Power construction behavior safety detection method and system

Abstract

The invention discloses an electric power construction behavior safety detection method and system, and relates to electric power safety. The method comprises the following steps: collecting historical electric power work ticket data, and constructing a work ticket risk feature library; according to the work ticket risk feature library, training a risk identification model based on multi-task learning; collecting text data of the current electric power work ticket, extracting work content and security measure fields, and generating a text feature vector; searching optimal parameter configuration in a parameter space of the risk identification model by using a discretization doliolaria optimization algorithm; the method comprises the following steps: collecting real-time image data of a working site, carrying out target detection and feature extraction on the image data, identifying position coordinates of working personnel and types and wearing state marks of safety protection appliances, encoding a detection result into structured data, and generating an image feature vector; aiming at the lack of a dynamic association verification mechanism between an electric power work ticket text specification and an actual behavior of a work site in the prior art, the static requirement of a work ticket is converted into a dynamic behavior supervision standard.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electric power safety, in particular to a power construction behavior safety detection method and system. BACKGROUND

[0002] With the continuous expansion of large-scale engineering projects and the continuous improvement of construction mechanization, the safety management of aerial work platforms has become a key link to protect the quality of engineering and the safety of workers. According to statistics, more than 70% of aerial work accidents are related to illegal operation, improper protection and insufficient platform stability control. The traditional safety management of aerial work mainly relies on the operation instruction book system and manual inspection, which constrains the operation behavior through pre-established written specifications, but this static management mode exposes many shortcomings in the dynamic and complex aerial work environment.

[0003] The operation instruction book, as the core system of aerial work safety management, specifies in detail the operation content, safety measures, operation steps and risk point pre-control requirements. However, there is a significant spatio-temporal separation between the text specifications in the operation instruction book and the actual execution on the aerial work site. On the one hand, the operation instruction book is developed before construction, and it is difficult to foresee all risk scenarios in the process of using a straight-arm aerial work platform for 24-meter high-altitude operation or a scissor lift for multi-person collaborative operation. On the other hand, there are subjective differences in the understanding and execution of the specifications by site workers, resulting in the phenomenon of "rules not followed, execution out of shape". Although some engineering enterprises have introduced video monitoring systems, most of them still remain in the post-backtracking stage, lacking real-time behavior recognition and risk warning capabilities.

[0004] In recent years, with the development of artificial intelligence technology, image recognition technology based on deep learning has begun to be applied to the field of aerial work safety monitoring. The existing technology identifies elements such as workers, safety protection tools, aerial work platforms and surrounding obstacles through target detection algorithms, to a certain extent, realizing the automatic monitoring of the operation site. However, these methods generally have the following technical defects:

[0005] Firstly, there is a lack of effective association mechanism between text specifications and image recognition. The existing system treats operation instruction book management and site monitoring as two independent modules, and fails to establish a dynamic mapping relationship from text requirements to behavior standards, resulting in the inability of image recognition results to accurately reflect whether they comply with the safety specifications of specific aerial work types (such as straight-line lifting operation of straight-arm platforms and large-scale collaborative operation of scissor platforms).

[0006] Secondly, the parameter configuration of the risk identification model lacks targeted optimization. Different types of overhead operations (such as using a straight-arm platform equipped with an active defense protection system for 460 kg heavy load operation, using a scissor lift for multi-person collaborative operation, and using an extension platform for operation on an inclined wall) have differentiated risk characteristics and monitoring priorities, but existing methods mostly use uniform model parameters, which cannot adapt to the needs of diversified operation scenarios. Even if some studies try to introduce parameter optimization, they face the problem of mismatch between optimization algorithms and parameter spaces.

[0007] In particular, in terms of parameter optimization, traditional continuous optimization algorithms (such as particle swarm optimization, genetic algorithm, etc.) will generate a large number of invalid intermediate solutions when dealing with discrete parameters of deep learning models (such as batch size, network layer number, etc.). For example, when the optimization algorithm generates a continuous value of 31.7 for the batch size, it needs to be converted to 32 by simple rounding. This rough discretization method disrupts the search logic of the optimization algorithm, causing the algorithm to fall into local optimum or slow down the convergence speed. As a new type of swarm intelligence algorithm, the Sine Cosine Algorithm (SCA) performs well in continuous optimization problems, but its position update mechanism is also designed based on continuous space. Direct application to discrete parameter optimization will result in low search efficiency and unstable solution quality.

[0008] In addition, the existing multi-task learning framework is not deep enough in the application of overhead operation safety detection. Overhead operation safety involves platform operation specifications, operation height control, protective equipment wearing, platform stability monitoring, and other risk dimensions. These dimensions have both independence and correlation. Simple single-task models cannot capture this complex risk structure, and existing multi-task learning methods lack task design and loss function optimization specific to overhead platform characteristics (such as reach control of straight-arm platforms and load distribution of scissor lifts).

