A construction site safety risk intelligent identification and early warning method based on communication engineering

Abstract

The application relates to the technical field of safety management of communication engineering construction, and discloses a construction site safety risk intelligent identification and early warning method based on communication engineering. The method comprises the following steps: acquiring visual monitoring data, environment and equipment operation parameter data and equipment safety state logic quantity; time aligning and space correlating the visual monitoring data, the environment and equipment operation parameter data, the equipment safety state logic quantity and process state information, and inputting the data into a multi-modal fusion model; obtaining a hazard source set and a prevention and control barrier set corresponding to a current construction process through a knowledge graph, performing causal reasoning based on real-time verification results of the prevention and control barrier, and generating a graded early warning in combination with a fused risk characteristic value. The application realizes deterministic verification through construction of a safety state machine based on internal electrical signals of equipment, solves false alarm and missed alarm by fusing internal and external data through gate modulation, outputs hierarchical early warning in combination with causal reasoning of the knowledge graph process and parameter closed-loop regulation and control, and self-evolves rules.

Description

Technical Field

[0001] This invention relates to the field of safety management technology for construction of communication engineering projects, and in particular to a method for intelligent identification and early warning of safety risks at construction sites of communication engineering projects. Background Technology

[0002] In the field of safety monitoring during telecommunications engineering construction, existing technologies mainly rely on external sensing methods such as visual monitoring, proximity sensing wristbands, and environmental sensors to detect risks. The criteria for judging electric shock risks depend on indirect evidence such as the strength of the power frequency electric field, which is prone to missed or false alarms due to sensor directional deviations, electromagnetic shielding, or changes in the hand posture of workers. Some solutions use multimodal data fusion for violation identification, but the fusion methods are mostly simple splicing or general weighting based on environmental parameters. They do not introduce the exact electrical isolation status inside the equipment (such as the opening and closing of circuit breakers, the position of grounding switches, and the RF power amplifier enable signal) to adjust the credibility of external sensing information, making it difficult to suppress false alarms triggered by visual or environmental data when the equipment is in a safe isolation state. In addition, existing safety analysis methods based on knowledge graphs are mostly geared towards static knowledge management or probabilistic risk classification, and do not bind the prevention and control barriers with real-time sensing data fields. They lack the ability to perform item-by-item causal verification based on the logic and gate conditions of the accident mechanism for the current process, and cannot accurately determine whether the causal conditions for the occurrence of an accident are complete, resulting in insufficient process-specificity and accuracy of early warning.

[0003] Therefore, there is an urgent need for an intelligent risk identification and early warning method that can directly acquire and utilize the internal electrical logic state of communication equipment for deterministic security verification, deeply integrate this deterministic state with multimodal perception data, and supplement it with process causal reasoning and rule self-evolution capabilities. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this invention provides an intelligent identification and early warning method for construction site safety risks based on communication engineering. It achieves deterministic verification by constructing a safety state machine through internal electrical signals of equipment, resolves false alarms and missed alarms by gating modulation and fusing internal and external data, and combines knowledge graph-based process causal reasoning and parameter closed-loop control to output hierarchical early warnings and self-evolving rules.

[0005] The objective of this invention can be achieved through the following technical solutions: This disclosure provides a method for intelligent identification and early warning of safety risks at construction sites based on communication engineering, including the following steps: S1. Multidimensional perception data acquisition: Acquire visual monitoring data, environmental and equipment operating parameter data of the construction site, and obtain equipment safety status logic quantities by parsing the electrical node logic signals deployed inside the communication equipment and constructing a safety operation state machine; S2. Spatiotemporal alignment and multimodal fusion: The visual monitoring data, environmental and equipment operating parameter data, equipment safety status logic quantities, and process status information transmitted from the work ticket system are time-aligned and spatially correlated, and then input into the multimodal fusion model to obtain fusion risk feature values; Among them, the multimodal fusion model generates a gating factor based on the logical quantity of the equipment safety status, and performs gating modulation on the coupled representation extracted from visual monitoring data and environmental and equipment operating parameter data—enhancing risk signals when the equipment is not reliably isolated, and suppressing false alarms when it is isolated; S3. Causal Reasoning and Hierarchical Early Warning of Construction Processes: By constructing a knowledge graph of construction safety, the set of hazards and control barriers corresponding to the current construction process are obtained. Causal reasoning is performed based on the real-time verification results of the control barriers, and hierarchical early warnings are generated by combining the fused risk feature values. The communication construction safety knowledge graph includes operation activity nodes, hazard source nodes, accident mechanism nodes, and prevention and control barrier nodes. The accident mechanism nodes define trigger condition combinations using logical AND gates. The prevention and control barrier nodes are bound to real-time verification data fields through verification AND relations. The inference engine periodically obtains real-time data of the bound fields and determines the satisfaction status of each prevention and control barrier. When all the trigger conditions defined by the accident mechanism node are met and the prevention and control barriers used to block each trigger condition are in an unsatisfied or unknown state, the accident mechanism is activated and an early warning is generated.

[0006] Furthermore, the logical quantity for obtaining the device security status includes: The circuit breaker auxiliary contacts, grounding switch position contacts, and RF power amplifier enable relay contacts of the communication equipment are accessed through the dry contact signal acquisition module. At the same time, the device operation status register is parsed through the Modbus / SNMP network management interface to map the circuit breaker opening and closing, grounding switch status, RF power amplifier enable signal and residual voltage value in real time, forming the original status logic quantity. Using the original state logic quantity as the driving source, a safe operation state machine is established for each type of critical communication equipment. The safe operation state machine contains multiple safe states, and state transitions are triggered by a specified logic signal sequence. When the logic signal sequence does not meet the preset order or a state bounce occurs, it is determined that the isolation has failed or the safe state has failed. The name of the current safe state of the state machine is merged with the original state logic quantity to form the equipment safe state logic quantity.

[0007] Furthermore, the key communication equipment includes RRU, base station power distribution unit, battery pack, antenna feeder surge arrester and coupler, and integrated cabinet power module. Each type of equipment has a separately adapted safety operation state machine established according to its power form and isolation requirements. Among them, the safety operation state machine of RRU includes four safety states: "normal operation", "RF shutdown", "feeder grounding" and "residual charge discharge completed". The state transition is jointly triggered by the RF power amplifier enable signal, circuit breaker auxiliary contact, grounding switch auxiliary contact, residual voltage detection value or preset discharge timing signal.

