object breaking. The preset rule is triggered, and the original value is multiplied by a coefficient of 1.2 for correction. The final safety risk prediction value = 0.7 x 1.2 = 0.84. Since 0.84 > 0.8, it is determined as high risk and triggers the highest level alarm.

[0229] The technical effect of S600 to S620 is to upgrade from single visual perception to "audio-visual combined" cooperative perception. In the case of visual obstruction, poor light or low model confidence, audio can provide a strong evidence, which significantly reduces the false negative and false positive.

[0230] Further, after outputting the safety risk prediction value, the method further comprises:

[0231] S700, when the prediction value exceeds the preset risk threshold, generating and sending an early warning information to a monitoring terminal.

[0232] In this step, after outputting the safety risk prediction value, the safety risk prediction value is compared with a plurality of preset risk thresholds, and corresponding early warning intervention mechanisms are started according to different grades.

[0233] In the embodiment of the application, at least two risk thresholds (such as a low risk threshold T low and a high risk threshold T high ), so as to divide the risk state into multiple grades (such as "normal", "low risk", "high risk"). When the safety risk prediction value exceeds the low risk threshold, an early warning information is automatically generated. The early warning information not only contains the risk level, but also contains the key entity, behavior relationship and occurrence position leading to the risk, and is sent to the large screen of the monitoring center or the mobile terminal of the on-site safety officer through the network.

[0234] In the embodiment of the present application, the risk threshold is not a constant, but a parameter that can be dynamically determined and optimized by combining data-driven and expert knowledge. Its determination basis and process are as follows:

[0235] (1) Initial setting based on historical data statistical analysis

[0236] Source: Collect and analyze a large amount of historical construction video data and its corresponding real safety event records confirmed by manual.

[0237] Method: Backtest these historical data on the trained lightweight graph neural network model to get the safety risk prediction value corresponding to each case. By statistical analysis (such as drawing the distribution curve of the prediction value under different real risk levels), select the critical point that can best distinguish the "safe" and "dangerous" states as the initial threshold.

[0238] For example: analysis found that 95% of the real high-risk events have model prediction values higher than 0.85, and 90% of the safe states have prediction values lower than 0.3. Then, T high =0.85 can be set as the high-risk threshold, and T low =0.3 as the low-risk warning threshold.

[0239] (2) Calibration by integrating industry standards and expert experience

[0240] Source: National standards (such as "Uniform Standard for Construction Safety" GB50870), industry safety standards, and domain experts (such as senior safety engineers).

[0241] Method: Submit the initial threshold obtained by data statistics to expert review. Experts will evaluate the severity of consequences for different risk levels according to their judgment of risk tolerance, and fine-tune the threshold.

[0242] For example: for "high-altitude falling" which has extremely serious consequences, even if its data statistical threshold may be 0.8, experts may suggest adjusting it to 0.7 to reflect the conservative safety principle of "better safe than sorry".

[0243] (3) Online adaptive optimization

[0244] Source: New feedback data generated in actual operation.

[0245] Method: Record the disposal results of each warning (such as real alarm and false alarm confirmed by safety personnel). When a certain amount of new data is accumulated, periodically or triggeredly re-execute the statistical analysis in step (1) to dynamically adjust and optimize the threshold, so that the system can adapt to the unique environment and work mode of the specific construction site.

[0246] The S700 achieves a complete automated closed loop from "risk perception -> analysis and reasoning -> decision-making -> proactive early warning". It changes the traditional passive mode that relies on manual inspections, and realizes early warning and real-time intervention for safety risks at construction sites, greatly improving safety management efficiency and response speed.

[0247] Furthermore, in this embodiment of the invention, the lightweight graph neural network is also configured to output a key inference path corresponding to the predicted value; the key inference path is used to identify the core entities and relationships that lead to the risk.

[0248] The key reasoning path is a structured dataset used to qualitatively and quantitatively reveal the root causes leading to high-risk predictions. It identifies the core entity sequences in the knowledge graph that contribute the most to the final risk, along with the core relationships connecting them. The generation of the key reasoning path relies on the attention mechanism or gradient information within the graph neural network.

