Automatic auditing method for medium-voltage power distribution safety supervision operation label image

Abstract

The invention is suitable for the technical field of safety supervision operation, and particularly relates to an automatic auditing method for a medium-voltage power distribution safety supervision operation label image. Compared with the prior art, the method comprises the following steps: acquiring a safety supervision operation label image; establishing a residual learning network model, and performing feature extraction by taking the safety supervision operation label image as the input of the residual learning network model to obtain a trunk feature; establishing a label feature module, and performing feature extraction by taking the safety supervision operation label image as the input of the label feature module to obtain potential features; and the label feature module compares the trunk features with the potential features, classifies and audits the safety supervision operation label images, and obtains an auditing result. According to the method, automatic auditing of label images for medium-voltage power distribution safety supervision operation is realized, potential safety hazards existing in the operation are found in time, data support and decision reference are provided, managers are helped to optimize the operation process, and the operation safety and efficiency are improved.

Description

Technical Field

[0001] The present invention is applicable to the technical field of safety supervision operations, and in particular relates to an automatic review method for medium-voltage power distribution safety supervision operation label images. Background Art

[0002] Medium voltage power distribution mainly refers to the power network with a voltage level of 10kV in the power grid, which mainly includes equipment such as overhead power distribution lines (cables), and is an important bridge connecting the main grid and load users. Safety and supervision operations refer to the use of advanced technical means such as Internet of Things technology and sensor technology to monitor and manage the operation site in industrial production in real time to ensure operation safety, improve work efficiency, and reduce management costs. Through the safety and supervision operation system, the location, behavior, environment, etc. of the operators can be monitored in real time, and the safety hazards in the operation can be discovered in time, providing data support and decision-making references to help managers optimize the operation process and improve operation safety and efficiency. The medium voltage power distribution safety and supervision operation image mainly refers to the use of drones to shoot the safety operation video images of the on-site staff repairing the 10kv medium voltage distribution cable in real time, and the recorded video images are framed for data collection, and then the labeled image data is obtained through data cleaning and data labeling. These manually or automatically labeled image data will have some problems, such as wrong labeling, missing labeling, etc. The problematic images are unqualified and need to be re-labeled until they are correctly labeled. However, in order to improve the quality of data annotation, how to quickly and accurately determine whether there are unqualified images in a group of annotated images is an urgent problem to be solved.

[0003] At present, the review method for annotated medium-voltage power distribution safety supervision operation images is manual review. Manual review is to use the crowdsourcing model, and multiple people separately review the labeled images of different scenes under the medium-voltage power distribution task. For example, one person is responsible for reviewing the labeled image of whether a helmet is worn, and one person is responsible for reviewing the labeled image of whether the medium-voltage distribution line is defective. Although the crowdsourcing model has advantages such as high accuracy and collective wisdom, it also has some disadvantages. These include difficulty in managing quality, difficulty in protecting intellectual property rights and data security, possible moral risks and low remuneration issues, difficulty in ensuring the professional level and reliability of participants, increased labor costs and low efficiency, etc. These shortcomings may affect the quality and completion of the review, and need to be considered and resolved when implementing the crowdsourcing model.

[0004] Therefore, there is an urgent need for a new automatic review method for medium-voltage distribution safety supervision operation label images to solve the above technical problems. Summary of the invention

[0005] The present invention proposes an automatic review method for label images of medium-voltage power distribution safety and supervision operations, aiming to conduct efficient and high-precision automatic review of label images in the field of safety and supervision operations and establish a comprehensive label data quality assessment system.

[0006] The automatic audit method comprises the following steps:

[0007] S1. Obtain the safety supervision operation label image;

[0008] S2. Establish a residual learning network model, use the safety supervision operation label image as the input of the residual learning network model to perform feature extraction, and obtain backbone features;

[0009] S3, establishing a label feature module, taking the safety supervision operation label image as input of the label feature module to perform feature extraction and obtain potential features;

[0010] S4. The label feature module classifies and reviews the safety supervision operation label image by comparing the main feature with the potential feature to obtain a review result.

