CN-122023837-A - Uranium hexafluoride leakage monitoring method based on machine vision

CN122023837ACN 122023837 ACN122023837 ACN 122023837ACN-122023837-A

Abstract

The invention relates to a uranium hexafluoride leakage monitoring method based on machine vision, which comprises the steps of obtaining an image to be detected containing leakage smoke in a uranium hexafluoride leakage environment, dividing the image to be detected into a smoke image and a background image based on an image division model, carrying out concentration estimation on smoke concentration reflected by the smoke image by adopting a twin neural network model, wherein an input module of the twin neural network model is used for receiving input data formed by the smoke image and a plurality of reference concentration sample images with different concentrations, a feature extraction module comprises two sub-networks which are identical in structure and share weights and are used for respectively extracting feature vectors of the smoke image and the reference concentration sample images, and an output module is used for selecting a sample image with the highest similarity with the smoke image to be detected from the reference concentration sample images by calculating Euclidean distance between the two feature vectors and determining the concentration value of the smoke image. The technical problem that the leakage concentration cannot be accurately obtained in real time in the prior art is solved.

Inventors

SUN SHUTANG
LI GUOQIANG
SUN HONGCHAO
WANG HAOQUAN
YAN JIN
ZHU YEMING
Lian Yiren
RONG YU
ZHANG ZHI
Niu Jiangyu
ZHANG JIANGANG

Assignees

中国辐射防护研究院

Dates

Publication Date: 20260512
Application Date: 20251226

Claims (10)

