Image recognition method for distribution network uninterruptible power operation

Abstract

The invention discloses an image recognition method for distribution network uninterruptible power operation. The method comprises the following steps: data acquisition; image processing; feature extraction and fault identification; monitoring and alarming in real time; and storing and analyzing data. The beneficial effects of the invention are that the system can solve the problems that the details of high-altitude operation cannot be effectively monitored, and the high-efficiency error prevention and cooperative communication cannot be realized in the original power-off-free operation of the distribution network, and timely eliminates the non-standard and violation conditions in the operation through providing an effective monitoring inspection and technical support means, thereby improving the working efficiency. Real-time guidance and help are provided for front-line operators; the working efficiency and the working quality are improved, the safety guarantee capability and the safety operation level of a distribution network non-power-cut operation site can be greatly improved, and safe and stable operation of a power grid is guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of electric power technology, and in particular to an image recognition method for non-stop power supply operation of a distribution network. Background Art

[0002] As the final link in the power system where the power network supplies power to users, the importance of the distribution network is self-evident. The efficiency of its operation and management directly affects the reliability and quality of the power supply of the power grid. Most distribution networks are built near industrial and residential areas, with the characteristics of dense line networks, scattered equipment distribution, complex terrain and environment, etc. Therefore, the failure rate of the distribution network is relatively high. Non-stop operation is a special operation method that directly performs maintenance, replaces parts, and handles defects without power outages for users. It is a technical measure to ensure the stable operation of the power supply system and an important way to improve the reliability of power supply.

[0003] Under the current situation where the requirements for power supply reliability management are becoming more and more stringent and users are more sensitive to power outages, non-stop power supply operations in the distribution network have become an important means to reduce the frequency and duration of power outages and reduce users' perception of power outages. In recent years, the company has actively explored various innovative technical means, focusing on improving the application of new technologies and informationization levels for non-stop power supply operations, so as to gradually promote the distribution network from power outage operations to non-stop power supply operations, and realize the comprehensive transformation from power outage maintenance to non-stop power supply maintenance. However, the distribution network lines are complex, with the characteristics of dense distribution facilities and narrow operating scope, which makes it easy for operators to cause safety accidents due to inadequate protective measures or illegal violations in the case of non-stop power supply operations.

[0004] At present, the safety of the distribution network non-stop operation site mainly relies on on-site supervision by staff or control by video recording. Although the management system is relatively complete, the risk level of distribution network non-stop operation is generally high, the types of tasks are numerous, the operation methods are diverse, the operation areas are complex and scattered, the operation process is short and fast, the abilities and qualities of practitioners are different, the management level of construction units is uneven, and it is difficult for on-site safety supervisors to achieve comprehensive monitoring. This often leads to inadequate or loopholes in the safety measures at the operation site (such as: on-site re-survey, protective wear, layout of isolation area warning signs, parking and operation of boom trucks, insulation shielding and standardization of operator actions, etc.).

[0005] With the rapid development of video surveillance technology and information technology, security monitoring systems based on video surveillance technology have also been widely used in non-stop power supply operation sites. This type of security monitoring system can effectively reduce the workload of on-site safety supervisors, expand the scope of monitoring, and to a certain extent improve the level of security monitoring at the operation site. The current common practice is to use a video surveillance system to monitor and record the operation site, and then play back the video and manually identify it after the operation is completed. This has a large workload, low efficiency, and is prone to omissions, and it cannot provide real-time alarms and corrections for violations of regulations on the site.

[0006] Therefore, it is necessary to propose an image recognition method for non-stop power supply operation in distribution network in view of the above technical solution. Summary of the invention

[0007] The object of the present invention is to provide an image recognition method for non-stop operation of distribution network.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A method for image recognition for non-stop power distribution network operation, the method steps are as follows:

[0010] S1, data collection;

[0011] S2, image processing;

[0012] S3, feature extraction and fault identification;

[0013] S4, real-time monitoring and alarm;

[0014] S5. Data storage and analysis.

[0015] In step S2, the image processing is performed using the YOLO algorithm.

[0016] The YOLO algorithm process is as follows:

[0017] S11, input the video stream image, divide each frame of the image into M×N cells, and extract W initial candidate frames from each cell;

[0018] S12, perform target recognition on the W initial candidate boxes respectively, and calculate the prediction confidence of the existence of the target in the W candidate boxes; if there is no target object in the candidate box, the candidate confidence value is zero, if there is a target object, the confidence is the ratio of the predicted box to the real box; and identify the target object category of the candidate box with the target object;

[0019] S13, the candidate box contains the confidence of the target object and the location information of the bounding box, that is, the output vector of the YOLO network is (X, Y, W, H, C), where X and Y represent the coordinates of the center point relative to the prediction box, W and H represent the length and width of the prediction box respectively, and the confidence C (confi-ence) represents the probability that the candidate box with the target object is a real sample;

[0020] S14, solving the occlusion problem of the head and body parts through the movement trajectory of the operator, and providing a virtual recognition and judgment result of the occluded parts;

[0021] S15. Give all the identification results: whether the person is wearing a safety helmet, insulating protective clothing, and insulating protective gloves.