[0011] Preferably, the residual learning network model includes a convolutional layer, a pooling layer and a channel attention module.

[0012] Preferably, the convolution layer includes four convolution blocks, each of which includes the channel attention module, and when the four convolution blocks perform feature extraction on the safety inspection operation label image, the number of cycles are 3, 4, 6 and 3 respectively.

[0013] Preferably, step S2 includes the following sub-steps:

[0014] Extract features of the safety inspection operation label image through the convolution layer, and perform global average pooling, one-dimensional convolution, and Sigmoid activation on the extracted features through the channel attention module to obtain a first feature;

[0015] The first features are averaged and batch normalized to obtain the backbone features.

[0016] Preferably, the label feature module includes a label feature pool, a label extractor, a global average pooling, a convolutional layer and an activation function.

[0017] Preferably, step S3 includes the following sub-steps:

[0018] The safety inspection operation label image is subjected to feature label extraction by the label extractor to obtain a first label feature;

[0019] Mapping the feature label with a specified label feature in the label feature pool to obtain a second label feature;

[0020] The second label feature is subjected to convolution processing, activation processing and average pooling processing respectively to obtain the potential feature.

[0021] Preferably, the residual learning network model is based on a ResNet-50 network model.

[0022] Compared with the prior art, the present invention obtains safety supervision operation label images; establishes a residual learning network model, uses the safety supervision operation label images as the input of the residual learning network model for feature extraction, and obtains the main features; establishes a label feature module, uses the safety supervision operation label images as the input of the label feature module for feature extraction, and obtains potential features; the label feature module compares the main features with the potential features, classifies and reviews the safety supervision operation label images, and obtains the review results. Automatic review of safety supervision operation label images for medium-voltage power distribution is achieved, safety hazards in the operation are discovered in a timely manner, data support and decision-making references are provided, and managers are helped to optimize the operation process and improve operation safety and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The present invention will be described in detail below in conjunction with the accompanying drawings. The above and other aspects of the present invention will become clearer and easier to understand through the detailed description made in conjunction with the following drawings. In the accompanying drawings:

[0024] Figure 1 It is a flowchart of an automatic review method for medium-voltage power distribution safety supervision operation label images provided by an embodiment of the present invention;

[0025] Figure 2 It is a schematic diagram of the flow structure of an automatic review method for medium-voltage power distribution safety supervision operation label images provided by an embodiment of the present invention;

[0026] Figure 3 It is a flow chart of a label feature module of an automatic review method for label images of medium-voltage power distribution safety supervision operations provided by an embodiment of the present invention;

[0027] Figure 4 It is a schematic diagram of the framework of the channel attention module of the automatic review method for medium-voltage power distribution safety supervision operation label images provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0029] Please refer to Figure 1-Figure 4 , the present invention provides an automatic review method for medium-voltage distribution safety supervision operation label images, and the automatic review method includes the following steps:

[0030] S1. Obtain the safety supervision operation label image.

[0031] S2. Establish a residual learning network model, take the safety supervision operation label image as the input of the residual learning network model for feature extraction, and obtain backbone features.

[0032] In the embodiment of the present invention, the residual learning network model includes a convolutional layer, a pooling layer, and a channel attention module (Efficient Channel Attention Module, ECA). Specifically, the ECA attention module improves the performance of the CNN model by highlighting valuable information in the feature map and suppressing irrelevant features. The mechanism of the ECA module is as Figure 4 shown,[[]]END]] Figure 4 is a schematic framework diagram of the channel attention module of the automatic review method for medium-voltage distribution safety supervision operation label images provided by the embodiment of the present invention. Each feature map F∈R (W×H×C) is first decomposed into a 1×1-dimensional space to obtain a feature F avg ∈R (1×1×C) , and then the input global average pooling (GAP) is used:

[0033]