[0008] Furthermore, the processing procedure of the multimodal fusion model includes: The skeletal key point sequence, seat belt double hook status sequence, indicator light status sequence and dangerous area intrusion event in the visual monitoring data are encoded by a time-series network. The environmental and equipment operating parameters are upgraded by a multilayer perceptron and linearly mapped after one-hot encoding of the equipment safety status logic quantity. The process status information is transformed by an embedding layer, unified into an embedding vector of the same dimension and superimposed with position encoding. Using visual embedding vectors as queries and environmental and equipment parameter embedding vectors as keys and values, a cross-attention mechanism is used to extract the coupled representation of personnel actions and equipment external states. The embedding vectors of equipment safety state logic quantities are passed through a fully connected layer and an activation function to generate safety suppression gating factors and unisolated enhancement gating factors, and the coupled representations are gated and modulated. The process embedding vectors are injected into each intermediate layer of the encoder, and finally mapped through the output layer to obtain 0-1 continuous fused risk feature values.

[0009] Furthermore, the gating modulation method is as follows: let the coupling representation of personnel action and equipment external state be C, the unisolated enhancement gating be G1, and the safety suppression gating be G2, then the output representation is C′=C⊙(1+βG1)⊙(1-λG2), where β and λ are preset weight coefficients; when the equipment is not reliably isolated, G1 approaches 1, and the risk representation is enhanced through the (1+βG1) term; when the equipment is safely isolated, G2 approaches 1, and the risk representation triggered by externally sensed data is suppressed through the (1-λG2) term.

[0010] Furthermore, the construction method of the communication construction safety knowledge graph is as follows: Define four types of entity nodes: work activities, hazard sources, accident mechanisms, and prevention and control barriers. Among them, the work activity node is refined to the process level and a perception parameter configuration table containing detection frame rate, confidence threshold, sensitivity level and sampling period is preset. Establish a system that connects operational activities to hazards, leads to hazard mechanisms, blocks accident mechanisms, and records the trigger condition indexes of the blocked mechanisms. Verify the connection between the hazard mechanisms and the specific sensing data fields. The verification results of the control barrier are updated periodically with three states: satisfied, not satisfied, and unknown. The unknown state is equivalent to not satisfied and participates in risk derivation according to the conservative safety strategy.

[0011] Furthermore, the generation of graded early warning includes: when the accident mechanism is activated by causal reasoning, the risk feature value R is weighted and fused with the pre-set severity score S of the accident mechanism node extracted from the knowledge graph, and the comprehensive risk score F = α·R + (1-α)·S is calculated. The F value is mapped to the preset threshold range to which it belongs to the early warning level from level I to level V, where α is an adjustable weight.

[0012] Furthermore, the process causal reasoning and hierarchical early warning also includes a dynamic adjustment step for sensing parameters: Based on the current process type, the inference engine reads parameters from the pre-set perception parameter configuration table of the corresponding job activity node in the knowledge graph, and sends parameter adjustment instructions to the visual perception module and sensor acquisition module through the edge computing node, so that the detection frame rate, model confidence threshold, alarm sensitivity and sensor sampling period take effect in the subsequent acquisition window.

[0013] Furthermore, it also includes: S4. Hidden Risk Chain Mining and Rule Solidification: Collect historical early warning records and near miss event data, and construct a heterogeneous graph containing personnel attribute nodes, equipment model nodes, environmental range nodes, work activity nodes and event nodes. Co-occurrence edges are established between nodes that appear together in the same operation or event. A heterogeneous graph Transformer model is adopted, which maps heterogeneous nodes to a unified dimension through a type-specialized projection matrix, stacks multi-head attention layers for message passing, and conducts supervised training with the probability of a real hidden danger occurring in this task as the prediction target. After training, attention weights are tracked for positive samples, edges with weights greater than a threshold are extracted to generate multi-hop paths, and high-frequency paths are clustered and statistically analyzed as candidate latent risk chains. After expert evaluation confirms the causal rationality, the candidate paths are compiled into new causal rules and added to the communication construction safety knowledge graph.

[0014] Furthermore, the method of compiling candidate paths into new causal rules is as follows: add a prevention and control barrier node under the corresponding operation activity node, add trigger conditions in the relevant accident mechanism node, and establish verification and relationship binding for the new barrier to the corresponding perception data field, synchronously write to the graph database, recompile the inference path table of the affected process, and update the inference engine rule table cache.

[0015] The beneficial effects of this invention are: This invention analyzes the dry contact signals and network management interface registers deployed inside communication equipment to construct a safety operation state machine adapted to the power form of various devices. Using the precise internal isolation state transition logic as the device safety state logic quantity, it elevates the electric shock risk judgment from external probabilistic sensing to deterministic internal verification, fundamentally eliminating false alarms caused by sensor directional deviations or shielding. The multimodal fusion model generates gating factors using the device safety state logic quantity, modulating the coupled representation of personnel actions and equipment states extracted from visual and environmental data—enhancing risk signals when there is no reliable isolation and suppressing false alarms when there is isolation. This achieves hierarchical and reliable fusion where internal deterministic state adjudication is based on external probabilistic perception, significantly improving the accuracy of risk identification. The communication system is based on the logic AND gate accident mechanism and verification AND relationship. The construction safety knowledge graph can perform causal verification of each prevention and control barrier according to the current process. It activates the accident mechanism only when all triggering conditions are met and the corresponding barriers are not satisfied. It also integrates risk feature values ​​and accident severity scores to output graded warnings, overcoming the shortcomings of traditional rule matching or probability classification, which can only output general warnings. At the same time, the perception parameter configuration table pre-set by the process nodes of the knowledge graph drives the dynamic adjustment of parameters such as visual detection frame rate and sensor sampling period, so that computing resources are focused on the highest risk object in the current process, forming a two-way closed loop of knowledge-driven perception and perception feedback reasoning. Furthermore, through the heterogeneous graph attention model, it automatically mines hidden risk chains from historical warnings and near misses, and after expert evaluation, compiles them into new causal rules that can be called in real time, realizing the self-growth and continuous evolution of the safety knowledge graph. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram illustrating the steps of an intelligent identification and early warning method for construction site safety risks based on communication engineering, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the steps of multi-dimensional sensing data acquisition provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the steps of spatiotemporal alignment and multimodal fusion provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the process causal reasoning and hierarchical early warning provided in the embodiments of the present invention; Figure 5This is a schematic diagram illustrating the steps of hidden risk chain mining and rule solidification provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This embodiment provides a method for intelligent identification and early warning of safety risks at construction sites based on communication engineering, such as... Figure 1 As shown, it includes the following steps: S1. Multi-dimensional sensing data acquisition: Acquiring visual monitoring data from the construction site, environmental and equipment operating parameter data, and obtaining equipment safety status logic quantities by parsing the electrical node logic signals deployed inside communication equipment and constructing a safety operation state machine, such as... Figure 2 As shown, it includes the following steps: S11. Deploy visual perception modules in the communication construction operation area, including installing anti-shake wide-angle cameras on the communication tower platform, deploying portable 4G / 5G surveillance cameras under the tower, and smart safety helmets worn by workers that integrate first-person view cameras and UWB / IMU positioning and attitude determination modules; the visual perception modules continuously collect video streams, run target detection models and multi-target trackers through edge computing nodes, and extract the key point sequence of personnel skeletons, the alternating posture of safety belt double hooks, the flashing pattern of equipment indicator lights, and the structured information of dangerous area intrusion to form visual monitoring data.

