Engineering supervision potential safety hazard early warning method and system based on edge calculation

Abstract

The invention provides a project supervision potential safety hazard early warning method and system based on edge calculation. The method comprises the following steps: acquiring multi-source heterogeneous sensing data acquired by a plurality of monitoring points in a target construction area; performing feature fuzzy extraction on the multi-source heterogeneous sensing data to identify a current construction condition and a dominant risk factor, and determining a to-be-executed potential safety hazard detection task according to the identified construction condition and risk factor; adaptively selecting a target model corresponding to the potential safety hazard detection task from a plurality of lightweight recognition models pre-deployed in the edge computing device for activation; and in the edge computing device, performing forward reasoning on the real-time multi-source heterogeneous sensing data of the corresponding monitoring point by using the activated target model, and generating a potential safety hazard identification result. By adopting the scheme of the invention, the edge-side adaptive model scheduling reasoning based on working condition sensing can be realized, so that the potential safety hazard early warning efficiency of project supervision in a resource-constrained environment is improved.

Description

Technical Field

[0001] This application relates to the field of safety early warning technology, and more specifically, to a method and system for early warning of safety hazards in engineering supervision based on edge computing. Background Technology

[0002] The main contents of engineering supervision work include continuous monitoring and risk control of the work behavior of personnel, the operating status of machinery and equipment, the execution of construction procedures, and environmental safety conditions at the construction site, so as to ensure that construction activities comply with safety regulations and technical standards. Engineering supervision safety early warning is the core link to ensure the safety of personnel and equipment at the construction site. Traditional early warning methods mainly rely on manual inspection and fixed sensors, which have problems such as large coverage blind spots and delayed response. In recent years, although automatic monitoring technology based on the Internet of Things and artificial intelligence has been applied, how to achieve real-time, accurate and intelligent early warning that can adapt to complex changes in the construction environment is still a key breakthrough that the industry urgently needs to make.

[0003] In existing technologies, real-time monitoring of engineering supervision safety is achieved by deploying multiple identification algorithms at the edge. The initial design aims to assist engineering supervisors in automatically identifying and alerting to key safety elements at construction sites, thereby reducing the intensity of manual supervision and increasing regulatory coverage. However, this typically employs a static model-calling strategy, which has significant drawbacks in resource-constrained edge computing environments. Regardless of the actual construction conditions and risk focus, the system often needs to continuously or alternately run all or a large number of pre-built models, resulting in the average consumption of limited computing, memory, and energy resources. Consequently, inference resources for immediate high-risk tasks are diluted, and overall system response latency increases, while a large number of unnecessary low-priority model operations continue to consume resources. This makes it difficult to meet the technical requirements of engineering supervision for rapid identification and differentiated supervision of key risks, and fails to achieve an optimal balance between computational efficiency and early warning accuracy. Therefore, how to achieve edge-side adaptive model scheduling and inference based on work condition awareness to improve the early warning efficiency of safety hazards in resource-constrained engineering supervision has become a challenge for the industry. Summary of the Invention

[0004] This application provides a method and system for early warning of safety hazards in engineering supervision based on edge computing. It can realize edge-side adaptive model scheduling and reasoning based on working condition perception, thereby improving the early warning efficiency of safety hazards in engineering supervision under resource-constrained environments.

[0005] Firstly, this application provides a method for early warning of safety hazards in engineering supervision based on edge computing, including: Acquire multi-source heterogeneous sensing data collected from multiple monitoring points within the target construction area; Feature fuzzy extraction is performed on multi-source heterogeneous sensing data to identify the current construction conditions and dominant risk factors. Based on the identified construction conditions and risk factors, the safety hazard detection tasks to be performed are determined. The system adaptively selects and activates a target model corresponding to the security hazard detection task from multiple lightweight identification models pre-deployed in edge computing devices; wherein the multiple lightweight identification models are designed with differentiated optimization for different construction stages or different risk types. In the edge computing device, the activated target model is used to perform forward reasoning on the real-time multi-source heterogeneous sensing data of the corresponding monitoring points to generate a safety hazard identification result. Based on the safety hazard identification results, it is determined whether there are safety hazards at the corresponding monitoring points, and when there are safety hazards, an early warning message is generated and pushed to the engineering supervision terminal.

[0006] Preferably, the multi-source heterogeneous sensing data specifically includes: video stream data collected by visual sensors deployed at monitoring points, and ambient sound and vibration spectrum data collected by acoustic sensors deployed at monitoring points.

[0007] Preferably, feature fuzzy extraction is performed on multi-source heterogeneous sensing data to identify the current construction conditions and dominant risk factors, specifically including: The collected video stream data is matched with preset typical construction scene templates using lightweight feature matching to obtain the scene matching degree; Based on the collected environmental acoustic and vibration spectrum data, identify the types of critical operations that are in progress; By integrating the scene matching degree with the identified key operation type, and using a pre-built lightweight fuzzy inference rule base, the membership degree of the current operation to different preset construction conditions is calculated. The preset construction condition with the highest membership degree is determined as the current construction condition, and the high-incidence risk factors in the historical statistics under this condition are taken as the dominant risk factors.