[0009] In summary, existing technologies cannot effectively convert from operation guidebook text specifications to overhead operation site behavior monitoring, cannot dynamically adjust monitoring strategies according to different operation platform types and operation heights, and have technical bottlenecks in the mismatch between algorithm mechanisms and parameter spaces during parameter optimization, which severely restricts the improvement of intelligent safety management level of large-scale mechanical projects. SUMMARY

[0010] To address the lack of dynamic correlation verification mechanism between power work ticket text specifications and actual behavior in the field in existing technologies, the present application provides a power construction behavior safety detection method and system, which constructs a dual verification mechanism of text analysis-image verification, and converts the static requirements of work tickets into dynamic behavior monitoring standards.

[0011] One aspect of the present application provides a power construction behavior safety detection method, comprising: S1, collecting historical power work ticket data, extracting the work ticket operation content, safety measures, operation steps and dangerous point pre-control measures field; constructing a work ticket risk feature library according to the extracted field, and each record in the risk feature library contains operation type code, risk feature vector and historical accident mark; S2, training a risk identification model based on multi-task learning according to the work ticket risk feature library, the risk identification model contains a shared feature extraction layer and multiple task output heads, and each output head corresponds to a type of operation risk assessment task; S3, collecting the text data of the current power work ticket, extracting the operation content and safety measures field, and generating a text feature vector; according to the text feature vector, searching for the optimal parameter configuration in the parameter space of the risk identification model by using the discretization sea squirt optimization algorithm, and obtaining the optimal model parameters matched with the operation type of the current work ticket; S4, collecting real-time image data of the operation site, performing target detection and feature extraction on the image data, identifying the position coordinates of the operation personnel, and the category and wearing state mark of the safety protection tool, encoding the detection results into structured data, and generating an image feature vector; S5, according to the image feature vector and the optimal model parameters, using the risk identification model for prediction to obtain the risk assessment result;

[0012] Further, S2, according to the work ticket risk feature library, training a risk identification model based on multi-task learning, the risk identification model contains a shared feature extraction layer and multiple task output heads, and each output head corresponds to a type of operation risk assessment task, including: extracting the risk feature vector and the historical accident mark from the work ticket risk feature library, and dividing the data set into three subsets of live operation, switching operation and maintenance operation according to the operation type code; according to the data subset, constructing a multi-task learning network architecture, wherein the shared feature extraction layer adopts a three-layer fully connected network, and three task output heads are set: operation sequence compliance output head, used for predicting whether the switching operation steps meet the standard operation procedure; safety distance detection output head, used for predicting whether the safety distance between the operation personnel and the live equipment meets the regulation requirement; protective tool specification output head, used for predicting whether the use of safety helmet and insulating gloves meets the specification requirement; setting a multi-task loss function, and using a batch gradient descent algorithm for model training to obtain a risk identification model based on multi-task learning.

[0013] Further, the multi-task loss function expression is as follows: ; wherein, the binary cross-entropy loss of operation sequence compliance, the multi-class cross-entropy loss of safety distance detection, the binary cross-entropy loss of protective tool specification, the task weight coefficient;

[0014] Further, S3, collect the text data of the current power work ticket, extract the job content and safety measures field, and generate a text feature vector; according to the text feature vector, search for the optimal parameter configuration in the parameter space of the risk identification model using the discretized sea cucumber optimization algorithm, obtain the optimal model parameters matched with the current work ticket operation type, including: the parameters of the initial discretized sea cucumber optimization algorithm, the parameters including: sea cucumber population size N, maximum iteration number and temperature parameters ; define the parameter search space, including: learning rate , batch size , and dropout rate ; construct a discrete index mapping table to map the continuous parameter search space to a discrete index; construct a fitness function F: ; wherein, is the accuracy of the model on the validation set, is the model complexity, is the balance coefficient; execute the discretized sea cucumber optimization process, and iteratively optimize until the termination condition is met, output the optimal parameter configuration .

[0015] In particular, the position updating mechanism of the traditional sea cucumber optimization algorithm (SSA) is based on continuous mathematical operations, and the leader position updating formula ) and the follower position updating formula Essentially, interpolation and extrapolation operations are performed in continuous space. However, the key hyperparameters of the risk identification model have obvious discrete characteristics: the batch size must be a positive integer (such as 16, 32, 64), the learning rate is usually selected from a predefined set (such as 0.001, 0.01, 0.1), and the dropout rate often uses fixed intervals (such as 0.1, 0.2, 0.3). Directly rounding the continuous optimization result will destroy the search logic of the algorithm, resulting in falling into a suboptimal solution.

