Work site safety detection methods, safety detection systems and safety fences

Abstract

The invention discloses a work site safety detection method, a safety detection system and a safety fence. The method comprises the following steps: using expanded data to train a feature extraction model, using the output of the feature extraction model to train a feature dimension reduction model, using the output of the feature dimension reduction model to train a classifier model, and applying each trained model to construction site detection, wherein the feature extraction model adopts a neural network based on a circular dynamic optimization algorithm, the feature dimension reduction model adopts an autoencoder neural network based on dual loss, and the classifier model adopts an extreme learning machine algorithm based on curvature-aware weight update; the feature extraction model of the invention can adjust the optimization strategy according to different training stages, simulate biological cycles to cope with different data characteristics, and improve the efficiency and accuracy of feature extraction; feature dimension reduction is performed by an autoencoder, and a method based on dual loss is adopted, so that the intrinsic structure of the data can be effectively mined and the representation ability of the reduced-dimensional data can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of safety detection, and in particular to a work site safety detection method, a safety detection system and a safety fence. Background Art

[0002] The traditional safety fence of the substation adopts a combination of ground stakes, poles, and fence warning tapes. The safety fence is installed and arranged according to the location of the specific maintenance content. The traditional safety fence isolates the working area (maintenance and test equipment) from the non-working area (live equipment) by finding suitable ground stakes to insert poles and then wrapping the fence warning tape around it.

[0003] At construction sites, ensuring the safety of workers is crucial because these environments are often highly dangerous, such as high voltage and complex mechanical operations. Workers must wear safety helmets and strictly abide by the regulations of safety fences to prevent any form of accidental injury or more serious accidents. However, in actual operations, monitoring whether workers comply with safety regulations, especially whether they wear safety equipment and whether they cross safety fences, often requires a lot of manpower and technical equipment for real-time monitoring. Existing safety monitoring methods rely on manual inspections or basic video surveillance systems, which are not only time-consuming and labor-intensive, but also have significant deficiencies in data collection, processing efficiency, and real-time response capabilities. These limitations increase the complexity of safety management and may affect the effective implementation of on-site safety measures.

[0004] The Chinese invention patent with application number CN202211371619.4 proposes a construction site work status monitoring method and system based on video linkage, including: obtaining real-time videos of multiple scenes at the construction site; fusing, identifying and analyzing the acquired real-time videos of multiple scenes to determine the real-time environmental status and personnel safety status of the construction site; and completing the monitoring of the construction site work status based on the obtained real-time environmental status and personnel safety status of the construction site.

[0005] The Chinese invention patent with application number CN202311856215.9 proposes a method, system and equipment for monitoring the distribution of construction site workers and detecting safety conditions, which involves the field of intelligent building safety. It uses artificial intelligence technology, namely the YOLOv5 network, to detect and classify targets in construction scene images, identify human bodies and human heads in construction scene images, and when identifying human heads, distinguish whether the human heads are wearing safety helmets. Then, according to the human bodies and human heads in the area blocks and whether the human heads are wearing safety helmets, the distribution of construction site workers is monitored and safety conditions are detected. In this way, artificial intelligence technology can be used to predict and evaluate risks, so that safety problems can be discovered at the first time and early warnings can be issued in time to ensure the personal safety of construction workers, make up for the defects caused by human management, and reduce the workload of supervisory staff.

[0006] The Chinese invention patent with application number CN202311649412.3 proposes a production equipment safety supervision method based on artificial intelligence. The equipment management QR code, monitoring camera (including edge algorithm box), sound and light alarm, voice announcer and controller are installed in the area or equipment where the equipment is located. The edge algorithm box is connected to the controller and remote monitoring platform signal through wireless signal or wired signal. The controller is connected to the sound and light alarm and voice announcer signal through wireless signal. The handheld mobile terminal is connected to the remote monitoring platform signal through wireless signal. The safety monitoring method includes the following steps: the first step: data reading, the second step: safety monitoring, the safety monitoring module simultaneously detects the inspection, inspection, maintenance and operation of the equipment, the third step: alarm, the fourth step: rectification and improvement, and the fifth step: performance assessment. This design not only comprehensively covers the safety management of enterprise equipment and facilities, and promptly checks the existing violations, but also eliminates the interference of personal relationships, urges the relevant responsible personnel to perform their duties safely, so as to take targeted measures, improve the level of refined management, and improve the efficiency of safety management.

[0007] The prior art has the following deficiencies:

[0008] 1. In the existing technology, data collection is limited by the specific conditions of on-site monitoring. It is often difficult to train a model with good generalization ability due to insufficient data volume, and insufficient diversity makes the model prone to overfitting.

[0009] 2. Traditional feature extraction methods fail to fully adapt to the variability of data, resulting in insufficient feature extraction efficiency and accuracy in complex scenarios.

[0010] 3. Existing feature dimensionality reduction technologies usually only focus on reducing data dimensions and ignore the mining of the intrinsic structure of the data, making the reduced data insufficient to support accurate classification and recognition.

[0011] 4. When dealing with complex data distribution, traditional classifiers lack an adaptive adjustment mechanism for data characteristics, especially in dynamically changing actual application scenarios, resulting in low classification accuracy. Summary of the invention

[0012] In view of the deficiencies in the prior art, the present invention provides a work site safety detection method, a safety detection system and a safety fence, which aim to solve the problems mentioned in the background technology.

[0013] To achieve the above object, the present invention provides the following technical solution: a method for detecting safety at a work site, comprising the following steps:

[0014] Step S1: Data collection and annotation: Collecting and annotating images of the construction site;

[0015] Step S2: data expansion; using a trained generative adversarial network algorithm based on hierarchical feature decomposition to generate samples of the collected construction site images to obtain an expanded image; the generative adversarial network based on hierarchical feature decomposition includes two components, a generator and a discriminator, and uses a hierarchical feature decomposition loss function;

[0016] Step S3: training feature extraction model;

[0017] A feature extraction model is constructed, the expanded image is stretched into a one-dimensional vector and then input into the feature extraction model, the feature extraction model is trained to obtain a trained feature extraction model, and the trained feature extraction model is used to extract features from the expanded image to obtain a feature map; the feature extraction model uses a neural network based on a circular dynamic optimization algorithm, the circular dynamic optimization algorithm adjusts the optimization strategy at different training stages to simulate various stages in the cell cycle, and different strategies are used in each stage to adjust the weight and bias of the neural network;

[0018] Step S4: training feature dimensionality reduction model;

[0019] Constructing a feature dimension reduction model, using a feature graph to train the feature dimension reduction model to obtain a trained feature dimension reduction model, and using the trained feature dimension reduction model to reduce the dimension of the feature graph to obtain a reduced-dimensional feature graph; the feature dimension reduction model uses an autoencoder neural network based on dual loss, and the autoencoder neural network based on dual loss optimizes the feature dimension reduction process by using the dual properties of the encoder and the decoder;

[0020] Step S5: training the classifier model;

[0021] A classifier model is constructed, and the classifier model is trained using a reduced-dimensionality feature map to obtain a trained classifier model, and the reduced-dimensionality feature map is classified using the trained classifier model to obtain a classification result; the classifier model adopts an extreme learning machine algorithm based on curvature-aware weight update, and the extreme learning machine algorithm based on curvature-aware weight update adopts a curvature-aware mechanism to analyze the local curvature changes of the input data and dynamically adjust the weight update rate of each data point;

[0022] Step S6: Construction site detection; the trained feature extraction model, the trained feature dimension reduction model and the trained classifier model are used to process the real-time construction site image.

