Intelligent safety violation behavior identification and early warning system applied to power plant

Abstract

The invention discloses an intelligent safety violation behavior recognition and early warning system applied to a power plant, and particularly relates to the field of behavior recognition and early warning, and the system comprises an image data collection module, an image data processing module, an image analysis algorithm module, an employee behavior recognition module, an employee violation behavior analysis module, and an employee violation behavior early warning module. The method comprises the following steps: acquiring an employee image in a working area in real time through a high-definition camera, denoising by using bilateral filtering, and enhancing the image through a histogram equalization method; then detecting the postures of the employees by using a Pose Net algorithm, and identifying behavior characteristics through model training; judging a safety behavior value according to the attitude data, comparing the safety behavior value with a preset threshold, and sending a corresponding instruction; if the instruction is a dangerous instruction, an unsafe behavior propagation value is calculated, and the value is obtained based on the individual susceptibility, the character influence and the relationship closeness coefficient of the two parties; and finally, judging the violation degree according to the propagation value and sending out an early warning notification of a corresponding level so as to take corresponding education and correction measures.

Description

Technical Field

[0001] The present invention relates to the field of behavior recognition and early warning technology, and more specifically, to an intelligent safety violation behavior recognition and early warning system applied to power plants. Background Art

[0002] In recent years, with the rapid development of national infrastructure construction, the new energy industry has been moving towards a new era of vigorous development. In this vibrant industry, the production site is undoubtedly a key area where safety accidents frequently occur. With the continuous expansion of the scale of production and construction, the process is becoming more and more complex. How to improve production site management, reduce the accident rate, and standardize operating procedures has become a difficult problem that many companies need to solve urgently. Against this background, the country's emphasis on production safety has been increasing, and various companies have also taken measures to ensure the safety of their employees, thereby ensuring the long-term development of the company.

[0003] In today's rapidly developing industrialization process, electricity, as an important pillar of the national economy, is of great importance for safe production; power plants, as the core places for electricity production, their operating efficiency and safety are directly related to the stability of energy supply and public safety of the entire society.

[0004] However, from the perspective of the actual needs of power plant safety production, there are still some shortcomings. For example, due to the complex production environment of power plants, the large number of equipment, and the existence of dangerous factors such as high temperature, high pressure, and strong electromagnetic fields, the workers' operating behaviors are often affected by multiple factors, which makes it easy to violate regulations and cause safety accidents. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides an intelligent safety violation behavior identification and early warning system applied to a power plant to solve the problems raised in the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] Image data acquisition module: used to use a high-definition camera to collect working images of employees in the working area, and transmit the collected image data to the image data processing module;

[0008] Image data processing module: used to perform image denoising and image enhancement processing on the collected images of employees working in the work area, and transmit the processed image data to the image analysis algorithm module;

[0009] Image analysis algorithm module: used to analyze the processed images to identify the behavioral characteristics of employees, and transmit the analyzed employee behaviors to the employee behavior recognition module;

[0010] Employee behavior recognition module: used to identify employee's violation behavior based on the transmitted behavior feature data, and transmit the identified instructions to the employee violation behavior analysis module;

[0011] Employee violation behavior analysis module: used to receive instructions from the employee behavior recognition module and build a model based on the instructions to calculate the unsafe behavior propagation value;

[0012] Employee violation behavior warning module: used to judge and warn of violation behaviors based on the unsafe behavior propagation value calculated in the employee violation behavior analysis module.

[0013] Preferably, in the image data acquisition module, a high-definition camera is installed in the working area to ensure that the field of view covers all key work links and avoids blind spots; configure camera parameters, including resolution, frame rate, and exposure time, to ensure that clear images can be captured under different lighting conditions; adjust the angle and direction of the camera to ensure full coverage of the employee activity area; start the camera and start real-time image acquisition to ensure the normal operation of the image data recording system; set the acquisition time period, and the acquisition time period is set to 15 minutes before the start of the work shift to 15 minutes after the end.

[0014] Preferably, in the image data processing module, bilateral filtering is used to perform image denoising on the collected image, and then a histogram equalization method is used to perform image enhancement on the denoised image.

