A power distribution network equipment operation state intelligent detection method based on power grid safety

Abstract

The present application relates to a kind of power distribution network equipment operating state intelligent detection method based on power grid safety, the method includes the following steps: step 1: input the picture of the power distribution station equipment to be detected, construct in convolutional neural network CNN, extract the feature information of picture;Step 2: while linearly capturing spatial dependence after convolution layer, the feature is batch normalization (Batch Normalization, BN), then using funnel activation function FReLU captures complex visual layout, adaptive obtains image local context, so that model has the ability of pixel-level modeling;Step 3: the image feature information extracted by CNN network is transmitted into feature fusion module BiFPN, and effective bidirectional cross-scale connection and weighted feature fusion are carried out, to improve the accuracy of small target;Step 4: finally output detection result and specific location information.

Description

TECHNICAL FIELD

[0001] The present application relates to a detection method, in particular to a power distribution network equipment operation state intelligent detection method based on power grid safety, and belongs to the image processing technical field in the computer vision direction. BACKGROUND

[0002] With the rapid development of the power system, the scale of the power grid is increasingly large, and the number of power distribution stations is also increasing. The power supply reliability of the power grid is related to the national economy and people's livelihood, and the operation state of the power distribution equipment directly determines the stability and safety of the power grid, so timely understanding the operation state of various power distribution equipment is an important guarantee for maintaining the stability and safety of power distribution. Common power distribution equipment includes high-voltage power distribution cabinets, transformers, power lines, circuit breakers, low-voltage switch cabinets, power distribution panels, switch boxes, control boxes, etc. If these equipment fails, it may cause a power outage accident in a local area, or even endanger people's lives and safety. In order to ensure the safe and reliable operation of the power distribution network, it is necessary to regularly detect and troubleshoot the power distribution equipment.

[0003] At present, the potential fault detection means of power equipment is still manual investigation, but due to the complexity of the power distribution network line, the wide distribution of equipment, etc., the workload of maintenance is increased, the manpower consumption is large, the pertinence is poor, there are overhauls or repairs, blind repairs, etc., and the running state of various equipment cannot be understood in real time.

[0004] The present application provides a power distribution equipment operation state detection method based on power grid safety, which can monitor, identify and warn the potential fault risk of the power distribution equipment, and eliminate the hidden danger in time, so as to eliminate the power distribution network accident in the bud. SUMMARY

[0005] The present application is aimed at the problems existing in the prior art, and provides a power distribution network equipment operation state intelligent detection method based on power grid safety. The technical scheme detects the operation state of various power distribution equipment in the power distribution station house, thereby ensuring the safety and stability of the power distribution network. By manually calibrating the operation state pictures of various equipment and combining the improved deep learning algorithm to extract the features of the detection equipment, the operation state of the power distribution equipment in the image to be processed can be accurately determined. There are many types of commonly used power distribution equipment, including complex power lines and small target switch cabinet indicator lights. The method disclosed in the present application mainly solves the following problems: on the one hand, the semantic features of the power distribution equipment are extracted from the complex irrelevant scene, and on the other hand, the operation state of the extracted power distribution equipment needs to be accurately determined.

[0006] In order to achieve the above object, the technical scheme of the present application is as follows: a power distribution network equipment operation state intelligent detection method based on power grid safety, characterized in that the method comprises the following steps:

[0007] Step 1: input the picture of the power distribution station equipment to be detected, construct a convolutional neural network CNN, and extract the feature information of the picture;

[0008] Step 2: after linearly capturing spatial dependence in the convolutional layer, normalize the features (Batch Normalization, BN), then use the funnel activation function FReLU to capture complex visual layouts and adaptively obtain the local context of the image, so that the model has the ability of pixel-level modeling; the structure is as shown in the accompanying Figure Two .

[0009] Step 3: the image feature information extracted by the CNN network is transmitted into the feature fusion module BiFPN for effective bidirectional cross-scale connection and weighted feature fusion to improve the accuracy of small targets; the structure is as shown in the accompanying Figure Three .

[0010] Step 4: finally output the detection result and specific position information.