[0023] Furthermore, the specific process of obtaining the trained generative adversarial network algorithm based on hierarchical feature decomposition is as follows:

[0024] Step S201: Initialize the generator and the discriminator The parameters are expressed as:

[0025] ;

[0026] In the formula, represents the standard normal distribution; Represents the parameters of the generator; represents the parameters of the discriminator; is the identity matrix; Indicates that the parameter is The discriminator of Indicates that the parameter is The generator of Represents a normal distribution with a mean of 0 and a variance of the identity matrix;

[0027] Step S202: Use the construction site image to pre-train the discriminator. The loss function of the discriminator is It is expressed as:

[0028] ;

[0029] In the formula, Represents the first samples, where the samples are images of the construction site; yes The corresponding annotation category; is the input batch size of the discriminator;

[0030] Step S203: Generator From random noise Generate a fake image, expressed as:

[0031] ;

[0032] ;

[0033] In the formula, Indicates the real image, i.e., the construction site image;

[0034] Discriminator For real images and fake images Evaluate, expressed as:

[0035] ;

[0036] ;

[0037] In the formula, Representation Discriminator The discrimination results of real images; Representation Discriminator The results of false image discrimination;

[0038] Calculating fake images The hierarchical feature deconstruction loss and adversarial loss of the generator are updated, and the hierarchical feature deconstruction loss is updated. and combat loss , expressed as:

[0039] ;

[0040] ;

[0041] In the formula, Indicates the extracted Layer characteristics; It is The weight coefficient of the layer feature; is the number of feature layers; Represents the expectation for the real image; represents the expectation of random noise;

[0042] Feature extraction function The features at a specific level in the real image or fake image are extracted through a combination of weighted convolutional layers, which can be expressed as:

[0043] ;

[0044] In the formula, It is Layer The weight of the convolution kernel; yes The corresponding bias term; represents the convolution operation, is the Sigmoid activation function; It is The number of convolution kernels of the layer; is the input of the function;

[0045] Define the output of the multilayer perceptron , expressed as:

[0046] ;

[0047] In the formula, It is The output of the hidden layer; is the weight from the hidden layer to the output layer; is the bias of the output layer, is the number of hidden layers;

[0048] Will Defined as ,in, and Respectively Layer weights and biases;

[0049] Update the discriminator and update the weights of the discriminator based on the evaluation results of the discriminator on real images and fake images; use the gradient descent method to update the parameters of the discriminator , expressed as:

[0050] ;

[0051] In the formula, Indicates parameter update symbol; represents the discriminator learning rate; represents the gradient operation; Denotes the discriminator loss function About the discriminator parameters The gradient of

[0052] Generator Parameters The update method is expressed as:

[0053] ;

[0054] In the formula, and They are the loss functions of the discriminator and the generator respectively; Represents the generator loss function About generator parameters The gradient of

[0055] Step S204: By analyzing the features in the fake image and feeding back the features to the generator, the parameters of the generator are updated again, which is expressed as:

[0056] ;

[0057] In the formula, is the learning enhancement coefficient; It is The reinforcement coefficient of the samples; is the target feature value of the real image; is the sample size; Indicates The generated features of samples;

[0058] gradient The calculation method is expressed as:

[0059] ;

[0060] Step S205: Repeat the iteration of steps S202 to S204 until a preset stop iteration condition is met.

[0061] Furthermore, the feature extraction model is trained to obtain the specific process of the trained feature extraction model as follows:

[0062] Step S301: Initialize a neural network, wherein the neural network adopts a 3-layer fully connected neural network, the number of hidden layers of the neural network is 2, the first hidden layer includes 100 neurons, the second hidden layer includes 50 neurons, and the activation function is a ReLU activation function;

[0063] Initialize the weights of the neural network and bias , and set the initial learning rate of the neural network , expressed as:

[0064] ;

[0065] ;

[0066] In the formula, represents the initial weights of the neural network; Represents the initial bias of the neural network; Indicates compliance with a specific distribution; The mean is 0 and the variance is Normal distribution of is the variance of the initial distribution;

[0067] Step S302: Using the accelerated gradient descent algorithm to update the parameters of the neural network, the updating method is expressed as:

[0068] ;

[0069] ;

[0070] In the formula, is the acceleration factor; The loss function for neural network training; is the gradient of the loss function with respect to the weight parameter; is the gradient of the loss function with respect to the bias parameter; For the The weight parameter of the iteration; For the The bias parameter for the iteration; For the The weight parameter of the iteration; For the The bias parameter for the iteration; is the symbol of partial derivative; is the learning rate during the growth phase;

[0071] Step S303: synthesize multiple weight parameters into new parameters, that is, define a synthesis function , expressed as:

[0072] ;

[0073] In the formula, yes The importance coefficient of for the number of

[0074] Importance coefficient Based on the previous stage The gradient size is dynamically calculated, The calculation method of the importance coefficient of each weight parameter is expressed as:

[0075] ;

[0076] In the formula, is the decay rate parameter;

[0077] Step S304: Setting performance evaluation function The feature extraction model output of the current iteration is used to determine whether it will enter the next iteration. The performance evaluation function is expressed as:

[0078] ;

[0079] In the formula, represents the performance evaluation function; Represents the loss of neural network training; is the performance threshold; when If the calculation is 1, then proceed to the next iteration, otherwise proceed to step S305;

[0080] Step 305: Use the perturbation method as the adjustment increment of the weight and bias, and the calculation method is expressed as:

[0081] ;

[0082] ;

[0083] In the formula, is the adjustment increment of the weight parameter; is the adjustment increment of the bias parameter; and are the amplitude and frequency parameters of the perturbation respectively;

[0084] and The setting method is expressed as:

[0085] ;

[0086] ;

[0087] In the formula, is the scaling factor of the perturbation amplitude;

[0088] Step S306: Constrain the weights of the neural network by weight pruning, which is expressed as:

[0089] ;

[0090] In the formula, represents the weight pruning function; is the gradient threshold; Represents the dot product operation;

[0091] Step S307: Adjusting the neural network learning rate , the adjustment method is expressed as:

[0092] ;

[0093] In the formula, is the attenuation coefficient; represents the initial learning rate; is the number of completed cycles;

[0094] Step S308: Repeat the iteration of steps S302 to S307 until a preset stop iteration condition is met.