[0020] It should be noted that the extraction of human skeleton keypoint sequences involves a top-down two-stage pose estimation pipeline used by the edge computing nodes. First, a lightweight human detection model, such as YOLOv8n, RTMDet-tiny, or other lightweight pedestrian detectors, is used to detect human instances in each frame of video and output bounding boxes. Then, a lightweight pose estimation model, such as RTMPose-tiny or a similar keypoint detection network, is run on each bounding box region to extract 17 or 26 human skeleton keypoint coordinates and confidence scores. If a single-stage pose estimation model, such as YOLOv8n-pose, is used, the model directly outputs the human bounding boxes and keypoints, which are then uniformly converted to the same keypoint sequence format in subsequent steps. Finally, a multi-object tracker assigns a unique ID to each human instance and concatenates the keypoint coordinates of the same ID in consecutive frames into a time-dimensioned skeleton keypoint sequence stream. Seatbelt double hook alternating posture extraction: A dedicated small target detection head is added to the high-resolution layer of the feature pyramid network of the general detection model to identify the rectangular frame position and opening / closing state of the seatbelt hooks on both sides of the waist, and simultaneously identify the safety rope, lifeline, tower anchor point or attachment point area; only when the hook is in the closed and locked state, and the hook frame meets the preset spatial overlap or distance threshold with the effective anchor point, lifeline or safety rope area, is it judged as "valid attachment". The double hook state sequence of the same person in consecutive frames is taken as a sliding window and fed into a lightweight temporal action recognition network (such as a lightweight version of SlowFast based on 3D convolution or a simplified version of TimeSformer based on Transformer), and outputs four action categories: "both hooks locked", "left hook unlocked and moved", "right hook unlocked and moved", and "both hooks disengaged". Among them, the alternation mode of "left hook unlocked - right hook locked" and "right hook unlocked - left hook locked" is judged as compliant double hook alternation, and "both hooks disengaged" lasting more than 3 seconds is directly marked as violation. Device indicator light blinking pattern extraction: The target detection model locates the light-emitting area of ​​the device panel indicator light and outputs the bounding box. For each indicator light area, multiple frames are continuously sampled, and the RGB color histogram is extracted frame by frame. The color classifier determines the current frame as "red", "green", "yellow" or "off". The obtained state sequence is sent to a one-dimensional temporal classification network (such as 1D-CNN or LSTM) to identify the blinking pattern as constant, slow, fast or intermittent. Finally, the color and blinking pattern are combined and mapped to the semantic description of the device's operating status through a preset device indicator light encoding rule table (such as "RUN indicator light green slow flash = normal operation, ALM indicator light red constant light = fault alarm, PWR indicator light green constant light = normal power supply").Dangerous area intrusion extraction: During the initialization phase, dangerous areas such as the tower fall radius area, the safe distance area for live equipment, and the manhole warning area are pre-marked in the video frame using polygons and associated with risk level attributes. Using the instance ID of each person and the frame-by-frame foot coordinates output by the multi-target tracker (DeepSORT), it is determined in real time whether the coordinates fall within the dangerous area polygon. When entering for the first time, an intrusion event containing the person ID, area type, and intrusion timestamp is recorded. When the foot point leaves and continues for more than the set de-shaking time, a departure event is recorded. At the same time, the posture data of the corresponding time period is extracted from the skeletal sequence stream and packaged and output together.

[0021] S12. Deploy IoT sensors on the bodies of workers and on the surfaces of communication equipment, including proximity induction wristbands worn by workers to collect power frequency electric field strength and calculate the relative distance change trend with charged bodies; attach wireless current clamps and infrared thermal imaging modules to the surfaces of RRUs (Remote Radio Units) and battery packs to collect feeder conduction current and temperature distribution; deploy temperature and humidity sensors in the base station room and battery pack area to collect ambient temperature and humidity data; deploy multi-gas sensors (O2, H2S, CO, CH4) and anemometers in the manhole; install ultrasonic anemometers and tilt sensors on the tower platform; package all collected data into environmental and equipment operating parameter data.

[0022] Understandably, the proximity induction wristband is implemented by comparing the real-time electric field strength with the pre-calibrated reference electric field strength of the charged body, and combining the hand motion acceleration and angular velocity output by the built-in IMU to perform attitude calculation, compensating for the directional deviation of the electric field measurement caused by wrist rotation, thereby continuously calculating the trend of the relative distance between the worker's wrist and the charged body.

[0023] S13. Deploy a protocol parsing gateway in the base station equipment room. This gateway connects to electrical nodes such as circuit breaker auxiliary contacts, grounding switch position contacts, and RF power amplifier enable relay contacts via a dry contact signal acquisition module. Simultaneously, it parses the equipment operating status register through a Modbus / SNMP network management interface, mapping in real-time safety-critical logic quantities such as circuit breaker opening / closing, grounding switch status, RF power amplifier enable signal, and residual voltage value, forming a signal mirror table. This signal mirror table maps the switching signals acquired from the dry contacts to the operating status register parsed by the network management interface, unifying them into original state logic quantities. The acquisition objects include circuit breaker opening / closing positions, grounding switch status, RF power amplifier enable signals, and residual voltage values—all directly related to the risk of electric shock. This process introduces dry contact signals, originally used only for equipment monitoring, into the field of construction safety, allowing the system to directly acquire the internal electrical states that determine the risk of electric shock, rather than relying on external probabilistic sensing.

[0024] S14. Using the original state logic quantity as the driving source, establish a safe operation state machine for each critical communication device, and stipulate that state transitions are triggered by a specified logic signal sequence; merge the newly added state machine state information with the original state logic quantity to jointly constitute the device safe state logic quantity (that is, the device safe state logic quantity includes two parts: the first part is the original state logic quantity obtained by parsing the dry contact signal and the network management interface register, and the second part is the state machine safe state derived by the safe operation state machine based on the original state logic quantity; the two are merged according to the device ID and timestamp to form a unified device safe state logic quantity).