[0015] Preferably, in the image analysis algorithm module, the posture pixel set of the employee is obtained using the state estimation algorithm Pose Net, and the steps of constructing the first model are specifically as follows:

[0016] Step B1: Use the convolutional layer to extract features from the processed image Q to obtain a feature map F. The specific calculation method of F is:

[0017] F = Conv(Q; β1), where F represents the feature map, Q represents the processed image, and β1 represents the parameters of the convolution layer;

[0018] Step B2: The feature map F is converted to the heat map K through the fully connected layer. The calculation method of the conversion is as follows:

[0019] K = γ(F(F; β2)), where K represents the heat map, F represents the feature map, β2 represents the parameter of the fully connected layer, and γ represents the activation function;

[0020] Step B3: Regress the coordinates of each key point (x i ,y i ); the specific calculation method is:

[0021] (x i,y i )=Regression(K;β3), where (x i ,y i ) represents the key point coordinates, K represents the heat map, β3 represents the parameter of the regression layer, and Regression represents the regression analysis;

[0022] Step B4: Train the constructed model.

[0023] Preferably, in the employee behavior recognition module, a minimum external matrix is ​​calibrated around the employee posture pixel set detected by the first model, and the employee posture pixel set is set to R = r1, r2, ..., r m ), the pixel set of the detected dangerous object is set to T = (t1, t2..., t m ), by traversing the pixel set P of the minimum circumscribed matrix of the human body, the safe behavior value of the pixel intersection of the set R and the dangerous object pixel set T is calculated. The calculation method of the safe behavior value U is specifically as follows:

[0024] Among them, U represents the safety behavior value, R represents the employee posture pixel set, and T represents the pixel set of dangerous objects.

[0025] Preferably, in the employee violation behavior analysis module, if the received instruction is a dangerous instruction, the AI ​​middle station constructs an unsafe behavior propagation model based on the small world network according to the received violation picture, and remotely deploys and updates the constructed model through the 5G customized network + WIFI hybrid network, and calculates the unsafe transmission behavior propagation value according to the model. The calculation method of the unsafe behavior propagation value is specifically as follows:

[0026] V g (s+1)=V g (s)+∑ h m[V h (s)-qV g (s)], where V g (s+1) represents the unsafe behavior propagation value of node g at time s+1, V g (s) represents the unsafe behavior propagation value of node g at time s, V h (s) represents the unsafe behavior propagation value of node h at time s, m represents the individual susceptibility coefficient, n represents the person influence coefficient, and q represents the closeness coefficient of the relationship between the two parties.

[0027] Preferably, in the employee violation behavior warning module, when the calculated unsafe behavior propagation value is greater than the preset unsafe behavior propagation value threshold, it is recorded as a major violation, and the AI ​​middle station immediately issues a major warning notification, and professionals conduct ideological education for the violators and suspend work for one day; if the calculated unsafe behavior propagation value is equal to the preset unsafe behavior propagation value threshold, the AI ​​middle station automatically records the employee as a medium violation, and the AI ​​middle station immediately issues a medium warning notification, and professionals conduct ideological education for the violators; if the calculated unsafe behavior propagation is less than the preset unsafe behavior threshold, the AI ​​middle station automatically records the employee's behavior at this time as a minor violation, and the AI ​​middle station issues a minor warning notification, and calls the employee's name to stay away from the violation area.

[0028] Technical effects and advantages of the present invention:

[0029] The present invention uses a high-definition camera to collect images of the working area of ​​the power plant in real time and transmits them to the data processing module. The data module uses bilateral filtering to denoise and histogram equalization to enhance the image, and then uses the Pose Net algorithm to analyze the posture characteristics of the employees and identify violations. The recognition results are used to calculate the propagation value of unsafe behavior and issue warnings based on preset thresholds; the present invention uses a high-definition camera to capture the operating behavior of workers in real time, and uses an AI algorithm for intelligent analysis to accurately identify violations and issue warning signals in a timely manner, thereby effectively avoiding the occurrence of safety accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a schematic diagram of module connection of the present invention. DETAILED DESCRIPTION

[0031] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0032] See also Figure 1 As shown, the present invention provides an intelligent safety violation behavior identification and early warning system for power plants, which includes: an image data acquisition module, an image data processing module, an image analysis algorithm module, an employee behavior identification module, an employee violation behavior analysis module, and an employee violation behavior early warning module.