[0008] Preferably, based on the identified construction conditions and risk factors, the specific safety hazard detection tasks to be performed include: Obtain at least one basic inspection task that needs to be performed periodically under the aforementioned construction conditions; Based on the aforementioned risk factors, at least one specific detection task requiring key monitoring is obtained from a predefined task mapping table; The at least one basic detection task and the at least one special detection task are merged and deduplicated to obtain a safety hazard detection task to be executed, wherein the safety hazard detection task includes at least one detection task.

[0009] Preferably, adaptively selecting and activating a target model corresponding to the security hazard detection task from multiple lightweight recognition models pre-deployed in edge computing devices specifically includes: Extract the task identifier for each detection task in the aforementioned safety hazard detection task; Based on the task identifier, a query is performed in the model index table stored locally on the edge computing device. The model index table records the mapping relationship between different task identifiers and the corresponding lightweight recognition model storage path. Based on the queried storage path, the corresponding model parameters are loaded from the local storage space of the edge computing device into memory, and the model loaded into memory that corresponds to the detection task is used as the target model for this activation.

[0010] Preferably, in the edge computing device, the activated target model is used to perform forward inference on the real-time multi-source heterogeneous sensing data of the corresponding monitoring points to generate a safety hazard identification result, specifically including: Based on the detection task corresponding to each activated target model, the corresponding type and time sequence of input data segments are extracted from the real-time multi-source heterogeneous sensing data; Each input data segment undergoes standardized preprocessing to meet the input requirements of the corresponding target model; Each standardized preprocessed input data segment is fed into the corresponding target model, and forward inference computation is performed in parallel on the computing unit of the edge computing device. The identification results of each target model, which represent the risk status under the corresponding detection task, are summarized to generate the safety hazard identification results.

[0011] Preferably, determining whether a safety hazard exists at a corresponding monitoring point based on the safety hazard identification results specifically includes: Each of the safety hazard identification results is compared with a preset threshold for the corresponding risk type. If any identification result exceeds its corresponding judgment threshold, it is determined that there is a specific type of hidden danger. Based on the risk level and location corresponding to all the out-of-limit identification results, a comprehensive hazard assessment conclusion is generated.

[0012] Secondly, this application provides an engineering supervision safety hazard early warning system based on edge computing, comprising: The acquisition module is used to acquire multi-source heterogeneous sensing data collected from multiple monitoring points within the target construction area; The processing module is used to perform feature fuzzy extraction on multi-source heterogeneous sensing data to identify the current construction conditions and dominant risk factors, and to determine the safety hazard detection tasks to be performed based on the identified construction conditions and risk factors. The processing module is also used to adaptively select and activate a target model corresponding to the security risk detection task from multiple lightweight recognition models pre-deployed in the edge computing device; The processing module is also used in the edge computing device to perform forward reasoning on the real-time multi-source heterogeneous sensing data of the corresponding monitoring points using the activated target model to generate a safety hazard identification result. The push module is used to determine whether there is a safety hazard at the corresponding monitoring point based on the safety hazard identification results, and to generate early warning information and push it to the engineering supervision terminal when there is a safety hazard.

[0013] Thirdly, this application provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described edge computing-based engineering supervision safety hazard early warning method.

[0014] Fourthly, this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned edge computing-based engineering supervision safety hazard early warning method.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: In this embodiment, firstly, feature fuzzy extraction is performed on multi-source heterogeneous sensing data to identify the current construction conditions and dominant risk factors. Feature fuzzy extraction enables flexible state perception of complex and dynamic construction sites with extremely low computational overhead, providing the system with a precise and lightweight decision context. This fundamentally solves the problems of rigid identification and misjudgment caused by environmental uncertainty in traditional methods. Secondly, based on the identified construction conditions and risk factors, a safety hazard detection task to be executed is determined. Determining the safety hazard detection task transforms abstract conditions and risk semantics into a specific, executable set of computational instructions, thereby bridging the perception layer and the execution layer and ensuring a high degree of alignment between the system's monitoring focus and the current actual high-risk points. Then, multiple lightweight identification modules pre-deployed in edge computing devices... The proposed solution adaptively selects and activates the target model corresponding to the safety hazard detection task. This adaptive model selection enables precise dynamic scheduling of computing resources. Based on the real-time decision context, it activates only a very small subset of models strongly relevant to the current task, thereby maximizing the conservation of limited memory and computing power on edge devices. This overcomes the resource waste and response latency bottlenecks caused by existing static or polling model invocation strategies. Finally, in the edge computing device, the activated target model performs forward inference on the real-time multi-source heterogeneous sensing data of the corresponding monitoring points to generate safety hazard identification results. This achieves high real-time performance, high targeting, accurate hazard identification, and closed-loop early warning under optimal resource allocation, enabling the entire system to maintain efficient concurrent processing capabilities and rapid emergency response speed even with limited computing power. In summary, this application's solution can realize edge-side adaptive model scheduling and inference based on working condition perception, thereby improving the safety hazard early warning efficiency of engineering supervision in resource-constrained environments. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating an application scenario of an edge computing-based engineering supervision safety hazard early warning method according to some embodiments of this application; Figure 2 This is an exemplary flowchart of an edge computing-based safety hazard early warning method for engineering supervision, as shown in some embodiments of this application. Figure 3 This is a schematic diagram of the feature fuzzy extraction process according to some embodiments of this application; Figure 4 This is a schematic diagram of the structure of an edge computing-based engineering supervision safety hazard early warning system according to some embodiments of this application; Figure 5 This is a schematic diagram of the structure of a computer device for implementing an edge computing-based engineering supervision safety hazard early warning method, according to some embodiments of this application. Detailed Implementation