[0011] A convolutional neural network (CNN) is mainly constructed as a backbone network, after linearly capturing spatial dependence in the convolutional layer, a funnel activation function (Funnel Rectified Linear Unit, FReLU) is added to solve the spatial insensitivity problem in the activation function, so that the regular (ordinary) convolution also has the ability to capture complex visual layouts, and the model has the ability of pixel-level modeling. Then add a feature fusion module, and use a bidirectional feature pyramid network (Bidirectional Feature Pyramid Network, BiFPN) to introduce weights to better balance the feature information of different scales.

[0012] In the backbone network, the funnel activation function is used to adaptively obtain the context information of the picture, and the spatial sensitivity of the convolutional layer is improved; at the same time, from the perspective of feature fusion, the weighted bidirectional feature pyramid network is used to connect the input and output nodes of the same layer across layers, realize higher level fusion, shorten the path of information transmission between upper and lower layers, and can maximize the retention of semantic information representation ability while retaining geometric detail features.

[0013] As an improvement of the present invention, in step 2, the main sub-layers in the Convolutional Neural Network (CNN) are convolutional layers and nonlinear activation layers. Currently, the most widely used activation is still the Rectified Linear Unit (ReLU). However, for image visual recognition tasks, the spatial insensitivity of activation is the main obstacle to achieving significant improvement in visual tasks. Therefore, the present invention adopts a conceptually simple but effective funnel activation, called Funnel Rectified Linear Unit (FReLU), which extends ReLU to 2D activation by adding negative overhead of spatial conditions.

[0014] The main form of ReLU is y = max(x, 0), where the condition depends on the pixel itself, and then max(·) is used to obtain the maximum value between x and the condition. FReLU, on the other hand, has the form y = max(x, T(x)), where T(·) is the 2D spatial condition. Simply put, the spatial condition in FReLU is achieved by comparing the output of the local convolution with the original data using a max comparison. The convolution here is a separable convolution, which essentially performs a non-linear processing on each channel of the layer, followed by normalization of the feature layer, keeping the input and output unchanged. Then, the output feature map is max-processed against the original input x. This solves the spatial insensitivity problem in activation functions, enabling even regular (ordinary) convolutions to capture complex visual layouts and giving the model pixel-level modeling capabilities.

[0015] (1) Funnel space conditions;

[0016] To implement spatial conditions, a parametric pooling window is used to create spatial dependencies. Specifically, the activation function is defined as follows:

[0017] f(x c,i,j ) = max(x c,i,j ,T(x c,i,j ))

[0018]

[0019] Where, x c,i,j The nonlinear activation f(·) is applied to the input pixel at position (i,j) in the c-th channel in two-dimensional space, and the function T(·) represents the funnel condition. Indicates x c,i,j The parameter pool window is centered. This represents the window coefficient shared in the same channel, and (·) represents the dot product.

[0020] (2) pixel-level modeling capability;

[0021] The definition of the spatial condition enables the network to produce the spatial condition in the nonlinear activation of each pixel. The network simultaneously performs nonlinear transformation and produces spatial dependence. Therefore, the pixel-level condition enables the network to have a pixel-level modeling capability, and the function max(·) provides each pixel with a choice of looking at the spatial context or not. Specifically, consider a network with n FReLU layers {F1, F2,... F n}, each FReLU layer F i has a k x k parameter window. For the sake of brevity, only the FReLU layers are analyzed, without considering the convolutional layers. Since the maximum choice is between 1 x 1 and k x k, each pixel after F1 has a set of activation functions {1, 1 + r} (r = k - 1). After F n layer, the set becomes {1, 1 + r, 1 + 2r,..., 1 + nr}, which provides each pixel with more choices, and if n is large enough, it can approximate any layout. With many different sizes of blocks, different sizes of blocks can approximate the shapes of oblique lines and arcs. The layout of objects in an image is usually not a simple horizontal or vertical layout, and they are usually in the shape of an oblique line or a circular arc, so the pixel-level modeling capability provided by the spatial condition can naturally extract the spatial structure of the object. The present application shows that it can better capture irregular and detailed object layouts in complex tasks.