[0095] Furthermore, the acceleration factor According to the decreasing speed and curvature setting of the loss function, the calculation method is:

[0096] ;

[0097] In the formula, is the preset maximum acceleration factor; is the minimum value function;

[0098] Decay rate parameter The calculation method is expressed as:

[0099] .

[0100] Furthermore, the feature graph is used to train the feature dimensionality reduction model, and the specific process of obtaining the trained feature dimensionality reduction model is as follows:

[0101] Step S401: Initialize the parameters of the autoencoder neural network;

[0102] Assume that the encoder parameters of the autoencoder neural network are , the decoder parameters of the autoencoder neural network are , use random initialization to initialize the weights and biases of the encoder and decoder of the autoencoder neural network;

[0103] Step S402: Input feature map , feature map Transformed into low-dimensional feature representation through encoder , and then the low-dimensional features are represented as The decoder reconstructs the output into the original dimension , expressed as:

[0104] ;

[0105] ;

[0106] In the formula, is the encoder function; is the decoder function;

[0107] Step S403: Calculate the dual loss, the dual loss Including reconstruction error and sparse self-expression error , expressed as:

[0108] ;

[0109] In the formula, is the regularization parameter of the dual loss;

[0110] Reconstruction error measure and The difference between is expressed as:

[0111] ;

[0112] In the formula, For the indivual ; For the indivual ;

[0113] Sparse self-expression error Compute low-dimensional feature representation and The error between the sparse linear combinations of is expressed as:

[0114] ;

[0115] In the formula, The feature map quantity; is the regularization parameter of the sparse self-expression error; is the self-expression matrix; For the A low-dimensional feature representation;

[0116] Step S404: Calculate using the chain rule based on the dual loss about and The gradient of , and the parameter update, is expressed as:

[0117] ;

[0118] ;

[0119] In the formula, is the learning rate of the autoencoder; The dual loss about The gradient of The dual loss about The gradient of

[0120] Step S405: Repeat the iteration of steps S402 to S404 until a preset stop iteration condition is met.

[0121] Furthermore, the self-expression matrix The construction method is expressed as:

[0122] ;

[0123] ;

[0124] In the formula, It includes all Matrix of for The transpose of is a hyperparameter of density; Represents Euclidean distance calculation; is the Softmax function.

[0125] Furthermore, the classifier model is trained using the dimension reduction feature map, and the specific process of obtaining the trained classifier model is as follows:

[0126] Step S501: Initialize the weights of the hidden layer of the extreme learning machine using random distribution and bias , expressed as:

[0127] ;

[0128] ;

[0129] In the formula, is the hidden layer weight; is the hidden layer bias; is the identity matrix; is a normal distribution;

[0130] Step S502: Input the reduced dimension feature map into the hidden layer of the extreme learning machine, and use the Sigmoid activation function to calculate the output of the hidden layer of the extreme learning machine , expressed as:

[0131] ;

[0132] In the formula, is the reduced dimension feature map of the input;

[0133] Step S503: Before calculating the weight of the output layer of the extreme learning machine, The curvature information adjusts the learning rate;

[0134] By calculating the local curvature of each reduced dimension feature map, a dynamic learning rate is assigned to each reduced dimension feature map , the local curvature is calculated using the second-order derivative of the reduced-dimensional feature map , the calculation method of the local curvature of the reduced dimension feature map is expressed as:

[0135] ;

[0136] In the formula, and Represents the dimension reduction feature map The first and second derivatives of ; Represents the local curvature of the reduced-dimensional feature map;

[0137] Dynamic Learning Rate It is expressed as:

[0138] ;

[0139] In the formula, Based on the local curvature The calculated dynamic learning rate;

[0140] It is calculated by the central difference method and is expressed as:

[0141] ;

[0142] In the formula, is the sampling interval;

[0143] It is calculated by the difference method and is expressed as:

[0144] ;

[0145] Step S504: Using permutation entropy The importance of is evaluated, and the calculation method of permutation entropy is expressed as:

[0146] ;

[0147] In the formula, yes No. 1 in the sequence The probability of a characteristic element appearing; express The permutation entropy of Indicates the number of features after feature dimensionality reduction;

[0148] The weight of the extreme learning machine is adjusted based on the permutation entropy, expressed as:

[0149] ;

[0150] In the formula, is the adjusted weight; is the original hidden layer weight;

[0151] Step S505: Use the least squares method to solve the weight of the output layer of the extreme learning machine, expressed as:

[0152] ;

[0153] In the formula, is the weight of the output layer; is the output of the hidden layer; express The transpose of is the target output vector.

[0154] A safety detection system for implementing a work site safety detection method, comprising:

[0155] The collection and annotation module is used to collect and annotate images of the construction site;

[0156] A data expansion module is used to generate samples of the collected construction site images using a trained generative adversarial network algorithm based on hierarchical feature decomposition to obtain an expanded image;

[0157] The feature extraction model training module is used to construct a feature extraction model, stretch the expanded image into a one-dimensional vector and then input it into the feature extraction model, train the feature extraction model to obtain a trained feature extraction model, and use the trained feature extraction model to extract features from the expanded image to obtain a feature map;

[0158] A feature dimension reduction model training module is used to construct a feature dimension reduction model, train the feature dimension reduction model using a feature graph to obtain a trained feature dimension reduction model, and reduce the dimension of the feature graph using the trained feature dimension reduction model to obtain a reduced dimension feature graph;

[0159] The classifier model training module is used to construct a classifier model, train the classifier model using the reduced dimension feature map to obtain a trained classifier model, and classify the reduced dimension feature map using the trained classifier model to obtain a classification result;

[0160] The on-site detection module is used to process real-time construction site images using the trained feature extraction model, the trained feature dimension reduction model and the trained classifier model.

[0161] A safety fence comprises: a camera, a radar, a solar panel, a power module, a voice and sound and light alarm module, a safety detection system and a retractable warning belt; the radar is used to monitor whether the heart rate of on-site workers is normal; the camera is used to photograph the behavior of on-site workers; the safety detection system is used to perform safety detection on the behavior of on-site workers photographed by the camera; the voice and sound and light alarm module is used to give an alarm according to the detection result of the safety detection system; the solar panel is used to convert light energy into electrical energy; the power module is used to store the electrical energy converted by the solar panel and to supply power to the camera, the radar, the voice and sound and light alarm module and the safety detection system.