[0017] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Figure 1 This diagram illustrates an application scenario of an edge computing-based safety hazard early warning method for engineering supervision, based on some embodiments of this application. The diagram includes monitoring equipment, an edge computing device, a communication network, and a terminal. The monitoring equipment is connected to the edge computing device via a data cable, and the terminal is connected to the edge computing device via the communication network. The edge computing device acquires real-time multi-source heterogeneous sensing data provided by various monitoring devices. After performing feature fuzzy extraction and preprocessing on the data, it sequentially executes construction condition identification, risk element extraction, safety hazard detection task determination, lightweight identification model adaptive selection and activation, and forward inference calculation. Finally, it generates a safety hazard identification result and early warning decision for the current construction scenario. When a safety hazard is identified, the edge computing device pushes the early warning information, risk location, on-site image fragments, and relevant confidence levels to the terminal in real time for viewing and handling by the supervising engineer, safety administrator, or project manager.

[0019] The monitoring equipment may include visual sensors, acoustic sensors, vibration sensors, and environmental parameter acquisition modules; the terminal may be, but is not limited to, a safety supervision mobile application, a web monitoring backend, a command center large screen, or a smart safety helmet terminal; the edge computing device may be an edge server, an industrial gateway, or an embedded industrial control computer with computing capabilities deployed on the project site.

[0020] refer to Figure 2 The figure is an exemplary flowchart of an edge computing-based safety hazard early warning method for engineering supervision, according to some embodiments of this application. This edge computing-based safety hazard early warning method for engineering supervision mainly includes the following steps: In step 101, multi-source heterogeneous sensing data collected from multiple monitoring points within the target construction area are acquired.

[0021] It should be noted that the monitoring point in this application refers to an intelligent Internet of Things terminal node that is deployed in a key location in the construction area, integrates visual and acoustic vibration sensing modules, and has communication capabilities. In this embodiment, the multi-source heterogeneous sensing data specifically includes: video stream data collected by visual sensors deployed at the monitoring point, and environmental sound and vibration spectrum data collected by acoustic sensors deployed at the monitoring point.

[0022] In step 102, feature fuzzy extraction is performed on the multi-source heterogeneous sensing data to identify the current construction conditions and dominant risk factors. Based on the identified construction conditions and risk factors, the safety hazard detection task to be performed is determined.

[0023] In this embodiment, reference Figure 3 As shown in the figure, this is a schematic diagram of the feature fuzzy extraction process in some embodiments of this application. In this embodiment, feature fuzzy extraction of multi-source heterogeneous sensing data to identify the current construction conditions and dominant risk factors can be achieved by the following steps: The collected video stream data is matched with preset typical construction scene templates using lightweight feature matching to obtain the scene matching degree; Based on the collected environmental acoustic and vibration spectrum data, identify the types of critical operations that are in progress; By integrating the scene matching degree with the identified key operation type, and using a pre-built lightweight fuzzy inference rule base, the membership degree of the current operation to different preset construction conditions is calculated. The preset construction condition with the highest membership degree is determined as the current construction condition, and the high-incidence risk factors in the historical statistics under this condition are taken as the dominant risk factors.

[0024] It should be noted that the core reason for the design of this application is to resolve the contradiction between resource constraints and the dynamic diversity of detection tasks in edge computing environments. By introducing a feature fuzzy extraction layer, the system constructs a low-overhead state perception and decision-making abstraction interface: First, lightweight feature matching and acoustic vibration analysis are used to achieve preliminary semanticization of multi-source data, compressing high-dimensional raw perception data into two low-dimensional features: scene matching degree and key operation type. Then, a lightweight fuzzy inference rule base is used to map discrete features into continuous membership degrees to different preset construction conditions, completing the transformation from deterministic input to probabilistic state representation. The beneficial effect of this step is that it provides a quantitative decision-making basis based on uncertainty inference for subsequent model scheduling, enabling the system to accurately activate the most relevant subset in the pre-deployed model pool according to the membership degree of real-time conditions. This not only avoids the computational and memory overhead caused by full model loading, but also ensures the pertinence and timeliness of hidden danger detection through a priori risk association mechanism. Finally, an adaptive early warning closed loop for resource perception is realized on the edge side, significantly improving the overall energy efficiency ratio of the system under limited computing power.