[0022] As an improvement of the present application, in step 3, the feature fusion module, the power distribution equipment is mainly concentrated on the distribution box, such as the press plate switch, the indicator light, etc. The pictures collected in the monitoring video system are all small targets. Due to low resolution and small size, it is difficult to detect tiny objects. The main reason for the poor performance of micro-target detection is the limitation of the network and the imbalance of the training data set.

[0023] The earliest neural network directly predicts from the high-level pyramid features extracted from the backbone, but this structure has no feature fusion, so the precision is relatively low. Then Feature Pyramid Networks (FPN) based on the idea of feature fusion was proposed. The idea of FPN is to establish a top-down path in the network structure for feature fusion, and use the fused feature map with higher semantic information for prediction, which can improve the accuracy to a certain extent.

[0024] Generally, FPN is simply to superimpose or add the feature map, such as using concat or shortcut connection, without distinguishing the feature map added at the same time. However, different input feature maps have different resolutions, and their contributions to the fusion input feature map are also different, so simply adding or superimposing them is not the best operation, and the precision is still limited. The present application adopts a Bidirectional Feature Pyramid Network (BiFPN) to realize the bidirectional fusion of deep and shallow features from top to bottom and from bottom to top, so as to enhance the transmission of feature information between different network layers, and adopts a simple and efficient weighted feature fusion mechanism, which uses the network to learn the weight of different input features, that is, weighted feature fusion, so as to balance the detection of different size targets and can obviously improve the detection effect of small targets and improve the detection precision. In general, the main idea of BiFPN is effective bidirectional cross-scale connection and weighted feature fusion. The specific structure is shown in FIG. Figure 2

[0025] The enhanced feature fusion formula is as follows:

[0026]

[0027] Where O is the output of the fusion feature, w i is a learnable weight, I i is the input feature, w i ≥0 is ensured by adding a Relu activation function after each w i , and a small value of learning rate ε = 0.0001 is set to avoid numerical instability. Similarly, the value of each normalized weight is also between 0 and 1, but since there is no Softmax operation here, the efficiency is much higher.

[0028] Compared with the prior art, the present application has the following advantages: for the identification of device state in power grid safety, a novel detection method is proposed; for the space insensitivity problem of the commonly used activation function in neural network, a funnel activation function is proposed, which can better capture irregular and detailed spatial information and adaptively obtain the local context ability of the image. For multi-class distribution equipment, a weighted bidirectional feature pyramid network is proposed to quickly perform multi-scale feature fusion and improve the accuracy of small targets. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 is a network structure schematic diagram;

[0030] Figure 2 ​A funnel activation structure schematic diagram;

[0031] Figure 3 A bidirectional feature pyramid structure schematic diagram. DETAILED DESCRIPTION

[0032] In order to deepen the understanding of the present application, the present embodiment will be described in detail below with reference to the accompanying drawings.

[0033] Embodiment 1: see Figures 1-3 A power grid safety-based power distribution network equipment operation state intelligent detection method, the method comprising the following steps:

[0034] Step 1: input the picture of the power distribution station equipment to be detected, construct a convolutional neural network CNN, and extract the feature information of the picture;

[0035] Step 2: after linearly capturing the spatial dependence in the convolutional layer, the features are batch normalized (Batch Normalization, BN), and then a funnel activation function FReLU is used to capture complex visual layouts and adaptively obtain the local context of the image, so that the model has the ability of pixel-level modeling;

[0036] Step 3: the image feature information extracted by the CNN network is transmitted to the feature fusion module BiFPN for effective bidirectional cross-scale connection and weighted feature fusion, improving the accuracy of small targets;

[0037] Step 4: finally output the detection result and specific position information.

[0038] Step 1: first, a convolutional neural network (Convolutional Neural Network, CNN) is constructed as a backbone network, and the network layer constructed contains four convolutional layers with sizes of 256x 256, 128x 128, 64x 64, and 32x32, to extract feature information of pictures of different sizes.