[0033] The image data acquisition module is used to use a high-definition camera to collect working images of employees in the working area, and transmit the collected image data to the image data processing module;

[0034] In the image data acquisition module, a high-definition camera is installed in the working area to ensure that the field of view covers all key work links and avoids blind spots; the camera parameters are configured, including resolution, frame rate, and exposure time, to ensure that clear images can be captured under different lighting conditions; the angle and direction of the camera are adjusted to ensure full coverage of the employee activity area; the camera is started to start real-time image acquisition to ensure the normal operation of the image data recording system; and the acquisition time period is set, which is set from 15 minutes before the start of the work shift to 15 minutes after the end of the work shift.

[0035] The image data processing module is used to perform image denoising and image enhancement processing on the collected images of employees working in the work area, and transmit the processed image data to the image analysis algorithm module;

[0036] In the image data processing module, the collected image is subjected to image denoising by using bilateral filtering, and the method of bilateral filtering denoising is specifically as follows:

[0037] Step A1: Initialize the weighted sum and weighted sum, and traverse each pixel point (a, b);

[0038] Step A2: The calculation method of spatial distance weight is as follows:

[0039] W d (a, b) is expressed as the spatial distance weight, σ d It is represented as the spatial standard deviation, (a, b) represents the pixel point, and (c, d) represents the central pixel; the weight decreases as the distance from c and d increases;

[0040] Step A3: The calculation method of spatial similarity weight is as follows:

[0041] W r (a, b) represents the weight at position (a, b), I(c, d) represents the image pixel value at position (c, d), I(a, b) represents the image pixel value at position (a, b), σ r It is expressed as the standard deviation that controls the width of the Gaussian function; the weight varies according to the grayscale difference between the neighborhood pixels and the central pixel;

[0042] Step A4: The calculation method of the comprehensive weight is as follows:

[0043] W(a, b) = W d (a, b)·W r (a, b), where W(a, b) represents the comprehensive weight, W d (a, b) is expressed as the spatial distance weight, W r (a, b) represents the weight at position 9a, b);

[0044] Step A5: Update the weighted sum and the weighted sum. The calculation method for updating the weighted sum is as follows:

[0045] I0(c, d) + =W(a, b)·I(c, d), where I0(c, d) + is represented as the updated weighted sum, W(a, b) represents the weight value at the coordinate (a, b), and I(c, d) represents the image pixel value at the position (c, d);

[0046] The calculation method for updating the weight sum is as follows:

[0047] in, is represented as the updated weight sum, and W(a, b) is represented as the weight value at the coordinate (a, b);

[0048] Step A6: Calculate the final filtered pixel value and perform normalization. The specific method for calculating the filtered pixel value is:

[0049] Among them, I1(c, d) represents the filtered pixel value, I0(c, d) + is expressed as the updated weighted sum, Expressed as the updated weight sum;

[0050] Step A7: Repeat steps A1 to A7 for all pixels in the image until all pixels are processed;

[0051] Through steps A1 to A7, bilateral filtering can effectively remove noise in the image;

[0052] The image after noise removal is enhanced using the histogram equalization method. The specific steps of image enhancement are as follows:

[0053] Step C1: Count the grayscale value of each pixel in the image and calculate the grayscale histogram of the image; the grayscale histogram reflects the frequency or probability of each grayscale level in the image;

[0054] Step C2: Calculate the cumulative distribution function based on the grayscale histogram. CDF is the integral of the grayscale histogram, that is, the ratio of the total number of pixels of each grayscale level and all grayscale levels below it to the total number of pixels in the image. The range of CDF is [0,1].