[0249] Attention-based approach: In an embodiment employing Graph Attention Networks (GAT), the key reasoning paths are extracted by quantifying the importance of different reasoning paths through path attention scores. The key reasoning paths are extracted by quantifying the importance of different reasoning paths through path attention scores. The score of the path is the continuous product of the attention weights corresponding to each relation edge on the path. For example, consider a reasoning path from person A to partition wall C: "Person A → [Operating] → Impact Drill B → [Action on] → Partition Wall C". The attention score of this path is calculated as follows: S path =α(Personnel A - Operating - Impact Drill B) × α(Impact Drill B - Acting on - Partition Wall C). Next, from all possible paths, the one or more paths with the highest path attention score will be selected as the key reasoning paths for this risk prediction.

[0250] Gradient-based approach: In embodiments employing Graph Convolutional Networks (GCNs), the importance of a risk is assessed by calculating the gradient of the final risk prediction relative to the input node and relation features. A larger absolute gradient value indicates a greater contribution of the node or relation to the risk, thus making it a component of the critical path.

[0251] A critical reasoning path can be represented as a sequence of (entity-relationship-entity) triples. For example:

[0252] [Personnel P1]--(Operating)-->[Impact Drill T1]--(Emitting)-->[Abnormal Noise];

[0253] [Person P1] --(is located in)-> [Unfixed partition wall board W1] --(below)-> [].

[0254] The technical effect of the lightweight graph neural network configured to output a key reasoning path corresponding to the predicted value is that the "black box" decision of the graph neural network is converted into a semantic chain understandable by humans, answering the question of "why is it determined as a risk?"; The safety officer can directly locate the risk source and key link according to the path prompt, realize fast and accurate disposal, and does not need to check the entire site.

[0255] In summary, the present application produces a significant synergistic effect through the dynamic weight adaptation mechanism, the lightweight graph neural network and the semantic knowledge graph, systematically solves the core technical contradictions of real-time risk identification in the edge computing environment, and achieves the following breakthrough technical effects:

[0256] First, the performance barriers of "high precision" and "low delay" in the edge computing scene are broken. The traditional scheme faces a dilemma in this scene: the complex model has high precision but cannot be real-time, and the lightweight model is fast but has insufficient precision. The present application dynamically binds the environmental interference factors and the model weight through the preset weight library, does not need online training during reasoning, and only needs to switch the weight efficiently, so that the "expert model" optimized for the current environment can be loaded and run instantly. This is essentially a "pre-training, slow execution" architectural innovation that fundamentally resolves the performance contradiction, making it possible to achieve millisecond-level, high-precision risk identification on resource-constrained edge devices.

[0257] Second, the adaptive robustness to respond to the variable environment of the real construction site is built. The performance of the traditional fixed model fluctuates sharply when the light, weather and operation surface height change. The present application introduces an intelligent closed loop of "perception-decision-adaptation": the system real-time perceives the image interference factors and uses them as control signals to drive the model weight to make optimal decisions and adapt. For example, when low light interference is detected, the weight set that has been trained in a large number of dark environment samples will be automatically loaded. This model has learned to suppress noise and enhance effective features. This makes the model evolve from a static and fragile program to a dynamic and environment-immune intelligent agent, thereby maintaining stable high precision in a variable environment and significantly reducing false positives and false negatives.

[0258] Third, the cognitive leap from "surface perception" to "deep semantic understanding" is achieved. The application discards the traditional end-to-end pixel classification paradigm and constructs a "pixel-entity-relation-risk" progressive cognitive chain. By constructing a knowledge graph, the semantic structure of the scene (who, where, what, and what) is first understood; a lightweight graph neural network performs relationship reasoning on this basis. The dynamic weight mechanism ensures that this high-order cognitive ability is still reliable in different physical environments. Therefore, it can discover potential risks such as "personnel under non-fixed board" that are composed of multiple entities and relationships, deep and have causal logic, realizing the qualitative change from "seeing" to "understanding" and even "judging".

[0259] Finally, the application successfully combines advanced deep learning technology with complex industrial site needs, and through systematic architectural innovation, realizes accurate, real-time, and robust intelligent perception and early warning of construction site safety risks in resource-constrained edge environments.

[0260] (Example Two)

[0261] The embodiment provides a video monitoring partition board installation risk system for realizing Figure 1 The method of the embodiment is shown. As Figure 2 The system comprises:

[0262] A monitoring device 1 is deployed in the partition board installation operation area to obtain a monitoring video stream in real time.