[0016] The present application introduces a probabilistic discrete mapping mechanism, which realizes intelligent conversion from continuous search space to discrete decision space through a probability distribution controlled by temperature parameter . This design maintains the advantages of SSA swarm intelligence search while solving the problem of discrete parameter optimization.

[0017] Further, S35, execute the discretized sea cucumber optimization process, including: initialize the population data matrix , wherein N is the population size, and each row represents the position data of an individual in three-dimensional continuous space .

[0018] Update the position data in each iteration: for the leader individual, according to the formula updating the continuous position data; for the follower individual, according to the formula calculating new position data; for each continuous value , calculating candidate discrete indexes: ; calculating mapping distance: ; according to the probability random mapping to generate the final discrete index; according to the discrete index, extracting the actual parameter value from the parameter candidate set; using the extracted parameter value to configure the model, evaluating the performance on the validation set, calculating the fitness value and updating the optimal solution record;

[0019] In particular, the hyperparameter optimization of deep learning models faces a fundamental contradiction: advanced swarm intelligence optimization algorithms (such as the Sine Algorithm) rely on mathematical operations in continuous space to achieve efficient search, while key parameters in practical applications (learning rate, batch size, dropout rate) must take discrete values. Traditional post-processing rounding methods not only destroy the search logic of optimization algorithms, but also may cause parameter configuration to deviate from the optimal region, creating safety hazards in high-risk power operation scenarios.

[0020] Based on temperature-controlled probabilistic discrete mapping, the present application realizes intelligent bridging of continuous search space and discrete parameter space by constructing a three-stage mechanism of continuous exploration-probabilistic mapping-discrete decision. That is, instead of simply quantizing continuous values to the nearest discrete point, the present application gives the algorithm the ability to intelligently select from the discrete candidate set through a probability function .

[0021] Specifically, when the algorithm searches to position c in the continuous space [0, 1], the system first calculates the distance between it and the nearest discrete index, and then determines the final discrete mapping result according to the probability distribution regulated by the temperature parameter . The closer the distance, the higher the mapping probability, but it is not absolutely determined. The introduction of the temperature parameter allows the algorithm to have different exploration characteristics at different optimization stages.

[0022] The core advantage of the Sine Algorithm lies in its chain-like cooperative search mechanism: the leader explores new areas through , and the follower achieves information transmission and local refinement through . If these operations are performed directly in discrete space, the position average will lose its meaning, the chain structure will be broken, and the algorithm will degenerate into a simple random search.

[0023] The introduction of the probabilistic mapping mechanism in the present application ensures that all the excellent properties of the algorithm in continuous space are preserved: the population still evolves in In space, the original formula evolves, the search trajectory remains continuous, and the information transmission mechanism works normally. Only when the fitness needs to be evaluated, the corresponding discrete parameter value is obtained through probability mapping.

[0024] In addition, the power system risk has nonlinear and catastrophic characteristics. Learning rate from 0.01 to 0.012 may have no effect, but if it is misadjusted to 0.1, it may cause model gradient explosion and completely lose risk identification ability. Probability mapping ensures that parameters can only take predefined safe values, fundamentally avoiding parameter avalanche risk.

[0025] Further, the annealing mechanism of the temperature parameter is highly consistent with the risk grading management concept of power operation. When dealing with low-risk operations such as routine maintenance, a higher temperature can be set to allow the algorithm to boldly explore and find more efficient parameter combinations; while dealing with high-voltage live operations, lower temperature makes the algorithm quickly converge to the most stable parameter configuration, reflecting the industry guideline that the higher the safety level, the smaller the fault tolerance space.

[0026] Further, S4, collects real-time image data of the operation site, detects and extracts features of the image data, identifies the position coordinates of the operation personnel, and the category and wearing state markers of safety protection tools, encodes the detection results into structured data, and generates an image feature vector, including: collecting real-time image data of the operation site, and preprocessing the image; using the YOLOv5 target detection model to identify the key targets in the image: detecting the operation personnel, outputting the boundary box coordinates and confidence ; detecting safety helmets and gloves, outputting the boundary box and confidence of the corresponding protective tools; detecting live equipment and operation equipment, outputting the equipment boundary box coordinates and category label; quantitatively analyzing the spatial relationship and wearing state of the target according to the boundary box coordinates to obtain the detection result; encoding the detection result into structured data to generate an image feature vector;

[0027] Further, according to the boundary box coordinates, the spatial relationship and wearing state of the target are quantitatively analyzed to obtain the detection result, including: extracting edge sampling point sets from the operation personnel boundary box and live equipment boundary box; calculating the Euclidean distance matrix between the point pairs based on the edge sampling point set; extracting the minimum value from the Euclidean distance matrix as the safety distance as the target spatial relationship detection result; dividing the operation personnel boundary box into functional areas according to a preset ratio; calculating the overlap degree of the protective tool boundary box and the corresponding human body functional area; generating a binary wearing state detection result according to the overlap degree threshold.