[0025] It should be noted that critical communication equipment refers to equipment that workers need to directly contact or operate at close range during construction or maintenance of communication engineering projects, and that contains unisolated electrical energy or residual charge in energy storage components. This includes RRUs, base station power distribution units, battery banks, antenna feeder surge arresters and couplers, and integrated cabinet power modules. Because the energy forms and isolation requirements of different equipment are fundamentally different—the core risk of RRUs is radio frequency energy, the core risk of battery banks is DC energy storage, and the core risk of base station power distribution units is AC / DC power distribution—each type of equipment requires a separate safety operating state machine adapted to its energy form and isolation procedures. Taking RRUs as an example, a finite state machine is constructed, including four safety states: "normal operation," "radio frequency shutdown," "feeder grounding," and "residual charge discharge completed." The state transition is triggered by the RF power amplifier enable signal, the circuit breaker auxiliary contact, the grounding switch auxiliary contact, the residual voltage detection value, or the preset discharge timing signal: when the RF power amplifier enable signal changes from high to low and the circuit breaker is in the open state, the state machine transitions from "normal operation" to "RF shutdown"; only when the RF power amplifier enable signal remains low and the grounding switch auxiliary contact is closed, the state machine transitions from "RF shutdown" to "feeder grounding"; when the residual voltage is lower than the preset safety threshold or the grounding holding time exceeds the preset discharge duration, the state machine transitions from "feeder grounding" to "residual charge discharge completed". If the above signal sequence does not meet the preset order, or if there is a state jump such as the RF power amplifier enable signal rising or the grounding switch abnormally opening, it is determined to be an isolation failure or a safety state failure. The state machine state is periodically refreshed synchronously. The current safety state name of the state machine (such as "normal operation", "RF shutdown", "feeder grounding", "residual charge discharge completed") is merged with the original state logic quantity formed by S13 to jointly constitute the equipment safety state logic quantity, which is output to S2 and S3 respectively for fusion risk characteristic calculation and barrier condition verification.

[0026] The above-mentioned S13 and S14, through the acquisition of the internal logic state of the device and the construction of the safety state machine, upgrade the electric shock risk criterion from indirect evidence of "external sensing of electric field" to direct evidence of "incomplete internal isolation", eliminating missed detections caused by sensor directionality or shielding, and realizing the transformation from probabilistic perception to deterministic safety verification.

[0027] S2. Spatiotemporal Alignment and Multimodal Fusion: Visual monitoring data, environmental and equipment operating parameter data, equipment safety status logic quantities, and process status information from the work order system are spatiotemporally aligned and spatially correlated, and then input into a multimodal fusion model to obtain fused risk characteristics, such as... Figure 3 As shown, it includes the following steps: S21. Visual monitoring data, environmental and equipment operating parameter data, equipment safety status logic quantities, and process status information (including current process, power outage permit identifier, personnel qualifications, etc.) transmitted from the work order system are uniformly time-aligned according to timestamps. Time alignment is performed using a preset length (e.g., 100ms) sliding window as the processing unit, and grid association is performed using tower section number, machine room number, and manhole number as spatial indexes. At the same time, a mapping relationship is established between personnel ID, equipment ID, UWB / IMU positioning coordinates, video screen coordinates, and equipment installation coordinates, so that personnel actions, personnel positions, target equipment operating parameters, and the equipment safety status logic quantities within the same time window can be bound to the same spatial entity or adjacent spatial entities.

[0028] Understandably, the work permit system is a digital platform for production scheduling and safety authorization in communication engineering construction management. It is used to initiate, approve, and record the type of work activity, planned time period, qualifications of participating personnel, hazard identification results, and permission status of various safety measures (such as "power outage confirmed" and "hot work approved") for each construction task. In this embodiment, the process status flow output by the work permit system is introduced as one of the input dimensions for multi-source data fusion. Together with other data flows, it completes spatiotemporal alignment and feature encoding, providing real-time input for the subsequent knowledge graph to instantiate safety preconditions based on real-time process and dynamically adjust the parameters of the perception model. This elevates general safety monitoring to process-level compliance reasoning synchronized with the construction process.

[0029] S22. The aligned data is fed into the constructed multimodal fusion model. After input embedding encoding, the coupling representation of personnel actions and equipment external state is extracted by cross attention. The coupling representation is generated by gating factor modulation using equipment safety state logic quantity. This includes enhancing risk signals when unreliable isolation and suppressing false alarms when isolated. Process embedding is injected to give context awareness. Finally, 0-1 continuous fusion risk feature value is output.

[0030] It should be noted that the construction of the multimodal fusion model includes: firstly, encoding the skeletal keypoint sequence in the visual monitoring data through a temporal graph convolutional network; encoding the seatbelt double hook state sequence through a lightweight temporal network; encoding the indicator light state sequence through a one-dimensional convolutional network; encoding the intrusion event in the hazardous area through an event embedding layer; upgrading the environmental and equipment operating parameters through a multilayer perceptron; linearly mapping the logical quantities of equipment safety status after one-hot encoding; transforming the current process name in the process status information through a word embedding layer; linearly mapping discrete permission states such as power outage permission marks after one-hot encoding; and embedding personnel qualification levels through a lookup table. All submodals are unified into 256-dimensional embedding vectors and superimposed with sinusoidal positional encoding to preserve temporal relationships. Then, using the visual embedding vector as the query and the environmental and equipment parameter embedding vectors as the key and value, an 8-head attention is set in a 64-dimensional subspace. The scaling dot product attention is calculated, and after concatenation, the spatiotemporal coupling representation of personnel actions and equipment external state is output through linear projection. During this process, the cross-attention mechanism automatically establishes a high attention association between action features such as "hand moving towards RRU feeder port" and "feeder conduction current value", which improves the two types of data from simple concatenation to semantic-level spatiotemporal coupling. Then, the embedding vector of equipment safety state logic quantity is generated by full connection layer and Sigmoid activation to generate safety suppression gating factor and unisolated enhancement gating factor, and the coupling representation is gated and modulated. Specifically, let the coupling representation of personnel actions and equipment external state be C, the unisolated enhancement gating be G1, and the safety suppression gating be G2, then the output representation can be expressed as C′=C⊙(1+βG1)⊙(1-λG2), where β and λ are learnable or preset weight coefficients. When the equipment is not reliably isolated, G1 approaches 1, thus enhancing the risk representation through the 1+βG1 term; when the equipment is safely isolated, G2 approaches 1, thus suppressing the risk representation related to electric shock triggered by external perception data such as vision and environment through the 1-λG2 term, realizing the rejection of external probabilistic risk inferences by internal deterministic safety status, and allowing the precise isolation status inside the equipment to directly regulate the risk judgment strength of the fusion model; at the same time, the process embedding vector is injected into each intermediate layer of the encoder through residual connections and learnable weights, so that each layer of features carries the process context, such as the automatic enhancement of the model's sensitivity to radio frequency related risk features under the "antenna feeder connection" process; finally, the global representation of the encoder is mapped to a 0-1 continuous fusion risk feature value through the output layer of a single neuron Sigmoid layer.