[0039] Step 2: after extracting the image features in each convolutional layer in step 1, since the extracted features have a value range of [0, 255], in order to eliminate the unit and scale difference between features, the features are uniformly batch normalized (Batch Normalization, BN) to fix the range between [-1, 1]. After processing, the features are processed by a funnel activation function (Funnel Rectified Linear Unit, FReLU) to solve the problem of spatial insensitivity in the activation function, so that the regular (ordinary) convolution also has the ability to capture complex visual layouts, and the model has the ability of pixel-level modeling. Finally, four layers of features are output.

[0040] Step 3: The four-layer features output in step 2 are used to perform bidirectional fusion of deep and shallow layer features by using a three-layer stacked bidirectional feature pyramid network (BiFPN) structure, by setting a top-down channel between convolutional layers to pass semantic information of high-level features, and a bottom-up channel to pass position information of low-level features, to enhance the transmission of feature information between different network layers, and use the network to learn the weights of different input features to perform differentiated fusion of input features of different layers, take into account detection of targets of different sizes, fully fuse multi-scale features, and improve the detection effect of small targets and improve the detection accuracy.

[0041] Step 4: Output the detection result and specific position information.

[0042] It should be noted that the above embodiments are not intended to limit the protection scope of the present application, and equivalent transformations or substitutions made on the basis of the above technical solutions all fall within the scope of protection of the claims of the present application.

Claims

1. A method for intelligent detection of the operating status of distribution network equipment based on power grid security, characterized in that, The method includes the following steps: Step 1: Input the image of the substation equipment to be detected, and construct a convolutional neural network (CNN) to extract the feature information of the image; Step 2: After linearly capturing spatial dependencies in the convolutional layers, the features are batch normalized (BN), and then the funnel activation function FReLU is used to capture complex visual layouts, adaptively acquiring local image context, enabling the model to perform pixel-level modeling; the funnel activation function is defined as: f(x c,i,j )=max(x c,i,j ,T(x c,i,j )) Where, x c,i,j The nonlinear activation f(·) is applied to the input pixel at position (i,j) in the two-dimensional space of the c-th channel, and the function T(·) represents the funnel condition. Indicates x c,i,j The parameter pool window is centered. This represents the window coefficient shared in the same channel, and (·) represents the dot product; Step 3: Input the image feature information extracted by the CNN network into the feature fusion module BiFPN to perform effective bidirectional cross-scale connection and weighted feature fusion, thereby improving the accuracy of small targets; The weighted feature fusion formula is as follows: Where O is the fusion feature output, w i It is a learnable weight, I i For the input features, w i ≥0 is achieved by each w i This is then ensured by adding a ReLU activation function, with the learning rate ε set to a small value of 0.0001, and the value of each normalized weight also between 0 and 1. Step 4: Finally, output the detection results and specific location information; The specific implementation process of step 1 is as follows: A convolutional neural network (CNN) was constructed as the backbone network. The constructed network layer contains four convolutional layers with sizes of 256x256, 128x128, 64x64, and 32x32, respectively, to extract feature information from images of different sizes. After extracting image features from each convolutional layer in step 1, since the extracted feature values ​​range from [0, 255], in order to eliminate the problem of differences in units and scales between features, the features are uniformly processed by batch normalization (BN) to fix their range between [-1, 1]. Then, the processed features are processed by the funnel activation function FReLU to solve the problem of spatial insensitivity in the activation function, so that regular convolutions can also capture complex visual layouts, and the model can have the ability to model pixels. Finally, four layers of features are output. The specific implementation process of step 2 is as follows: The four-layer features output in step 2 are used in a three-layer stacked bidirectional feature pyramid network (BiFPN) structure. By setting up top-down paths between convolutional layers to transmit semantic information of high-level features and bottom-up paths to transmit positional information of low-level features, bidirectional fusion of deep and shallow features is achieved, enhancing the transmission of feature information between different network layers. Furthermore, the network is used to learn the weights of different input features, enabling discriminative fusion of input features from different layers. This approach takes into account the detection of targets of different sizes, fully integrates multi-scale features, improves the detection effect of small targets, and enhances detection accuracy.

Images