Risk target identification method and device for electric power operation, and computer program product

Abstract

The invention discloses a risk target identification method and device for electric power operation, and a computer program product. The method comprises the following steps: predicting a video type to which a differentiated video of a power operation scene belongs by using a first prediction model to obtain a plurality of first type conditional probabilities of the differentiated video; performing gradient extraction on the video frame image of the differentiated video based on the first type of conditional probability to obtain a video image information gradient, and performing regularization processing on the video image information gradient to calibrate a risk target in the differentiated video; predicting a risk type to which the risk target belongs by using a second prediction model to obtain a plurality of second type conditional probabilities of the risk target; and determining the risk type corresponding to the maximum second type conditional probability as the actual risk type of the risk target. According to the method and the device, the technical problem that potential safety hazards are easily caused due to relatively low accuracy when risk targets in an electric power operation scene are classified and identified in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of risk management and control in power operation scenarios, and in particular to a method and device for identifying risk targets in power operations, and a computer program product. Background Art

[0002] When the power industry is performing different types of operations, classifying and identifying risk targets in the operation scene through monitoring videos helps to eliminate risks and issue early warnings, and achieve safe power operations. At present, a major problem with classifying and identifying risk targets in power operation scenes is that the accuracy rate is not high enough. If misidentification occurs, it will not only increase the workload, but also may miss the opportunity to eliminate risks and issue early warnings.

[0003] There are several solutions in the prior art for risk target identification in surveillance videos: 1) Classify and identify risk objects in regional scenes based on drone videos, use drones to collect surveillance videos and perform multi-scale enhancement, build a risk target identification model based on convolutional neural networks, and complete the classification and identification of risk objects in scene surveillance videos. This research result uses the probability of correctly identifying risk targets as an evaluation indicator. The experimental results show that the minimum probability of correctly identifying risk targets by this method is 0.969, but this method does not divide the types of surveillance videos. When applied to risk target identification in differentiated videos, the accuracy needs to be improved; 2) Realize real-time target recognition based on the improved YOL0v3-tiny algorithm. First, improve the YOL0v3-tiny network structure and complete it in surveillance videos. Target detection. This research result uses mean average precision as the evaluation index. The experimental results show that the mean average precision of this method reaches 0.458. This method does not take into account that the video images obtained in different work scenes have differentiated fuzzy edges, resulting in a large gap between the recognition accuracy and the ideal value; 3) Based on the automatic recognition method of intruder targets in the monitoring video of dangerous work areas, the danger level of the work area is divided and the dangerous area is determined; the monitoring video data is integrated and the moving targets in the dangerous area are identified using algorithms such as the three-frame difference method. This research result uses the intrusion detection accuracy in different scenarios as the evaluation index. The experimental results show that the minimum accuracy of this method is 89.2%. This method does not process the offset between the center of the blurred image and the actual center, so it cannot obtain a higher accuracy.

[0004] With regard to the problem that the above-mentioned related technologies have low accuracy in classifying and identifying risk targets in power operation scenarios, which easily leads to safety hazards, no effective solution has been proposed so far. Summary of the invention

[0005] The embodiments of the present invention provide a method and device for identifying risk targets in power operations, and a computer program product, so as to at least solve the technical problem in the related art that the accuracy rate is low when classifying and identifying risk targets in power operation scenarios, which may easily lead to safety hazards.

[0006] According to one aspect of an embodiment of the present invention, a method for identifying risk targets in power operations is provided, comprising: using a first prediction model to predict the video type to which a differentiated video of a power operation scene belongs, and obtaining a plurality of first-type conditional probabilities of the differentiated video, wherein the first prediction model is obtained by training using a plurality of groups of first training data in a deep learning manner, and each of the plurality of groups of first training data comprises: a sample differentiated video, a sample video type corresponding to the sample differentiated video, the differentiated video refers to different video data of the power operation scene with different conditional features, the first-type conditional probability refers to the probability that the differentiated video belongs to any of the video types, and the sum of the probabilities of the plurality of first-type conditional probabilities is 1; performing gradient extraction on a video frame image of the differentiated video based on the first-type conditional probability, and obtaining a video image information gradient of the differentiated video, wherein the video image information gradient The degree is used to reflect the feature changes of the video frame image; the video image information gradient of the differentiated video is regularized to calibrate the risk target in the differentiated video, wherein the risk target refers to a specific target that needs to be identified and classified in the differentiated video; the risk type to which the risk target calibrated in the differentiated video belongs is predicted using a second prediction model to obtain multiple second-type conditional probabilities of the risk target, wherein the second prediction is obtained by training with multiple groups of second training data in a deep learning manner, and each of the multiple groups of second training data includes: a sample risk target, a sample risk type corresponding to the sample risk target, and the second-type conditional probability refers to the probability that the risk target belongs to any of the risk types; the risk type corresponding to the largest second-type conditional probability among multiple second-type conditional probabilities is determined as the actual risk type of the risk target.

[0007] Optionally, gradient extraction is performed on the video frame image of the differentiated video based on the first type of conditional probability to obtain the video image information gradient of the differentiated video, including: arbitrarily selecting multiple frames of the video frame image from the differentiated video, and preprocessing the multiple frames of the video frame image; using an image gradient calculation method to calculate the gradient values ​​of the multiple frames of the video frame image, wherein the gradient value is used to reflect the change in pixel intensity in the video frame image; and the gradient values ​​of the multiple frames of the video frame image are merged according to the time series of the multiple frames of the video frame image to obtain the video image information gradient of the differentiated video.

[0008] Optionally, after performing gradient extraction on the video frame image of the differentiated video based on the first type of conditional probability to obtain the video image information gradient of the differentiated video, the risk target identification method for power operations further includes: determining, based on the video image information gradient, that the video frame image whose gradient value is lower than a gradient threshold is a blurred video frame image, wherein the blurred video frame image is an image in which the risk target in the video frame image has blurred edges; determining a center processing box based on a blur center and size of a blurred area in the blurred video frame image, wherein the blurred area refers to an area in the video frame image where the edge of the risk target is blurred, and the center processing box is used to perform image enhancement processing on the blurred area; obtaining an offset between the blurred center and the actual center, wherein the actual center refers to the center position of the risk target in a clear state, and the clear state refers to a state in which the blurriness of the video frame image is lower than a blurriness threshold, wherein the offset is calculated by normalization; performing an image enhancement operation on the blurred area using the center processing box based on the offset to make the blurred video frame image reach the clear state.

[0009] Optionally, obtaining the offset between the blur center and the actual center includes: determining the offset by using a first formula according to coordinate values ​​of the blur center and the actual center, wherein the first formula is: (h, j) represents the coordinates of the pixel point in the video frame image, h represents the row coordinate of the pixel point, j represents the column coordinate of the pixel point, Δh represents the offset in the row direction, j represents the offset in the column direction, m h represents the coordinates of the blur center in the row direction, m j represents the coordinates of the blur center in the column direction, R h represents the coordinates of the actual center in the row direction, R j represents the coordinates of the actual center in the column direction, and ζ represents the adjustment coefficient.

[0010] Optionally, the image enhancement operation is performed on the blurred area using the central processing box according to the offset so that the blurred video frame image reaches the clear state, including: adjusting the position of the central processing box according to the offset to obtain an adjustment result; when the adjustment result indicates that the center of the central processing box is aligned with the blur center, performing the image enhancement operation on the blurred area so that the blurred video frame image reaches the clear state.

[0011] Optionally, the video image information gradient of the differentiated video is regularized to calibrate the risk target in the differentiated video, including: regularizing the video image information gradient of the differentiated video to obtain a regularized gradient; obtaining a central processing box in the image enhancement process of the differentiated video; obtaining a gradient offset value in the video image information gradient, wherein the gradient offset value corresponds to the offset between the blur center of the blurred area in the video frame image and the actual center of the risk target; dividing the boundary of the central processing box according to the gradient offset value and the regularized gradient to obtain a regularized boundary box, so as to use the regularized boundary box to calibrate the risk target in the differentiated video.

[0012] Optionally, using a second prediction model to predict the risk type to which the risk target marked in the differentiated video belongs to, and obtaining multiple second-type conditional probabilities of the risk target, includes: using a second formula to determine the second-type conditional probability that the risk target in the video frame image belongs to any of the risk types, and obtaining multiple second-type conditional probabilities, wherein the second formula is: N represents the second type conditional probability that the risk target belongs to any of the risk types, X represents any one of the multiple risk targets, B represents any one of the multiple risk types, C(X) represents the first probability that the risk target in the video frame image is X, C(X|B) represents the second probability that the risk type to which the risk target belongs is B, and V represents the filter coefficient.