[0162] Furthermore, it also includes hooks or strong magnets, which are used to fix the safety fence.

[0163] Compared with the existing technology, the present invention has the following beneficial effects:

[0164] (1) The present invention adopts a generative adversarial network based on hierarchical feature decomposition for data expansion, which effectively solves the problem of insufficient training samples caused by limited monitoring image data and complex scenes; the generator and discriminator generate sufficiently realistic images through adversarial learning to enhance the diversity and quantity of model training sets; through data expansion technology, the coverage and diversity of the data set can be significantly improved without increasing the actual monitoring equipment and manpower costs, effectively reducing the model overfitting phenomenon and improving the generalization ability of the model.

[0165] (2) The feature extraction model adopted in the present invention is implemented through a three-layer fully connected neural network, and adopts a circular dynamic optimization algorithm. The optimization strategy is adjusted according to different training stages, and the biological cycle is simulated to cope with different data characteristics, thereby improving the efficiency and accuracy of feature extraction. The circular dynamic optimization enables the model to adapt to new data characteristics in a short time, reduces the risk of overfitting, and ensures the efficiency and stability of the feature extraction process.

[0166] (3) The present invention performs feature dimensionality reduction through an autoencoder and adopts a dual loss-based method to minimize the difference between input and output, and enhances the feature representation of the code through a sparse self-expression layer, effectively mining the intrinsic structure of the data and improving the representation ability of the reduced-dimensional data; feature dimensionality reduction not only improves the processing speed, but also improves the representation quality of the data through a deeper level of data structure analysis, which contributes to more accurate classification and recognition.

[0167] (4) The present invention adopts a curvature-aware extreme learning machine algorithm as a classifier model, and dynamically adjusts the weight update rate by analyzing the local curvature changes of the input data, thereby optimizing the adaptability to complex data distribution and improving the classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0168] Figure 1 The present invention is a flow chart of the method.

[0169] Figure 2 It is a schematic diagram of the safety fence structure of the present invention. DETAILED DESCRIPTION

[0170] The present invention provides a technical solution:

[0171] like Figure 1 As shown, a method for safety detection at a work site comprises the following steps:

[0172] Step S1: data collection and labeling; collect construction site images and store the collected construction site images in jpg format; label the collected construction site images, and the labeling categories include: wearing a safety helmet, not wearing a safety helmet, crossing the safety fence, and not crossing the safety fence.

[0173] Data collection, acquisition, labeling and preprocessing are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor model generalization ability and affect the accuracy of the model.

[0174] Step S2: data expansion: using the trained generative adversarial network algorithm based on hierarchical feature decomposition to generate samples of the collected construction site images to obtain expanded images.

[0175] The generative adversarial network based on hierarchical feature decomposition includes two components: a generator and a discriminator; the generator is responsible for generating sufficiently realistic image data, while the discriminator attempts to distinguish the generated image from the real image; in view of the problem that the traditional generative adversarial network has insufficient generalization ability in specific tasks such as seat belt and helmet recognition, the present invention adopts a hierarchical feature deconstruction loss function, which not only considers the global image quality, but also focuses on optimizing the feature area of ​​the generated image, i.e., the fake image, to ensure that the features of the seat belt and helmet can be effectively learned and reproduced.

[0176] Step S3: training a feature extraction model; constructing a feature extraction model, stretching the expanded image into a one-dimensional vector and inputting it into the feature extraction model, training the feature extraction model to obtain a trained feature extraction model, and using the trained feature extraction model to extract features from the expanded image to obtain a feature map.

[0177] The traditional gradient descent method relies on global gradient information to update weights. It is inspired by the periodic metabolic process of biological cells. Different mechanisms are used at different stages to optimize the response and adaptability of organisms. A neural network based on a circular dynamic optimization algorithm is used for feature extraction. The circular dynamic optimization algorithm adjusts the optimization strategy at different training stages to simulate the various stages of the cell cycle. Different strategies are used in each stage to adjust the weights and biases of the neural network, so that the algorithm can adapt to changes in training data more flexibly, reduce the risk of overfitting, and effectively escape from the local optimal solution.

[0178] Step S4: training a feature dimensionality reduction model; constructing a feature dimensionality reduction model, using a feature graph to train the feature dimensionality reduction model to obtain a trained feature dimensionality reduction model, and using the trained feature dimensionality reduction model to reduce the dimension of the feature graph to obtain a reduced-dimensional feature graph.

[0179] An autoencoder neural network based on dual loss is used as a feature dimensionality reduction model. The autoencoder neural network based on dual loss optimizes the feature dimensionality reduction process by utilizing the duality property of the encoder and the decoder. The encoder part compresses the input feature vector into a low-dimensional representation, and the decoder part attempts to reconstruct the input feature from the low-dimensional representation. The traditional autoencoder uses the reconstruction error as the optimization target. The autoencoder of the present invention not only minimizes the difference between the input and the reconstructed output, but also forces more structural information to be revealed in the encoded feature representation through a sparse self-expression layer, thereby realizing in-depth mining of the intrinsic structural characteristics of the data.

[0180] Step S5: training a classifier model; constructing a classifier model, using the reduced-dimensionality feature graph to train the classifier model to obtain a trained classifier model, and using the trained classifier model to classify the reduced-dimensionality feature graph to obtain a classification result.

[0181] The extreme learning machine algorithm based on curvature-aware weight update is adopted as the classifier model. To address the problem of insufficient generalization ability of traditional extreme learning machines under complex data distribution, a curvature-aware mechanism is adopted to dynamically adjust the weight update rate of each data point by analyzing the local curvature changes of the input data, so as to improve the generalization ability and classification accuracy of the classifier model.

[0182] Step S6: Construction site detection; the trained feature extraction model, the trained feature dimension reduction model and the trained classifier model are used to process the real-time construction site image.

[0183] Among them, the specific process of obtaining the trained generative adversarial network algorithm based on hierarchical feature decomposition is:

[0184] Step S201: Initialize the generator and the discriminator The parameters are expressed as:

[0185] (1);

[0186] In the formula, represents the standard normal distribution; Represents the parameters of the generator; represents the parameters of the discriminator; is the identity matrix; Indicates that the parameter is The discriminator of Indicates that the parameter is The generator of Represents a normal distribution with mean 0 and variance the identity matrix.

[0187] Step S202: Use the construction site images to pre-train the discriminator to improve its classification accuracy in the initial stage. In this stage, the loss function of the discriminator is It is expressed as:

[0188] (2);

[0189] In the formula, Represents the first samples (construction site images); yes The corresponding annotation category; is the input batch size of the discriminator.

[0190] Step S203: Generator From random noise Generate a fake image, expressed as:

[0191] (3);

[0192] (4);

[0193] In the formula, Represents a real image (construction site image).