[0028] In particular, traditional object detection algorithms use rectangular bounding boxes as output, a design stemming from considerations of computational efficiency and ease of annotation. However, this abstraction presents several problems in power operation scenarios. First, there is an irreconcilable contradiction between the dynamic complexity of the human body contour and the static rectangular bounding box. Workers' postures vary greatly during tasks—limbs outstretched when climbing, arms extended forward when operating, and bending over during maintenance. These posture changes cause the actual space occupied by the human body to far exceed or deviate from the bounding box, inevitably leading to significant errors in distance calculations based on the box's center or corners. Second, the irregular shape of power equipment exacerbates this problem. The cylindrical body of a transformer, the protruding operating handle of a switchgear, and the slender structure of busbars—these complex shapes, when forcibly compressed into rectangular bounding boxes, lose a significant amount of crucial spatial information.

[0029] Therefore, this application proposes an edge sampling point set + Euclidean distance matrix to expand the target representation from four corner points to dozens of edge points. Specifically, uniform sampling is performed on the four edges of each bounding box, with a typical sampling density of 10-20 points per edge. Euclidean distance matrix The algorithm extends distance calculation from one-to-one to many-to-many. For m sampling points of workers and n sampling points of live equipment, the algorithm calculates the distance between all possible point pairs: Then extract the minimum value from the matrix. As a safety distance, this global search strategy ensures that the system can capture the most dangerous distance information regardless of the object's deformation or the angle from which it approaches. Furthermore, the dynamic nature of power operations requires the monitoring system to respond to attitude changes in real time. Edge sampling points automatically adjust their positions as the bounding box updates, ensuring that the target's critical contours are always covered.

[0030] Furthermore, in step S5, based on the image feature vector and the optimal model parameters, a risk identification model is used to make predictions and obtain risk assessment results. This includes: configuring the risk identification model with the optimal model parameters to establish a parameterized model instance; using the image feature vector as input data, and according to the parameterized model instance: performing a nonlinear transformation on the input features through a shared feature extraction layer; inputting the transformed features in parallel into multiple task output heads; extracting the corresponding prediction probability distribution from each output head; performing threshold judgment and category mapping based on the prediction probability distribution; and generating risk assessment results through weighted aggregation based on the threshold judgment and category mapping results.

[0031] Another aspect of the present application also provides a power construction behavior safety detection system, comprising: a historical data processing module, configured to collect historical power work ticket data, and extract operation content, safety measures, operation steps and dangerous point pre-control measures fields in the work ticket; construct a work ticket risk feature library according to the extracted fields, and each record in the risk feature library contains operation type code, risk feature vector and historical accident label; a model training module connected with the historical data processing module, configured to train a risk identification model based on multi-task learning according to the work ticket risk feature library, and the risk identification model contains a shared feature extraction layer and multiple task output heads, and each output head corresponds to a type of operation risk assessment task; a parameter optimization module, configured to collect text data of a current power work ticket, extract operation content and safety measures fields, and generate a text feature vector; search for optimal parameter configuration in a parameter space of the risk identification model by using a discretization sea squirts optimization algorithm according to the text feature vector, and obtain optimal model parameters matched with an operation type of the current work ticket; an image processing module, configured to collect real-time image data of a work site, perform target detection and feature extraction on the image data, identify operation personnel position coordinates, safety protection tool categories and wearing state labels, encode the detection results into structured data, and generate an image feature vector; and a risk assessment module connected with the model training module, the parameter optimization module and the image processing module respectively, configured to perform prediction by using the risk identification model according to the image feature vector and the optimal model parameters, and obtain a risk assessment result.

[0032] Compared with the prior art, the present application has the following advantages:

[0033] In view of the problems in the prior art that there is a lack of dynamic correlation verification mechanism between power work ticket text specifications and actual behaviors in work sites, and that a large number of invalid candidate solutions are generated in the process of processing a risk identification model discrete parameter space by using a traditional continuous optimization algorithm, resulting in the inability to quickly adapt to real-time safety monitoring requirements of different operation types, the present application provides a power construction behavior safety detection method and system, which can automatically convert static safety requirements in a work ticket into dynamic monitoring parameters matched with a current operation scene by using a discretization sea squirts optimization algorithm, establish a closed-loop detection mechanism of text specification-parameter optimization-image verification, ensure the parameter search efficiency, realize accurate identification and real-time early warning of multi-dimensional risk elements such as operation sequence, safety distance and protection tools, and effectively solve the problem that specifications and behavior execution are disconnected in power construction safety supervision. BRIEF DESCRIPTION OF DRAWINGS

[0034] The present application will be further described in the form of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting, and in these embodiments, the same reference numbers represent the same structures, wherein:

