Method for continuously improving potential safety hazard identification capability of third-party construction

Abstract

The invention discloses a method for continuously improving the potential safety hazard identification capability of third-party construction, and belongs to the technical field of safety monitoring and artificial intelligence. The method comprises the following steps: acquiring multi-dimensional dynamic data such as a video and a sensor of a construction site, performing real-time hidden danger identification based on the data to determine a current hidden danger state, predicting a future risk upgrading probability through a risk evolution prediction model, and then fusing the current hidden danger state with the future risk upgrading probability to obtain a risk upgrading result. According to the method, the problems of alarm lag and low accuracy in the prior art are solved, the conversion from passive response to active early warning is realized by introducing risk evolution prediction and a dynamic alarm decision mechanism, and the risk evolution prediction and dynamic alarm decision mechanism is improved. And the system is endowed with a self-evolution capability through a data closed loop, so that the accuracy, the foresight and the intelligent level of safety management and control are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of safety monitoring technology and artificial intelligence technology, specifically a method for continuously improving the ability of third parties to identify construction safety hazards. Background Technology

[0002] As the nation's energy arteries and critical infrastructure, the safe and stable operation of oil and gas pipelines is of paramount importance. Among the many factors threatening pipeline safety, accidental damage caused by third-party construction (such as excavation, drilling, and piling) is the leading cause of pipeline leaks and even explosions. Therefore, timely and effective monitoring and early warning of third-party construction operations along pipelines is a core aspect of pipeline operation and maintenance.

[0003] Currently, to prevent such risks, the main technical means commonly used in the industry include traditional manual inspections on foot or by vehicle, and the deployment of fixed-point video surveillance systems in critical pipe sections or high-risk areas. Manual inspections, as the most basic protective measure, can, to some extent, detect and stop illegal construction activities. However, manual inspections are not only costly and inefficient, but their cyclical nature also makes them unable to effectively monitor sudden, short-term hazardous operations, resulting in blind spots and time delays.

[0004] To compensate for the shortcomings of manual inspections, the industry has gradually introduced technologies such as video surveillance and drone patrols. While these technologies have increased the coverage and frequency of monitoring, their practical application has also revealed new challenges. The massive amounts of unstructured video data place a huge burden on manual monitoring in the background, and traditional image analysis algorithms, such as simple motion detection, are prone to generating numerous false alarms due to irrelevant factors in the environment, such as animal activity, changes in lighting, and weather conditions, severely interfering with the normal work of security personnel. Even with the addition of vibration and sound sensors to assist in judgment, the lack of a deep understanding of complex construction scenarios makes it difficult to effectively distinguish between normal production activities and potentially threatening high-risk construction behaviors, resulting in consistently low alarm accuracy.

[0005] More importantly, the aforementioned existing technological solutions are essentially "passive response" monitoring. Their core logic lies in identifying explicit risks that have already occurred, such as an excavator encroaching on a pipeline's legally protected area or having already begun excavation work. However, these technologies lack the ability to effectively predict the dynamic evolution of risks. A construction activity that currently appears to be in a compliant location and of low risk may rapidly escalate into a major safety hazard in the near future due to minor changes in operation, unexpected movement of construction machinery, or unforeseen geological conditions. Existing technologies cannot provide forward-looking warnings for this, often triggering alarms only after the risk is imminent or has already caused damage, thus missing the optimal time for intervention.

[0006] Furthermore, the capabilities of existing systems are typically static. Their built-in alarm rules and analytical models become fixed after deployment, unable to learn and optimize from massive amounts of daily monitoring data and real-world, manually confirmed hazard events. This lack of effective data feedback loops and continuous learning mechanisms leads to a stagnation in the system's identification and early warning capabilities, making it unable to adapt to the constantly changing risk patterns brought about by new construction machinery and processes, and hindering the continuous improvement of its overall safety performance. Summary of the Invention

[0007] To address the aforementioned issues, this invention provides a method for continuously improving the ability of third parties to identify construction safety hazards. This method introduces a risk evolution prediction mechanism and constructs a data-driven closed-loop self-optimization system, thereby achieving accurate, forward-looking, and intelligent management and control of safety hazards.