[0194] Discriminator For real images (Construction site images) and fake images Evaluate, expressed as:

[0195] (5);

[0196] (6);

[0197] In the formula, Representation Discriminator The discrimination results of real images; Representation Discriminator The results of distinguishing fake images.

[0198] Among them, calculate the fake image The hierarchical feature deconstruction loss and adversarial loss of the generator are updated, and the hierarchical feature deconstruction loss is updated. and combat loss , expressed as:

[0199] (7);

[0200] (8);

[0201] In the formula, Indicates the extracted Layer characteristics; It is The weight coefficient of the layer feature; is the number of feature layers; Represents the expectation for the real image; represents the expectation of random noise.

[0202] Feature extraction function The features at a specific level in the image data (real image or fake image) are extracted through a combination of weighted convolutional layers, which can be expressed as:

[0203] (9);

[0204] In the formula, It is Layer The weight of the convolution kernel; yes The corresponding bias term; represents the convolution operation, is the Sigmoid activation function; It is The number of convolution kernels of the layer; is the input of the function.

[0205] Define the output of the multilayer perceptron , expressed as:

[0206] (10);

[0207] In the formula, It is The output of the hidden layer; is the weight from the hidden layer to the output layer; is the bias of the output layer, is the number of hidden layers.

[0208] Will Defined as ,in, and Respectively The weights and biases of the layer.

[0209] Among them, the discriminator is updated, and the weight of the discriminator is updated according to the evaluation results of the discriminator on the real image and the fake image; specifically, the parameters of the discriminator are updated using the gradient descent method , expressed as:

[0210] (11);

[0211] In the formula, Indicates parameter update symbol; represents the discriminator learning rate; represents the gradient operation; Denotes the discriminator loss function About the discriminator parameters In this embodiment, Set to 0.01.

[0212] Similarly, the generator parameters The update method is expressed as:

[0213] (12);

[0214] In the formula, and They are the loss functions of the discriminator and the generator respectively; Represents the generator loss function About generator parameters In this embodiment, Set to 0.01.

[0215] Step S204: by means of feature reinforcement learning, that is, by analyzing the features in the fake image (such as the helmet) and feeding back the features to the generator, the generator can generate these features more accurately in subsequent training iterations, enhance the generator's feature generation capability, and update the generator parameters again, which is expressed as:

[0216] (13);

[0217] In the formula, is the learning enhancement coefficient; It is The reinforcement coefficient of the samples; is the target feature value of the real image; is the sample size; Indicates The generated features of samples; in this embodiment, Set to 0.4.

[0218] Among them, the gradient The calculation method is expressed as:

[0219] (14);

[0220] In the formula, It can be calculated using the chain rule.

[0221] Step S205: Repeat the iterations of steps S202 to S204 until a preset condition for stopping the iteration is met. In this embodiment, the preset condition for stopping the iteration is reaching a preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0222] Among them, the specific process of training the feature extraction model to obtain the trained feature extraction model is:

[0223] Step S301: Initialize the neural network. The neural network adopts a 3-layer fully connected neural network. The number of hidden layers of the neural network is 2. The first hidden layer includes 100 neurons, the second hidden layer includes 50 neurons, and the activation function is the ReLU activation function.

[0224] Specifically, initialize the weights of the neural network and bias , and set the initial learning rate of the neural network , expressed as:

[0225] (15);

[0226] (16);

[0227] In the formula, represents the initial weights of the neural network; Represents the initial bias of the neural network; Indicates compliance with a specific distribution; The mean is 0 and the variance is Normal distribution of is the variance of the initial distribution; in this embodiment, Set to 0.01.

[0228] Step S302: Using the accelerated gradient descent algorithm to update the parameters of the neural network, the updating method is expressed as:

[0229] (17);

[0230] (18);

[0231] In the formula, is the acceleration factor; The loss function for neural network training; is the gradient of the loss function with respect to the weight parameter; is the gradient of the loss function with respect to the bias parameter; For the The weight parameter of the iteration; For the The bias parameter for the iteration; For the The weight parameter of the iteration; For the The bias parameter for the iteration; is the symbol of partial derivative; is the learning rate in the growth phase; in this embodiment, the loss function Using cross entropy loss, Set to 0.01.

[0232] Among them, the acceleration factor According to the decline rate and curvature of the loss function, in order to avoid over-adjustment, the calculation method is:

[0233] (19);

[0234] In the formula, is the preset maximum acceleration factor; is the minimum value function.

[0235] Step S303: synthesize multiple weight parameters into new parameters through specific operations to improve the generalization ability of the network; define the synthesis function , expressed as:

[0236] (20);

[0237] In the formula, yes The importance coefficient of for The number of

[0238] Among them, the importance coefficient Based on the previous stage The gradient size is dynamically calculated, The calculation method of the importance coefficient of each weight parameter is expressed as:

[0239] (twenty one);

[0240] In the formula, is the decay rate parameter.

[0241] Among them, the decay rate parameter The calculation method is expressed as:

[0242] (twenty two).

[0243] Step S304: Setting performance evaluation function , to decide whether the feature extraction model output of the current iteration will enter the next iteration. The performance evaluation function is expressed as:

[0244] (twenty three);

[0245] In the formula, represents the performance evaluation function; Represents the loss of neural network training; is the performance threshold; in this embodiment, Set to 0.5; when If the calculation is 1, the next iteration is entered, that is, the growth phase is entered; otherwise, the pre-mitotic phase is entered, that is, step S305.

[0246] Step 305: Use the perturbation method as the adjustment increment of the weight and bias, and the calculation method is expressed as:

[0247] (twenty four);

[0248] (25);

[0249] In the formula, is the adjustment increment of the weight parameter; is the adjustment increment of the bias parameter; and are the amplitude and frequency parameters of the perturbation respectively; in this embodiment, Set to 5, Set to 3.14.

[0250] in, and The setting method is expressed as:

[0251] (26);

[0252] (27);

[0253] In the formula, is the scaling factor of the perturbation amplitude; in this embodiment, Set to 0.95.

[0254] Step S306: Constrain the weights of the neural network by weight pruning, which is expressed as:

[0255] (28);

[0256] In the formula, represents the weight pruning function; is the gradient threshold; represents a dot product operation; in this embodiment, Set it to 0.01, that is, only keep the parameters whose weight is greater than 0.01.

[0257] Step S307: Adjusting the neural network learning rate , so that the learning rate decreases, and the adjustment method is expressed as:

[0258] (29);

[0259] In the formula, is the attenuation coefficient; represents the initial learning rate; is the number of completed cycles; in this embodiment, Set to 0.95.