[0013] According to another aspect of an embodiment of the present invention, a risk target identification device for power operation is also provided, including: a first prediction module, used to use a first prediction model to predict the video type to which the differentiated video of the power operation scene belongs, and obtain multiple first-type conditional probabilities of the differentiated video, wherein the first prediction model is obtained by training using multiple groups of first training data in a deep learning manner, and each of the multiple groups of first training data includes: a sample differentiated video, a sample video type corresponding to the sample differentiated video, the differentiated video refers to different video data of the power operation scene with different conditional features, the first-type conditional probability refers to the probability that the differentiated video belongs to any of the video types, and the sum of the probabilities of multiple first-type conditional probabilities is 1; a gradient extraction module, used to perform gradient extraction on the video frame image of the differentiated video based on the first-type conditional probability, and obtain the video image information gradient of the differentiated video, wherein the video image information gradient is used to invert reflecting the feature changes of the video frame image; a regularization processing module, used to regularize the video image information gradient of the differentiated video to calibrate the risk target in the differentiated video, wherein the risk target refers to a specific target that needs to be identified and classified in the differentiated video; a second prediction module, used to predict the risk type to which the risk target calibrated in the differentiated video belongs using a second prediction model, and obtain multiple second-type conditional probabilities of the risk target, wherein the second prediction is obtained by training with multiple groups of second training data in a deep learning manner, and each of the multiple groups of second training data includes: a sample risk target, a sample risk type corresponding to the sample risk target, and the second-type conditional probability refers to the probability that the risk target belongs to any of the risk types; a first determination module, used to determine that the risk type corresponding to the largest second-type conditional probability among multiple second-type conditional probabilities is the actual risk type of the risk target.

[0014] Optionally, the gradient extraction module includes: a preprocessing unit, used to arbitrarily select multiple frames of the video frame images from the differentiated video and preprocess the multiple frames of the video frame images; a calculation unit, used to calculate the gradient values ​​of the multiple frames of the video frame images using an image gradient calculation method, wherein the gradient value is used to reflect the change of pixel intensity in the video frame image; a fusion unit, used to fuse the gradient values ​​of the multiple frames of the video frame images according to the time series of the multiple frames of the video frame images to obtain the video image information gradient of the differentiated video.

[0015] Optionally, the risk target identification method for power operation further includes: a second determination module, which is used to perform gradient extraction on the video frame image of the differentiated video based on the first type of conditional probability, and then determine, based on the video image information gradient of the differentiated video, that the video frame image with a gradient value lower than a gradient threshold is a blurred video frame image, wherein the blurred video frame image is an image in which the risk target in the video frame image has blurred edges; a third determination module, which is used to determine a center processing frame according to a blur center and size of a blurred area in the blurred video frame image, wherein the blurred area refers to an area in the video frame image where the edge of the risk target is blurred, and the center processing frame is used to perform image enhancement processing on the blurred area; an acquisition module, which is used to acquire an offset between the blur center and an actual center, wherein the actual center refers to the central position of the risk target in a clear state, and the clear state refers to a state in which the blur of the video frame image is lower than the blur threshold, wherein the offset is calculated by normalization; an image enhancement module, which is used to perform an image enhancement operation on the blurred area using the center processing frame according to the offset, so that the blurred video frame image reaches the clear state.

[0016] Optionally, the acquisition module includes: a first determination unit, configured to determine the offset using a first formula according to the coordinate values ​​of the blur center and the actual center, wherein the first formula is: (h, j) represents the coordinates of the pixel point in the video frame image, h represents the row coordinate of the pixel point, j represents the column coordinate of the pixel point, Δh represents the offset in the row direction, j represents the offset in the column direction, m h represents the coordinates of the blur center in the row direction, m j represents the coordinates of the blur center in the column direction, R h represents the coordinates of the actual center in the row direction, R j represents the coordinates of the actual center in the column direction, and ζ represents the adjustment coefficient.

[0017] Optionally, the image enhancement module includes: an adjustment unit, used to adjust the position of the central processing box according to the offset to obtain an adjustment result; an image enhancement unit, used to perform the image enhancement operation on the blurred area when the adjustment result indicates that the center of the central processing box is aligned with the blurred center, so as to make the blurred video frame image reach the clear state.

[0018] Optionally, the regularization processing module includes: a regularization processing unit, used to perform regularization processing on the video image information gradient of the differentiated video to obtain a regularized gradient; a first acquisition unit, used to obtain a central processing box in the image enhancement processing process of the differentiated video; a second acquisition unit, used to obtain a gradient offset value in the video image information gradient, wherein the gradient offset value corresponds to the offset between the blur center of the blurred area in the video frame image and the actual center of the risk target; a boundary division unit, used to perform boundary division on the central processing box according to the gradient offset value and the regularization gradient to obtain a regularized boundary box, so as to use the regularized boundary box to calibrate the risk target in the differentiated video.

[0019] Optionally, the second prediction module includes: a second determination unit, configured to determine the second type conditional probability that the risk target in the video frame image belongs to any of the risk types using a second formula, to obtain a plurality of second type conditional probabilities, wherein the second formula is: N represents the second type conditional probability that the risk target belongs to any of the risk types, X represents any one of the multiple risk targets, B represents any one of the multiple risk types, C(X) represents the first probability that the risk target in the video frame image is X, C(X|B) represents the second probability that the risk type to which the risk target belongs is B, and V represents the filter coefficient.

[0020] According to another aspect of an embodiment of the present invention, a risk target identification system for electric power operations is further provided. The risk target identification system for electric power operations uses any one of the risk target identification methods for electric power operations described above.

[0021] According to another aspect of an embodiment of the present invention, a computer-readable storage medium is further provided, wherein the computer-readable storage medium includes a stored program, wherein the program executes any one of the above-mentioned risk target identification methods for power operations.

[0022] According to another aspect of an embodiment of the present invention, a processor is further provided, wherein the processor is used to run a program, wherein when the program is run, any one of the above-mentioned risk target identification methods for power operations is executed.

[0023] According to another aspect of an embodiment of the present invention, a computer program product is provided, comprising computer instructions, which, when executed by a processor, perform any one of the above-mentioned methods for identifying risk targets of power operations.

[0024] In an embodiment of the present invention, a first prediction model is used to predict the video type to which a differentiated video of an electric power operation scene belongs, and a plurality of first type conditional probabilities of the differentiated video are obtained, wherein the first prediction model is obtained by training using a plurality of groups of first training data in a deep learning manner, and each of the plurality of groups of first training data includes: a sample differentiated video, and a sample video type corresponding to the sample differentiated video, wherein the differentiated video refers to different video data of an electric power operation scene with different conditional characteristics, and the first type conditional probability refers to the probability that the differentiated video belongs to any video type, and the sum of the probabilities of the plurality of first type conditional probabilities is 1; based on the first type conditional probability, a gradient extraction is performed on a video frame image of the differentiated video to obtain a video image information gradient of the differentiated video, wherein the video image information gradient is used to reflect the video frame The method comprises the following steps: the first step is to regularize the video image information gradient of the differentiated video to calibrate the risk target in the differentiated video, wherein the risk target refers to a specific target that needs to be identified and classified in the differentiated video; the second prediction model is used to predict the risk type to which the risk target calibrated in the differentiated video belongs, and a plurality of second-type conditional probabilities of the risk target are obtained, wherein the second prediction is obtained by training with multiple groups of second training data in a deep learning manner, and each group in the multiple groups of second training data includes: a sample risk target, a sample risk type corresponding to the sample risk target, and the second-type conditional probability refers to the probability that the risk target belongs to any risk type; and the risk type corresponding to the largest second-type conditional probability among the multiple second-type conditional probabilities is determined as the actual risk type of the risk target. Through the technical solution provided by the present invention, conditional probability calculation is performed through a graph convolutional neural network, blurred edges in video images are captured, image gradients are extracted using a normalization processing method, and at the same time, information gradients are regularized, and the type probability scores of risk targets in video images are determined according to the regularized gradients, thereby achieving the technical effect of accurately classifying and identifying risk targets in risk-differentiated videos of power operation scenarios, improving the safety and efficiency of power operation scenarios, and further solving the technical problem in the related art that the accuracy rate is low when classifying and identifying risk targets in power operation scenarios, which easily leads to safety hazards. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0026] Figure 1 It is a hardware structure block diagram of a mobile terminal of a method for identifying risk targets of electric power operations according to an embodiment of the present invention;

[0027] Figure 2is a flow chart of a method for identifying risk targets of power operations according to an embodiment of the present invention;

[0028] Figure 3 is a statistical diagram of the number of risk-differentiated samples at a current detection operation site according to an embodiment of the present invention;

[0029] Figure 4 is a statistical diagram of the number of risk-differentiated samples at a transformer installation operation site according to an embodiment of the present invention;

[0030] Figure 5 is a statistical diagram of the number of risk-differentiated samples at an electric energy meter calibration operation site according to an embodiment of the present invention;

[0031] FIG6( a ) is a statistical diagram of identification results of a current detection operation site under different methods according to an embodiment of the present invention;

[0032] FIG6( b ) is a statistical diagram of identification results of a transformer installation operation site under different methods according to an embodiment of the present invention;

[0033] FIG6( c ) is a statistical diagram of recognition results of an electric energy meter calibration operation site under different methods according to an embodiment of the present invention;

[0034] FIG. 7( a ) is a statistical diagram of the classification and recognition accuracy when the risk target is “personnel” according to various methods of the embodiments of the present invention;

[0035] FIG. 7( b ) is a statistical diagram of the classification and recognition accuracy of the risk target when the risk target is “car” according to various methods of the embodiments of the present invention;

[0036] FIG. 7( c ) is a statistical diagram of the classification and recognition accuracy when the risk target is an “insulator” according to various methods of the embodiments of the present invention;

[0037] Figure 8 is a schematic diagram of a risk target identification device for power operations according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings 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 should fall within the scope of protection of the present invention.