[0008] To achieve the above objectives, this invention provides a method for continuously improving the ability of third parties to identify construction safety hazards, which is implemented through the following technical solution:

[0009] This method includes the following steps:

[0010] S1: Obtain multi-dimensional dynamic data from the construction site.

[0011] In one embodiment, this step is achieved by constructing a three-dimensional monitoring network consisting of a fixed sensing layer, a mobile sensing layer, and a distributed sensing layer. The fixed sensing layer can be deployed in key locations such as valve chambers and pressure regulating stations; the mobile sensing layer can utilize drones or inspection vehicles; and the distributed sensing layer can deploy vibration or strain sensors near pipelines. Through this network, comprehensive video image data and sensor data reflecting the construction site status can be collected, providing a foundation for subsequent integrated analysis.

[0012] S2: Based on the multi-dimensional dynamic data, perform real-time hazard identification to determine the real-time hazard status, and predict the risk evolution trend of the construction site.

[0013] This step is one of the core elements of this invention, and it comprises two aspects:

[0014] First, real-time hazard identification is performed to determine the real-time hazard status. In one embodiment, the system performs target detection and behavior analysis on the collected video image data to identify key targets such as construction machinery and personnel, as well as their behaviors. This is then cross-validated using sensor data to confirm the authenticity and intensity of the construction activities, ultimately forming a real-time hazard status characterizing the current overall construction situation.

[0015] Secondly, the evolution trend of risks at the construction site is predicted. This is a key innovation of the invention, giving the method a forward-looking nature. In one embodiment, this prediction is achieved through a pre-defined risk evolution prediction model. This model considers not only the current real-time hazard status but also historical hazard data related to the construction site to predict the probability that the current situation will evolve into a high-level risk event within a pre-defined time window. The probability P of this risk escalation is... risk The following formula can be used for calculation:

[0016] P risk (t+Δt|S t H loc )=f pred (S t H loc );

[0017] Among them, S t H represents the real-time hazard status at time t. loc f is a vector of historical data related to the construction site. pred For the risk evolution prediction model, t is the current time, and Δt is the future time increment within the preset time window.

[0018] S3: Generate dynamic graded alarms based on the real-time hazard status and the predicted risk evolution trend.

[0019] This step enables the alarm mechanism to be dynamic and intelligent. In one embodiment, alarm generation is a dual evaluation process:

[0020] First, the system will assess the current real-time potential hazard status based on preset pipeline attributes, construction attributes, and environmental attributes, and determine a basic alarm level.

[0021] Then, the system integrates the predicted risk evolution trend obtained in step S2 and dynamically adjusts the basic alarm level to obtain the final alarm level. For example, even if the basic alarm level is not high, if the predicted probability of future risk escalation is extremely high, the system will still raise the final alarm level. This final alarm level A final It can be determined by the following formula:

[0022] A final =g final (L base ,P risk )

[0023] Among them, L base For the basic alarm level, P risk g represents the probability of the risk escalating. final This is a pre-defined decision function used to fuse two inputs to output the final decision.

[0024] In addition, the method of the present invention may also include step S4: establishing a closed loop of data throughout the entire lifecycle.

[0025] To continuously improve the system's capabilities, this invention also provides a closed-loop optimization mechanism. In one embodiment, the system stores the predicted results of each real-time hazard status and risk evolution trend, along with subsequent manual review conclusions and actual handling results, into a hazard database. This accumulated data is used to iteratively optimize the model within the system. For example, it uses samples with real labels to incrementally train or fine-tune the identification model used for real-time hazard identification; simultaneously, it uses sample pairs containing predicted results and actual handling results to update the parameters of the risk evolution prediction model by minimizing the prediction loss function, making its prediction capabilities increasingly accurate.