[0055] Step C3: Use CDF to convert the grayscale value of the original image. The specific calculation method of the conversion is:

[0056] Among them, Q represents the new gray value, CDF minExpressed as the minimum value of CDF, CDF max It is represented as the maximum value of CDF, CDF(e) is represented as the original gray value, L is represented as the number of gray levels of the output image, and the new gray level refers to the number of gray levels of the output image;

[0057] Step C4: Map the converted grayscale values ​​to the new grayscale levels to obtain the enhanced image; this process redistributes the grayscale values ​​of the original image according to the distribution of CDF, so that the grayscale histogram of the output image tends to be balanced.

[0058] The image analysis algorithm module is used to perform image analysis on the processed image to identify the behavioral characteristics of the employee, and transmit the analyzed employee behavior to the employee behavior recognition module;

[0059] In the image analysis algorithm module, the posture pixel set of the employee is obtained using the state estimation algorithm Pose Net, and the steps of constructing the first model are specifically as follows:

[0060] Step B1: Use the convolutional layer to extract features from the processed image Q to obtain a feature map F. The specific calculation method of F is:

[0061] F = Conv(Q; β1), where F represents the feature map, Q represents the processed image, and β1 represents the parameters of the convolution layer;

[0062] Step B2: The feature map F is converted to the heat map K through the fully connected layer. The calculation method of the conversion is as follows:

[0063] K = γ(F(F; β2)), where K represents the heat map, F represents the feature map, β2 represents the parameter of the fully connected layer, and γ represents the activation function;

[0064] Step B3: Regress the coordinates of each key point (x i ,y i ); the specific calculation method is:

[0065] (x i ,y i )=Regression(K;β3), where (x i ,y i ) represents the key point coordinates, K represents the heat map, β3 represents the parameter of the regression layer, and Regression represents the regression analysis;

[0066] Step B4: Train the constructed model. The specific method of model training is as follows:

[0067] Step B41: The input image passes through the convolution layer, the fully connected layer and the regression layer to obtain the predicted heat map and key point coordinates;

[0068] Step B42: Calculate the total loss function based on the prediction results and the true label; the calculation method of the total loss function is specifically as follows:

[0069] P = P0 + δP1, where P represents the total loss function, P0 represents the mean square error loss of the heat map, P1 represents the loss of the key point coordinates, and δ represents the weight coefficient;

[0070] P0=∑ i ∥K i -G i ∥ 2 , where P0 represents the mean square error loss function of the heat map, K i Represented as a predicted heat map, G i Represented as a true heat map;

[0071] Among them, P1 represents the loss function of the key point coordinates, x j ,y j ) represents the predicted key point coordinates, Represented as real key point coordinates;

[0072] Step B43: Update the model parameters by gradient descent method to minimize the loss function;

[0073] Step B44: Repeat the above process until the model converges.

[0074] The trained Pose Net model is used to predict the employee's posture pixel set in the processed image.

[0075] The employee behavior recognition module is used to identify the employee's illegal behavior based on the transmitted behavior feature data, and transmit the identified instructions to the employee illegal behavior analysis module;

[0076] In the employee behavior recognition module, a minimum external matrix is ​​calibrated around the employee posture pixel set detected by the first model, and the employee posture pixel set is set to R = (r1, r2, ..., r m ), the pixel set of the detected dangerous object is set to T = (t1, t2..., t m ), by traversing the pixel set P of the minimum circumscribed matrix of the human body, the safe behavior value of the pixel intersection of the set R and the dangerous object pixel set T is calculated. The calculation method of the safe behavior value U is specifically as follows:

[0077] Among them, U represents the safety behavior value, R represents the pixel set of employee posture, and T represents the pixel set of dangerous objects;

[0078] The calculated safety behavior value is compared with the preset safety behavior threshold. If the calculated safety behavior value is greater than the preset safety behavior threshold, a safety instruction is issued; if the calculated safety behavior value is less than the preset safety behavior threshold, a danger instruction is output; and the issued danger instruction is transmitted to the employee violation behavior analysis module.