[0039] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0040] As introduced in the background technology, the accuracy of classifying and identifying risk targets in power operation scenarios in related technologies is low, which easily leads to safety hazards. In view of the above defects, in the embodiments of the present invention, a method and device for identifying risk targets in power operation, a computer program product, a system for identifying risk targets in power operation, a computer-readable storage medium, a processor and a computer program product are provided.

[0041] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0042] The method embodiments provided in the embodiments of the present invention can be executed in a mobile terminal, a computer terminal or a similar computing device. Taking running on a mobile terminal as an example, Figure 1 1 is a hardware structure block diagram of a mobile terminal for a method for identifying risk targets of power operations according to an embodiment of the present invention. Figure 1 As shown, the mobile terminal may include one or more ( Figure 1 Only one is shown in the figure) a processor 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA) and a memory 104 for storing data, wherein the mobile terminal may also include a transmission device 106 and an input / output device 108 for communication functions. It can be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the mobile terminal. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations shown.

[0043] The memory 104 can be used to store computer programs, for example, software programs and modules of application software, such as the computer program corresponding to the risk target identification method of the power operation in the embodiment of the present invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, the above method is implemented. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories can be connected to the mobile terminal via a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof. The transmission device 106 is used to receive or send data via a network. The above-mentioned specific examples of the network may include a wireless network provided by a communication provider of the mobile terminal. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, referred to as NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0044] According to an embodiment of the present invention, a method embodiment of a method for identifying risk targets in power operations is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in an order different from that shown here.

[0045] Figure 2 is a flow chart of a method for identifying risk targets of power operations according to an embodiment of the present invention. Figure 2 As shown, the method comprises the following steps:

[0046] Step S202: Use the first prediction model to predict the video type to which the differentiated video of the power operation scene belongs, and obtain multiple first-type conditional probabilities of the differentiated video, wherein the first prediction model is trained using multiple groups of first training data in a deep learning manner, and each of the multiple groups of first training data includes: a sample differentiated video, and a sample video type corresponding to the sample differentiated video. The differentiated video refers to different video data of the power operation scene with different conditional characteristics. The first-type conditional probability refers to the probability that the differentiated video belongs to any video type, and the sum of the probabilities of the multiple first-type conditional probabilities is 1.

[0047] It should be noted that the above-mentioned differentiated videos refer to video data that are classified into different types according to the different characteristics and conditions of the power operation scene. These videos may come from the same scene but have different risk characteristics, or from completely different power operation environments. Each video may contain specific risk targets. Therefore, differentiated videos can be understood as multiple videos, each of which is processed and classified separately due to its different content or risk attributes.

[0048] In this embodiment, the first prediction model is used to predict the video type of the differentiated video of the power operation scene, which can help identify different types of power operation scenes, thereby improving the accuracy of identifying potential risk targets, and improving the conditional probabilities of multiple first types of differentiated videos. The type of video can be more accurately determined, and the characteristics of different power operation scene videos can be identified by learning a large amount of training data through a deep information model, thereby improving the accuracy of differentiated video classification, thereby providing a more reliable basis for the identification of risk targets, helping to improve the safety and efficiency of power operation scenes and reduce the occurrence of accidents.

[0049] Specifically, a graph convolutional neural network (i.e., the first prediction model) is used to obtain the conditional probability of differentiated video types. The core element in the graph convolutional network is the loss function, which can calculate the gap between the input value and the true value in the network. The smaller the gap, the more similar the obtained output value is to the actual situation. The main way to classify and identify risk-differentiated videos in power operation scenes is to divide the corresponding video types. Only when the corresponding scenes are correctly divided into their corresponding types can the objects in the scenes be identified. The graph convolutional network uses the following loss function for analysis: Where q(·) is the mean square error loss function, a is the hinge loss function, which is used to measure the gap between the predicted value and the actual value of the classification model, and is particularly suitable for classification models. t is the recognition output value of the graph convolutional neural network, u is the input video data in the network, y and i are the learning rate parameters and bias of the neural network, which are set to 0.05 and 0.7 respectively, p(·) is the interval function, w is the number of risky videos in the power operation scene, and r e is the true type of the e-th video data, a(·) is the hinge loss function, which is expressed as a variable based on p(·), "c+" is a function that takes a positive value, [s] is the confidence of the differentiated video type classification, and only when this value is high enough can it indicate that the loss of video classification is small. It is measured by q(·) and a(·), and the conditional probability of the classification of differentiated videos is calculated according to the cross loss. The output probability of different differentiated video types in the graph convolutional neural network is determined. The output probability (i.e., the first type conditional probability) contains the following formula, which is expressed as: Indicates the confidence of the video output result of type e, which is the probability value after normalization. By calculating the logarithm of the true type and the conditional distribution probability, the conditional probability of each type of video is obtained, which reflects the classification confidence of GCN for each type of video. Among them, a represents the conditional probability of video classification, which is calculated by the cross entropy loss function, reflecting the classification confidence of the classification model for each video type, and s e is the output of the video of type e in the graph convolutional network, and also represents the confidence of the output result of this type, d e is the conditional distribution probability of type e video, f and d are the real distribution probability of this type of video and the conditional distribution probability of classification. The loss function in the graph convolutional network is used to measure the distribution of differentiated video types and divide them according to different types of real attributes. Based on this, the edge information of power operation scenes in different differentiated videos can be extracted.

[0050] It should be noted that the above-mentioned recognition output value refers to the output result obtained by the graph convolutional neural network (GCN) after training, analyzing and processing the input video data. The specific form of the recognition output value depends on the architecture and training method of the GCN. For example, it can be a vector in which each element represents the probability of a video type or risk level, or it can be an image containing the location information of the target.

[0051] It should be noted that the above r e Indicates the true type of the e-th video data, which is used to compare with the recognition output value of the graph convolutional neural network to calculate the loss function. It is the same concept and is known that r e It has the following characteristics: e is a discrete value representing a predefined video type, such as 1, 2, 3, etc. e It is known information and is used to train the GCN model. During the training process, GCN will use the input video data and the corresponding r e value, and continuously adjust the network parameters to reduce the value of the loss function. After the model training is completed, r e It can be used to evaluate the performance of the model, such as calculating classification accuracy, recall rate and other indicators.

[0052] It should be noted that the confidence of the above-mentioned differentiated video type division can be understood as the reliability or confidence level of the video type division results of the graph convolutional neural network. The confidence of differentiated video type division refers to the confidence of the division results, which is used to indicate the degree of match between the results of GCN video type division and its true type. The confidence of differentiated video type division is an important indicator that can be used to evaluate the performance of the GCN model and help understand the output results of the model.

[0053] It should be noted that the output probability means that after GCN analyzes the input video data, it outputs the probability value of each video type. The probability value range is between 0 and 1, and the sum of the probability values ​​of all video types is 1. The higher the probability value, the more likely GCN believes that the video type will appear; the lower the probability value, the less likely GCN believes that the video type will appear.

[0054] Step S204, performing gradient extraction on the video frame image of the differentiated video based on the first type conditional probability to obtain the video image information gradient of the differentiated video, wherein the video image information gradient is used to reflect the feature change of the video frame image.

[0055] In this embodiment, gradient extraction can help identify feature changes in differentiated videos, thereby identifying potential risk targets. By extracting gradients from video frame images, changes in pixel intensity in video frame images can be captured, thereby identifying key features and edges in the image. Changes in the video can be analyzed more accurately, thereby discovering potential safety risks in a timely manner, thereby improving the safety and efficiency of power operations.

[0056] Step S206 , regularizing the video image information gradient of the differentiated video to calibrate the risk target in the differentiated video, wherein the risk target refers to a specific target in the differentiated video that needs to be identified and classified.

[0057] In this embodiment, regularizing the video image information gradient of the differentiated video can help calibrate risk targets, that is, specific targets that need to be identified and classified. Regularizing the video image information gradient can help reduce the overfitting phenomenon of the model during the training process, improve the recognition accuracy and classification effect of the target, enhance the robustness and adaptability of the system, and accurately calibrate the risk targets, so that risk management and early warning can be more effectively carried out, thereby helping the monitoring system to quickly and accurately identify non-potential risks in power operations, take corresponding measures in time to avoid accidents, and ensure the safety and smooth progress of power operations.

[0058] Step S208, using the second prediction model to predict the risk type to which the risk target marked in the differentiated video belongs, to obtain multiple second type conditional probabilities of the risk target, wherein the second prediction is obtained by training using multiple groups of second training data in a deep learning manner, and each of the multiple groups of second training data includes: a sample risk target, a sample risk type corresponding to the sample risk target, and the second type conditional probability refers to the probability that the risk target belongs to any risk type.