[0026] This invention provides a method for continuously improving the ability of third parties to identify construction safety hazards. It has the following beneficial effects:

[0027] 1. This invention significantly improves the reliability of hazard identification by establishing a cross-validation mechanism based on multi-dimensional data, effectively eliminating the problem of rampant false alarms caused by misjudgments from a single information source. In existing technologies, harmless shadow movements or animal activity are easily misjudged as construction threats based solely on video footage. However, in this invention, the system compares visually identified excavator activity with vibration sensor data at that location. Only when the spectral characteristics of the vibration signal match the pattern of large machinery operation is it confirmed as a genuine hazard. This rigorous logical verification frees security personnel from a deluge of invalid alarms, allowing them to focus on truly noteworthy events.

[0028] 2. The most significant contribution of this invention lies in its transformation of the safety monitoring paradigm from the traditional "post-event response" to "pre-event prediction" by introducing a risk evolution prediction model. The system no longer merely detects a violation that has already occurred; instead, it proactively calculates the probability that seemingly compliant operations will evolve into high-risk events in the future, based on subtle dynamics of the current scenario and historical data. This provides managers with valuable golden intervention time, enabling them to take action before potentially destructive consequences materialize, achieving truly proactive and preventative safety management.

[0029] 3. The innovative dynamic hierarchical alarm logic of this invention enables alarm levels to accurately reflect the true urgency of an event. Traditional alarm systems often rely on rigid, spatially distance-based rules, while the fusion decision function designed in this invention organically combines the assessment of the current state (basic alarm level) with the prediction of future risks. A construction activity that is currently outside the safe distance but rapidly approaching the pipeline will be immediately assigned the highest level alarm due to its extremely high probability of future risk, thus ensuring that the most dangerous dynamic threats receive the highest priority attention and handling immediately.

[0030] 4. This invention, by introducing vectors containing historical information about the construction location as input to the prediction model, endows the system with the ability to conduct scenario-based and differentiated risk assessments. For areas with soft geology or past subsidence, or special pipe sections containing high-pressure, high-value pipelines, the system assigns higher risk weights during risk prediction. This analytical approach, combining static historical environmental risks with dynamic real-time construction behavior, makes risk assessment no longer one-size-fits-all, but allows for personalized and refined adjustments based on the specific circumstances of different locations, greatly improving the accuracy of the assessment.

[0031] 5. This invention designs a complete, full-lifecycle data closed-loop optimization process, ensuring that the system's performance can continuously evolve during use. Every piece of data from the generation of an alarm to its final manual confirmation as a real hazard or a false alarm is used as valuable training samples to feed back into the identification and prediction model. This adaptive learning mechanism enables the system to continuously learn from experience, automatically adapting to new types of construction machinery, constantly changing operating methods, and various complex environmental backgrounds, thus ensuring its long-term effectiveness and advanced nature. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0033] Figure 2 This is a flowchart of the method of the present invention;

[0034] Figure 3 This is a schematic diagram of the internal logic of the present invention;

[0035] Figure 4 This is a schematic diagram of the decision logic of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0037] See attached document Figure 1 The present invention provides a system for continuously improving the ability of third parties to identify construction safety hazards. The system can be physically built on a hybrid computing architecture that is "edge-cloud" collaborative, and aims to realize distributed data processing and centralized intelligent analysis.

[0038] Logically, the system may include a front-end perception layer (a), an edge processing unit (b), and a cloud platform (c).

[0039] The front-end perception layer (a) is the data source for the entire system, and its core task is to capture comprehensive, multimodal information from the construction site. This front-end perception layer (a) integrates various types of hardware devices, specifically including: fixed sensing devices deployed at key locations along the pipeline, such as high-definition dome cameras with infrared night vision and 360-degree rotation capabilities; mobile sensing devices performing large-scale periodic inspections, such as drones equipped with high-definition cameras and lidar; and distributed sensing devices directly deployed above or near the pipeline, such as high-sensitivity vibration sensors for monitoring mechanical operations and fiber optic strain sensors for capturing minute deformations in the pipeline. These devices together form a three-dimensional monitoring network, ensuring the comprehensiveness and diversity of the raw data.