[0035] Figure 1 This is an exemplary flowchart of a safety detection method for power construction activities according to some embodiments of this application;

[0036] Figure 2 This is an exemplary flowchart illustrating the construction of a multi-task learning risk identification model according to some embodiments of this application;

[0037] Figure 3 This is an exemplary flowchart illustrating the acquisition of optimal parameter configuration according to some embodiments of this application;

[0038] Figure 4 This is an exemplary flowchart illustrating the construction of image feature vectors according to some embodiments of this application. Detailed Implementation

[0039] The methods and systems provided in the embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0040] like Figure 1 As shown, S1 involves collecting historical power work order data and extracting fields for work content, safety measures, work steps, and hazard pre-control measures from the work orders; constructing a work order risk feature library based on the extracted fields, where each record contains a work type code, a risk feature vector, and historical accident markers; S2 involves training a multi-task learning-based risk identification model based on the work order risk feature library, where the risk identification model includes a shared feature extraction layer and multiple task output heads, each corresponding to a type of work risk assessment task; S3 involves collecting the text data of the current power work order and extracting the work content and safety measures fields. S1. Generate text feature vectors; based on the text feature vectors, use the discretized sea squirt optimization algorithm to search for the optimal parameter configuration in the parameter space of the risk identification model to obtain the optimal model parameters that match the current work order type; S4. Collect real-time image data of the work site, perform target detection and feature extraction on the image data, identify the location coordinates of the workers, as well as the type and wearing status of safety protective equipment, encode the detection results into structured data, and generate image feature vectors; S5. Based on the image feature vectors and the optimal model parameters, use the risk identification model to make predictions and obtain the risk assessment results.

[0041] Specifically, S1 involves collecting historical power work order data and extracting fields related to work content, safety measures, work steps, and hazard prevention and control measures from the work orders; based on the extracted fields, a work order risk feature database is constructed.

[0042] S11. Perform text preprocessing on the extracted fields of work content, safety measures, work steps and hazard point control measures, including removing special characters, word segmentation and stop word filtering, to obtain preprocessed text data;

[0043] S12, the TF-IDF algorithm is used to extract features from the preprocessed text data, calculate the weight value of each word in different fields, and construct an initial feature matrix;

[0044] S13, the initial feature matrix is processed by principal component analysis dimension reduction, and the principal components whose cumulative variance contribution rate reaches the preset threshold are retained to generate a risk feature vector after dimension reduction;

[0045] S14, match the pre-defined job type dictionary based on the keywords in the job content field, and determine the job type code corresponding to each work ticket;

[0046] S15, query the historical accident database, associate the corresponding accident record according to the work ticket number, when there is an associated accident, set the historical accident mark as the accident type code, and when there is no associated accident, set the historical accident mark as the safety mark;

[0047] S16, combine the job type code, risk feature vector and historical accident mark into a structured record, and store it in the work ticket risk feature library.

[0048] As shown in Figure 2 S2, according to the work ticket risk feature library, a risk identification model based on multi-task learning is trained, and the risk identification model includes a shared feature extraction layer and multiple task output heads, each output head corresponds to a type of job risk assessment task, including:

[0049] The risk feature vector and the historical accident mark are extracted from the work ticket risk feature library, and the data set is divided into three subsets of live working, switching operation and maintenance operation according to the job type code;

[0050] According to the data subset, a multi-task learning network architecture is constructed, wherein the shared feature extraction layer adopts a three-layer fully connected network, and three task output heads are set:

[0051] Operation sequence compliance output head, used to predict whether the switching operation steps meet the standard operation procedure;

[0052] Safety distance detection output head, used to predict whether the safety distance between the worker and the live equipment meets the regulation requirements;

[0053] Protective equipment specification output head, used to predict whether the use of safety helmet and insulating gloves meets the specification requirements;

[0054] Set the multi-task loss function, and use the batch gradient descent algorithm to train the model to obtain a risk identification model based on multi-task learning.

[0055] Among them, the multi-task loss function is set, including:

[0056] S31, define the multi-task loss function expression as:

[0057] ; wherein, is the binary cross-entropy loss of operation sequence compliance, is the multi-class cross-entropy loss of safety distance detection, is the binary cross-entropy loss of protective equipment specification, is the task weight coefficient;

[0058] S32, for three binary classification tasks, a unified binary cross-entropy loss calculation method is used:

[0059] ; wherein, N is the number of task samples, is the true label, is the logit value output by the model, is the sigmoid activation function;

[0060] S34, according to the importance level of power construction safety, set the task weight coefficient: ; wherein, the safety distance detection weight is the highest, because it is directly related to personal safety;

[0061] S35, add L2 regularization term to prevent overfitting: ; wherein, β=0.0001 is the regularization coefficient, and W is the model parameter set.