[0025] It should also be noted that the scene matching degree in this application is a numerical indicator used to quantify the similarity between real-time video scenes and preset typical construction scenes; the key operation type is a classification label used to describe the main mechanical or manual operation categories identified based on acoustic and vibration data in the current construction area; the membership degree is a quantitative value used to represent the probability that the current multi-source sensing data belongs to a certain preset construction condition; and the dominant risk element is used to refer to the set of safety hazard types that are most likely to occur under the current construction condition based on historical statistical data.

[0026] In specific implementation, the collected video stream data is matched with preset typical construction scene templates using lightweight feature matching. The scene matching degree can be obtained by the following method: key frames are extracted from the collected video stream data at fixed intervals and fed into a pre-trained lightweight convolutional neural network model for feature extraction, generating a feature vector for that frame. Then, the feature vector is compared with the feature vector of each template in the preset typical construction scene template library using cosine similarity calculation to obtain a set of similarity values. This set of values ​​is normalized and output as a scene matching degree vector. It should be further noted that the pre-trained lightweight convolutional neural network model adopts a proprietary network based on the MobileNetV2 architecture, and its core parameters are configured as follows: the model input is red, green, and blue (Red, Green, Blue) pixels downsampled to 224x224 pixels and normalized. The model uses a blue (RGB) image tensor. The main structure includes an initial standard convolutional layer (3x3 kernel, 2 stride, 32 output channels) and 15 inverse residual bottleneck modules. The expansion factor of the key modules is set to 6, and the number of intermediate channels is dynamically adjusted according to the layer depth, ranging from 96 to 320. A ReLU6 activation function is used uniformly to ensure numerical stability under low-precision computation. At the end of the network, a global average pooling layer compresses the feature map into a 1280-dimensional feature vector. A fully connected layer is then passed through a random deactivation layer with a dropout rate of 0.2, outputting a logical value equal to the number of preset typical construction scene categories. The model is initialized with weights pre-trained on a large construction site image dataset and fine-tuned using a professional dataset containing construction elements such as safety helmets, machinery, and support structures. The final model size is controlled within 8.7MB, supports INT8 quantization deployment, and has a single-frame inference latency of less than 45 milliseconds on edge computing devices, achieving a balance between accuracy and efficiency. This provides a reliable and lightweight feature extraction foundation for scene matching calculation in real-time video streams.

[0027] In addition, in specific implementation, the identification of the ongoing key operation type based on the collected environmental acoustic and vibration spectrum data can be achieved in the following way: Mel frequency cepstral coefficient feature extraction is performed on the synchronously collected environmental sound data, and wavelet packet energy feature extraction is performed on the vibration spectrum data. The fused feature vector is then input into a pre-trained lightweight classification model, such as a gradient boosting decision tree. This classification model outputs a probability distribution of the preset key operation type, and the type with the highest probability is determined as the currently identified key operation type. The reason for this is that acoustic and vibration signals can effectively characterize the physical characteristics of specific operations such as drilling, pouring, and cutting, thereby supplementing the deficiencies of pure visual information. The fusion of the scene matching degree and the identified key operation type, using a pre-set lightweight fuzzy inference rule base, can be used to calculate the membership degree of the current operation to different preset construction conditions. This can be achieved in the following way: A lightweight fuzzy inference rule base is established, which defines the association between different combinations of scene matching degree and key operation type and each preset construction condition in the form of "if-then". The previously obtained scene matching degree vector and key operation type probability distribution are then used as input, and the triangular membership rule in fuzzy inference is applied. The membership function and rule triggering mechanism calculate a set of membership values ​​corresponding to each preset construction condition. This is done by using fuzzy logic to handle the uncertainty and fuzziness in sensor information, achieving soft decision-making and flexible fusion of multi-source information. It should be further explained that the lightweight fuzzy inference rule base is a set of decision logic abstracted from expert experience to handle uncertainty. Its technical principle is: first, the precise input values ​​(such as scene matching degree and the probability of key operation types) are transformed into membership degrees to different fuzzy linguistic values ​​through a membership function; then, a set of membership values ​​is generated from several "such as..." The rule base consisting of "effect-then" type rules is activated, for example, "If the membership degree of the scenario matching 'foundation pit' is high and the membership degree of the operation type 'excavation' is high, then the membership degree of the work condition 'earthwork stage' is high". The system aggregates the results of all triggered rules through fuzzy synthesis operation (commonly using the minimum-maximum method). Finally, the output fuzzy set is defuzzified, for example, by the centroid method, to calculate the precise membership degree value belonging to each preset construction work condition. The entire reasoning process is implemented through efficient methods such as matrix operations, with low computational load, thus enabling real-time and flexible judgment of complex construction states on resource-constrained edge devices.

[0028] In addition, in specific implementation, determining the preset construction condition with the highest membership degree as the current construction condition and using the historically high-incidence risk factors under this condition as the dominant risk factors can be achieved in the following way: compare the membership degree values ​​of all preset construction conditions obtained by calculation, and formally determine the preset construction condition with the highest membership degree as the current construction condition. Then, the system queries the historical hidden danger database bound to the current construction condition, and directly uses the top several risks with the highest statistical frequency in the database as the dominant risk factors identified this time. The reason for doing this is based on the domain prior knowledge that "the current condition determines the main risk", which efficiently transforms the condition identification results into specific and targeted monitoring targets, thereby guiding the subsequent model scheduling.