[0031] S3. Process Causal Reasoning and Graded Early Warning: By constructing a communication construction safety knowledge graph, the set of hazards and barrier conditions corresponding to the current construction process are obtained. Matching reasoning is performed based on the real-time verification results of the barrier conditions, and graded early warnings are generated by combining fused risk features. Figure 4 As shown, it includes the following steps: S31. Construct a knowledge graph for communication construction safety, defining four types of ontology nodes: work activities, hazard sources, accident mechanisms, and prevention and control barriers. Establish containment, cause, blockage, and verification relationships. The verification relationship binds the prevention and control barrier to the corresponding real-time verification data field (real-time verification data fields include at least one of visual monitoring data, environmental and equipment operating parameter data, equipment safety status logic quantities, and process status information transmitted from the work ticket system). The graph is stored in a graph database, and the causal reasoning paths for each process are pre-compiled and constructed offline by domain experts.

[0032] Specifically, constructing a knowledge graph for communication construction safety includes: First, defining the ontology structure of the knowledge graph, which includes four types of nodes and attributes: Operation activity nodes are refined to the process level, with attributes including the operation activity name, the relevant specialty, the standard operating procedure number, and a perception parameter configuration table (including the target frame rate, confidence threshold, sensitivity level, and sensor sampling period for each detection item), with instantiations such as "antenna feeder splicing," "battery pack replacement," and "manhole fiber optic cable splicing"; Hazard source nodes describe the inherent energy or substances in the operation, with attributes including the hazard source name, energy type, and location, with instantiations such as "radio frequency energy" (tower platform), "DC bus residual voltage" (equipment room power cabinet), and "manhole oxygen deficiency"; Accident mechanism nodes formalize the causal relationship of accidents as a combination of necessary conditions for a logic AND gate, with attributes including accident type and a list of triggering conditions. The instantiation includes the "RF burn" accident mechanism (triggering conditions: "human body parts exposed to RF field" AND "RF energy release") and the "electric shock" accident mechanism (triggering conditions: "personnel contact with live conductor" AND "equipment not reliably isolated" AND "protective equipment failure"). The prevention and control barrier node describes the measures to block a certain condition in the accident mechanism. The attributes include barrier name, barrier type and target. Instantiations include "RF shutdown and grounding" (physical isolation, blocking "RF energy release"), "alternating use of double hook safety belts throughout the process" (personal protection, blocking the risk of falling from height due to "personnel losing fall protection"), and "insulating gloves and voltage detection procedures" (personal protection, blocking "protective equipment failure", verified by the detection status of the worker's insulating gloves in visual monitoring data and the voltage detection signature in the work ticket system). Secondly, four types of relationship edges are established to form reasoning paths: including relationships connecting work activities to hazards, causing relationships connecting hazards to accident mechanisms, blocking relationships connecting accident mechanisms to control barriers and recording the trigger condition index of the blocking, and verifying relationships connecting control barriers to specific sensing data fields. For example, "RF shutdown and grounding" is verified in the current state of the RRU state machine in the equipment safety status logic quantity; "alternating use of double hook safety belts throughout the process" is verified in the alternating state of double hooks in visual monitoring data; and "forced ventilation to O2>19.5%" is verified in the O2 concentration of the manhole in the environmental and equipment operating parameter data. The graph is stored in a graph database, indexed by work activity nodes, and a complete reasoning path table for each process from hazard source, accident mechanism, trigger condition, control barrier to sensing verification field is pre-compiled for subsequent online reasoning calls by S32.

[0033] S32. The inference engine instantiates a set of hazard source, accident mechanism, and prevention barrier nodes in the knowledge graph based on the process name passed in by the work order. It periodically verifies each barrier through AND relationships, obtains the corresponding perception data fields, and determines the satisfaction status. When all necessary triggering conditions in the logical AND gates defined by the accident mechanism node are determined to be true by real-time data, and the prevention barriers used to block the corresponding triggering conditions are all in a "not satisfied" or "unknown" state, the accident mechanism is activated (if any necessary triggering condition is not met, or any corresponding prevention barrier is verified as "satisfied," then the accident mechanism is not activated or is blocked), generating a structured early warning containing causal paths and evidence snapshots. Simultaneously, it dynamically adjusts the visual detection frame rate and threshold according to the preset parameters of the process, including the following steps: Process instantiation: When the work ticket system inputs the current process information, the inference engine uses the process name as the query key to retrieve the work activity node in the knowledge graph database built by S31, and traverses the relationship edges of containment, cause, and blockage step by step to instantiate the set of hazard sources, accident mechanisms and prevention and control barriers corresponding to the current process, and generate the process inference path table.

[0034] Barrier condition verification: The inference engine periodically verifies each barrier in real time through verification relationships, obtaining the corresponding fields and judgment rules from the visual monitoring data, environmental and equipment operating parameter data, or equipment safety status logic quantities bound to the barrier: For the barrier "RF shutdown and grounding", the current state of the RRU state machine in the equipment safety status logic quantity is satisfied if it is in "feeder grounding" or "residual charge discharge completed", otherwise it is not satisfied; For the barrier "double hook safety belt alternating throughout the process", the alternation of double hooks is compliant within a 30-second window in the visual monitoring data and there is no event where both hooks are disengaged for more than 3 seconds, then it is satisfied; For the barrier "forced ventilation to O2>19.5%", the O2 concentration in the manhole is greater than 19.5% in the environmental and equipment operating parameter data. The verification results of each barrier are updated every 100ms with three values: "satisfied", "not satisfied", and "unknown".

[0035] "Satisfied" indicates that the corresponding prevention and control barrier has been verified as effective by real-time data; "Not Satisfied" indicates that the corresponding prevention and control barrier has failed or not been executed; and "Unknown" indicates that the bound sensing data field is temporarily unavailable due to sensor failure, communication interruption, or insufficient data quality. When matching accident mechanisms, "Unknown" is treated as "Not Satisfied" according to a conservative safety strategy and participates in risk deduction, while generating a sensing link anomaly prompt.