[0059] In this embodiment, the risk type to which the risk target calibrated by the second prediction model belongs is predicted, which helps to further improve the recognition accuracy and reliability. The second prediction model trained by deep learning can predict multiple second-type conditional probabilities of risk targets based on sample risk targets and corresponding sample risk types in multiple groups of second training data, that is, the probability that the risk target belongs to any risk type. It can be used to evaluate the risk level of risk targets in the video, which helps to take corresponding preventive or response measures for targets of different risk levels. By identifying and analyzing risk types in advance, accidents can be effectively prevented, monitoring and safety resources can be allocated more effectively, and the safety of power workers and equipment can be guaranteed, which helps to improve the safety and efficiency of power operations.

[0060] Step S210: Determine the risk type corresponding to the largest second type conditional probability among the plurality of second type conditional probabilities as the actual risk type of the risk target.

[0061] In this embodiment, the risk type corresponding to the largest second-type conditional probability among multiple second-type conditional probabilities is determined to be the actual risk type, so that the highest risk that may occur can be identified, so that corresponding risk control measures can be taken in time to reduce the probability of occurrence of potential risk events and ensure the safe conduct of power operations. By identifying the actual risk type, risk management strategies and measures can be formulated in a targeted manner to improve the safety and efficiency of power operations.

[0062] As can be seen from the above, in an embodiment of the present invention, a first prediction model is used to predict the video type to which the differentiated video of the electric power operation scene belongs, and multiple first-type conditional probabilities of the differentiated video are obtained, wherein the first prediction model is obtained by training using multiple groups of first training data in a deep learning manner, and each of the multiple groups of first training data includes: a sample differentiated video, a sample video type corresponding to the sample differentiated video, and the differentiated video refers to different video data of the electric power operation scene with different conditional characteristics. The first-type conditional probability refers to the probability that the differentiated video belongs to any video type, and the sum of the probabilities of multiple first-type conditional probabilities is 1; based on the first-type conditional probability, the video frame image of the differentiated video is gradient extracted to obtain the video image information gradient of the differentiated video, wherein the video image information gradient is used to reflect the feature changes of the video frame image; the video image information gradient of the differentiated video is regularized to calibrate the risk targets in the differentiated video, wherein the risk targets refer to the differentiated videos that need to be identified and The specific goal of classification is to predict the risk type to which the risk target marked in the differentiated video belongs by using the second prediction model, and obtain multiple second type conditional probabilities of the risk target, wherein the second prediction is obtained by training with multiple groups of second training data in a deep learning manner, and each group in the multiple groups of second training data includes: sample risk targets, sample risk types corresponding to the sample risk targets, and the second type conditional probability refers to the probability that the risk target belongs to any risk type; the risk type corresponding to the largest second type conditional probability among multiple second type conditional probabilities is determined as the actual risk type of the risk target, and the conditional probability calculation is achieved through the graph convolutional neural network, and the blurred edges in the video image are captured, and the image gradient is extracted by normalization processing. At the same time, the information gradient is regularized, and the type probability score of the risk target in the video image is determined according to the regularized gradient, thereby achieving the technical effect of accurately classifying and identifying risk targets in the differentiated video of the risk of power operation scenes, and improving the safety and efficiency of the power operation scenes.

[0063] The technical solution provided by the embodiment of the present invention solves the technical problem in the related art that the accuracy rate is low when classifying and identifying risk targets in power operation scenarios, which easily leads to safety hazards.

[0064] According to the above embodiment of the present invention, gradient extraction is performed on the video frame image of the differentiated video based on the first type of conditional probability to obtain the video image information gradient of the differentiated video, including: arbitrarily selecting multiple video frame images from the differentiated video, and preprocessing the multiple video frame images; using an image gradient calculation method to calculate the gradient values ​​of the multiple video frame images, wherein the gradient values ​​are used to reflect the change of pixel intensity in the video frame image; and the gradient values ​​of the multiple video frame images are fused according to the time series of the multiple video frame images to obtain the video image information gradient of the differentiated video.

[0065] In this embodiment, by fusing the gradient values ​​of multiple video frame images, the video image information gradient of the differentiated video can be obtained, which is then used to identify possible risk targets in power operations. By analyzing and identifying the video image information gradient, the safety and efficiency of power operations can be improved, and it can help the system pre-process the video frame images and calculate the gradient values, thereby more accurately capturing the pixel intensity changes in the video frame images.

[0066] Optionally, the above preprocessing may include but is not limited to: grayscale, denoising, and normalization operations.

[0067] Specifically, the method and process of obtaining pixel gradient values ​​are as follows: 1) Randomly select images, randomly select several frames of images from the video to ensure the diversity and representativeness of the samples, and then the selected image frames need to cover different work scenes and time periods for comprehensive analysis and evaluation; 2) Preprocess the image, preprocess the selected image frames, the preprocessing step can improve the image quality and make the subsequent gradient calculation more accurate; 3) Calculate the pixel gradient value, use the image gradient calculation method to calculate the gradient value of each pixel, the gradient value reflects the change of pixel intensity, the gradient value at the edge is usually larger, and the gradient value in the flat area is smaller; 4) Obtain the edge information of the image, by calculating the image The gradient is used to obtain the edge information of the risk target in the image. The gradient calculation can be performed in the horizontal and vertical directions respectively, and then the results are merged to obtain the total gradient strength; 5) Determine the image blur and count the distribution of gradient values ​​in each frame. If the gradient value of a frame is small as a whole, it means that there is less edge information in the image and there may be blur; 6) Filter and process blurred images. According to the distribution of gradient values, filter out image frames with blur, and perform edge enhancement on the filtered blurred images to improve the clarity of the image; 7) Risk target identification. Perform risk target identification on the processed image and accurately locate and classify risk targets based on clear edge information.

[0068] Optionally, the above-mentioned image gradient calculation method may include a variety of classic image gradient calculation methods, among which the most common image gradient calculation methods include: Sobel operator, Prewitt operator, Laplacian operator and Canny edge detection, etc.

[0069] According to the above embodiment of the present invention, after performing gradient extraction on the video frame image of the differentiated video based on the first type of conditional probability to obtain the video image information gradient of the differentiated video, the risk target identification method for power operation also includes: determining, based on the video image information gradient, that a video frame image whose gradient value is lower than a gradient threshold is a blurred video frame image, wherein the blurred video frame image is an image in which the risk target in the video frame image has blurred edges; determining a center processing frame based on the blur center and size of the blurred area in the blurred video frame image, wherein the blurred area refers to an area in the video frame image where the edge of the risk target is blurred, and the center processing frame is used to perform image enhancement processing on the blurred area; obtaining an offset between the blurred center and the actual center, wherein the actual center refers to the center position of the risk target in a clear state, and the clear state refers to a state in which the blurriness of the video frame image is lower than the blurriness threshold, wherein the offset is calculated by normalization; performing an image enhancement operation on the blurred area using the center processing frame according to the offset to make the blurred video frame image reach a clear state.

[0070] In this embodiment, by calculating the video image information gradient, the system can identify blurred video frame images with gradient values ​​lower than a preset threshold. Through gradient detection and image enhancement operations, the accuracy and efficiency of risk target identification can be effectively improved, and risk monitoring and management can be better performed. The blurred video frame images can be processed to improve image quality and make the areas with blurred edges of risk targets clear, which is convenient for subsequent analysis and processing. The normalized offset calculation method enables the method to adapt to images with different degrees of blur, and has good adaptability and generalization ability, thereby more accurately identifying and locating risk targets.

[0071] It should be noted that since the video is composed of a series of images, and the images obtained in different power operation scenes are differentiated, it is inevitable that blurred edges will be generated during the capture process. In order to achieve differentiated video classification and recognition in power operation scenes, the information target corresponding to the conditional probability in the above-mentioned conditional probability (corresponding differentiated video) is first to extract the features of the input video data, and then obtain the feature representation through graph convolutional neural network processing. The difference between the network output and the true value is calculated through the mean square error and hinge loss function. The conditional distribution probability is calculated to obtain the output confidence of each video type. The conditional probability calculation obtains the conditional probability of each video type, and determines the information target. According to the output conditional probability, the classification result and confidence of each video data are determined. The classification result of each input video and its corresponding conditional probability can be determined to achieve differentiated video classification and recognition in the operation scene. The edge information gradient is calculated to determine whether there is blur in the video. If so, it needs to be enhanced before recognition.

[0072] The calculation is as follows: Among them, (h, j) is a group of image pixels in the video, g(h, j) is the pixel grayscale value of the midpoint of the image, and c(d e ) is based on d e The gradient function, Δ 2 c(h, j) is the information gradient at that location, l h，j 、z h，j 、x h，j is the color gradient of colors l, z, x in (h, j), k l , k z , k x are the representative features of colors l, z, x, l, z, x represent the gradient features of red, green and blue in the image respectively, which are used to calculate the edge information gradient in the image and identify and process the blurred image area in the video. b is the total number of image pixels and n is the blurriness of the image.