[0029] In this embodiment, the safety hazard detection task to be performed can be determined by the following steps based on the identified construction conditions and risk factors: Obtain at least one basic inspection task that needs to be performed periodically under the aforementioned construction conditions; Based on the aforementioned risk factors, at least one specific detection task requiring key monitoring is obtained from a predefined task mapping table; The at least one basic detection task and the at least one special detection task are merged and deduplicated to obtain a safety hazard detection task to be executed, wherein the safety hazard detection task includes at least one detection task.

[0030] It should be noted that the basic detection tasks are a set of standard safety hazard detection items that need to be periodically executed to ensure basic safety under specific construction conditions; the special detection tasks are a set of safety hazard detection items that are used to conduct key and targeted monitoring of specific dominant risk factors that have been identified; and the safety hazard detection tasks are a set of final instructions to be executed, formed by merging the basic detection tasks and special detection tasks, used to guide the target recognition model in the edge computing device to activate and infer.

[0031] In practical implementation, firstly, the system maintains a mapping table between construction conditions and basic inspection tasks. This table predefines a list of inspection items that must be periodically executed for each preset construction condition. Then, the current construction condition determined by the feature fuzzy extraction step is used as the query key to retrieve the corresponding basic inspection task list from this mapping table. Secondly, the system also maintains a mapping table between risk factors and special inspection tasks. This table predefines the targeted inspection items that need to be activated when a certain dominant risk factor appears. In implementation, the system uses the identified dominant risk factor as the query key to retrieve the corresponding special inspection task list from the above mapping table. Then, the basic inspection task list obtained in the first step and the special inspection task list obtained in the second step are combined. During the combination operation, the system compares the unique identifiers of all inspection tasks and automatically removes duplicate inspection tasks, forming a final task set without duplicate elements. The reason for this is to unify the scheduling and resource integration of periodic routine monitoring and temporary key monitoring, avoiding repeated inspections of the same hidden danger.

[0032] In step 103, a target model corresponding to the security hazard detection task is adaptively selected and activated from multiple lightweight identification models pre-deployed in the edge computing device; wherein, the multiple lightweight identification models are designed with differentiated optimization for different construction stages or different risk types.

[0033] In this embodiment, the adaptive selection of the target model corresponding to the security hazard detection task from multiple lightweight recognition models pre-deployed in edge computing devices for activation is achieved through the following steps: Extract the task identifier for each detection task in the aforementioned safety hazard detection task; Based on the task identifier, a query is performed in the model index table stored locally on the edge computing device. The model index table records the mapping relationship between different task identifiers and the corresponding lightweight recognition model storage path. Based on the queried storage path, the corresponding model parameters are loaded from the local storage space of the edge computing device into memory, and the model loaded into memory that corresponds to the detection task is used as the target model for this activation.

[0034] It's important to note that under the condition of severely limited edge computing resources, the core reason for adaptive model selection is to achieve precise allocation of computing resources and maximize energy efficiency. Its technical advantage lies in constructing a resource-aware dynamic model scheduling mechanism: based on the results of prior operational conditions and risk identification, the system can select and activate the subset most relevant to the current context from the pre-deployed model pool, rather than loading all models. The benefits of this are twofold: first, it significantly reduces the peak memory usage of parallel model loading, avoiding system instability caused by memory overflow; second, it reduces the ineffective load on computing units, performing forward inference only on activated models, thereby saving computation cycles and energy consumption, directly improving system response speed and endurance; finally, this mechanism makes it possible to deploy a large pool of expert models covering multiple stages and risks on limited hardware resources. Through dynamic, on-demand model instantiation, the system achieves an optimal balance between edge-side resource constraints and complex detection requirements while ensuring the integrity of the detection task.

[0035] It should also be noted that the task identifier in this application is a digital code used to uniquely identify a specific security hazard detection task; the model index table is a mapping table used to record the location relationship between the security hazard detection task identifier and its corresponding lightweight identification model file in the storage system.

[0036] In practice, the system first reads a set of security hazard detection tasks, which contains one or more detection tasks. By traversing this set, it extracts the predefined task identifiers that serve as unique identifiers for each detection task and collects these identifiers into a list to be processed. Next, it accesses a structured file in the local storage of the edge computing device, namely the model index table. The file's content uses the task identifier as the primary key and the complete path of the model file in the local storage space as the corresponding value. Then, it queries and matches each task identifier in the list of identifiers to be processed against the aforementioned model index table to retrieve the model file storage path mapped to each identifier. Finally, based on each model file storage path obtained from the query, the system reads the corresponding model parameter file from the local storage through the edge computing device's file system interface. It then calls the model loading interface of the edge computing framework to parse and initialize these model parameters into the device's memory space, forming an executable computation graph instance. Finally, it manages these successfully loaded and ready model instances in a unified manner.