[0022] The steps of the semi-supervised learning recognition algorithm of the YOLO algorithm are:

[0023] S21. Design the operator identification network and specific network structure. The network is a cascade network structure. This process is from coarse to fine. This has significantly improved the efficiency of capturing unsafe behaviors of operators who do not wear safety equipment, and has a better application effect on small target objects of long-range monitoring cameras.

[0024] S22, selecting a large public data set and some real construction site image sets to train the network model offline, determining and fine-tuning the parameters of the network model, and then obtaining a generalized model;

[0025] S23. Apply semi-supervised learning to online learning, enhance the generalization ability of the target monocular camera in some specific scenarios, and analyze the motion trajectory of the operator to solve the occlusion problem of the head and body.

[0026] Step S1 includes data acquisition, including laser radar, high-precision inertial sensor (IMU) and camera as data acquisition sources, and judges the distance between people and charged bodies through human key point recognition system, charged body recognition system, data fusion system and spatial position judgment system, and gives active warning.

[0027] Among them, the human limb key point position recognition module of the human body key point recognition system quickly recognizes the position of human limb key points based on the improved convolutional pose machine network.

[0028] The specific steps of the improved convolutional pose machine network to quickly identify the key points of human limbs are as follows:

[0029] S31, performing a convolution operation on the acquired original camera image and the Gaussian function template to generate a central image;

[0030] S32, the center graph is pooled into a small center graph and sent to the second, third and fourth stage networks;

[0031] S33, the original camera image is sent to the second, third and fourth stage networks after being operated by MultiBlock_1;

[0032] S34, the original camera image is sent to the second and third stage networks after being operated by MultiBlock_0;

[0033] S35, the characteristic data after the second stage network operation is sent to the third and fourth stage networks;

[0034] S36. In the fourth stage network, the feature data is merged and then subjected to the ConvBlock operation to output 64*64*18 feature data.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. The present invention can solve the problem that the original distribution network non-stop power supply operation cannot effectively monitor the details of high-altitude operations, efficiently prevent errors and collaborative communication. By providing an effective monitoring and auditing and technical support means, it can timely eliminate irregularities and violations in operations, provide real-time guidance and assistance to front-line operators; improve work efficiency and work quality, and can greatly improve the safety assurance capability and safe operation level of the distribution network non-stop power supply operation site, and ensure the safe and stable operation of the power grid.

[0037] 2. The present invention can greatly promote the supervision of the non-stop operation process of the distribution network to move towards visualization, networking and intelligence, deepen the deep integration and extensive application of advanced technologies such as the Internet of Things, big data, and state perception with the non-stop operation technology of the distribution network, effectively integrate the infrastructure resources of the power system and various information resources, and improve the information level and safe operation level of the distribution network operation and maintenance work. It can greatly promote the promotion and application of non-stop intelligent operation of the distribution network, significantly reduce power grid operation failures, improve power supply reliability and the safe operation level of the power grid, and have good social benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the identification method flow of the present invention;

[0039] Figure 2 It is a schematic diagram of the YOLO algorithm flow of the present invention;

[0040] Figure 3 This is a schematic diagram of the improved convolutional pose machine network structure of the present invention; DETAILED DESCRIPTION

[0041] The embodiments of the present invention are described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0042] The following is combined with Figures 1 to 3 The present invention is described in further detail.

[0043] like Figure 1-3 As shown, an embodiment of the present invention is: a method for image recognition for non-stop power distribution network operation, the method steps are as follows:

[0044] S1, data collection;

[0045] S2, image processing;

[0046] S3, feature extraction and fault identification;

[0047] S4, real-time monitoring and alarm;

[0048] S5. Data storage and analysis.

[0049] In step S2, the image processing is performed using the YOLO algorithm.

[0050] The YOLO algorithm process is as follows:

[0051] S11, input the video stream image, divide each frame of the image into M×N cells, and extract W initial candidate frames from each cell;

[0052] S12, perform target recognition on the W initial candidate boxes respectively, and calculate the prediction confidence of the existence of the target in the W candidate boxes; if there is no target object in the candidate box, the candidate confidence value is zero, if there is a target object, the confidence is the ratio of the predicted box to the real box; and identify the target object category of the candidate box with the target object;

[0053] S13, the candidate box contains the confidence of the target object and the location information of the bounding box, that is, the output vector of the YOLO network is (X, Y, W, H, C), where X and Y represent the coordinates of the center point relative to the prediction box, W and H represent the length and width of the prediction box respectively, and the confidence C (confi-ence) represents the probability that the candidate box with the target object is a real sample;

[0054] S14, solving the occlusion problem of the head and body parts through the movement trajectory of the operator, and providing a virtual recognition and judgment result of the occluded parts;

[0055] S15. Give all the identification results: whether the person is wearing a safety helmet, insulating protective clothing, and insulating protective gloves.