[0062] As shown in Figure 3 S3, collect the text data of the current power work ticket, extract the operation content and safety measures field, and generate a text feature vector;

[0063] According to the text feature vector, use the discrete Ctenophore optimization algorithm to search for the optimal parameter configuration in the parameter space of the risk identification model, obtain the optimal model parameters matched with the current work ticket operation type, including:

[0064] S31, collect the text data of the current power work ticket, use regular expressions to extract the operation content and safety measures field, and perform word segmentation and stop word removal processing on the extracted text;

[0065] S32, use the jieba word segmentation tool for Chinese word segmentation, remove stop words, and retain power professional terms; use the Word2Vec model to convert the word segmentation results into 300-dimensional word vectors, perform average pooling operation on the word vector matrix of each field to generate field-level feature vectors; concatenate the field-level feature vectors of the operation content and safety measures to form a 256-dimensional text feature vector;

[0066] S33, initialize the discrete Ciona optimization algorithm parameters:

[0067] Set the Ciona population size N = 30, the maximum number of iterations , temperature parameters ;

[0068] Define the parameter search space, including the learning rate , batch size , dropout rate ;

[0069] Initialize the data structure of the discrete Ciona optimization algorithm: create a parameter candidate set: learning rate set contains 10 discrete values, batch size set BS = {16, 24, 32, 48, 64} contains 5 discrete values, dropout rate set DR = {0.1, 0.2, 0.3, 0.4, 0.5} contains 5 discrete values; Construct a discrete index mapping table to map the continuous space [0, 1] to the discrete index {0, 1,..., D-1};

[0070] S34, construct the fitness function: ; Wherein, accuracy is the accuracy of the model on the validation set, complexity is the model complexity, is the balance coefficient;

[0071] S35, execute the discrete Ciona optimization process:

[0072] Initialize the population data matrix , where N is the population size, and each row represents the position data of an individual in the three-dimensional continuous space ;

[0073] Update the position data in each iteration:

[0074] For the leader individual, update the continuous position data according to the formula ;

[0075] For the follower individual, calculate the new position data according to the formula ;

[0076] Map the continuous position to the discrete parameter using the probability boundary softening mechanism: for each continuous value , calculate the candidate discrete index: ; Calculate the mapping distance: ; According to the probability , randomly map to generate the final discrete index;

[0077] Extract the actual parameter value from the parameter candidate set according to the discrete index: learning rate: wherein, denotes a set of learning rate candidates, is the corresponding discrete index; batch size: wherein, denotes a set of learning rate candidates, is the corresponding discrete index; dropout rate: wherein, denotes a set of learning rate candidates, is the corresponding discrete index; update global optimal solution;

[0078] S36, iteratively optimize until the termination condition is met, output the optimal parameter configuration , apply the parameter configuration to the risk identification model.

[0079] As Figure 4 shown, S4, collect real-time image data of the work site, perform target detection and feature extraction on the image data, identify the position coordinates of the workers, and identify the category and wearing state markers of the safety protection tools, encode the detection results into structured data, and generate an image feature vector:

[0080] S41, collect real-time image data of the work site, and pre-process the image:

[0081] Uniformly scale the original image to 640x480 resolution;

[0082] Perform histogram equalization processing to enhance the contrast of the image;

[0083] Apply a Gaussian filter to remove image noise, with a filter kernel size of 3x3;

[0084] S42, use a YOLOv5 target detection model to identify key targets in the image:

[0085] Detect workers, output bounding box coordinates and confidence ;

[0086] Detect safety helmets, insulating gloves, insulating shoes, and other protective tools, and output the bounding boxes and confidence of each type of protective tool;

[0087] Detect live equipment and work equipment, and output the equipment bounding box coordinates and category label;

[0088] S43, calculate the spatial relationship data between the workers and the live equipment:

[0089] Extract the center point coordinates of the worker bounding box ;

[0090] Edge sampling is performed on the work personnel bounding box, and M=20 points are uniformly sampled on the four edges to obtain a personnel edge point set , denotes the coordinates of the i-th sampling point.

[0091] Edge sampling is performed on the charged equipment bounding box, and N=20 points are uniformly sampled on the four edges to obtain an equipment edge point set , denotes the coordinates of the j-th sampling point.

[0092] The minimum edge distance between the personnel and the charged equipment is calculated: ;

[0093] According to the image depth information or the calibration parameters, the pixel distance is converted into the actual physical distance ;

[0094] S44, analyze the protective equipment wearing state data:

[0095] The head region of the personnel is defined as the upper 1 / 3 region of the bounding box.

[0096] The intersection area of the safety helmet bounding box and the head region is calculated and the union area ;

[0097] The intersection-to-union ratio is calculated: ;

[0098] When , the wearing state is marked as 1, otherwise it is marked as 0.