[0040] Edge processing unit b is deployed close to the front-end perception layer a, serving as a preprocessing and rapid response node for data entering the cloud. Edge processing unit b receives the raw data stream from the front-end perception layer a and performs preliminary computational tasks. These tasks include preprocessing video image data such as denoising and compression, and real-time filtering of sensor time-series data to reduce the computational load on the cloud platform. More importantly, edge processing unit b can embed a lightweight recognition model for performing extremely low-latency preliminary judgments. For example, upon detecting a clear, extremely high-risk intrusion, it can directly trigger a local audible and visual alarm device, achieving instantaneous response.

[0041] The cloud platform c is the core computing and decision-making hub of the method of this invention, responsible for performing complex intelligent analysis and global optimization. This cloud platform c is deployed on a server cluster and integrates multiple collaborative software functional modules.

[0042] In one specific embodiment, the cloud platform c includes: a data management module c1, an intelligent analysis engine c2, an alarm decision module c3, and a model optimization module c4.

[0043] The data management module c1 is responsible for receiving and processing data uploaded from various edge processing units b. It performs spatiotemporal alignment on the pre-processed multi-dimensional dynamic data, that is, unifies data from different sources and timestamps under the same spatiotemporal benchmark, and stores it in a structured manner in the hidden danger event database, providing a unified and clean data foundation for subsequent analysis and model training.

[0044] The intelligent analysis engine c2 is the functional carrier that executes the core algorithm of this invention. This engine integrates an identification model for real-time hazard identification and a prediction model for predicting risk evolution trends. It calls upon data prepared by the data management module c1 to first determine the real-time hazard status at the construction site. Then, based on this status and historical data, it uses its built-in prediction model to calculate the probability of future risk escalation. The specific calculation principles and formulas involved in this engine will be explained in detail in the subsequent description of the methodological process.

[0045] The alarm decision module c3 receives the real-time hazard status and future risk probability from the intelligent analysis engine c2. This module has a built-in dual evaluation decision mechanism. First, it evaluates the current real-time hazard status based on a preset static and dynamic rule base to obtain a basic alarm level. Then, it merges this basic alarm level with the predicted future risk probability, dynamically adjusts it through a decision function, and finally generates an accurate and forward-looking final alarm level, and pushes the structured alarm information to the corresponding management terminal.

[0046] The model optimization module c4 is crucial for continuously improving the capabilities of this invention. This module periodically extracts new data samples labeled with actual handling results from the hazard database of the data management module c1. Using these samples, it can automatically or semi-automatically retrain and fine-tune the parameters of the identification and prediction models in the intelligent analysis engine c2, forming a complete closed loop of "data acquisition – analysis and alarm – result feedback – model optimization." This ensures that the entire system's identification and prediction capabilities can adapt to environmental changes and new risk patterns, achieving continuous self-evolution.

[0047] The following will describe in detail the specific process of a method for continuously improving the ability of third parties to identify construction safety hazards in an embodiment of the present invention. This method can be executed by the cloud platform c in the aforementioned system architecture, and works in collaboration with the front-end perception layer a and the edge processing unit b.

[0048] Reference Figures 2 to 4 In this embodiment, step S1 of the method, namely acquiring multi-dimensional dynamic data of the construction site, has been described in the aforementioned system architecture. Its core lies in capturing data through a three-dimensional monitoring network and performing preliminary processing through an edge processing unit. After the data is uploaded to the cloud platform c, the data management module c1 first performs in-depth preprocessing. Specifically, for video image data, the system applies spatiotemporal alignment technology to unify images from different fixed cameras and mobile drones into the same geographic coordinate system, generating a dynamic panoramic view of the construction area; for sensor data, precise timestamp synchronization and normalization are performed to facilitate fusion with visual data.