[0079] The employee violation behavior analysis module is used to receive instructions issued by the employee behavior recognition module, and build a model based on the instructions to calculate the unsafe behavior propagation value;

[0080] In the employee violation behavior analysis module, if the received instruction is a dangerous instruction, the AI ​​center constructs an unsafe behavior propagation model based on the small-world network according to the received violation picture, and remotely deploys and updates the constructed model through the 5G customized network + WIFI hybrid network, and calculates the unsafe transmission behavior propagation value according to the model. The calculation method of the unsafe behavior propagation value is as follows:

[0081] V g (s+1)=V g (s)+Σ h m[V h (s)-qV g (s)], where V g (s+1) represents the unsafe behavior propagation value of node g at time s+1, V g (s) represents the unsafe behavior propagation value of node g at time s, V h (s) represents the unsafe behavior propagation value of node h at time s, m represents the individual susceptibility coefficient, n represents the person's influence coefficient, and q represents the closeness coefficient of the relationship between the two parties;

[0082] The calculation method of the character influence coefficient is as follows:

[0083] Among them, n represents the influence coefficient of the character, U represents the number of times the unsafe behavior spreads during the simulation time, and S max It is expressed as the total time of simulation running;

[0084] The calculation method of individual susceptibility coefficient is as follows:

[0085] Where m represents the individual susceptibility coefficient, V g (s) represents the unsafe behavior propagation value of node g at time s;

[0086] The calculation method of the closeness coefficient of the relationship between the two parties is as follows:

[0087] q=z∑ gU, where q represents the coefficient of the closeness of the relationship between the two parties, z represents the adjustment rate of the number of unsafe behavior propagation times during the simulation time, and U represents the number of times the unsafe behavior propagates during the simulation time;

[0088] In the communication dynamics, individual susceptibility is a key variable; for the communication of unsafe behaviors, individual susceptibility refers to the characteristics or tendencies of individuals to accept and adopt unsafe behaviors; in the actual working environment, there are significant differences in the degree of acceptance of unsafe behaviors among different employees; some employees have weak safety awareness and are easily influenced by others, making them more likely to accept and implement unsafe behaviors; on the contrary, some employees have strong safety awareness, are able to recognize the potential hazards of unsafe behaviors, and actively avoid the negative impact of these behaviors;

[0089] An important factor in the spread of unsafe behavior is the closeness of the relationship between the communicator and the person being communicated, that is, whether the two parties can establish a trusting relationship. The closer the relationship between the two parties, the higher the degree of trust, and the faster the information exchange;

[0090] The calculated unsafe behavior propagation value is transmitted to the employee violation behavior warning module;

[0091] The employee violation behavior warning module is used to judge and warn the violation behavior according to the unsafe behavior propagation value calculated in the employee violation behavior analysis module;

[0092] In the employee violation behavior warning module, when the calculated unsafe behavior propagation value is greater than the preset unsafe behavior propagation value threshold, it is recorded as a major violation, and the AI ​​center immediately issues a major warning notification, and professionals conduct ideological education for the violators and suspend work for one day; if the calculated unsafe behavior propagation value is equal to the preset unsafe behavior propagation value threshold, the AI ​​center automatically records the employee as a medium violation, and the AI ​​center immediately issues a medium warning notification, and professionals conduct ideological education for the violators; if the calculated unsafe behavior propagation is less than the preset unsafe behavior threshold, the AI ​​center automatically records the employee's behavior at this time as a minor violation, and the AI ​​center issues a minor warning notification, and calls the employee's name to stay away from the violation area.

[0093] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. The intelligent safety violation behavior identification and early warning system applied to power plants is characterized by: include: Image data acquisition module: used to use a high-definition camera to collect working images of employees in the working area, and transmit the collected image data to the image data processing module; Image data processing module: used to perform image denoising and image enhancement processing on the collected images of employees working in the work area, and transmit the processed image data to the image analysis algorithm module; Image analysis algorithm module: used to analyze the processed images to identify the behavioral characteristics of employees, and transmit the analyzed employee behaviors to the employee behavior recognition module; Employee behavior recognition module: used to identify employee's violation behavior based on the transmitted behavior feature data, and transmit the identified instructions to the employee violation behavior analysis module; Employee violation behavior analysis module: used to receive instructions from the employee behavior recognition module and build a model based on the instructions to calculate the unsafe behavior propagation value; Employee violation behavior warning module: used to judge and warn of violation behaviors based on the unsafe behavior propagation value calculated in the employee violation behavior analysis module.