[0263] A data processing module 2 is configured to construct a current knowledge graph representing the state of the construction site based on the monitoring video stream. The module further comprises:

[0264] A visual perception submodule is used to identify and track entities and identify their state attributes.

[0265] A relationship reasoning and graph generation submodule is used to determine the relationship between entities based on spatial attributes and behavior logic and assemble the knowledge graph.

[0266] A sensor module 3 is used to collect image interference factor data of the current environment.

[0267] A data preprocessing module 4 is coupled to the sensor module 3 and is configured to normalize and vectorize the multi-source heterogeneous original image interference factor data collected by the sensor module to form a unified image interference factor vector.

[0268] A weight management module 5 is coupled to the data preprocessing module 4 and pre-stores a weight library; the weight management module is configured to dynamically select a corresponding weight set from the weight library according to the image interference factor vector and load the weight set to the lightweight graph neural network inference module.

[0269] A lightweight graph neural network inference module 6 is coupled to the data processing module and the weight management module and is configured to take the current knowledge graph as input and output a safety risk prediction value of the construction site.

[0270] Further, the system further comprises:

[0271] An early warning module is coupled to the lightweight graph neural network inference module and is configured to generate and send early warning information to a monitoring terminal when the safety risk prediction value exceeds a preset risk threshold.

[0272] It should be noted that the system embodiment and the corresponding method embodiment are based on the same inventive concept. Therefore, the related technical features, implementation details and the technical effects that can be achieved in the method embodiment are also applicable to the system embodiment.

[0273] (Embodiment Three)

[0274] The embodiment provides an installation risk system based on a video monitoring partition board, which is used for realizing Figure 1 The method of the illustrated embodiment comprises:

[0275] A plurality of monitoring devices are deployed at different positions of a collaborative partition board installation work area and are used for acquiring monitoring video streams in real time from different perspectives.

[0276] The collaborative partition board installation work area is composed of a global work area and a plurality of local work sub-areas allocated to different monitoring devices. The global work area is the overall physical space covering all partition board installation work activities. The local work sub-area is an exclusive monitoring range allocated to a specific monitoring device in the global work area and delimited by a virtual fence.

[0277] A plurality of edge processing units, each of which is coupled to one of the monitoring devices and is configured to perform the following operations:

[0278] Constructing a local knowledge graph based on the video stream acquired by the monitoring device connected to the edge processing unit.

[0279] Acquiring image interference factor data from the perspective of the monitoring device connected to the edge processing unit and forming an image interference factor vector.

[0280] Based on the image interference factor vector, a weight set is dynamically selected from a pre-stored weight library and loaded into a lightweight graph neural network.

[0281] The lightweight graph neural network is used to process the local knowledge graph and output a local risk prediction value.

[0282] A central decision-making module, coupled to all the edge processing units, is configured to fuse all the local risk prediction values ​​to generate a global security risk prediction value. The fusion of all the local risk prediction values ​​is achieved through a fusion algorithm, which includes, but is not limited to, the maximum value fusion method, the weighted average fusion method, or the evidence theory fusion method. When using the weighted average fusion method, the weight of each local risk prediction value can be determined through any one or a combination of the following methods:

[0283] (1) The weights of each monitoring device are dynamically adjusted based on the historical recognition accuracy of each device. The higher the accuracy, the greater the weight of the device. Specifically, the accuracy weight of each monitoring device is equal to the historical recognition accuracy of that device divided by the sum of the historical recognition accuracy of all devices.

[0284] (2) The image quality score is weighted based on the current image interference factor vector. The image quality score is calculated based on parameters such as light intensity, weather conditions, and image clarity. The output of the device with higher image quality is given priority. Specifically, the quality weight of each monitoring device is equal to the image quality score of that device divided by the sum of the image quality scores of all devices.

[0285] (3) Consider the relative spatial distance between the monitoring equipment and the core risk area to ensure that the observation data closest to the risk source receives the highest weight. The relative spatial distance can be Euclidean distance, and the distance weight W for the i-th device is given by the distance weight W. distance-i =(1 / (1+D i )) / Σ(1 / (1+D j )), D i Let D be the relative spatial distance between the i-th device and the core risk area. j Let Σ(1 / (1+D) be the relative spatial distance between the i-th device and the core risk area. j )) represents all 1 / (1+D j The summation is performed, where i and j both range from 1 to m, and m is the number of monitoring devices.