[0034] The ECA mechanism solves the computational overhead problem of cross-channel interaction in SENet by introducing local cross-channel interaction (local CCI). Local CCI can achieve computational overhead at a relatively low cost while allowing each channel in a small local group to be interdependent with each other channel. First, the global parameter space described by C×C is decomposed into smaller local spaces defined by k×C, where k is the size of the predefined local region, and k < C. Therefore, the local attention-based CCI can be expressed as:

[0035]

[0036] where are k adjacent channels in

[0037]

[0038] In addition, the above-mentioned shared local cross-channel interactions are implemented by a one-dimensional convolution kernel with k layers. In this way, ECA can be expressed as:

[0039]

[0040] In an embodiment of the present invention, the convolution layer includes four convolution blocks, each of which includes the channel attention module. When the four convolution blocks perform feature extraction on the safety inspection operation label image, the number of cycles is 3, 4, 6 and 3 respectively. Taking the image of the safety inspection operation label as input, the size is 1×112×112, and the feature map sizes after processing by the four convolution blocks are 64×56×56, 128×28×28, 256×14×14 and 512×7×7 respectively. The channel attention module performs two convolutions on each residual block to enhance key features while suppressing irrelevant features.

[0041] In this embodiment of the present invention, step S2 includes the following sub-steps:

[0042] The safety inspection operation label image is subjected to feature extraction through the convolution layer, and the extracted features are subjected to global average pooling processing, one-dimensional convolution processing, and Sigmoid activation processing through the channel attention module to obtain the first feature; specifically, the channel attention module includes a series of global average pooling, 1×1 convolution and Sigmoid activation functions, and its task is to calculate the attention weight and refine the feature map to form, and this operation is repeated until the last convolution block.

[0043] The first features are averaged and batch normalized to obtain the backbone features.

[0044] In an embodiment of the present invention, the residual learning network model is based on the ResNet-50 network model. Specifically, feature extraction is an important process in image recognition, and its purpose is to extract key features from the input image. In order to improve the prediction accuracy and generalization ability, an improved CNN based on ECA and ResNet-50 is used as a feature extractor to generate 512 feature embedding vectors for classification and metric learning. The ECA module is introduced in ResNet-50, which enables the model to learn channel information in the feature map in an adaptive and efficient manner without computational overhead.

[0045] The residual learning network model solves the gradient vanishing problem or the degradation problem that repeated multiplication in back-propagation may lead to infinitesimal gradients. After the residual block, an average pooling operation is applied and the 512-dimensional feature embedding is fed into the binary head and edge head respectively for multi-task training.

[0046] Specifically, the present invention constructs a residual deep learning network of the convolutional layer by using label data such as helmet wearing, work clothes wearing, and insulating gloves wearing in the field of safety and supervision operations. Using the residual deep learning network of safety and supervision operations, the label images in the field of safety and supervision operations are efficiently and accurately automatically reviewed, a comprehensive label data quality assessment system is established, and a systematic quality check of the accuracy and compliance of the label data is achieved.

[0047] S3. Establish a label feature module, use the safety supervision operation label image as input of the label feature module to perform feature extraction, and obtain potential features.

[0048] In an embodiment of the present invention, the label feature module includes a label feature pool, a label extractor, a global average pooling, a convolutional layer and an activation function, which is mainly used to extract the label field of the input image and then obtain the known features in the safety and supervision operation field.

[0049] In the embodiment of the present invention, Figure 3 As shown, Figure 3 1 is a flow chart of a label feature module of an automatic review method for label images of medium-voltage power distribution safety supervision operations provided by an embodiment of the present invention. In the figure, σ is a sigmoid activation function, and step S3 includes the following sub-steps:

[0050] The safety inspection operation label image is subjected to feature label extraction by the label extractor to obtain a first label feature; if the label extractor fails to extract the image label, the image will be classified as "missing label".

[0051] Mapping the feature label with a specified label feature in the label feature pool to obtain a second label feature;

[0052] The second label feature is subjected to convolution processing, activation processing and average pooling processing respectively to obtain the potential feature.

[0053] S4. The label feature module classifies and reviews the safety supervision operation label image by comparing the main feature with the potential feature to obtain a review result.