[0036] Accident Mechanism Matching and Risk Derivation: Accident mechanism nodes are defined using logical AND gates to define combinations of trigger conditions. When all barriers corresponding to the trigger conditions are not met, the accident mechanism is activated, and the corresponding accident type is at risk of occurrence; if any barrier is met, it is blocked. For example, the trigger conditions for the "radio frequency burn" accident mechanism are "human body parts exposed to radio frequency field" AND "radio frequency energy release". When visual monitoring data detects that personnel have entered a pre-marked radio frequency exposure area or the RRU feeder close-range operation area, and the equipment safety status logic quantity indicates that the RRU is not in the "feeder grounded" or "residual charge discharge completed" state, i.e., the barrier "radio frequency shutdown and grounding" verification is not met, the accident mechanism is activated; if the barrier is met, even if personnel enter the radio frequency exposure area or the RRU close-range operation area, it is still determined that the accident mechanism is blocked, and no such warning is generated.

[0037] Causal warning information generation: When the accident mechanism is activated, structured warning information is generated, including: risk type, complete causal path in the format of "work activity - includes - hazard source - leads to - accident mechanism - blocked by - prevention and control barrier", barrier name and its real-time data value not met, five-level warning level that integrates risk feature value and accident severity assessment, and evidence snapshot consisting of visual frame sequence before and after triggering, instantaneous value of equipment safety status logic quantity and instantaneous value of environmental parameter, which is pushed in real time through MQTT protocol.

[0038] It should be noted that the comprehensive assessment of the five-level early warning level includes: the fusion risk feature value R output by S22 is only used to participate in the early warning level calculation after the accident mechanism in S32 is activated by causal reasoning; the fusion risk feature value R is used to characterize the current multimodal perceived risk intensity, and the knowledge graph reasoning result is used to determine whether the accident mechanism meets the causal conditions. The pre-set severity score S of the currently activated accident mechanism node is extracted from the knowledge graph (a continuous value between 0 and 1, pre-calibrated by domain experts during knowledge graph construction based on the severity of the accident consequences, such as S=0.9 for radiofrequency burns and S=0.7 for oxygen deficiency in manholes). The fusion risk feature value R output from step S22 is weighted and fused with the severity score S to calculate the comprehensive risk score F = α·R + (1-α)·S, where α is an adjustable weight. The preset threshold intervals [0,0.2), [0.2,0.4), [0.4,0.6), [0.6,0.8), and [0.8,1.0] to which the F value belongs are mapped to warning levels from Level I (ignored, only leaving a trace in the background without actively pushing on-site alarms) to Level V (critical).

[0039] Dynamic adjustment of perception parameters: The inference engine drives the adjustment of visual detection parameters based on the current process type. Each work activity node in the knowledge graph has a pre-set perception parameter configuration table, such as for "Antenna Feeder Connection": safety belt double hook detection frame rate 25fps, confidence threshold 0.85, near-electric intrusion detection frame rate 15fps with high sensitivity, ground enclosure detection frame rate 5fps with degraded sensitivity; for "Manhole Fiber Optic Cable Splicing": manhole entry / exit counting frame rate 10fps, overstay threshold 5 minutes, multi-gas sensor sampling period increased to 50ms, ground enclosure detection frame rate 10fps. The inference engine synchronously reads the configuration table during process instantiation and sends parameter adjustment commands to the visual perception module in S11 and the sensor acquisition module in S12 through the edge computing node model management interface. This ensures that the detection frame rate, model confidence threshold, alarm sensitivity, and sensor sampling period take effect within subsequent acquisition windows, thus focusing computing resources and alarm strategies on the highest-risk object in the current process, forming a two-way closed loop of knowledge graph-driven perception and perception result feedback inference.

[0040] S4. Hidden Risk Chain Mining and Rule Solidification: Historical warning records and near miss event data are collected, a heterogeneous graph is constructed, and a graph neural network is used to extract hidden risk chains. After review, these chains are compiled into new causal rules and added to the knowledge graph. Figure 5 As shown, it includes the following steps: S41. Offline collection of the reviewed early warning records and near miss event ledgers from S32 to construct a heterogeneous graph. The heterogeneous graph includes five types of heterogeneous nodes: personnel attribute nodes, equipment model nodes, environmental range nodes, work activity nodes, and event nodes. Among them, personnel attribute, equipment model, environmental range, and work activity nodes are used to represent risk factors, and event nodes are used to represent each historical work record, review early warning record, or near miss event record. Co-occurrence edges are established between nodes that appear together in the same work or event. The HGT model is used to map heterogeneous nodes to a unified dimension using a type-specialized projection matrix, and multi-head attention layers are stacked for message passing. The probability of a real hidden danger occurring in this work is used as the prediction target. Positive samples are verified early warnings and near miss events, and negative samples are false alarm records and no abnormal work records. The training is supervised by binary cross-entropy loss.

[0041] Specifically, heterogeneous graph construction: Define five classes of heterogeneous nodes, and construct an initial feature vector for each class of nodes based on its attributes: Personnel attribute node: includes skill level (beginner / intermediate / advanced, one-hot encoded), years of service (continuous value, normalized to [0,1]), and fatigue index (from IMU data and work duration statistics of the S1 smart safety helmet; if missing, the weighted value of the cumulative work duration and continuous work duration for the day is used). This node ID is associated with the personnel number in the work ticket system to ensure traceability.

[0042] Device model node: Includes device type (RRU, battery pack, power distribution unit, etc., embedded vector representation), device voltage level or RF power level (scalar), and service life (scalar). Device model information comes from the communication asset management system, and the corresponding relationship has been recorded when sensors are attached to the surface of the S1 devices.

[0043] Environmental interval nodes: Continuous environmental parameters are discretized into intervals, such as temperature divided into four levels: "<0℃", "0-25℃", "25-35℃", and ">35℃", and humidity divided into three levels: "<40%", "40%-80%", and ">80%". Each environmental interval node corresponds to an interval label, and its feature is its one-heat encoding.

[0044] Job activity nodes: The job activity nodes defined in the S31 knowledge graph are directly reused, with the process-level job activity name as the node identifier, and the feature is its pre-trained process embedding vector (which is of the same origin as the process embedding in S22).

[0045] Event Node: Each historical construction operation record or near miss record is identified by a unique number (work ticket number or near miss log number). The feature vector is composed of the meta-attributes of the event (event type, time of occurrence, site number, etc.) through one-hot encoding and normalization, and is used as the target node for binary classification prediction in the training phase.

[0046] Definition of Heterogeneous Graph Edges: If personnel, equipment, environmental ranges, and work activities co-occur in the same construction work record (indexed by the work ticket number) or the same near miss record, then an undirected edge is established between the corresponding nodes. Simultaneously, undirected co-occurrence edges are established between the event node corresponding to that record and the aforementioned four types of nodes, so that each historical event exists in the graph as a star-shaped connected subgraph with the event node as the center and the four types of attribute nodes as the periphery. No manual weights are assigned to edges during the initial graph construction phase; all co-occurrence edges are treated with equal weight. The correlation strength between risk elements is not given by the initial edge weights but is automatically learned through type-specialized attention weights during HGT model training. The heterogeneous graph constructed in this way maps each historical event to a connected subgraph in the graph, and the multi-hop paths between nodes implicitly encode the combination relationships of risk elements.