[0073] In addition, by randomly selecting continuous images, based on conditional probability and gradient function, images with blur are screened, and the edge information gradient in the image is calculated to determine whether the image is blurry. The specific standard is the size of the gradient value. By calculating the gradient value in the image (i.e., the rate of change of the pixel), it can be determined whether the image is clear. A small gradient value indicates that the change in the image is not obvious, and blur may exist. A large gradient value indicates that there are obvious edges and details in the image, and the image is clear. If blur exists in the selected video image, it needs to be processed. Processing is performed according to the offset between the blur center and the actual center, and the center processing frame of the differentiated video image is set.

[0074] It should be noted that the above-mentioned center processing box is an auxiliary tool for positioning and adjusting the image when processing a blurred image. For example, the center processing box helps to locate the center position of the blurred area in the image so that processing can be performed on these areas.

[0075] According to the above embodiment of the present invention, obtaining the offset between the blur center and the actual center includes: determining the offset using a first formula according to the coordinate values ​​of the blur center and the actual center, wherein the first formula is: (h, j) represents the coordinates of the pixel in the video frame image, h represents the row coordinate of the pixel, j represents the column coordinate of the pixel, Δh represents the offset in the row direction, j represents the offset in the column direction, m h represents the coordinates of the blur center in the row direction, m j represents the coordinates of the blur center in the column direction, R h Represents the coordinates of the actual center in the row direction, R j represents the coordinate of the actual center in the column direction, and ζ represents the adjustment coefficient.

[0076] In this embodiment, the offset between the blurred center and the actual center is determined. Calculating the offset can help identify the exact position of the risky target. By analyzing the pixel coordinates of the risky target in the video frame image, the position of the target can be determined and corresponding measures can be taken to perform risk management, thereby improving the safety and efficiency of power operations.

[0077] The offset is calculated in a normalized manner as follows: Among them, Q and W are the width and height of the central processing box, ζ is taken as ζ∈[-0.5, 0.5], (Δh, Δj) is the offset of normalization processing, and the gradient value of its pixel is obtained by selecting and processing random images. This value is used as a judgment indicator to realize the identification of risky targets in differentiated videos.

[0078] According to the above embodiment of the present invention, an image enhancement operation is performed on the blurred area using the central processing frame according to the offset so that the blurred video frame image can reach a clear state, including: adjusting the position of the central processing frame according to the offset to obtain an adjustment result; when the adjustment result indicates that the center of the central processing frame is aligned with the blur center, an image enhancement operation is performed on the blurred area so that the blurred video frame image can reach a clear state.

[0079] In this embodiment, the blurred video frame image is processed to improve the clarity of the image. By adjusting the position of the central processing box and performing image enhancement operations, risk targets that may exist in power operations can be effectively identified and analyzed, which helps to improve the ability to identify potential risks and enhance work safety and efficiency.

[0080] Specifically, by calculating the offset between the blurred center and the actual center, the center processing box can help adjust the image so that the blurred area is aligned with the actual center to reduce blur. The center processing box can be used as a benchmark to help enhance edge information and improve image clarity during image processing. The center processing box does not exist in the image itself, but is obtained through calculation and setting. It is a tool for auxiliary processing to help locate and adjust the blurred area in the image. Q and W represent the width and height of the center processing box respectively. The parameters are used to define the size of the center processing box to help locate and adjust the blurred area in the image. In practical applications, Q and W are usually known or can be obtained through settings, and can be set according to the specific size of the image and as needed.

[0081] According to the above embodiment of the present invention, the video image information gradient of the differentiated video is regularized to calibrate the risk target in the differentiated video, including: regularizing the video image information gradient of the differentiated video to obtain a regularized gradient; obtaining a central processing box in the image enhancement process of the differentiated video; obtaining a gradient offset value in the video image information gradient, wherein the gradient offset value corresponds to the offset between the blur center of the blurred area in the video frame image and the actual center of the risk target; dividing the boundary of the central processing box according to the gradient offset value and the regularized gradient to obtain a regularized boundary box, so as to calibrate the risk target in the differentiated video using the regularized boundary box.

[0082] In this embodiment, the image information gradient of the differentiated video is regularized and a central processing frame is obtained, and the boundary of the central processing frame is divided by combining the gradient offset value and the regularized gradient to obtain a regularized boundary frame, so as to accurately calibrate the risk targets in the differentiated video, thereby improving the recognition accuracy of risk targets, timely discovering and handling potential safety risks, and ensuring the safe conduct of power operations.

[0083] Specifically, the regularized gradient identifies the risk differentiation video of the power operation scene. To identify the risk target of the differentiated video in different power operation scenes, it is necessary to locate and divide the target scene before identification. The processing frame is constructed by the edge gradient, wherein the processing frame refers to the central processing frame of the differentiated video image. When performing specific target identification, the processing frame can be directly calibrated. This time, the gradient offset value is used as the basis, and the boundary is divided according to the regularization principle to control it within a certain scale range. The specific operation process is as follows: Among them, (G, S) is the set regularized bounding box, (D A , D S ) is the coordinate of the upper left corner of the bounding box, θ(G A ) and θ(G S) is the relative coordinate of the center of the edge gradient bounding box in the recognition image, F A With F S F is the coordinate size constraint value corresponding to the upper left corner of the ruler bounding box. L With F J H is the coordinate size constraint value corresponding to the lower right corner of the ruler bounding box. L With H J is the width and height of the bounding box where the image is located, G A , G S , G L and G J is the relative value of the coordinate size of the upper left corner and the lower right corner of the bounding box, and K is the offset between the bounding box and the image.

[0084] It should be noted that the above-mentioned regularized bounding box takes into account the gradient offset value and the regularization principle. It dynamically adjusts the size and position of the bounding box according to the information gradient and offset of the image to ensure that the bounding box can accurately cover the risk target while avoiding including irrelevant information or missing the key features of the target. In this way, the regularized bounding box can more accurately adapt to the risk targets in different power operation scene videos and improve the accuracy and efficiency of recognition.

[0085] According to the above embodiment of the present invention, the risk type to which the risk target marked in the differentiated video belongs is predicted by using the second prediction model to obtain multiple second-type conditional probabilities of the risk target, including: using the second formula to determine the second-type conditional probability that the risk target in the video frame image belongs to any risk type, and obtaining multiple second-type conditional probabilities, wherein the second formula is: N represents the second type conditional probability that the risk target belongs to any risk type, X represents any one of multiple risk targets, B represents any one of multiple risk types, C(X) represents the first probability that the risk target in the video frame image is X, C(X|B) represents the second probability that the risk type to which the risk target belongs is B, and V represents the filter coefficient.

[0086] In this embodiment, by calculating the second type of conditional probability, the specific risk type to which the risk target in the video frame image belongs can be determined more accurately, which helps the system to accurately identify and classify different types of risk targets, improve the ability to identify and prevent potential risks in power operations, and effectively reduce accident risks and ensure the safety of power workers.

[0087] Specifically, the boundary of the image is demarcated, and the probability of the risk target category is calculated according to its specific width and height, and finally the type of the risk target in the video where the image is located is obtained: Z(X)=C(X)*V, (i.e., the second type conditional probability), where Z(X) is the accuracy of the X boundary determination. By determining the type of randomly selected video images, the recognition boundary is determined under the corresponding probability score, thereby realizing the identification of risk targets in the risk-differentiated video of the power operation scene.

[0088] The embodiment is described below by means of comparative experimental tests. In the above, a method for classifying and identifying videos with differentiated risks in power operation scenarios is constructed by means of a graph convolutional neural network. To verify that the method can achieve effective classification and identification of differentiated risks in power operation scenarios, a comparative experimental test is used for demonstration. Scheme 1 in the prior art is used as control method 1 and scheme 2 in the prior art is used as control method 2, respectively, and compared with the present embodiment to verify the effectiveness of different methods. To ensure the authenticity of this experimental test, the power operation scenarios of the actual power system are divided and classified according to three operation scenarios, and the transactions with operation risks in the scenarios are calibrated and photographed. Table 1 shows the test data collection method:

[0089] Table 1

[0090]

[0091] As shown in Table 1, this time the classification is carried out according to three power operation scenes, and different scenes contain image data of illegal intrusion and normal operation. The acquisition method of each group of images is differentiated, and they are acquired by database expansion and real shooting respectively. Among them, the data acquisition of personnel and cars mainly relies on the real shooting method. Since this embodiment is the classification and recognition of risk-differentiated images in the operation scene, the statistics of the number of samples are only analyzed by the type of illegal intrusion, and data statistics are performed through accurate and good data, and divided into training set and test set according to the ratio of 4:1.