[0049] In this embodiment, step S2 of the method, namely, real-time hazard identification and risk evolution trend prediction based on multi-dimensional dynamic data, is executed by the intelligent analysis engine c2. This step constitutes the core technology of the present invention. First, in order to determine the real-time hazard status, the recognition model in the engine analyzes the preprocessed video data. This recognition model can adopt a deep learning network based on the YOLOv8 architecture and incorporating an attention mechanism. This network has been trained on a large number of construction scene samples and can accurately identify various targets, including but not limited to excavators, pile drivers, and construction workers, and track their movement trajectories. Subsequently, the system performs multi-source data cross-validation. For example, when the recognition model detects an excavator operating within a pipeline protection zone in the video, the system immediately queries the vibration sensor data corresponding to that geographical location. Only when the amplitude and spectral characteristics of the vibration signal match the pattern of large machinery operation is the event confirmed as a high-confidence real-time hazard and encapsulated into a structured real-time hazard state vector S. t The vector S t It contains complete scene information at the current time t, and its specific structure can be represented as follows:

[0050] S t ={O t ,L t V t E t}

[0051] Among them, O t The target object set includes the category, quantity, and behavioral state of all identified targets; L t This is a set of spatial locations, recording the three-dimensional coordinates of each target and its relative distance to the pipeline; Vt This is a set of sensor readings, containing readings of key physical quantities such as vibration and strain at this moment; E t It is a set of environmental parameters, encompassing background information such as weather, lighting, and construction type.

[0052] Next, the intelligent analytics engine c2 initiates the prediction of risk evolution trends. This process is driven by a built-in risk evolution prediction model f. pred Completed, this model can employ a Long Short-Term Memory (LSTM) network architecture, as it excels at handling and predicting long-term dependencies in time series data. The prediction model f pred The input includes not only the real-time hazard state vector S at the current moment t It also includes a historical data vector H extracted from a database of potential hazards and related to the current construction location. loc The vector H loc It encapsulates static and semi-static information about the location, including geological conditions, past hazard records, and pipeline maintenance history. The model, through comprehensive analysis of these two vectors, predicts future events within a predetermined time window T. horizon Within this context, the probability P of the current construction status evolving into a high-level risk event (such as pipeline damage or leakage) is... risk This prediction process can be precisely described by the following formula:

[0053] P risk (t+Δt|S t H loc )=f pred (S t H loc ), where 0 < Δt ≤ T h orizom

[0054] This formula reveals one of the core innovations of this method: by learning historical patterns, it enables quantitative prediction of future risks.

[0055] In this embodiment, step S3 of the method, namely generating dynamic hierarchical alarms, is executed by the alarm decision module c3. This module first determines the alarm decision based on the current real-time hazard status vector S. t The information is used to calculate a basic alarm level L using a three-dimensional dynamic threshold model. base This model is a rule-based and multi-dimensional parameter query system that comprehensively evaluates the importance of the pipeline itself (by C). pipe Vector definition, including pressure, material, etc.), compliance of construction activities and potential damage (by C) work Vector definition, including construction method, whether reporting is required, etc.) and the vulnerability of the surrounding environment (by C). envThe vector definition (including population density, important facilities, etc.) is calculated as L. base =g base (C pipe C work C env ).

[0056] Subsequently, the alarm decision module c3 executes a more critical fusion decision step. It will adjust the aforementioned basic alarm level L... base The probability P of future risks predicted by the intelligent analysis engine c2 risk The combination is performed through a pre-defined decision function g. final To determine the final alarm level A final The decision function A final =g final (L base ,P risk The inherent logic of L is that even base The assessment is at a lower level (e.g., compliant construction and distance is still within a safe range), but if P risk This indicates a very high probability that the event will evolve into a high-risk event in the near future (e.g., predicting that geological subsidence will lead to pipeline stress exceeding limits), and the decision function g final A will also be final It is directly upgraded to the highest level. This mechanism ensures that the alarm system can not only respond to the "present" but also warn of the "future," achieving a balance between accuracy and foresight in alarms.