[0047] Heterogeneous Graph Transformer Training: The Heterogeneous Graph Transformer (HGT) model is employed, utilizing its type-specialized projection mechanism to handle heterogeneous nodes. Specifically, independent linear projection matrices are set for personnel attribute nodes, equipment model nodes, environmental range nodes, work activity nodes, and event nodes, mapping the original features of each type of node to a unified 256-dimensional hidden representation space. The model consists of three stacked HGT layers, each containing a multi-head attention (8 heads) and a feedforward network. During attention calculation, type-specialized Key, Query, and Value projection matrices are used for different types of node pairs (e.g., personnel-equipment, equipment-environment, environment-work activity, etc.), and edge types are introduced into the attention weight calculation (the edge type in this heterogeneous graph is singular but scalable). After three layers of message passing, the final representation of the event node corresponding to each historical work record, review warning record, or near miss event record is taken. This representation is then passed through a fully connected layer and a sigmoid function to output the predicted probability of a real hidden danger occurring in the corresponding work for that record.

[0048] Training sample construction: Each historical work record or near miss record is considered a sample. Positive samples are labeled "real hazard occurred" (including records in S32 early warnings that have been manually verified as "true," and records in the near miss log), while negative samples are labeled "no hazard occurred" (including records verified as "false alarms," ​​and randomly selected work records showing no safety anomalies). Supervised training is performed using a binary cross-entropy loss function.

[0049] S42. After training, attention weights are tracked for positive samples, and edges with weights greater than a threshold are extracted to generate multi-hop paths. High-frequency paths are clustered and statistically analyzed as candidate latent risk chains. After expert evaluation confirms their causal rationality, the rule converter compiles the candidate paths into new causal rules: a prevention and control barrier node is added under the corresponding work activity, trigger conditions are added in the relevant accident mechanism node, and a verification and relationship binding is established for the new barrier to the corresponding perception data field. The data is synchronously written to the graph database, the inference path table of the affected process is recompiled, and the inference engine rule table cache is updated. This completes the automatic mining of unknown risk coupling patterns from historical data and solidifies them into a self-growing closed loop that can call knowledge in real time.

[0050] Specifically, after training convergence, for each positive sample (a work record where a potential hazard has been confirmed), the attention weight distribution from the event node to various attribute nodes via multiple hops is tracked using the attention weight matrix of each layer of the HGT model. The specific extraction method is as follows: An attention weight threshold θ (e.g., 0.15) is set. From the HGT computation graph corresponding to each positive sample, all edges with attention weights greater than θ are extracted, and a depth-first traversal is performed along these edges to generate all paths with a length not exceeding 4 hops. Frequently occurring path patterns in all positive samples are clustered and statistically sorted, and paths with a frequency and support higher than a preset threshold are selected as candidate latent risk chains.

[0051] For example, if the path "service life < 1 year — [co-occurrence] → certain type of battery pack — [co-occurrence] → ambient humidity > 80% — [co-occurrence] → real hidden danger event" appears frequently in multiple positive samples, and this path rarely appears in negative samples, then it is listed as a candidate hidden risk chain.

[0052] Expert Assessment and Knowledge Compilation: Each quarter, a list of candidate hidden risk chains is submitted to a review panel composed of telecommunications construction safety experts and knowledge engineers. Experts assess the causal rationality of each path based on construction safety regulations, physical mechanisms, and field experience. If a reasonable hidden risk chain is confirmed, it is compiled into a new causal rule in the knowledge graph by a rule converter. The compilation process includes: If the elements involved in the implicit risk chain already have corresponding nodes in the existing knowledge graph (e.g., the work activity "battery pack replacement" already exists), then a new prevention and control barrier node is added under this work activity, such as "additional insulation testing for battery replacement in high humidity environment". A risk condition corresponding to the failure of this barrier is added to the trigger condition list of the corresponding accident mechanism node, such as "additional insulation testing not performed in high humidity environment". The newly added prevention and control barrier is connected to this new risk condition through a blocking relationship, and is bound to the environmental humidity reading, the worker's length of service or skill level field, and the special permit confirmation field in the work ticket system through verification. Verification is performed on the humidity reading in the environmental and equipment operating parameter data and the worker's skill level field in the work ticket system. The verification logic is defined as "when the environmental humidity > 80% and the worker's length of service < 2 years, confirmation that additional insulation testing has been performed (through work ticket special permit confirmation) is required".

[0053] If the implicit risk chain reveals a completely new hazard source or accident coupling relationship, then add the corresponding hazard source node and accident mechanism node according to the ontology definition in S31, and establish a complete causal reasoning link in the above manner.

[0054] The rule converter automatically writes the aforementioned changes into the graph database according to the knowledge graph ontology structure defined in S31 (four types of nodes: work activities, hazard sources, accident mechanisms, and prevention barriers, and four types of relation edges: contain, cause, block, and verify). It also recompiles the causal reasoning path table for the affected processes and synchronously updates the rule table cache read by the S32 online reasoning engine. After precompilation, the new rules are stored in the reasoning path table in a unified format along with the existing rules. The S32 reasoning engine can perform matching and reasoning without distinguishing the rule source, achieving a seamless closed loop from offline discovery to online reasoning of the hidden risk chain. After completing this closed loop, subsequent online reasoning can automatically call the newly added causal rules for real-time barrier verification without manual intervention in the reasoning process.

[0055] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.