[0092] Figure 3 is a statistical diagram of the number of risk-differentiated samples at a current detection operation site according to an embodiment of the present invention, Figure 4 is a statistical chart of the number of risk-differentiated samples at a transformer installation operation site according to an embodiment of the present invention, Figure 5 is a statistical diagram of the number of risk-differentiated samples at the electric energy meter calibration operation site according to an embodiment of the present invention. The specific number of samples is as follows: Figure 3 , Figure 4 , Figure 5As shown in the figure, the data of risk test samples included in different scenarios are different, among which the number included in the current detection operation site is larger, and the number included in the transformer installation operation site is smaller. According to the training set and the test set, since the recognition and classification process of the traditional method is a known situation, the new method is trained through the training samples before the overall test, and a comparative test is performed after completing 3 recognition trainings. The MATLAB test platform is selected for training, and three groups of recognition methods are connected for comparison, and different methods are connected for recognition.

[0093] Table 2 shows the quantitative data of risk target identification results in different power operation scenarios:

[0094] Table 2

[0095]

[0096] Figure 6(a) is a statistical diagram of the recognition results of the current detection operation site under different methods according to an embodiment of the present invention, Figure 6(b) is a statistical diagram of the recognition results of the transformer installation operation site under different methods according to an embodiment of the present invention, and Figure 6(c) is a statistical diagram of the recognition results of the electric energy meter verification operation site under different methods according to an embodiment of the present invention. The recognition results are shown in Figures 6(a), 6(b), and 6(c). According to the data in Figures 6(a), 6(b), 6(c) and Table 2, this embodiment identifies risk samples in different power operation scenarios, and the difference with the actual number of experimental groups is less than 1, and the risk targets in different scenarios can be identified. However, the recognition results of the two groups of traditional methods are significantly different from the actual number of samples, indicating that there is recognition loss.

[0097] In order to further verify the classification and recognition effect of the method in this application, all the above sample data are processed uniformly, and statistics are directly performed on personnel, cars and insulators. The risk samples in different scenarios are classified and identified respectively to verify the classification and recognition accuracy of different methods. Table 3 shows the quantitative data of the recognition accuracy results of different risk targets by different methods:

[0098] Table 3

[0099]

[0100] FIG7(a) is a statistical diagram of the classification and recognition accuracy of the risk target when the risk target is “personnel” according to the various methods of the embodiments of the present invention, FIG7(b) is a statistical diagram of the classification and recognition accuracy of the risk target when the risk target is “car” according to the various methods of the embodiments of the present invention, and FIG7(c) is a statistical diagram of the classification and recognition accuracy of the risk target when the risk target is “insulator” according to the various methods of the embodiments of the present invention. As shown in FIG7(a), FIG7(b), FIG7(c) and Table 3, different types of risk data samples are shuffled and classified and recognized according to the power operation scenarios in which they are located. Under the application of the method in this paper, accurate recognition can still be achieved, and the minimum accuracy is 0.989. The recognition results of the two groups of traditional methods after scrambling are more different from the actual situation, and the accuracy is lower than 0.98. Based on the above, this embodiment can effectively identify risk transactions in different scenarios and can be classified according to different scenarios, which has specific practical application value.

[0101] The above experiment uses current detection scenario, transformer installation scenario and electric energy meter inspection scenario as monitoring video test types, and identifies risk targets in different scenarios respectively. The experimental results show that the maximum error value of the number of risk target identification in this embodiment is 1, and the minimum value of the correct identification probability of risk targets is 0.989, which proves that the method can accurately realize the classification and identification of differentiated risk videos in power operation scenarios. The power operation scenario contains a variety of risk targets. When monitoring the transmission line, it is necessary to classify and identify the risk targets in the differentiated scenario monitoring videos. Based on the above embodiments and experiments, it can be seen that the method of classifying and identifying differentiated risk videos for power operation scenarios based on graph convolutional neural networks divides the differentiated operation site monitoring video types, and then identifies the illegal intrusion risk targets, and the correct identification probability index value of the risk target is improved.

[0102] It should be noted that the above-mentioned risk targets refer to specific targets that need to be identified and classified in the work scene video, which are usually related to safety and risk management.

[0103] Through the technical solution provided by the above-mentioned embodiment of the present invention, the confidence is determined by the loss function in the graph convolutional neural network, the conditional probability calculation is performed on the video type of the power operation scene, and the actual type of the differentiated video is determined; the fuzzy edges in the video image are captured by conditional probability, the edge information offset is calculated under normalization processing, and the image gradient in the power operation scene video is extracted; the information gradient is regularized, and the regularized bounding box corresponding to the video image is designed. The type probability score of the risk target in the video image is determined according to the regularized gradient, and the classification of the risk differentiated video of the power operation scene and the identification of the risk target are realized. The maximum error value of the number of risk target identification is 1, and the minimum value of the probability of correct identification of the risk target is 0.989, which can accurately realize the classification and identification of the risk differentiated video of the power operation scene.

[0104] That is, the above-mentioned technical solution provided by the embodiment of the present invention comprehensively considers various factors, adopts appropriate deep learning prediction models, and advanced image processing methods, which can accurately realize the classification of risk-differentiated videos of power operation scenarios and the identification of risk targets, timely discover and warn of potential safety risks, and reduce accidents and losses in power operations.

[0105] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0106] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods of each embodiment of the present application.

[0107] According to an embodiment of the present invention, there is also provided a risk target identification device for electric power operations for implementing the risk target identification method for electric power operations. Figure 8 is a schematic diagram of a risk target identification device for power operation according to an embodiment of the present invention. Figure 8 As shown, the device includes: a first prediction module 801, a gradient extraction module 803, a regularization processing module 805, a second prediction module 807 and a first determination module 809. The risk target identification device for power operation is described below.

[0108] The first prediction module 801 is used to predict the video type of the differentiated video of the power operation scene using the first prediction model, and obtain multiple first-type conditional probabilities of the differentiated video, wherein the first prediction model is obtained by training with multiple groups of first training data in a deep learning manner, and each of the multiple groups of first training data includes: a sample differentiated video, and a sample video type corresponding to the sample differentiated video. The differentiated video refers to different video data of the power operation scene with different conditional characteristics. The first-type conditional probability refers to the probability that the differentiated video belongs to any video type, and the sum of the probabilities of multiple first-type conditional probabilities is 1.

[0109] The gradient extraction module 803 is used to perform gradient extraction on the video frame image of the differentiated video based on the first type of conditional probability to obtain the video image information gradient of the differentiated video, wherein the video image information gradient is used to reflect the feature change of the video frame image.

[0110] The regularization processing module 805 is used to perform regularization processing on the video image information gradient of the differential video to calibrate the risk target in the differential video, wherein the risk target refers to a specific target that needs to be identified and classified in the differential video.

[0111] The second prediction module 807 is used to use the second prediction model to predict the risk type to which the risk target marked in the differentiated video belongs, and obtain multiple second-type conditional probabilities of the risk target, wherein the second prediction is obtained by training with multiple groups of second training data in a deep learning manner, and each of the multiple groups of second training data includes: a sample risk target, a sample risk type corresponding to the sample risk target, and the second-type conditional probability refers to the probability that the risk target belongs to any risk type.

[0112] The first determination module 809 is used to determine that the risk type corresponding to the largest second type conditional probability among multiple second type conditional probabilities is the actual risk type of the risk target.

[0113] It should be noted here that the above-mentioned first prediction module 801, gradient extraction module 803, regularization processing module 805, second prediction module 807 and first determination module 809 correspond to steps S202 to S210 in the above-mentioned embodiment, and the five modules are the same as the instances and application scenarios implemented by the corresponding steps, but are not limited to the contents disclosed in the above-mentioned embodiment.

[0114] As can be seen from the above, in the scheme recorded in the above-mentioned embodiment of the present invention, firstly, the first prediction module can be used to use the first prediction model to predict the video type to which the differentiated video of the electric power operation scene belongs, and obtain multiple first-type conditional probabilities of the differentiated video, wherein the first prediction model is obtained by training using multiple groups of first training data in a deep learning manner, and each of the multiple groups of first training data includes: a sample differentiated video, a sample video type corresponding to the sample differentiated video, and the differentiated video refers to different video data of the electric power operation scene with different conditional characteristics. The first-type conditional probability refers to the probability that the differentiated video belongs to any video type, and the sum of the probabilities of multiple first-type conditional probabilities is 1; then, the gradient extraction module can be used to perform gradient extraction on the video frame image of the differentiated video based on the first-type conditional probability to obtain the video image information gradient of the differentiated video, wherein the video image information gradient is used to reflect the feature changes of the video frame image; then, the regularization processing module can be used to perform regularization processing on the video image information gradient of the differentiated video to calibrate the risk target in the differentiated video, wherein the risk target refers to the differentiated video. a specific target that needs to be identified and classified in the video; secondly, the second prediction module can use the second prediction model to predict the risk type to which the risk target marked in the differentiated video belongs, and obtain multiple second type conditional probabilities of the risk target, wherein the second prediction is obtained by training with multiple groups of second training data in a deep learning manner, and each of the multiple groups of second training data includes: a sample risk target, a sample risk type corresponding to the sample risk target, and the second type conditional probability refers to the probability that the risk target belongs to any risk type; finally, the first determination module can be used to determine that the risk type corresponding to the largest second type conditional probability among multiple second type conditional probabilities is the actual risk type of the risk target, so as to achieve conditional probability calculation through graph convolutional neural network, capture blurred edges in video images, extract image gradients by normalization processing, and at the same time, regularize the information gradient, and determine the type probability score of the risk target in the video image according to the regularized gradient, thereby achieving the technical effect of accurately classifying and identifying risk targets in differentiated videos of power operation scene risks, and improving the safety and efficiency of power operation scenes.