[0099] S45, encode the detection result into a structured data matrix:

[0100] A 12-dimensional feature vector is created: [number of personnel, average x coordinate of personnel, average y coordinate of personnel, minimum safety distance, safety helmet wearing rate, insulating glove wearing rate, insulating shoe wearing rate, number of charged equipment, work area, personnel density, equipment proportion, average detection confidence].

[0101] The coordinate values are normalized and mapped to the [0, 1] interval.

[0102] The wearing rate data is 0-1 encoded.

[0103] S46, feature enhancement is performed on the structured data:

[0104] The first 8 principal components are extracted using principal component analysis (PCA).

[0105] The principal components are spliced with the original 12-dimensional features to form a 20-dimensional image feature vector.

[0106] L2 normalization is performed on the feature vector to ensure that the vector length is 1.

[0107] In S5, the risk identification model is used to make predictions based on the image feature vector and the optimal model parameters, resulting in a risk assessment result, including:

[0108] S51, load the obtained optimal model parameters Configure the network parameters of the risk identification model;

[0109] S52, input the image feature vector into the configured risk identification model:

[0110] The image feature vector is transformed by the shared feature extraction layer;

[0111] The transformed features are input into three task output heads;

[0112] S53, obtain the prediction results from each task output head:

[0113] The operation sequence compliance output head outputs a binary classification probability value ;

[0114] The safe distance detection output head outputs a multi-classification probability distribution , where k is the number of distance levels;

[0115] The protective equipment specification output head outputs a binary classification probability value ;

[0116] S54, post-processing and risk level determination of the prediction results:

[0117] When , it is marked that the operation sequence is at risk of violation;

[0118] According to the class corresponding to the maximum probability , the safe distance risk level is determined;

[0119] When , it is marked that the protective equipment is not worn in a standard manner;

[0120] S55, generate a risk assessment report by integrating the prediction results of each task:

[0121] Calculate the comprehensive risk score: , where is the safe distance risk level, ;

[0122] According to the comprehensive risk score, the risk level is divided into low risk , medium risk , and high risk ;

[0123] The structured assessment results include each risk indicator and a comprehensive risk level.

[0124] The application created by the present application and its implementation are described above, which is not limited, and can be realized in other specific forms without departing from the spirit or essential characteristics of the application. The embodiment shown in the drawing is only one of the embodiments of the application, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired thereby, without departing from the spirit of the present application, similar structural forms and embodiments can be designed without creative design, which shall belong to the protection scope of the present application. In addition, the word "comprises" does not exclude other elements or steps, and the word "one" before the element does not exclude the inclusion of "multiple" elements. The words "first", "second" and the like are used to indicate names, and do not mean any particular order.

Claims

1. A safety detection method for power construction activities, characterized in that, include: S1. Collect historical power work order data and extract the fields of work content, safety measures, work steps, and hazard pre-control measures from the work order; construct a work order risk feature library based on the extracted fields. Each record in the risk feature library contains a work type code, a risk feature vector, and historical accident markers. S2, based on the work ticket risk feature library, train a risk identification model based on multi-task learning. The risk identification model includes a shared feature extraction layer and multiple task output heads, each of which corresponds to a type of work risk assessment task. S3: Collect the text data of the current power work order, extract the work content and safety measures fields, and generate a text feature vector; Based on the text feature vector, the optimal parameter configuration is searched in the parameter space of the risk identification model using the discretized tunicate optimization algorithm to obtain the optimal model parameters that match the current work order operation type. S4 collects real-time image data of the work site, performs target detection and feature extraction on the image data, identifies the location coordinates of the workers, as well as the type and wearing status of safety protective equipment, encodes the detection results into structured data, and generates image feature vectors; S5. Based on the image feature vector and the optimal model parameters, a risk identification model is used to make predictions and obtain risk assessment results.

2. The safety detection method for power construction activities according to claim 1, characterized in that: S2, based on the work order risk feature library, train a risk identification model based on multi-task learning, including: Risk feature vectors and historical accident markers are extracted from the work order risk feature database. The dataset is divided into three subsets based on the work type code: live-line work, switching operation, and maintenance work. Based on the data subset, a multi-task learning network architecture is constructed, in which the shared feature extraction layer adopts a three-layer fully connected network with three task output heads: The operation sequence compliance output header is used to predict whether the switching operation steps comply with the standard operating procedures. The safety distance detection output head is used to predict whether the safety distance between the operator and the live equipment meets the requirements of the regulations. The protective equipment specification output head is used to predict whether the use of safety helmets and insulating gloves meets the specification requirements; By setting a multi-task loss function and using the batch gradient descent algorithm for model training, a risk identification model based on multi-task learning is obtained.