[0286] In a preferred embodiment, the weight of each local risk prediction value can be the product of accuracy weight, quality weight, and distance weight.

[0287] The core risk area refers to the spatial range within the construction site with the highest probability of risk occurrence and the greatest potential hazard. This can be identified using real-time risk hotspot detection technology based on spatial location clustering.

[0288] Algorithm optimization strategy:

[0289] The monitoring area is divided into uniform grids using a grid-based clustering algorithm, or the optimized DBSCAN algorithm is used to speed up neighborhood queries through spatial indexing to avoid complex distance matrix calculations, reducing the computational complexity from O(n²) to O(nlogn).

[0290] Input feature simplification:

[0291] Only using two-dimensional plane coordinates (x, y) and risk prediction values as clustering inputs can discard redundant feature dimensions and focus on spatial risk distribution pattern recognition.

[0292] Real-time density statistics:

[0293] The distribution density of risk prediction values is calculated in real time on the spatial grid;

[0294] The sliding window mechanism is used to balance real-time performance and statistical stability;

[0295] Quickly identify high-risk density areas: one or more grid areas with the highest risk density value are determined as high-risk density areas, i.e. risk core areas. The risk density risk_density = Σ(risk_values) / grid_area. Where risk_values refers to the set of local risk prediction values of all risk entities falling within the current grid cell, Σ(risk_values) represents the summation of risk prediction values of all risk entities within the current grid cell, and grid_area represents the projected area of a single grid cell in the actual physical space, with units of square meters.

[0296] In this embodiment, each edge processing unit processes the video stream of the monitoring device connected to it in the same or similar manner as the processing of a single video stream in the aforementioned method embodiment (e.g. the flow corresponding to Figure 1 The core implementation principle, models (such as lightweight target detection, attribute recognition, graph neural network models) and processing logic used in the series of steps of local knowledge graph construction, interference factor vector formation, weight set dynamic selection and loading, and local risk prediction value generation performed independently by each edge processing unit are consistent with steps S200 to S500 in the aforementioned method embodiment. The difference is that the system achieves multi-angle and high-efficiency coverage of the collaborative working area through parallel distributed computing of multiple edge processing units, and finally the central decision module completes information fusion and global decision-making.

[0297] It should be noted that the embodiment is not a simple repetition of the single device method. The core progress is to solve the three technical bottlenecks of inherent visual blind area, high false alarm rate and limited complex risk identification ability of single device system through distributed sensing and central decision. The system significantly improves the reliability of the alarm through the cross verification mechanism of multiple source information, and realizes the deep sensing of the composite risk which cannot be identified by the single device through the global situation fusion. Therefore, the multi-device cooperative architecture brings new technical effects and has technical advantages that the single device system cannot achieve.

[0298] (Eighth embodiment)

[0299] The embodiment of the present application also provides an electronic device, comprising: at least one processor; and a memory connected with the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are configured to execute the method described in the embodiment of the present application.

[0300] (Fifth embodiment)

[0301] The embodiment of the present application also provides a computer readable storage medium, which stores computer executable instructions, and the computer instructions are used to execute the method described in the embodiment of the present application.

[0302] It should be understood that the various forms of flow shown above can be used to reorder, add or delete steps. For example, each step described in the present application can be executed in parallel, sequentially or in different order, as long as the desired results of the technical solutions disclosed in the present application can be achieved, which is not limited herein.