[0115] The technical solution provided by the embodiment of the present invention solves the technical problem in the related art that the accuracy rate is low when classifying and identifying risk targets in power operation scenarios, which easily leads to safety hazards.

[0116] In an optional embodiment, the gradient extraction module includes: a preprocessing unit, which is used to arbitrarily select multiple frames of video frame images from the differentiated video and preprocess the multiple frames of video frame images; a calculation unit, which is used to calculate the gradient values ​​of the multiple frames of video frame images using an image gradient calculation method, wherein the gradient value is used to reflect the change of pixel intensity in the video frame image; and a fusion unit, which is used to fuse the gradient values ​​of the multiple frames of video frame images according to the time series of the multiple frames of video frame images to obtain the video image information gradient of the differentiated video.

[0117] In an optional embodiment, the risk target identification method for power operations also includes: a second determination module, which is used to perform gradient extraction on the video frame image of the differentiated video based on the first type of conditional probability, and then determine that the video frame image with a gradient value lower than the gradient threshold is a blurred video frame image according to the video image information gradient after obtaining the video image information gradient of the differentiated video, wherein the blurred video frame image is an image in which the risk target has blurred edges in the video frame image; a third determination module, which is used to determine the center processing frame according to the blur center and size of the blurred area in the blurred video frame image, wherein the blurred area refers to the area in the video frame image where the edge of the risk target is blurred, and the center processing frame is used to perform image enhancement processing on the blurred area; an acquisition module, which is used to obtain the offset between the blurred center and the actual center, wherein the actual center refers to the central position of the risk target in a clear state, and the clear state refers to the state where the blurriness of the video frame image is lower than the blurriness threshold, wherein the offset is calculated by normalization; an image enhancement module, which is used to perform image enhancement operation on the blurred area using the center processing frame according to the offset, so that the blurred video frame image reaches a clear state.

[0118] In an optional embodiment, the acquisition module includes: a first determination unit, configured to determine the offset using a first formula according to the coordinate values ​​of the blur center and the actual center, wherein the first formula is: (h, j) represents the coordinates of the pixel in the video frame image, h represents the row coordinate of the pixel, j represents the column coordinate of the pixel, Δh represents the offset in the row direction, j represents the offset in the column direction, m h represents the coordinates of the blur center in the row direction, m j represents the coordinates of the blur center in the column direction, P h Represents the coordinates of the actual center in the row direction, R j represents the coordinate of the actual center in the column direction, and ζ represents the adjustment coefficient.

[0119] In an optional embodiment, the image enhancement module includes: an adjustment unit, used to adjust the position of the central processing box according to an offset to obtain an adjustment result; an image enhancement unit, used to perform an image enhancement operation on the blurred area when the adjustment result indicates that the center of the central processing box is aligned with the blurred center, so as to make the blurred video frame image clear.

[0120] In an optional embodiment, the regularization processing module includes: a regularization processing unit, which is used to perform regularization processing on the video image information gradient of the differentiated video to obtain a regularized gradient; a first acquisition unit, which is used to obtain a central processing box in the process of image enhancement processing on the differentiated video; a second acquisition unit, which is used to obtain a gradient offset value in the video image information gradient, wherein the gradient offset value corresponds to the offset between the blur center of the blurred area in the video frame image and the actual center of the risk target; a boundary division unit, which is used to perform boundary division on the central processing box according to the gradient offset value and the regularization gradient to obtain a regularized boundary box, so as to calibrate the risk target in the differentiated video using the regularized boundary box.

[0121] In an optional embodiment, the second prediction module includes: a second determination unit, used to determine the second type conditional probability that the risk target in the video frame image belongs to any risk type by using a second formula, to obtain multiple second type conditional probabilities, wherein the second formula is: N represents the second type conditional probability that the risk target belongs to any risk type, X represents any one of multiple risk targets, B represents any one of multiple risk types, C(X) represents the first probability that the risk target in the video frame image is X, C(X|B) represents the second probability that the risk type to which the risk target belongs is B, and V represents the filter coefficient.

[0122] According to another aspect of an embodiment of the present invention, a risk target identification system for electric power operations is further provided. The risk target identification system for electric power operations uses any one of the above-mentioned risk target identification methods for electric power operations.

[0123] According to another aspect of an embodiment of the present invention, a processor is further provided, and the processor is used to run a program, wherein when the program is run, any one of the above-mentioned risk target identification methods for power operations is executed.

[0124] According to another aspect of an embodiment of the present invention, a computer program product is provided, comprising computer instructions, which, when executed by a processor, execute any one of the above-mentioned risk target identification methods for power operations.

[0125] According to another aspect of an embodiment of the present invention, a computer-readable storage medium is further provided, the computer-readable storage medium comprising a stored program, wherein the program executes any one of the above-mentioned risk target identification methods for power operations.

[0126] Optionally, in this embodiment, the computer-readable storage medium may be located in any one of the computer terminals in a computer terminal group in a computer network, or in any one of the communication devices in a communication device group.

[0127] Optionally, in this embodiment, the computer-readable storage medium is configured to store program code for executing the following steps: using a first prediction model to predict the video type to which the differentiated video of the electric power operation scene belongs, and obtaining multiple first-type conditional probabilities of the differentiated video, wherein the first prediction model is obtained by training using multiple groups of first training data in a deep learning manner, and each of the multiple groups of first training data includes: a sample differentiated video, a sample video type corresponding to the sample differentiated video, the differentiated video refers to different video data of the electric power operation scene with different conditional characteristics, the first-type conditional probability refers to the probability that the differentiated video belongs to any video type, and the sum of the probabilities of multiple first-type conditional probabilities is 1; based on the first-type conditional probability, a gradient extraction is performed on the video frame image of the differentiated video to obtain a video image information gradient of the differentiated video, wherein, The video image information gradient is used to reflect the feature changes of the video frame image; the video image information gradient of the differentiated video is regularized to calibrate the risk target in the differentiated video, wherein the risk target refers to the specific target that needs to be identified and classified in the differentiated video; the second prediction model is used to predict the risk type to which the risk target calibrated in the differentiated video belongs, and multiple second type conditional probabilities of the risk target are obtained, wherein the second prediction is obtained by training with multiple groups of second training data in a deep learning manner, and each of the multiple groups of second training data includes: a sample risk target, a sample risk type corresponding to the sample risk target, and the second type conditional probability refers to the probability that the risk target belongs to any risk type; the risk type corresponding to the largest second type conditional probability among the multiple second type conditional probabilities is determined as the actual risk type of the risk target.

[0128] Optionally, in this embodiment, the computer-readable storage medium is configured to store program codes for executing the following steps: arbitrarily selecting multiple video frame images from the differentiated video, and preprocessing the multiple video frame images; using an image gradient calculation method to calculate gradient values ​​of the multiple video frame images, wherein the gradient values ​​are used to reflect changes in pixel intensity in the video frame images; and fusing the gradient values ​​of the multiple video frame images according to a time series of the multiple video frame images to obtain a video image information gradient of the differentiated video.

[0129] Optionally, in this embodiment, a computer-readable storage medium is configured to store program codes for executing the following steps: determining, based on a gradient of video image information, that a video frame image whose gradient value is lower than a gradient threshold is a blurred video frame image, wherein the blurred video frame image is an image in which the risk target in the video frame image has blurred edges; determining a center processing box based on a blur center and a size of a blurred area in the blurred video frame image, wherein the blurred area refers to an area in the video frame image where the edge of the risk target is blurred, and the center processing box is used to perform image enhancement processing on the blurred area; obtaining an offset between the blurred center and the actual center, wherein the actual center refers to the center position of the risk target in a clear state, and the clear state refers to a state in which the blurriness of the video frame image is lower than a blurriness threshold, wherein the offset is calculated by normalization; performing an image enhancement operation on the blurred area using the center processing box according to the offset to make the blurred video frame image reach a clear state.

[0130] Optionally, in this embodiment, the computer-readable storage medium is configured to store a program code for executing the following steps: determining the offset using a first formula according to the coordinate values ​​of the blur center and the actual center, wherein the first formula is: (h, j) represents the coordinates of the pixel in the video frame image, h represents the row coordinate of the pixel, j represents the column coordinate of the pixel, Δh represents the offset in the row direction, j represents the offset in the column direction, m h represents the coordinates of the blur center in the row direction, m j represents the coordinates of the blur center in the column direction, R h Represents the coordinates of the actual center in the row direction, R j represents the coordinate of the actual center in the column direction, and ζ represents the adjustment coefficient.