[0037] Additionally, it should be noted that the differentiated optimization design in this application is carried out under a unified lightweight constraint framework, targeting the specific data patterns and performance requirements of different construction stages or risk types, and involves targeted model architecture selection, data representation customization, compression strategy adjustment, and task post-processing rule adaptation. This design is based on the characteristics of the target task: for example, for edge protection detection with extremely high real-time requirements, a simplified YOLO-Fastest architecture with INT8 full integer quantization is adopted; for structural crack identification that requires fine discernment, a lightweight attention module is introduced into the ShuffleNetV2 backbone network, and mixed precision quantization is used to retain the accuracy of key layers; for acoustic and vibration signal anomaly detection, a one-dimensional convolutional network is designed and targeted knowledge distillation training is performed. All optimizations are based on the measured inference latency and memory usage of edge devices as hard constraints. By adjusting parameters such as network depth, width, input resolution, and post-processing threshold, the optimal balance between accuracy and efficiency of each specialized model in its specific task scenario is achieved. Thus, at the system level, parallel, accurate, and real-time monitoring of multiple types of hidden dangers is completed with an acceptable total resource overhead. This is only an example and is not intended to limit the specific scope of the invention.

[0038] In step 104, in the edge computing device, the activated target model is used to perform forward inference on the real-time multi-source heterogeneous sensing data of the corresponding monitoring points to generate a safety hazard identification result.

[0039] In this embodiment, in the edge computing device, the activation of the target model to perform forward inference on the real-time multi-source heterogeneous sensing data of the corresponding monitoring points and generate the safety hazard identification result can be achieved through the following steps: Based on the detection task corresponding to each activated target model, the corresponding type and time sequence of input data segments are extracted from the real-time multi-source heterogeneous sensing data; Each input data segment undergoes standardized preprocessing to meet the input requirements of the corresponding target model; Each standardized preprocessed input data segment is fed into the corresponding target model, and forward inference computation is performed in parallel on the computing unit of the edge computing device. The identification results of each target model, which represent the risk status under the corresponding detection task, are summarized to generate the safety hazard identification results.

[0040] It should be noted that the forward inference computation in this application refers to the mathematical operation process on the computing hardware of the edge computing device, in which the preprocessed input data is propagated layer by layer according to the computation graph defined by the target model to obtain the original output tensor of the model; the security hazard identification result refers to a structured data set that can systematically characterize the existence and status of various risks at the current monitoring point.

[0041] In practical implementation, firstly, the data distributor can extract raw data corresponding to the sensor type and within a specified time window from the shared real-time multi-source heterogeneous sensing data buffer based on the data requirements declared by each model in the currently activated target model set. For example, model A needs the latest 5 seconds of video frames, and model B needs the vibration spectrum of the most recent 2 minutes. Secondly, for each obtained input data segment, the standardized preprocessing program chain associated with the target model corresponding to that data segment is called. This program chain sequentially performs operations such as image scaling and normalization, audio resampling and Mel spectrogram conversion, sensor data denoising and feature scaling, ultimately converting the raw input data segment into a standard tensor format that fully meets the input requirements of the model. Then, each standardized input tensor is sent to its corresponding preprocessing module. The system loads target model instances into memory and utilizes the parallel computing framework provided by the edge computing device's operating system to submit these model inference tasks as independent threads or computational tasks to the device's central processing unit, graphics processing unit, or neural processing unit for computation. Each model independently completes its forward inference computation, generating raw, unparsed output tensors. Finally, the system calls the post-processing parsing function corresponding to each target model to decode the obtained output tensors. This parsing process is performed according to the task type. For example, it extracts bounding box positions and class confidence from the output of the object detection model, and class labels and probabilities from the output of the classification model. Furthermore, the system collects and encapsulates these parsed result entries with timestamps and task identifiers according to a predefined structure (such as a list of JSON objects).

[0042] In step 105, based on the safety hazard identification results, it is determined whether there is a safety hazard at the corresponding monitoring point, and if there is a safety hazard, an early warning message is generated and pushed to the engineering supervision terminal.

[0043] In this embodiment, determining whether a safety hazard exists at a corresponding monitoring point based on the safety hazard identification result can be achieved through the following steps: Each of the safety hazard identification results is compared with a preset threshold for the corresponding risk type. If any identification result exceeds its corresponding judgment threshold, it is determined that there is a specific type of hidden danger. Based on the risk level and location corresponding to all the out-of-limit identification results, a comprehensive hazard assessment conclusion is generated.

[0044] It should be noted that the judgment threshold in this application is a preset critical value used to quantify the identification results of different types of safety hazards; the comprehensive hazard judgment conclusion is used to integrate all confirmed specific types of hazard information to form a structured output of the final general judgment on the overall safety status of the monitoring point.