[0054] Compared with the prior art, the present invention obtains safety supervision operation label images; establishes a residual learning network model, uses the safety supervision operation label images as the input of the residual learning network model for feature extraction, and obtains the main features; establishes a label feature module, uses the safety supervision operation label images as the input of the label feature module for feature extraction, and obtains potential features; the label feature module compares the main features with the potential features, classifies and reviews the safety supervision operation label images, and obtains the review results. Automatic review of safety supervision operation label images for medium-voltage power distribution is achieved, safety hazards in the operation are discovered in a timely manner, data support and decision-making references are provided, and managers are helped to optimize the operation process and improve operation safety and efficiency.

[0055] To better reflect the proposed method, the ECA-ResNet50 network algorithm and the residual learning network model after adding the label feature module are experimentally compared, and the training accuracy, verification accuracy, training loss value and verification loss value of the two algorithms are compared and analyzed. The recognition effects of the two algorithms are shown in Table 1.

[0056] Table 1: Comparison between the proposed network model and the existing ECA-ResNet50 network model

[0057]

[0058] As shown in Table 1, the method proposed in the present invention has improved both the training accuracy and the verification accuracy by 0.0562 and 0.0516 respectively, and reduced the training loss and the verification loss by 0.0147 and 0.0150 respectively. The addition of the label feature module better fits the training data set, increases the robustness and accuracy of the algorithm, and improves the recognition accuracy.

[0059] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0060] The embodiments of the present invention are described above in conjunction with the accompanying drawings. What is disclosed is only the preferred embodiment of the present invention. However, the present invention is not limited to the above-mentioned specific implementation manner. The above-mentioned specific implementation manner is only illustrative rather than restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms and equivalent changes without departing from the scope of protection of the purpose of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. An automatic review method for label images of medium voltage power distribution safety supervision operations, characterized in that: The automatic audit method comprises the following steps: S1. Obtain the safety supervision operation label image; S2. Establish a residual learning network model, use the safety supervision operation label image as the input of the residual learning network model to perform feature extraction, and obtain backbone features; S3, establishing a label feature module, taking the safety supervision operation label image as input of the label feature module to perform feature extraction and obtain potential features; S4. The label feature module classifies and reviews the safety supervision operation label image by comparing the main feature with the potential feature to obtain a review result.

2. The automatic review method for medium voltage power distribution safety supervision operation label images according to claim 1 is characterized in that: The residual learning network model includes a convolutional layer, a pooling layer and a channel attention module.

3. The automatic review method for medium voltage power distribution safety supervision operation label images according to claim 2 is characterized in that: The convolution layer includes four convolution blocks, each of which includes the channel attention module. When the four convolution blocks perform feature extraction on the safety inspection operation label image, the number of cycles are 3, 4, 6 and 3 respectively.

4. The automatic review method for medium voltage power distribution safety supervision operation label images according to claim 3 is characterized in that: Step S2 includes the following sub-steps: Extract features of the safety inspection operation label image through the convolution layer, and perform global average pooling, one-dimensional convolution, and Sigmoid activation on the extracted features through the channel attention module to obtain a first feature; The first features are averaged and batch normalized to obtain the backbone features.

5. The automatic review method for medium voltage power distribution safety supervision operation label image according to claim 1, characterized in that: The label feature module includes a label feature pool, a label extractor, a global average pooling, a convolution layer and an activation function.

6. The automatic review method for medium voltage power distribution safety supervision operation label image according to claim 5, characterized in that: Step S3 includes the following sub-steps: The safety inspection operation label image is subjected to feature label extraction by the label extractor to obtain a first label feature; Mapping the feature label with a specified label feature in the label feature pool to obtain a second label feature; The second label feature is subjected to convolution processing, activation processing and average pooling processing respectively to obtain the potential feature.

7. The automatic review method for medium voltage power distribution safety supervision operation label image according to claim 1, characterized in that: The residual learning network model is based on the ResNet-50 network model.

Images