2. According to claim 1, the intelligent safety violation behavior identification and early warning system applied to power plants is characterized by: In the image data acquisition module, a high-definition camera is installed in the working area to ensure that the field of view covers all key working links and avoids blind spots; Configure camera parameters, including resolution, frame rate, and exposure time, to ensure that clear images can be captured under different lighting conditions; Adjust the angle and direction of the camera to ensure full coverage of the employee activity area; Start the camera and start real-time image acquisition to ensure the normal operation of the image data recording system; set the acquisition time period, which is set from 15 minutes before the start of the work shift to 15 minutes after the end of the work shift.

3. The intelligent safety violation behavior identification and early warning system for power plants according to claim 1 is characterized in that: In the image data processing module, bilateral filtering is used to denoise the collected image, and the image after denoising is enhanced by histogram equalization. The method of using bilateral filtering is specifically as follows: Step A1: Initialize the weighted sum and weighted sum, and traverse each pixel point (a, b); Step A2: The calculation method of spatial distance weight is as follows: W d (a, b) is expressed as the spatial distance weight, σ d It is represented as the spatial standard deviation, (a, b) represents the pixel point, and (c, d) represents the central pixel; the weight decreases as the distance from c and d increases; Step A3: The calculation method of spatial similarity weight is as follows: W r (a, b) represents the weight at position (a, b), I(c, d) represents the image pixel value at position (c, d), I(a, b) represents the image pixel value at position (a, b), σ r It is expressed as the standard deviation that controls the width of the Gaussian function; the weight varies according to the grayscale difference between the neighborhood pixels and the central pixel; Step A4: The calculation method of the comprehensive weight is as follows: W(a, b) = W d (a, b)·W r (a, b), where W(a, b) represents the comprehensive weight, W d (a, b) is expressed as the spatial distance weight, W r (a, b) represents the weight at position (a, b); Step A5: Update the weighted sum and the weighted sum. The calculation method for updating the weighted sum is as follows: I0(c, d) + =W(a, b).I(c, d), where I0(c, d) + is represented as the updated weighted sum, W(a, b) represents the weight value at the coordinate (a, b), and I(c, d) represents the image pixel value at the position (c, d); The calculation method for updating the weight sum is as follows: in, is represented as the updated weight sum, and W(a, b) is represented as the weight value at the coordinate (a, b); Step A6: Calculate the final filtered pixel value and perform normalization. The specific method for calculating the filtered pixel value is: Among them, I1(c, d) represents the filtered pixel value, I0(c, d) + is expressed as the updated weighted sum, Expressed as the updated weight sum; Step A7: Repeat steps A1 to A7 for all pixels in the image until all pixels are processed; Through steps A1 to A7, bilateral filtering can effectively remove noise in the image.

4. The intelligent safety violation behavior identification and early warning system for power plants according to claim 1 is characterized in that: In the image analysis algorithm module, the posture pixel set of the employee is obtained using the state estimation algorithm Pose Net, and the steps of constructing the first model are specifically as follows: Step B1: Use the convolutional layer to extract features from the processed image Q to obtain a feature map F. The specific calculation method of F is: F = Conv(Q; β1), where F represents the feature map, Q represents the processed image, and β1 represents the parameters of the convolution layer; Step B2: The feature map F is converted to the heat map K through the fully connected layer. The calculation method of the conversion is as follows: K = γ(F(F; β2)), where K represents the heat map, F represents the feature map, β2 represents the parameter of the fully connected layer, and γ represents the activation function; Step B3: Regress the coordinates of each key point (x i ,y i ); the specific calculation method is: (x i ,y i )=Regression(K;β3), where (x i ,y i ) represents the key point coordinates, K represents the heat map, β3 represents the parameter of the regression layer, and Regression represents the regression analysis; Step B4: Train the constructed model.