[0045] In practice, firstly, the system internally sets up a risk type-judgment threshold mapping table, which records the numerical judgment threshold corresponding to each risk type. During implementation, the system iterates through each result in the safety hazard identification results, retrieves the corresponding judgment threshold from the mapping table based on the risk type identifier recorded in that result, and compares the quantified value (such as the confidence score) in the result with the threshold. Secondly, after the comparison is completed, an empty list is created to record exceeding items. The system checks each comparison result; if the value of a certain identification result exceeds the judgment threshold retrieved from the mapping table, the system determines that the current monitoring point contains the risk type corresponding to that result. The system identifies and records the type, confidence level, and location of each specific type of hazard in a list, which is then used as the hazard type list. The system reads this list and, for each hazard, retrieves its preset risk level from a pre-defined risk level database based on its risk type. All hazards are then categorized and sorted according to their risk levels. Combined with the location information of each hazard, a summary report is generated, including the overall risk level, main risk types, and risk point distribution. This is done to perform a global assessment and information aggregation of scattered individual hazards, providing supervisors with a clear and comprehensive basis for decision-making.

[0046] It should also be noted that after the system determines that a safety hazard exists and generates a comprehensive hazard assessment conclusion, the specific implementation of the early warning information generation and push steps is as follows: First, the comprehensive hazard assessment conclusion and the associated original data evidence are standardized and encapsulated to generate a structured early warning message body. This message body is usually in JSON format and includes fields such as early warning level, hazard type, location description, confidence level, recommended measures, and data summary. Then, the system calls a lightweight message push client through the network module built into the edge computing device. This client is based on communication protocols suitable for IoT scenarios such as MQTT or HTTP / 2 and asynchronously publishes the encapsulated early warning message to a message broker center deployed on a cloud server or local server. The message broker center is responsible for terminal online status management, message routing, and priority queuing. Finally, the engineering supervision terminals (such as mobile APP or web backend) that have subscribed to the corresponding construction area topic receive the early warning message in real time through a long connection. After the terminal client parses the message content, it triggers different intensities of audio and visual prompts according to the early warning level and displays the location and details on the graphical interface, thus completing the closed loop from edge-side hazard identification to supervisor perception.

[0047] On the other hand, in this embodiment, this application provides an engineering supervision safety hazard early warning system based on edge computing, referring to... Figure 4The figure is a schematic diagram of the structure of an edge computing-based engineering supervision safety hazard early warning system according to some embodiments of this application. The edge computing-based engineering supervision safety hazard early warning system 400 includes: an acquisition module 401, a processing module 402, and a push module 403, which are described below: The acquisition module 401 is used to acquire multi-source heterogeneous sensing data collected from multiple monitoring points within the target construction area; The processing module 402 is used to perform feature fuzzy extraction on multi-source heterogeneous sensing data to identify the current construction conditions and dominant risk factors, and to determine the safety hazard detection tasks to be performed based on the identified construction conditions and risk factors. The processing module 402 is further configured to adaptively select and activate a target model corresponding to the security hazard detection task from multiple lightweight recognition models pre-deployed in the edge computing device; The processing module 402 is also used in the edge computing device to perform forward reasoning on the real-time multi-source heterogeneous sensing data of the corresponding monitoring points using the activated target model to generate a safety hazard identification result. The push module 403 is used to determine whether there is a safety hazard at the corresponding monitoring point based on the safety hazard identification result, and to generate early warning information and push it to the engineering supervision terminal when there is a safety hazard.

[0048] In addition, this application also provides a computer device, which includes a memory and a processor. The memory stores code, and the processor is configured to acquire the code and execute the above-described edge computing-based engineering supervision safety hazard early warning method.

[0049] In this embodiment, reference Figure 5 The figure is a schematic diagram of the structure of a computer device implementing an edge computing-based engineering supervision safety hazard early warning method according to some embodiments of this application. The edge computing-based engineering supervision safety hazard early warning method in the above embodiments can... Figure 5 The computer device shown is used to implement this, and the computer device 500 includes at least one processor 501, a communication bus 502, a memory 503, and at least one communication interface 504.

[0050] Processor 501 can be a general-purpose central processing unit (CPU) or an application-specific integrated circuit (ASIC).

[0051] The communication bus 502 can be used to transmit information between the aforementioned components.

[0052] Memory 503 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital versatile optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. Memory 503 may exist independently and be connected to processor 501 via communication bus 502. Memory 503 may also be integrated with processor 501.

[0053] The memory 503 stores program code for executing the solution of this application, and its execution is controlled by the processor 501. The processor 501 executes the program code stored in the memory 503. The program code may include one or more software modules. The edge computing-based engineering supervision safety hazard early warning method in the above embodiments can be implemented by the processor 501 and one or more software modules in the program code in the memory 503.

[0054] Communication interface 504 uses any transceiver-like device to communicate with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0055] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single-CPU) processor or a multi-core (multi-CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0056] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0057] In addition, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned edge computing-based engineering supervision safety hazard early warning method.