[0303] The above specific embodiments do not constitute a limitation on the scope of protection of the present application. Those skilled in the art should understand that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent replacement and improvement within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A method for assessing installation risks of a video surveillance partition based on a wallboard, characterized by, The method comprises the following steps: S100, acquiring a monitoring video stream of a partition plate installation work area in real time; S200, constructing a current knowledge graph representing a construction site state based on the monitoring video stream, specifically comprising: S201, identifying an entity of interest from the monitoring video stream, the entity of interest including a construction worker, a partition plate, an installation tool, and a support structure; S202, identifying state attributes and spatial location attributes of each entity of interest; S203, determining relationships between entities of interest based on the spatial location attributes and a preset construction behavior logic to form the current knowledge graph, specifically comprising the following sub-steps: S2031, spatial relationship calculation: based on entity spatial location attributes provided by a visual geometry analysis module, through geometric operation and threshold comparison, quantitatively calculating static and dynamic spatial correlations between entities; S2032, behavior relationship reasoning: combining the calculated spatial relationships with state attributes provided by a lightweight attribute recognition model, inputting into a preset construction behavior logic rule base for matching and reasoning, thereby inferring high-level semantic behavior relationships; S2033, knowledge graph assembly and update: taking the identified entities as graph nodes, taking various relationships determined in steps S2031 and S2032 as edges connecting the nodes, and injecting timestamp information, finally generating or updating a structured, dynamic knowledge graph representing the current construction site state; S300, acquiring image interference factor data of the current environment and forming an image interference factor vector; S400, dynamically selecting a corresponding weight set from a preset weight library according to the image interference factor vector and loading it into a lightweight graph neural network; the preset weight library stores a plurality of weight sets, each weight set being associated with one or more image interference factor vectors as indexes; the dynamic selection is completed by matching the current image interference factor vector with the indexes in the weight library; S500, inputting the current knowledge graph into the lightweight graph neural network loaded with the weight set, and outputting a safety risk prediction value of the construction site.

2. The method of claim 1, wherein, The image interference factor data comprises at least one of the following: installation height of a monitoring device; environmental light intensity of a shooting area; shooting pitch angle of a monitoring device; current weather condition; height information of an installation work surface.

3. The method of claim 1, wherein, The specific way of dynamic selection is to find a weight set corresponding to an index with the smallest Euclidean distance from the current image interference factor vector in the preset weight library.

4. The method of claim 3, wherein, If the current image interference factor vector is between multiple indexes in the preset weight library, the specific way of dynamic selection is to generate a corresponding weight set based on interpolation calculation of the current image interference factor vector and adjacent indexes.

5. The method of claim 1, wherein, The initial feature vector of a node of the lightweight graph neural network is composed of the state attributes and spatial location attributes of the corresponding entity of interest; in the message passing process, the lightweight graph neural network uses different trainable parameters for information propagation according to different edge relationship types.

6. The method of claim 1, wherein, The method further comprises: acquiring an audio stream of the construction site; identify whether there is an abnormal audio event related to a security risk from the audio stream; the generation of the security risk prediction value integrates the output result of the lightweight graph neural network and the identification result of the abnormal audio event.

7. A video monitoring based installation risk system for a wallboard, characterized by The system is used to implement the method of any one of claims 1 to 6, and the system comprises: a monitoring device deployed in a partition board installation operation area, configured to obtain a monitoring video stream in real time; a data processing module configured to construct a current knowledge graph representing the state of the construction site based on the monitoring video stream; a sensor module for collecting image interference factor data of the current environment; a data preprocessing module coupled to the sensor module and configured to convert the image interference factor data collected by the sensor module into an image interference factor vector; a weight management module coupled to the data preprocessing module and pre-storing a weight library; the weight management module is configured to dynamically select a corresponding weight set from the weight library according to the image interference factor vector and load it into a lightweight graph neural network inference module; a lightweight graph neural network inference module coupled to the data processing module and the weight management module, configured to input the current knowledge graph and output a security risk prediction value of the construction site.

8. A video monitoring based installation risk system for a wallboard, characterized by The system is used to implement the method of any one of claims 1 to 6, and the system comprises: a plurality of monitoring devices deployed at different positions in the cooperative partition board installation operation area, configured to obtain monitoring video streams from different perspectives in real time; a plurality of edge processing units, each coupled to one of the monitoring devices and configured to perform the following operations: construct a local knowledge graph based on the video stream obtained by the monitoring device connected to the edge processing unit; obtain image interference factor data from the perspective of the monitoring device connected to the edge processing unit and form an image interference factor vector; dynamically select a weight set from a pre-stored weight library according to the image interference factor vector and load it into a lightweight graph neural network; process the local knowledge graph using the lightweight graph neural network to output a local risk prediction value; a central decision module coupled to all the edge processing units and configured to fuse all the local risk prediction values to generate a global security risk prediction value.

Images