3. The safety detection method for power construction activities according to claim 1, characterized in that: The multi-task loss function expression is as follows: ;in, The binary cross-entropy loss is for compliance with the order of operations. For multi-class cross-entropy loss in safe distance detection, To protect against the standard binary cross-entropy loss of the equipment, This represents the task weighting coefficient.

4. The safety detection method for power construction activities according to claim 2 or 3, characterized in that: S3 utilizes a discretized tunicate optimization algorithm to search for the optimal parameter configuration in the parameter space of the risk identification model, including: Define the parameter search space, including: learning rate Batch size and dropout rate ; Construct a discrete index mapping table to map the continuous parameter search space to discrete indexes; Construct the fitness function F: ;in, The accuracy of the model on the validation set. For model complexity, This is the balance coefficient; Perform a discretized optimization process for the tunicate, iteratively optimizing until the termination condition is met, and output the optimal parameter configuration. .

5. The safety detection method for power construction activities according to claim 4, characterized in that: Performing the discretized tunicate optimization process includes: Initialize population data matrix Where N is the population size, and each row represents an individual in a three-dimensional continuous space. Location data in the middle; Update the position data in each iteration: For individual leaders, according to the formula Update continuous location data; For individual followers, according to the formula Calculate new location data; For each continuous value Calculate candidate discrete indices: ; Calculate the mapping distance: ; According to probability Perform random mapping to generate the final discrete index; Extract actual parameter values ​​from the parameter candidate set based on discrete indexes; Configure the model using the extracted parameter values, evaluate performance on the validation set, calculate fitness values, and update the optimal solution record.

6. The safety detection method for power construction activities according to claim 4, characterized in that: S4 generates image feature vectors, including: Collect real-time image data of the work site and preprocess the images; Identify key targets in images using the YOLOv5 object detection model: Inspection personnel, output bounding box coordinates and confidence level ; Detect safety helmets and insulating gloves, and output the corresponding protective equipment bounding box and confidence level; Detects live equipment and operating equipment, and outputs the bounding box coordinates and category labels of the equipment; The spatial relationship and wearing status of the target are quantitatively analyzed based on the bounding box coordinates to obtain the detection results; The detection results are encoded into structured data to generate image feature vectors.

7. The safety detection method for power construction activities according to claim 6, characterized in that: The spatial relationships and wearing status of the target are quantitatively analyzed based on the bounding box coordinates to obtain the detection results, including: Extract edge sampling point sets from the bounding boxes of the workers and the live equipment, respectively; Calculate the Euclidean distance matrix between point pairs based on the edge sampling point set; The minimum value is extracted from the Euclidean distance matrix and used as the safety distance, which is then used as the result of target spatial relationship detection. The boundary box of the workers is divided into functional areas according to a preset ratio; Calculate the overlap between the boundary frame of the protective equipment and the corresponding functional areas of the human body; Binary wearing status detection results are generated based on the overlap threshold.

8. The safety detection method for power construction activities according to any one of claims 5 to 7, characterized in that: S5, using a risk identification model to make predictions, obtains risk assessment results, including: Configure the parameters of the risk identification model using the optimal model parameters and establish a parameterized model instance; Using image feature vectors as input data, based on a parameterized model instance: The input features are transformed nonlinearly through a shared feature extraction layer; The transformed features are input into multiple task output heads in parallel; Extract the corresponding predicted probability distribution from each output head; Perform threshold judgment and category mapping based on the predicted probability distribution; Based on the threshold judgment and category mapping results, a risk assessment result is generated through weighted aggregation.

9. A safety detection system for power construction activities, characterized in that, include: The historical data processing module is used to collect historical power work order data and extract fields such as work content, safety measures, work steps, and hazard pre-control measures from the work orders; it constructs a work order risk feature library based on the extracted fields, and each record in the risk feature library contains a work type code, a risk feature vector, and historical accident markers; The model training module, connected to the historical data processing module, is used to train a risk identification model based on multi-task learning according to the work ticket risk feature library. The risk identification model includes a shared feature extraction layer and multiple task output heads, each of which corresponds to a type of work risk assessment task. The parameter optimization module is used to collect the text data of the current power work order, extract the work content and safety measures fields, and generate a text feature vector; based on... The text feature vectors are used to search for the optimal parameter configuration in the parameter space of the risk identification model using the discretized sea squirt optimization algorithm, so as to obtain the optimal model parameters that match the current work order operation type. The image processing module is used to acquire real-time image data of the work site, perform target detection and feature extraction on the image data, identify the location coordinates of the workers, as well as the type and wearing status of safety protective equipment, encode the detection results into structured data, and generate image feature vectors. The risk assessment module is connected to the model training module, parameter optimization module, and image processing module, respectively. It is used to make predictions based on the image feature vector and the optimal model parameters using the risk identification model to obtain the risk assessment result.

Images