[0260] Step S308: Repeat the iterations of steps S302 to S307 until a preset condition for stopping iteration is met. In this embodiment, the preset condition for stopping iteration is reaching a preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0261] Among them, the feature map is used to train the feature dimensionality reduction model, and the specific process of obtaining the trained feature dimensionality reduction model is as follows:

[0262] Step S401: Initialize the parameters of the autoencoder neural network; specifically, the encoder parameters of the autoencoder neural network are set to , the decoder parameters of the autoencoder neural network are , the weights and biases of the encoder and decoder of the autoencoder neural network are initialized by random initialization.

[0263] Step S402: Input feature map , feature map The encoder transforms the low-dimensional feature representation into a low-dimensional space , and then the low-dimensional features are represented as The decoder reconstructs the output into the original dimension , expressed as:

[0264] (30);

[0265] (31);

[0266] In the formula, is the encoder function; Decoder function.

[0267] Step S403: Calculate the dual loss, the dual loss Including reconstruction error and sparse self-expression error , expressed as:

[0268] (32);

[0269] In the formula, is the regularization parameter of the dual loss.

[0270] Reconstruction error measure and The difference between is expressed as:

[0271] (33);

[0272] In the formula, For the indivual ; For the indivual .

[0273] Sparse self-expression error Compute low-dimensional feature representation and The error between the sparse linear combinations of is expressed as:

[0274] (34);

[0275] In the formula, The feature map quantity; is the regularization parameter of the sparse self-expression error; is the self-expression matrix; For the A low-dimensional feature representation.

[0276] Among them, the self-expression matrix The construction method is expressed as:

[0277] (35);

[0278] (36);

[0279] In the formula, It includes all Matrix of for The transpose of is a hyperparameter of density; Represents Euclidean distance calculation; is the Softmax function.

[0280] Step S404: Calculate using the chain rule based on the dual loss about and The gradient of , and the parameter update, is expressed as:

[0281] (37);

[0282] (38);

[0283] In the formula, is the learning rate of the autoencoder; The dual loss about The gradient of The dual loss about In this embodiment, Set to 0.01.

[0284] Step S405: Repeat the iterations of step S402 to step S404 until a preset condition for stopping iteration is met; in this embodiment, the preset condition for stopping iteration is reaching a preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0285] Among them, the specific process of using the reduced dimension feature map to train the classifier model and obtain the trained classifier model is:

[0286] Step S501: Initialize the weights of the hidden layer of the extreme learning machine using random distribution and bias , expressed as:

[0287] (39);

[0288] (40);

[0289] In the formula, is the hidden layer weight; is the hidden layer bias; is the identity matrix; is a normal distribution.

[0290] Step S502: Input the reduced dimension feature map into the hidden layer of the extreme learning machine, and use the Sigmoid activation function to calculate the output of the hidden layer of the extreme learning machine , expressed as:

[0291] (41);

[0292] In the formula, is the reduced dimension feature map of the input.

[0293] Step S503: Before calculating the weight of the output layer of the extreme learning machine, Specifically, by calculating the local curvature of each reduced-dimensionality feature map, a dynamic learning rate is assigned to each reduced-dimensionality feature map. , so as to achieve sensitive response to complex data features; in this embodiment, the local curvature The calculation uses the second-order derivative of the reduced-dimensional feature map, assuming that the reduced-dimensional feature map Approximately to its local linear expansion, the calculation method of the local curvature of the reduced dimensionality feature map is expressed as:

[0294] (42);

[0295] In the formula, and Represents the dimension reduction feature map The first and second derivatives of ; Represents the local curvature of the reduced feature map.

[0296] Among them, the dynamic learning rate It is expressed as:

[0297] (43);

[0298] In the formula, Based on the local curvature The calculated dynamic learning rate.

[0299] in, It is calculated by the central difference method and is expressed as:

[0300] (44);

[0301] In the formula, is the sampling interval.

[0302] in, It is calculated by the difference method and is expressed as:

[0303] (45).

[0304] Step S504: Using permutation entropy The importance of The influence during the training process, optimizing the classification boundary, and the calculation method of permutation entropy are expressed as:

[0305] (46);

[0306] In the formula, yes No. 1 in the sequence The probability of a characteristic element appearing; express The permutation entropy of Indicates the number of features after feature dimensionality reduction.

[0307] The weight of the extreme learning machine is adjusted based on the permutation entropy, expressed as:

[0308] (47);

[0309] In the formula, is the adjusted weight; are the original hidden layer weights.

[0310] Step S505: Use the least squares method to quickly solve the weight of the output layer of the extreme learning machine, expressed as:

[0311] (48);

[0312] In the formula, is the weight of the output layer; is the output of the hidden layer; express The transpose of is the target output vector.

[0313] A safety detection system for implementing a work site safety detection method, comprising:

[0314] The collection and annotation module is used to collect and annotate images of the construction site; the data expansion module is used to generate samples of the collected construction site images using the trained generative adversarial network algorithm based on hierarchical feature decomposition to obtain an expanded image; the feature extraction model training module is used to build a feature extraction model, stretch the expanded image into a one-dimensional vector and then input it into the feature extraction model, train the feature extraction model to obtain a trained feature extraction model, and use the trained feature extraction model to extract features from the expanded image to obtain a feature map; the feature dimension reduction model training module is used to build a feature dimension reduction model, train the feature dimension reduction model using the feature map to obtain a trained feature dimension reduction model, and use the trained feature dimension reduction model to reduce the dimension of the feature map to obtain a reduced dimension feature map; the classifier model training module is used to build a classifier model, train the classifier model using the reduced dimension feature map to obtain a trained classifier model, and use the trained classifier model to classify the reduced dimension feature map to obtain a classification result; the on-site detection module is used to use the trained feature extraction model, the trained feature dimension reduction model and the trained classifier model to process real-time construction site images.

[0315] like Figure 2As shown, a safety fence includes: a camera 1, a radar 2, a solar panel 3, a power module, a voice and sound and light alarm module 5, a safety detection system 6 and a retractable warning belt 7; the radar 2 is used to monitor whether the heart rate of the on-site workers is normal; the camera 1 is used to shoot the behavior of the on-site workers; the safety detection system 6 is used to perform safety detection on the behavior of the on-site workers shot by the camera 1; the voice and sound and light alarm module 5 is used to alarm according to the detection result of the safety detection system 6; the solar panel 3 is used to convert light energy into electrical energy; the power module is used to store the electrical energy converted by the solar panel 3, and to power the camera 1, the radar 2, the voice and sound and light alarm module 3 and the safety detection system 6; wherein the power module is arranged inside the safety fence, and the power module can also be used with an external power supply; the fence can be formed by pulling out the retractable warning belt 7.

[0316] Among them, hooks or strong magnets are also included, and the hooks or strong magnets are used to fix the safety fence; the safety fence can be hung on or magnetically attracted to the pillar for fixation through the hooks and strong magnets.