[0057] Finally, this embodiment also includes a closed-loop optimization step, which is executed by the model optimization module c4. This includes the entire lifecycle data of each alarm event from generation to completion of handling, including the initial S... t Predicted P risk The final A final The actual results y (whether the event is a real hidden danger and whether it ultimately leads to an accident), as well as subsequent manual intervention confirmation, will be fully recorded in the hidden danger event database. The model optimization module c4 will periodically call upon this valuable new data to iterate on the two core models within the intelligent analysis engine c2. For the identification model, the module adopts supervised learning, using samples with real labels for fine-tuning. For the prediction model f... pred The module then minimizes the prediction loss function Loss pred To update its network parameters, the loss function can be in the form of binary cross-entropy:

[0058]

[0059] Where N is the total number of samples used for this round of optimization, y iLet P be the true result label for the i-th sample. risk,i Let be the predicted risk probability of the model for the i-th sample. Through this closed-loop optimization, the overall performance of the system can be continuously and specifically improved in practical applications.

[0060] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for continuously improving the ability of third parties to identify construction safety hazards, characterized in that, Includes the following steps: S1: Acquire multi-dimensional dynamic data from the construction site; S2: Based on the multi-dimensional dynamic data, perform real-time hazard identification to determine the real-time hazard status, and predict the risk evolution trend of the construction site; S3: Generate dynamic graded alarms based on the real-time hazard status and the predicted risk evolution trend.

2. The method for continuously improving the ability of third parties to identify construction safety hazards according to claim 1, characterized in that, In step S1, multi-dimensional dynamic data of the construction site is acquired, including: A three-dimensional monitoring network consisting of a fixed sensing layer, a mobile sensing layer, and a distributed sensing layer is used to collect video image data and sensor data.

3. The method for continuously improving the ability of third parties to identify construction safety hazards according to claim 1, characterized in that, In step S2, real-time hazard identification is performed to determine the real-time hazard status, specifically including: Target detection and behavior analysis are performed on the video image data in the multi-dimensional dynamic data; and cross-validation is performed in combination with sensor data to determine the real-time hazard status.

4. The method for continuously improving the ability of third parties to identify construction safety hazards according to claim 3, characterized in that, In step S2, predicting the risk evolution trend of the construction site specifically includes: Based on a preset risk evolution prediction model, and combined with the real-time hazard status and historical hazard data, the probability of risk escalation within a preset time window is predicted.

5. The method for continuously improving the ability of third parties to identify construction safety hazards according to claim 4, characterized in that, The probability P of risk escalation risk Calculated using the following formula: P risk (t+Δt|S t ,H loc )=f pred (S t ,H loc ); Among them, S t H represents the real-time hazard status at time t. loc f is a vector of historical data related to the construction site. pred This refers to the risk evolution prediction model.

6. The method for continuously improving the ability of third parties to identify construction safety hazards according to claim 1, characterized in that, In step S3, generating dynamic hierarchical alarms specifically includes: A basic alarm level is determined based on preset pipeline attributes, construction attributes, and environmental attributes; By integrating the predicted results of the risk evolution trend, the basic alarm level is dynamically adjusted to obtain the final alarm level.

7. The method for continuously improving the ability of third parties to identify construction safety hazards according to claim 6, characterized in that, The final alarm level A final Determined by the following formula: A final =g final (L base ,P risk ); Among them, L base For the basic alarm level, P risk g represents the probability of the risk escalating. final This is a pre-defined decision function.

8. The method for continuously improving the ability of third parties to identify construction safety hazards according to claim 1, characterized in that, The method further includes step S4: S4: Establish a closed-loop data system for the entire lifecycle, store the real-time hazard status, the predicted results of the risk evolution trend, the conclusions of manual review, and the handling results into the hazard database, and use the data in the hazard database to iteratively optimize the risk evolution prediction model and the identification model used for real-time hazard identification.

9. The method for continuously improving the ability of third parties to identify construction safety hazards according to claim 8, characterized in that, The iterative optimization of the risk evolution prediction model specifically includes: The parameters of the risk evolution prediction model are updated by minimizing the prediction loss function using sample pairs formed by the prediction results and the treatment results in the hazard database.

10. The method for continuously improving the ability of third parties to identify construction safety hazards according to claim 8, characterized in that, The iterative optimization of the identification model used for real-time hazard identification specifically includes: The identification model is incrementally trained or fine-tuned using samples with real labels from the hazard database.

Images