[0131] Optionally, in this embodiment, the computer-readable storage medium is configured to store program codes for executing the following steps: adjusting the position of the central processing box according to the offset to obtain an adjustment result; when the adjustment result indicates that the center of the central processing box is aligned with the blur center, performing an image enhancement operation on the blurred area to make the blurred video frame image clear.

[0132] Optionally, in this embodiment, the computer-readable storage medium is configured to store program code for executing the following steps: regularizing the video image information gradient of the differentiated video to obtain a regularized gradient; obtaining a central processing box in the image enhancement processing of the differentiated video; obtaining a gradient offset value in the video image information gradient, wherein the gradient offset value corresponds to the offset between the blur center of the blurred area in the video frame image and the actual center of the risk target; dividing the boundary of the central processing box according to the gradient offset value and the regularized gradient to obtain a regularized boundary box, so as to calibrate the risk target in the differentiated video using the regularized boundary box.

[0133] Optionally, in this embodiment, the computer-readable storage medium is configured to store a program code for executing the following steps: determining a second type conditional probability that a risk target in a video frame image belongs to any risk type by using a second formula to obtain a plurality of second type conditional probabilities, wherein the second formula is: N represents the second type conditional probability that the risk target belongs to any risk type, X represents any one of multiple risk targets, B represents any one of multiple risk types, C(X) represents the first probability that the risk target in the video frame image is X, C(X|B) represents the second probability that the risk type to which the risk target belongs is B, and V represents the filter coefficient.

[0134] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0135] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0136] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0137] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0138] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0139] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-0nly Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk, etc., various media that can store program codes.

[0140] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for identifying risk targets in power operations, characterized in that: include: A first prediction model is used to predict the video type to which the differentiated video of the electric power operation scene belongs, and a plurality of first-type conditional probabilities of the differentiated video are obtained, wherein the first prediction model is obtained by training with multiple groups of first training data in a deep learning manner, and each of the multiple groups of first training data includes: a sample differentiated video, and a sample video type corresponding to the sample differentiated video, the differentiated video refers to different video data of the electric power operation scene with different conditional features, the first-type conditional probability refers to the probability that the differentiated video belongs to any of the video types, and the sum of the probabilities of the multiple first-type conditional probabilities is 1; Performing gradient extraction on the video frame image of the differentiated video based on the first type of conditional probability to obtain a video image information gradient of the differentiated video, wherein the video image information gradient is used to reflect feature changes of the video frame image; Regularizing the video image information gradient of the differentiated video to calibrate risk targets in the differentiated video, wherein the risk targets refer to specific targets in the differentiated video that need to be identified and classified; The risk type to which the risk target marked in the differentiated video belongs is predicted using a second prediction model to obtain a plurality of second-type conditional probabilities of the risk target, wherein the second prediction model is obtained by training with a plurality of sets of second training data in a deep learning manner, and each of the plurality of sets of second training data includes: a sample risk target and a sample risk type corresponding to the sample risk target, and the second-type conditional probability refers to the probability that the risk target belongs to any of the risk types; Determine that the risk type corresponding to the largest second-type conditional probability among multiple second-type conditional probabilities is the actual risk type of the risk target.

2. The method for identifying risk targets of power operations according to claim 1, characterized in that: Performing gradient extraction on the video frame image of the differentiated video based on the first type conditional probability to obtain the video image information gradient of the differentiated video includes: arbitrarily selecting a plurality of video frame images from the differential video, and preprocessing the plurality of video frame images; Calculating the gradient values ​​of multiple frames of the video frame images using an image gradient calculation method, wherein the gradient values ​​are used to reflect the change of pixel intensity in the video frame images; The gradient values ​​of multiple frames of the video frame images are fused according to the time sequence of the multiple video frame images to obtain the video image information gradient of the differential video.

3. The method for identifying risk targets of power operations according to claim 1, characterized in that: After performing gradient extraction on the video frame image of the differentiated video based on the first type conditional probability to obtain the video image information gradient of the differentiated video, the method further includes: Determining, according to the video image information gradient, that the video frame image whose gradient value is lower than the gradient threshold is a blurred video frame image, wherein the blurred video frame image is an image in which the risk target in the video frame image has blurred edges; Determine a central processing frame according to the blur center and size of a blur area in the blurry video frame image, wherein the blur area refers to an area in the video frame image where the edge of the risk target is blurred, and the central processing frame is used to perform image enhancement processing on the blur area; Obtaining an offset between the blur center and the actual center, wherein the actual center refers to the central position of the risk target in a clear state, and the clear state refers to a state in which the blurriness of the video frame image is lower than a blurriness threshold, wherein the offset is calculated by normalization; The image enhancement operation is performed on the blurred area using the central processing frame according to the offset, so as to make the blurred video frame image reach the clear state.

4. The method for identifying risk targets of power operations according to claim 3 is characterized in that: Obtaining the offset between the blur center and the actual center, including: The offset is determined by using a first formula according to the coordinate values ​​of the blur center and the actual center, wherein the first formula is: (h, j) represents the coordinates of the pixel point in the video frame image, h represents the row coordinates of the pixel point, j represents the column coordinates of the pixel point, Δh represents the offset in the row direction, j represents the offset in the column direction, mh represents the coordinates of the blur center in the row direction, m j represents the coordinates of the blur center in the column direction, R h represents the coordinates of the actual center in the row direction, R j represents the coordinates of the actual center in the column direction, and ζ represents the adjustment coefficient.

5. The method for identifying risk targets of power operations according to claim 3, characterized in that: Performing an image enhancement operation on the blurred area using the central processing frame according to the offset so as to make the blurred video frame image reach the clear state includes: Adjusting the position of the central processing frame according to the offset to obtain an adjustment result; When the adjustment result indicates that the center of the central processing frame is aligned with the blur center, the image enhancement operation is performed on the blurred area to make the blurred video frame image reach the clear state.

6. The method for identifying risk targets of power operations according to claim 1, characterized in that: Regularizing the video image information gradient of the differentiated video to calibrate the risk target in the differentiated video includes: Regularizing the video image information gradient of the differential video to obtain a regularized gradient; Acquire a central processing frame in the process of performing image enhancement processing on the differentiated video; Acquire a gradient offset value in the video image information gradient, wherein the gradient offset value corresponds to an offset between a blur center of a blur region in the video frame image and an actual center of the risk target; The central processing box is divided into boundaries according to the gradient offset value and the regularized gradient to obtain a regularized bounding box, so as to calibrate the risk target in the differential video using the regularized bounding box.

7. The method for identifying risk targets of power operations according to claim 1, characterized in that: The risk type to which the risk target marked in the differentiated video belongs is predicted using the second prediction model to obtain a plurality of second type conditional probabilities of the risk target, including: The second type conditional probability that the risk target in the video frame image belongs to any of the risk types is determined by using a second formula to obtain a plurality of second type conditional probabilities, wherein the second formula is: N represents the second type conditional probability that the risk target belongs to any of the risk types, X represents any one of the multiple risk targets, B represents any one of the multiple risk types, C(X) represents the first probability that the risk target in the video frame image is X, C(X|B) represents the second probability that the risk type to which the risk target belongs is B, and V represents the filter coefficient.

8. A risk target identification device for power operations, characterized in that: include: A first prediction module is used to predict the video type to which the differentiated video of the electric power operation scene belongs by using a first prediction model, and obtain multiple first-type conditional probabilities of the differentiated video, wherein the first prediction model is obtained by training with multiple groups of first training data in a deep learning manner, and each of the multiple groups of first training data includes: a sample differentiated video, and a sample video type corresponding to the sample differentiated video, the differentiated video refers to different video data of the electric power operation scene with different conditional features, the first-type conditional probability refers to the probability that the differentiated video belongs to any of the video types, and the sum of the probabilities of multiple first-type conditional probabilities is 1; a gradient extraction module, configured to perform gradient extraction on the video frame image of the differentiated video based on the first type of conditional probability, to obtain a video image information gradient of the differentiated video, wherein the video image information gradient is used to reflect a feature change of the video frame image; A regularization processing module, used for performing regularization processing on the video image information gradient of the differentiated video to calibrate the risk target in the differentiated video, wherein the risk target refers to a specific target that needs to be identified and classified in the differentiated video; A second prediction module is used to predict the risk type to which the risk target marked in the differentiated video belongs by using a second prediction model, and obtain multiple second-type conditional probabilities of the risk target, wherein the second prediction is obtained by training with multiple sets of second training data in a deep learning manner, and each of the multiple sets of second training data includes: a sample risk target and a sample risk type corresponding to the sample risk target, and the second-type conditional probability refers to the probability that the risk target belongs to any of the risk types; The first determination module is used to determine that the risk type corresponding to the largest second-type conditional probability among multiple second-type conditional probabilities is the actual risk type of the risk target.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein the program executes the risk target identification method for power operations according to any one of claims 1 to 7.

10. A computer program product comprising computer instructions, characterized in that When the computer instructions are executed by a processor, the risk target identification method for power operations described in any one of claims 1 to 7 is performed.

Images