[0317] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for safety detection at a work site, characterized in that: The steps include: Step S1: Data collection and annotation: Collecting and annotating images of the construction site; Step S2: data expansion; using a trained generative adversarial network algorithm based on hierarchical feature decomposition to generate samples of the collected construction site images to obtain an expanded image; the generative adversarial network based on hierarchical feature decomposition includes two components, a generator and a discriminator, and uses a hierarchical feature decomposition loss function; Step S3: training feature extraction model; A feature extraction model is constructed, the expanded image is stretched into a one-dimensional vector and then input into the feature extraction model, the feature extraction model is trained to obtain a trained feature extraction model, and the trained feature extraction model is used to extract features from the expanded image to obtain a feature map; the feature extraction model uses a neural network based on a circular dynamic optimization algorithm, the circular dynamic optimization algorithm adjusts the optimization strategy at different training stages to simulate various stages in the cell cycle, and different strategies are used in each stage to adjust the weight and bias of the neural network; Step S4: training feature dimensionality reduction model; Constructing a feature dimension reduction model, using a feature graph to train the feature dimension reduction model to obtain a trained feature dimension reduction model, and using the trained feature dimension reduction model to reduce the dimension of the feature graph to obtain a reduced-dimensional feature graph; the feature dimension reduction model uses an autoencoder neural network based on dual loss, and the autoencoder neural network based on dual loss optimizes the feature dimension reduction process by using the dual properties of the encoder and the decoder; Step S5: training the classifier model; Construct a classifier model, use the reduced dimension feature map to train the classifier model to obtain a trained classifier model, use the trained classifier model to classify the reduced dimension feature map to obtain a classification result; The classifier model adopts an extreme learning machine algorithm based on curvature-aware weight update, and the extreme learning machine algorithm based on curvature-aware weight update adopts a curvature-aware mechanism to analyze the local curvature changes of the input data and dynamically adjust the weight update rate of each data point; Step S6: Construction site detection; using the trained feature extraction model, the trained feature dimension reduction model and the trained classifier model to process the real-time construction site image; The specific process of training the classifier model using the reduced dimension feature map to obtain the trained classifier model is as follows: Step S501: Initialize the weights of the hidden layer of the extreme learning machine using random distribution and bias , expressed as: ； ； In the formula, is the hidden layer weight; is the hidden layer bias; is the identity matrix; is a normal distribution; Step S502: Input the reduced dimension feature map into the hidden layer of the extreme learning machine, and use the Sigmoid activation function to calculate the output of the hidden layer of the extreme learning machine , expressed as: ； In the formula, is the reduced dimension feature map of the input; Step S503: Before calculating the weight of the output layer of the extreme learning machine, The curvature information adjusts the learning rate; By calculating the local curvature of each reduced dimension feature map, a dynamic learning rate is assigned to each reduced dimension feature map , the local curvature is calculated using the second-order derivative of the reduced-dimensional feature map , the calculation method of the local curvature of the reduced dimension feature map is expressed as: ； In the formula, and Represents the dimension reduction feature map The first and second derivatives of ; Represents the local curvature of the reduced-dimensional feature map; Dynamic Learning Rate It is expressed as: ； In the formula, Based on the local curvature The calculated dynamic learning rate; It is calculated by the central difference method and is expressed as: ； In the formula, is the sampling interval; Indicates A dimension-reduced feature map; Indicates A dimension-reduced feature map; It is calculated by the difference method and is expressed as: ； In the formula, Indicates A dimension-reduced feature map; Step S504: Using permutation entropy The importance of is evaluated, and the calculation method of permutation entropy is expressed as: ； In the formula, yes No. 1 in the sequence The probability of a characteristic element appearing; express The permutation entropy of Indicates the number of features after feature dimensionality reduction; The weight of the extreme learning machine is adjusted based on the permutation entropy, expressed as: ； In the formula, is the adjusted weight; is the original hidden layer weight; Step S505: Use the least squares method to solve the weight of the output layer of the extreme learning machine, expressed as: ； In the formula, is the weight of the output layer; is the output of the hidden layer; express The transpose of is the target output vector.

2. A method for detecting safety at a work site according to claim 1, characterized in that: The specific process of obtaining a trained generative adversarial network algorithm based on hierarchical feature decomposition is as follows: Step S201: Initialize the generator and the discriminator The parameters are expressed as: ； In the formula, represents the standard normal distribution; Represents the parameters of the generator; represents the parameters of the discriminator; is the identity matrix; Indicates that the parameter is The discriminator of Indicates that the parameter is The generator of Represents a normal distribution with a mean of 0 and a variance of the identity matrix; Step S202: Use the construction site image to pre-train the discriminator. The loss function of the discriminator is It is expressed as: ； In the formula, Represents the first samples, where the samples are images of the construction site; yes The corresponding annotation category; is the input batch size of the discriminator; Step S203: Generator From random noise Generate a fake image, expressed as: ； ； In the formula, Indicates the real image, i.e., the construction site image; Discriminator For real images and fake images Evaluate, expressed as: ； ； In the formula, Representation Discriminator The discrimination results of real images; Representation Discriminator The results of false image discrimination; Calculating fake images The hierarchical feature deconstruction loss and adversarial loss of the generator are updated, and the hierarchical feature deconstruction loss is updated. and combat loss , expressed as: ； ； In the formula, Indicates the extracted Layer characteristics; It is The weight coefficient of the layer feature; is the number of feature layers; Represents the expectation for the real image; represents the expectation of random noise; Feature extraction function The features at a specific level in the real image or fake image are extracted through a combination of weighted convolutional layers, which can be expressed as: ； In the formula, It is Layer The weight of the convolution kernel; yes The corresponding bias term; represents the convolution operation, is the Sigmoid activation function; It is The number of convolution kernels of the layer; is the input of the function; Define the output of the multilayer perceptron , expressed as: ； In the formula, It is The output of the hidden layer; is the weight from the hidden layer to the output layer; is the bias of the output layer, is the number of hidden layers; Will Defined as ,in, and Respectively Layer weights and biases; Update the discriminator and update the weights of the discriminator based on the evaluation results of the discriminator on real images and fake images; use the gradient descent method to update the parameters of the discriminator , expressed as: ； In the formula, Indicates parameter update symbol; represents the discriminator learning rate; represents the gradient operation; Denotes the discriminator loss function About the discriminator parameters The gradient of Generator Parameters The update method is expressed as: ； In the formula, and They are the loss functions of the discriminator and the generator respectively; Represents the generator loss function About generator parameters The gradient of Step S204: By analyzing the features in the fake image and feeding back the features to the generator, the parameters of the generator are updated again, which is expressed as: ； In the formula, is the learning enhancement coefficient; It is The reinforcement coefficient of the samples; is the target feature value of the real image; is the sample size; Indicates The generated features of samples; gradient The calculation method is expressed as: ； Step S205: Repeat the iteration of steps S202 to S204 until a preset stop iteration condition is met.