[0056] The steps of the semi-supervised learning recognition algorithm of the YOLO algorithm are:

[0057] S21. Design the operator identification network and specific network structure. The network is a cascade network structure. This process is from coarse to fine. This has significantly improved the efficiency of capturing unsafe behaviors of operators who do not wear safety equipment, and has a better application effect on small target objects of long-range monitoring cameras.

[0058] S22, selecting a large public data set and some real construction site image sets to train the network model offline, determining and fine-tuning the parameters of the network model, and then obtaining a generalized model;

[0059] S23. Apply semi-supervised learning to online learning, enhance the generalization ability of the target monocular camera in some specific scenarios, and analyze the motion trajectory of the operator to solve the occlusion problem of the head and body.

[0060] Step S1 includes data acquisition, including laser radar, high-precision inertial sensor (IMU) and camera as data acquisition sources, and judges the distance between people and charged bodies through human key point recognition system, charged body recognition system, data fusion system and spatial position judgment system, and gives active warning.

[0061] Among them, the human limb key point position recognition module of the human body key point recognition system quickly recognizes the position of human limb key points based on the improved convolutional pose machine network.

[0062] The specific steps of the improved convolutional pose machine network to quickly identify the key points of human limbs are as follows:

[0063] S31, performing a convolution operation on the acquired original camera image and the Gaussian function template to generate a central image;

[0064] S32, the center graph is pooled into a small center graph and sent to the second, third and fourth stage networks;

[0065] S33, the original camera image is sent to the second, third and fourth stage networks after being operated by MultiBlock_1;

[0066] S34, the original camera image is sent to the second and third stage networks after being operated by MultiBlock_0;

[0067] S35, the characteristic data after the second stage network operation is sent to the third and fourth stage networks;

[0068] S36. In the fourth stage network, the feature data is merged and then subjected to the ConvBlock operation to output 64*64*18 feature data.

[0069] Compared with the prior art, the present invention has the following beneficial effects:

[0070] The present invention can solve the problem that the original distribution network non-stop operation cannot effectively monitor the details of high-altitude operations, efficiently prevent errors and collaborative communication. By providing an effective monitoring and auditing and technical support means, it can timely eliminate irregularities and violations in operations, provide real-time guidance and assistance to front-line operators, improve work efficiency and work quality, and greatly improve the safety assurance capability and safety operation level of the distribution network non-stop operation site, ensuring the safe and stable operation of the power grid.

[0071] like Figure 3 As shown in the figure, the improved network has four stages. In addition to MultiBlock_1, the small center graph and the second and third stages directly connected to them, the input sources of the third and fourth stages are the outputs from the first and second stages respectively. This four-stage jump connection design structure simplifies the original network without losing too much of the original convolution pose machine detection performance, effectively reducing the number of network iterations, the number of network parameters and the size of the network model to meet the real-time requirements of non-stop operations.

[0072] The present invention can greatly promote the supervision of the non-stop operation process of the distribution network to move towards visualization, networking and intelligence, deepen the deep integration and wide application of advanced technologies such as the Internet of Things, big data, and state perception with the non-stop operation technology of the distribution network, effectively integrate the infrastructure resources of the power system and various information resources, and improve the informatization level and safe operation level of the distribution network operation and maintenance work. It can greatly promote the promotion and application of intelligent non-stop operation of the distribution network, significantly reduce power grid operation failures, improve power supply reliability and the safe operation level of the power grid, and have good social benefits.

[0073] In view of the actual situation of non-stop power supply operation in distribution network, typical non-stop power supply operation projects were selected as research objects. According to the relevant standards, specifications and practical experience of non-stop power supply operation in distribution network, the unsafe conditions of personnel, power equipment, operating vehicles and tools, surrounding environment, process management and other aspects during the operation were analyzed, the existing risk factors were identified, the key points of risk monitoring on-site of non-stop power supply operation in distribution network were proposed, and a method for automatic detection of multi-target features that can be applied to video surveillance was established.

[0074] Through on-site photography and simulation tests, an image feature library of on-site monitoring targets (such as power equipment, operators, operating vehicles and tools, environmental factors, etc.) for distribution network non-stop operations is established, and target recognition and positioning technology based on image feature matching is studied to achieve automatic detection of specific targets such as insulating protective equipment (safety helmets, insulating clothing, insulating gloves) and key parts of operators (head, hands) during distribution network non-stop operations.