5. The intelligent safety violation behavior identification and early warning system for power plants according to claim 4 is characterized in that: The specific method of model training is: Step B41: The input image passes through the convolution layer, the fully connected layer and the regression layer to obtain the predicted heat map and key point coordinates; Step B42: Calculate the total loss function based on the prediction results and the true label; the calculation method of the total loss function is specifically as follows: P = P0 + δP1, where P represents the total loss function, P0 represents the mean square error loss of the heat map, P1 represents the loss of the key point coordinates, and δ represents the weight coefficient; P0=∑ i ||K i -G i || 2 , where P0 represents the mean square error loss function of the heat map, K i Represented as a predicted heat map, G i Represented as a true heat map; Among them, P1 represents the loss function of the key point coordinates, (x j ,y j ) represents the predicted key point coordinates, Represented as real key point coordinates; Step B43: Update the model parameters by gradient descent method to minimize the loss function; Step B44: Repeat the above process until the model converges; The trained Pose Net model is used to predict the employee's posture pixel set in the processed image.

6. According to claim 1, the intelligent safety violation behavior identification and early warning system for power plants is characterized by: In the employee behavior recognition module, a minimum external matrix is ​​calibrated around the employee posture pixel set detected by the first model, and the employee posture pixel set is set to R = (r1, r2, ..., r m ), the pixel set of the detected dangerous object is set to T = (t1, t2..., t m ), by traversing the pixel set P of the minimum circumscribed matrix of the human body, the safe behavior value of the pixel intersection of the set R and the dangerous object pixel set T is calculated. The calculation method of the safe behavior value U is specifically as follows: Among them, U represents the safety behavior value, R represents the employee posture pixel set, and T represents the pixel set of dangerous objects.

7. The intelligent safety violation behavior identification and early warning system for power plants according to claim 1 is characterized in that: In the employee violation behavior analysis module, if the received instruction is a dangerous instruction, the AI ​​center constructs an unsafe behavior propagation model based on the small-world network according to the received violation picture, and remotely deploys and updates the constructed model through the 5G customized network + WIFI hybrid network, and calculates the unsafe transmission behavior propagation value according to the model. The calculation method of the unsafe behavior propagation value is as follows: V g (s+1)=V g (s)+∑ h m[V h (s)-qV g (s)], where V g (s+1) represents the unsafe behavior propagation value of node g at time s+1, V g (s) represents the unsafe behavior propagation value of node g at time s, V h (s) represents the unsafe behavior propagation value of node h at time s, m represents the individual susceptibility coefficient, n represents the person's influence coefficient, and q represents the closeness coefficient of the relationship between the two parties; The calculated unsafe behavior propagation value is transmitted to the employee violation behavior warning module.

8. The intelligent safety violation behavior identification and early warning system for power plants according to claim 7 is characterized in that: The calculation method of the character influence coefficient is as follows: Among them, n represents the influence coefficient of the character, U represents the number of times the unsafe behavior spreads during the simulation time, and S max It is expressed as the total time of simulation running; The calculation method of individual susceptibility coefficient is as follows: Where m represents the individual susceptibility coefficient, V g ( s ) It is represented as the unsafe behavior propagation value of node g at time s; The calculation method of the closeness coefficient of the relationship between the two parties is as follows: q=z∑ g U, where q represents the closeness coefficient of the relationship between the two parties, z represents the adjustment rate of the number of unsafe behavior propagation times within the simulation time, and U represents the number of times the unsafe behavior propagates within the simulation time.

9. The intelligent safety violation behavior identification and early warning system for power plants according to claim 1 is characterized in that: In the employee violation behavior warning module, when the calculated unsafe behavior propagation value is greater than the preset unsafe behavior propagation value threshold, it is recorded as a major violation, and the AI ​​center immediately issues a major warning notification, and professionals conduct ideological education for the violators and suspend work for one day; if the calculated unsafe behavior propagation value is equal to the preset unsafe behavior propagation value threshold, the AI ​​center automatically records the employee as a medium violation, and the AI ​​center immediately issues a medium warning notification, and professionals conduct ideological education for the violators; if the calculated unsafe behavior propagation is less than the preset unsafe behavior threshold, the AI ​​center automatically records the employee's behavior at this time as a minor violation, and the AI ​​center issues a minor warning notification, and calls the employee's name to stay away from the violation area.

Images