3. A method for detecting safety at a work site according to claim 2, characterized in that: The specific process of training the feature extraction model to obtain the trained feature extraction model is as follows: Step S301: Initialize a neural network, wherein the neural network adopts a 3-layer fully connected neural network, the number of hidden layers of the neural network is 2, the first hidden layer includes 100 neurons, the second hidden layer includes 50 neurons, and the activation function is a ReLU activation function; Initialize the weights of the neural network and bias , and set the initial learning rate of the neural network , expressed as: ； ； In the formula, represents the initial weights of the neural network; Represents the initial bias of the neural network; Indicates compliance with a specific distribution; The mean is 0 and the variance is Normal distribution of is the variance of the initial distribution; Step S302: Using the accelerated gradient descent algorithm to update the parameters of the neural network, the updating method is expressed as: ； ； In the formula, is the acceleration factor; The loss function for neural network training; is the gradient of the loss function with respect to the weight parameter; is the gradient of the loss function with respect to the bias parameter; For the The weight parameter of the iteration; For the The bias parameter for the iteration; For the The weight parameter of the iteration; For the The bias parameter for the iteration; is the symbol of partial derivative; is the learning rate during the growth phase; Step S303: synthesize multiple weight parameters into new parameters, that is, define a synthesis function , expressed as: ； In the formula, yes The importance coefficient of for the number of Importance coefficient Based on the previous stage The gradient size is dynamically calculated, The calculation method of the importance coefficient of each weight parameter is expressed as: ； In the formula, is the decay rate parameter; Step S304: Setting performance evaluation function The feature extraction model output of the current iteration is used to determine whether it will enter the next iteration. The performance evaluation function is expressed as: ； In the formula, represents the performance evaluation function; Represents the loss of neural network training; is the performance threshold; when If the calculation is 1, then proceed to the next iteration, otherwise proceed to step S305; Step 305: Use the perturbation method as the adjustment increment of the weight and bias, and the calculation method is expressed as: ； ； In the formula, is the adjustment increment of the weight parameter; is the adjustment increment of the bias parameter; and are the amplitude and frequency parameters of the perturbation respectively; and The setting method is expressed as: ； ； In the formula, is the scaling factor of the perturbation amplitude; Step S306: Constrain the weights of the neural network by weight pruning, which is expressed as: ； In the formula, represents the weight pruning function; is the gradient threshold; Represents the dot product operation; Step S307: Adjusting the neural network learning rate , the adjustment method is expressed as: ； In the formula, is the attenuation coefficient; represents the initial learning rate; is the number of completed cycles; Step S308: Repeat the iteration of steps S302 to S307 until a preset stop iteration condition is met.

4. A method for detecting safety at a work site according to claim 3, characterized in that: Acceleration factor According to the decreasing speed and curvature setting of the loss function, the calculation method is: ； In the formula, is the preset maximum acceleration factor; is the minimum value function; Decay rate parameter The calculation method is expressed as: 。 5. A method for detecting safety at a work site according to claim 4, characterized in that: The feature graph is used to train the feature dimensionality reduction model. The specific process of obtaining the trained feature dimensionality reduction model is as follows: Step S401: Initialize the parameters of the autoencoder neural network; Assume that the encoder parameters of the autoencoder neural network are , the decoder parameters of the autoencoder neural network are , use random initialization to initialize the weights and biases of the encoder and decoder of the autoencoder neural network; Step S402: Input feature map , feature map Transformed into low-dimensional feature representation through encoder , and then the low-dimensional features are represented as The decoder reconstructs the output into the original dimension , expressed as: ； ； In the formula, is the encoder function; is the decoder function; Step S403: Calculate the dual loss, the dual loss Including reconstruction error and sparse self-expression error , expressed as: ； In the formula, is the regularization parameter of the dual loss; Reconstruction error measure and The difference between is expressed as: ； In the formula, For the indivual ; For the indivual ; Sparse self-expression error Computing low-dimensional feature representations and The error between the sparse linear combinations of is expressed as: ； In the formula, The feature map quantity; is the regularization parameter of the sparse self-expression error; is the self-expression matrix; For the A low-dimensional feature representation; Step S404: Calculate using the chain rule based on the dual loss about and The gradient of , and the parameter update, is expressed as: ； ； In the formula, is the learning rate of the autoencoder; The dual loss about The gradient of The dual loss about The gradient of Step S405: Repeat the iteration of steps S402 to S404 until a preset stop iteration condition is met.

6. A method for detecting safety at a work site according to claim 5, characterized in that: Self-expression matrix The construction method is expressed as: ； ； In the formula, It includes all Matrix of for The transpose of is a hyperparameter of density; Represents Euclidean distance calculation; is the Softmax function.

7. A safety detection system for implementing a work site safety detection method according to any one of claims 1 to 6, characterized in that: include: The collection and annotation module is used to collect and annotate images of the construction site; A data expansion module is used to generate samples of the collected construction site images using a trained generative adversarial network algorithm based on hierarchical feature decomposition to obtain an expanded image; The feature extraction model training module is used to construct a feature extraction model, stretch the expanded image into a one-dimensional vector and then input it into the feature extraction model, train the feature extraction model to obtain a trained feature extraction model, and use the trained feature extraction model to extract features from the expanded image to obtain a feature map; A feature dimension reduction model training module is used to construct a feature dimension reduction model, train the feature dimension reduction model using a feature graph to obtain a trained feature dimension reduction model, and reduce the dimension of the feature graph using the trained feature dimension reduction model to obtain a reduced dimension feature graph; The classifier model training module is used to construct a classifier model, train the classifier model using the reduced dimension feature map to obtain a trained classifier model, and classify the reduced dimension feature map using the trained classifier model to obtain a classification result; The on-site detection module is used to process real-time construction site images using the trained feature extraction model, the trained feature dimension reduction model and the trained classifier model.

8. A safety fence, characterized in that: include: Camera, radar, solar panel, power module, voice and sound and light alarm module, safety detection system as claimed in claim 7 and retractable warning belt; The radar is used to monitor whether the heart rate of the on-site workers is normal; the camera is used to capture the behavior of the on-site workers; The safety detection system is used to conduct safety detection on the behaviors of field workers captured by the camera; the voice and sound and light alarm modules are used to give alarms based on the detection results of the safety detection system; the solar panels are used to convert light energy into electrical energy; The power module is used to store the electrical energy converted by the solar panels and to power the cameras, radars, voice and sound and light alarm modules and safety detection systems.

9. A safety fence according to claim 8, characterized in that: It also includes hooks or strong magnets, which are used to fix the safety fence.

Images