[0075] According to the distribution network non-stop operation procedures, for typical violations in the preparation and operation process of distribution network non-stop operations, the characteristics of typical violations are extracted through on-site photography and simulation tests, and the violation judgment method based on image recognition technology and the distribution network non-stop operation risk warning mechanism are studied. Computers are used instead of human eyes to realize intelligent monitoring of the operation site, automatically identify personnel wearing insulating protective equipment (safety helmets, insulating clothing, insulating gloves), insulating shielding objects and shielding tools (insulators, crossarms, conductor shielding covers, insulator shielding covers, crossarm shielding covers, etc.), and warn of violations.

[0076] Through on-site field experiments, this project can realize intelligent management and control of the entire process of non-stop power supply operations; the human body key point monitoring system can determine whether a start-up meeting and weather detection are carried out before the operation to ensure that all preparatory work before the operation is completed; the operator wearing insulating protective equipment inspection system can ensure that the operator wears a full set of insulating protective equipment before non-stop power supply operations. When the system detects that the operator is not wearing the corresponding insulating protective equipment, the system will prompt him to wear the corresponding insulating protective equipment through voice; the safety distance monitoring and early warning system can ensure that the operator maintains a safe distance from the live body when performing non-stop power supply operations. When the working distance is close to the safety distance, the system will give a voice prompt to ensure the safety of the operator.

[0077] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A method for image recognition for non-stop power distribution network operation, characterized in that: The method steps are: S1, data collection; S2, image processing; S3, feature extraction and fault identification; S4, real-time monitoring and alarm; S5. Data storage and analysis.

2. The image recognition method for non-stop power distribution network operation according to claim 1, characterized in that: In step S2, the image processing is performed using the YOLO algorithm.

3. The image recognition method for non-stop power distribution network operation as claimed in claim 2, characterized in that: The YOLO algorithm process is as follows: S11, input the video stream image, divide each frame of the image into M×N cells, and extract W initial candidate frames from each cell; S12, performing target recognition on the W initial candidate frames respectively, and calculating the prediction confidence of the existence of the target in the W candidate frames; If there is no target object in the candidate box, the candidate confidence value is zero. If there is a target object, the confidence is the ratio of the predicted box to the real box. And the target object category is identified for the candidate box with the target object. S13, the candidate box contains the confidence of the target object and the location information of the bounding box, that is, the output vector of the YOLO network is (X, Y, W, H, C), where X and Y represent the coordinates of the center point relative to the prediction box, W and H represent the length and width of the prediction box respectively, and the confidence C represents the probability that the candidate box containing the target object is a real sample; S14, solving the occlusion problem of the head and body parts through the movement trajectory of the operator, and providing a virtual recognition and judgment result of the occluded parts; S15. Give all the identification results, including whether the wearer is wearing a safety helmet, insulating protective clothing, and insulating protective gloves.

4. The image recognition method for non-stop power distribution network operation as claimed in claim 3, characterized in that: The steps of the semi-supervised learning recognition algorithm of the YOLO algorithm are: S21. Design operators identify the network and specific network structure; S22, selecting a large public data set and some real construction site image sets to train the network model offline, determining and fine-tuning the parameters of the network model, and then obtaining a generalized model; S23. Apply semi-supervised learning to online learning, enhance the generalization ability of the target monocular camera in some specific scenarios, and analyze the motion trajectory of the operator to solve the occlusion problem of the head and body.

5. The image recognition method for non-stop operation of distribution network as claimed in claim 1, characterized in that: Step S1 includes data acquisition, including laser radar, high-precision inertial sensor and camera as data acquisition sources, judging the distance between people and charged bodies through human key point recognition system, charged body recognition system, data fusion system and spatial position judgment system, and giving active warning.

6. The image recognition method for non-stop power distribution network operation according to claim 5, characterized in that: Among them, the human limb key point position recognition module of the human body key point recognition system quickly recognizes the position of human limb key points based on the improved convolutional pose machine network.

7. The image recognition method for non-stop operation of distribution network as claimed in claim 6, characterized in that: The specific steps of the improved convolutional pose machine network to quickly identify the key points of human limbs are as follows: S31, performing a convolution operation on the acquired original camera image and the Gaussian function template to generate a center image; S32, the center graph is pooled into a small center graph and sent to the second, third and fourth stage networks; S33, the original camera image is sent to the second, third and fourth stage networks after being operated by MultiBlock_1; S34, the original camera image is sent to the second and third stage networks after being operated by MultiBlock_0; S35, the characteristic data after the second stage network operation is sent to the third and fourth stage networks; S36. In the fourth stage network, the feature data is merged and then subjected to the ConvBlock operation to output 64*64*18 feature data.

Images