Claims

1. A method for intelligent identification and early warning of safety risks at construction sites based on communication engineering, characterized in that, Includes the following steps: Multidimensional perception data acquisition: acquire visual monitoring data, environmental and equipment operating parameter data of the construction site, and obtain equipment safety status logic quantities by parsing the electrical node logic signals deployed inside the communication equipment and constructing a safety operation state machine; Spatiotemporal alignment and multimodal fusion: Visual monitoring data, environmental and equipment operating parameter data, equipment safety status logic quantities, and process status information transmitted from the work ticket system are time-aligned and spatially correlated, and then input into a multimodal fusion model to obtain fusion risk feature values; Among them, the multimodal fusion model generates a gating factor based on the logical quantity of the equipment safety status, and performs gating modulation on the coupled representation extracted from visual monitoring data and environmental and equipment operating parameter data—enhancing risk signals when the equipment is not reliably isolated, and suppressing false alarms when it is isolated; Process-based causal reasoning and hierarchical early warning: By constructing a communication construction safety knowledge graph, the set of hazards and control barriers corresponding to the current construction process are obtained. Causal reasoning is performed based on the real-time verification results of the control barriers, and hierarchical early warning is generated by combining the fused risk feature values. The communication construction safety knowledge graph includes operation activity nodes, hazard source nodes, accident mechanism nodes, and prevention and control barrier nodes. The accident mechanism nodes define trigger condition combinations using logical AND gates. The prevention and control barrier nodes are bound to real-time verification data fields through verification AND relations. The inference engine periodically obtains real-time data of the bound fields and determines the satisfaction status of each prevention and control barrier. When all the trigger conditions defined by the accident mechanism node are met and the prevention and control barriers used to block each trigger condition are in an unsatisfied or unknown state, the accident mechanism is activated and an early warning is generated. The logical quantities for obtaining the device security status include: The circuit breaker auxiliary contacts, grounding switch position contacts, and RF power amplifier enable relay contacts of the communication equipment are accessed through the dry contact signal acquisition module. At the same time, the device operation status register is parsed through the Modbus / SNMP network management interface to map the circuit breaker opening and closing, grounding switch status, RF power amplifier enable signal and residual voltage value in real time, forming the original status logic quantity. Using the original state logic quantity as the driving source, a safe operation state machine is established for each type of critical communication equipment. The safe operation state machine contains multiple safe states, and state transitions are triggered by a specified logic signal sequence. When the logic signal sequence does not meet the preset order or a state bounce occurs, it is determined that the isolation has failed or the safe state has failed. The name of the current safe state of the state machine is merged with the original state logic quantity to form the equipment safe state logic quantity. The processing steps of the multimodal fusion model include: The skeletal key point sequence, seat belt double hook status sequence, indicator light status sequence and dangerous area intrusion event in the visual monitoring data are encoded by a time-series network. The environmental and equipment operating parameters are upgraded by a multilayer perceptron and linearly mapped after one-hot encoding of the equipment safety status logic quantity. The process status information is transformed by an embedding layer, unified into an embedding vector of the same dimension and superimposed with position encoding. Using visual embedding vectors as queries and environmental and equipment parameter embedding vectors as keys and values, a cross-attention mechanism is used to extract the coupled representation of personnel actions and equipment external states. The embedding vectors of equipment safety state logic quantities are passed through a fully connected layer and an activation function to generate safety suppression gating factors and unisolated enhancement gating factors, and the coupled representations are gated and modulated. The process embedding vectors are injected into each intermediate layer of the encoder, and finally mapped through the output layer to obtain 0-1 continuous fused risk feature values.

2. The method according to claim 1, characterized in that, The key communication equipment includes RRU, base station power distribution unit, battery pack, antenna feeder surge arrester and coupler, and integrated cabinet power module. Each type of equipment has a separately adapted safety operation state machine based on its power form and isolation requirements. Among them, the safety operation state machine of RRU includes four safety states: "normal operation", "RF shutdown", "feeder grounding" and "residual charge discharge completed". The state transition is jointly triggered by the RF power amplifier enable signal, circuit breaker auxiliary contact, grounding switch auxiliary contact, residual voltage detection value or preset discharge timing signal.

3. The method according to claim 1, characterized in that, The gating modulation method is as follows: let the coupling representation of personnel action and equipment external state be C, the unisolated enhancement gating be G1, and the safety suppression gating be G2, then the output representation is C′=C⊙(1+βG1)⊙(1-λG2), where β and λ are preset weight coefficients; when the equipment is not reliably isolated, G1 approaches 1, and the risk representation is enhanced through the (1+βG1) term; when the equipment is safely isolated, G2 approaches 1, and the risk representation triggered by external sensing data is suppressed through the (1-λG2) term.

4. The method according to claim 1, characterized in that, The knowledge graph for communication construction safety is constructed as follows: Define four types of entity nodes: work activities, hazard sources, accident mechanisms, and prevention and control barriers. Among them, the work activity node is refined to the process level and a perception parameter configuration table containing detection frame rate, confidence threshold, sensitivity level and sampling period is preset. Establish a system that connects operational activities to hazards, leads to hazard mechanisms, blocks accident mechanisms, and records the trigger condition indexes of the blocked mechanisms. Verify the connection between the hazard mechanisms and the specific sensing data fields. The verification results of the control barrier are updated periodically with three states: satisfied, not satisfied, and unknown. The unknown state is equivalent to not satisfied and participates in risk derivation according to the conservative safety strategy.

5. The method according to claim 1, characterized in that, The generation of graded early warning includes: when the accident mechanism is activated by causal reasoning, the risk feature value R is weighted and fused with the pre-set severity score S of the accident mechanism node extracted from the knowledge graph, and the comprehensive risk score F = α·R + (1-α)·S is calculated. The F value is mapped to the preset threshold range to which it belongs to the early warning level from level I to level V, where α is an adjustable weight.

6. The method according to claim 1, characterized in that, The process causal reasoning and hierarchical early warning also includes a dynamic adjustment step for sensing parameters: Based on the current process type, the inference engine reads parameters from the pre-set perception parameter configuration table of the corresponding job activity node in the knowledge graph, and sends parameter adjustment instructions to the visual perception module and sensor acquisition module through the edge computing node, so that the detection frame rate, model confidence threshold, alarm sensitivity and sensor sampling period take effect in the subsequent acquisition window.

7. The method according to claim 1, characterized in that, Also includes: Hidden risk chain mining and rule solidification: Collect historical early warning records and near miss event data to construct a heterogeneous graph containing personnel attribute nodes, equipment model nodes, environmental range nodes, work activity nodes and event nodes, and establish co-occurrence edges between nodes that appear together in the same operation or event; A heterogeneous graph Transformer model is adopted, which maps heterogeneous nodes to a unified dimension through a type-specialized projection matrix, stacks multi-head attention layers for message passing, and conducts supervised training with the probability of a real hidden danger occurring in this task as the prediction target. After training, attention weights are tracked for positive samples, edges with weights greater than a threshold are extracted to generate multi-hop paths, and high-frequency paths are clustered and counted as candidate latent risk chains. After experts have confirmed the causal rationality, the candidate paths are compiled into new causal rules and added to the communication construction safety knowledge graph.

8. The method according to claim 7, characterized in that, The method of compiling candidate paths into new causal rules is as follows: add a prevention and control barrier node under the corresponding operation activity node, add trigger conditions in the relevant accident mechanism node, and establish verification and relationship binding for the new barrier to the corresponding perception data field, synchronously write to the graph database, recompile the inference path table of the affected process, and update the inference engine rule table cache.

Images