[0058] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0059] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for early warning of safety hazards in engineering supervision based on edge computing, characterized in that, include: Acquire multi-source heterogeneous sensing data collected from multiple monitoring points within the target construction area; Feature fuzzy extraction is performed on multi-source heterogeneous sensing data to identify the current construction conditions and dominant risk factors. Based on the identified construction conditions and risk factors, the safety hazard detection tasks to be performed are determined. The system adaptively selects and activates a target model corresponding to the security hazard detection task from multiple lightweight identification models pre-deployed in edge computing devices; wherein the multiple lightweight identification models are designed with differentiated optimization for different construction stages or different risk types. In the edge computing device, the activated target model is used to perform forward reasoning on the real-time multi-source heterogeneous sensing data of the corresponding monitoring points to generate a safety hazard identification result. Based on the safety hazard identification results, it is determined whether there are safety hazards at the corresponding monitoring points, and when there are safety hazards, an early warning message is generated and pushed to the engineering supervision terminal.

2. The method as described in claim 1, characterized in that, The multi-source heterogeneous sensing data specifically includes: video stream data collected by visual sensors deployed at monitoring points, and ambient sound and vibration spectrum data collected by acoustic sensors deployed at monitoring points.

3. The method as described in claim 1, characterized in that, Feature fuzzy extraction is performed on multi-source heterogeneous sensing data to identify the current construction conditions and dominant risk factors, specifically including: The collected video stream data is matched with preset typical construction scene templates using lightweight feature matching to obtain the scene matching degree; Based on the collected environmental acoustic and vibration spectrum data, identify the types of critical operations that are in progress; By integrating the scene matching degree with the identified key operation type, and using a pre-built lightweight fuzzy inference rule base, the membership degree of the current operation to different preset construction conditions is calculated. The preset construction condition with the highest membership degree is determined as the current construction condition, and the high-incidence risk factors in the historical statistics under this condition are taken as the dominant risk factors.

4. The method as described in claim 1, characterized in that, Based on the identified construction conditions and risk factors, the specific tasks to be performed for safety hazard detection include: Obtain at least one basic inspection task that needs to be performed periodically under the aforementioned construction conditions; Based on the aforementioned risk factors, at least one specific detection task requiring key monitoring is obtained from a predefined task mapping table; The at least one basic detection task and the at least one special detection task are merged and deduplicated to obtain a safety hazard detection task to be executed, wherein the safety hazard detection task includes at least one detection task.

5. The method as described in claim 1, characterized in that, Adaptively selecting and activating a target model corresponding to the security hazard detection task from multiple lightweight recognition models pre-deployed in edge computing devices specifically includes: Extract the task identifier for each detection task in the aforementioned safety hazard detection task; Based on the task identifier, a query is performed in the model index table stored locally on the edge computing device. The model index table records the mapping relationship between different task identifiers and the corresponding lightweight recognition model storage path. Based on the queried storage path, the corresponding model parameters are loaded from the local storage space of the edge computing device into memory, and the model loaded into memory that corresponds to the detection task is used as the target model for this activation.

6. The method as described in claim 1, characterized in that, In edge computing devices, the activated target model is used to perform forward inference on real-time multi-source heterogeneous sensing data of corresponding monitoring points to generate safety hazard identification results, specifically including: Based on the detection task corresponding to each activated target model, the corresponding type and time sequence of input data segments are extracted from the real-time multi-source heterogeneous sensing data; Each input data segment undergoes standardized preprocessing to meet the input requirements of the corresponding target model; Each standardized preprocessed input data segment is fed into the corresponding target model, and forward inference computation is performed in parallel on the computing unit of the edge computing device. The identification results of each target model, which represent the risk status under the corresponding detection task, are summarized to generate the safety hazard identification results.

7. The method as described in claim 1, characterized in that, Based on the safety hazard identification results, determining whether a corresponding monitoring point has a safety hazard specifically includes: Each of the safety hazard identification results is compared with a preset threshold for the corresponding risk type. If any identification result exceeds its corresponding judgment threshold, it is determined that there is a specific type of hidden danger. Based on the risk level and location corresponding to all the out-of-limit identification results, a comprehensive hazard assessment conclusion is generated.

8. A safety hazard early warning system for engineering supervision based on edge computing, characterized in that, include: The acquisition module is used to acquire multi-source heterogeneous sensing data collected from multiple monitoring points within the target construction area; The processing module is used to perform feature fuzzy extraction on multi-source heterogeneous sensing data to identify the current construction conditions and dominant risk factors, and to determine the safety hazard detection tasks to be performed based on the identified construction conditions and risk factors. The processing module is also used to adaptively select and activate a target model corresponding to the security risk detection task from multiple lightweight recognition models pre-deployed in the edge computing device; The processing module is also used in the edge computing device to perform forward reasoning on the real-time multi-source heterogeneous sensing data of the corresponding monitoring points using the activated target model to generate a safety hazard identification result. The push module is used to determine whether there is a safety hazard at the corresponding monitoring point based on the safety hazard identification results, and to generate early warning information and push it to the engineering supervision terminal when there is a safety hazard.

9. A computer device comprising a memory and a processor, the memory storing code, characterized in that, The processor is configured to acquire the code and execute the edge computing-based engineering supervision safety hazard early warning method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the edge computing-based engineering supervision safety hazard early warning method as described in any one of claims 1 to 7.

Images