1. The uranium hexafluoride leakage monitoring method based on machine vision is characterized by comprising the following steps of: obtaining an image to be detected containing leakage smoke in a uranium hexafluoride leakage environment; Identifying a smoke area and a background area in the image to be detected based on an image segmentation model, so as to divide the smoke area and the background area into a smoke image and a background image; adopting a twin neural network model to carry out concentration estimation on the smoke concentration reflected by the smoke map; the twin neural network model comprises an input module, a feature extraction module and an output module; the input module is used for receiving input data composed of the smoke graph and a plurality of reference concentration sample graphs with different concentrations; The characteristic extraction module comprises two subnetworks which are identical in structure and share weight values and are used for respectively extracting characteristic vectors of the smoke graph and the reference concentration sample graph; The output module selects a sample graph with the highest similarity with the smoke graph to be detected from the reference concentration sample graphs by calculating Euclidean distance between the smoke graph feature vector and each reference concentration sample graph feature vector, so as to further determine the concentration value of the smoke graph.
2. The method of claim 1, wherein the obtaining of the plurality of reference concentration sample maps having different concentrations comprises: the method comprises the steps of collecting a large number of clear smoke images and corresponding parameter data through experiments or simulation, wherein the parameter data comprise gray scale, atmospheric illumination intensity and distance parameters from a camera to smoke; The concentration category labeling of the smoke graph sample comprises the following steps: And calculating a concentration value corresponding to the smoke graph sample by using an aerosolization equation, wherein the calculation formula is as follows: , wherein P, B, A, S respectively represents the gray value of a pixel point in a smoke region of the smoke image, the gray value of a pixel point corresponding to a background region, the atmospheric illumination intensity and the distance from a camera to smoke, and m is a dimming coefficient for quantifying the smoke concentration; Mapping the concentration value of each smoke graph sample to a plurality of divided concentration class labels to obtain a data set; And selecting a plurality of images which have different concentrations and can cover all concentration categories from the data set as the reference concentration sample graph.
3. The method of claim 2, wherein the training of the twin neural network model comprises: According to the concentration category labels of the reference concentration sample graphs, respectively matching the smoke graph samples in the data set with each reference concentration sample graph to generate a positive sample combination and a negative sample combination; The sample of the positive sample combination consists of a reference concentration sample graph and a smoke pattern book consistent with the concentration type label thereof, the sample of the negative sample combination consists of a reference concentration sample graph and a smoke pattern sample inconsistent with the concentration type label thereof, and corresponding labels are respectively added to the positive and negative sample combinations to form a training set; and inputting the training set into a twin neural network model, wherein in each training process, the reference concentration sample graph and the smoke graph sample are respectively input into two sub-networks, so that the model learns the difference of the two sub-networks.
4. The method of claim 3, wherein the training of the twin neural network model further comprises: performing iterative computation on characteristic distances between smoke graph samples and reference concentration sample graphs in each round of training set samples, and performing back propagation according to a contrast loss function to complete parameter training of the whole network; the contrast loss function is as follows: , Wherein L is a contrast loss function value, y is a label of a training set sample, and Euclidean distance , 、 The feature vectors of the smoke graph sample and the reference concentration sample graph in the ith training are respectively shown, n is the training round, a is a set similarity threshold, and when d exceeds the threshold a, the loss is 0.
5. The method of claim 1, wherein each sub-network employs a pre-training model, the structure of which comprises ResNet a residual network, and a spatial pyramid pooling layer is arranged before a full-connection layer in the ResNet residual network, so as to ensure that input samples are consistent in size to adapt to smog map input with different sizes.
6. The method of claim 1, wherein the accuracy and efficiency of the twin neural network model is assessed using average absolute error and single frame average time consumption as an evaluation index.
7. The method of claim 6, wherein the calculating of the mean absolute error comprises: , Wherein N represents the number of test samples, For the true label value of the nth test sample i.e. true concentration value, For the n-th test sample concentration value predicted by the network, MAE is used to measure the average error between the predicted concentration value and the true label value.
8. The method of claim 1, wherein the image segmentation model is constructed by training a lightweight Residual U-Net through a training set; The light-weight Residual U-Net is characterized in that double convolution in the U-Net is replaced by MobileNetV reverse Residual blocks, and is matched with depth separable up-sampling to realize the light weight of a network, an input image is firstly processed through a convolution layer and then passes through Residual blocks of different levels to generate a series of feature images, the feature images are processed through a series of decoders of an up-sampling module, after each up-sampling decoder, the feature images are spliced with corresponding feature images of an encoder from a down-sampling path, and finally 1X 1 convolution check pixel-level semantic mapping is carried out.
9. The method of claim 8, wherein CBAM modules are added after each decoder in the decoder portion of the U-Net, the CBAM modules being modules that combine spatial attention mechanisms and channel attention mechanisms.
10. The method of claim 8, wherein the obtaining of the training set comprises: intercepting an image from a uranium hexafluoride leakage site video, marking a smoke area in the image by using Labelme, and manufacturing a dataset; Adjusting the images in the dataset to the same size, and keeping the input image as a single-channel gray scale image; normalizing the image pixel values increases the diversity of the data by data enhancement.

Description

Uranium hexafluoride leakage monitoring method based on machine vision Technical Field The invention relates to the technical field of accident monitoring, in particular to a uranium hexafluoride leakage monitoring method based on machine vision. Background In the prior art, a monitoring scheme for machine vision based on video file analysis is widely applied to various industrial processes, and effective monitoring of the position, state, quality and production safety state of a target to be monitored is realized. However, in the field of gas leakage detection, in particular uranium hexafluoride leakage accidents in nuclear fuel recycling facilities, there is currently a lack of solutions that enable an effective monitoring of the concentration of leakage fumes based on machine vision. The conventional uranium hexafluoride leakage monitoring generally arranges a sensor at a fixed position for local point detection, such as a hydrogen fluoride gas alarm instrument, a radioactive aerosol monitor instrument and other devices commonly used at present, when leakage occurs, the sensor identifies leakage substances to judge leakage, that is, the existing scheme only focuses on the whole state of leakage gas, that is, whether leakage exists or not, and for the leakage degree, that is, the concentration judgment of leakage smoke, a refined evaluation method is lacking. Therefore, the accuracy and the refinement degree of the results of the existing monitoring scheme are difficult to meet the customization requirement of the nuclear emergency scheme. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a uranium hexafluoride leakage monitoring method based on machine vision, which aims at solving the technical problems that the accuracy and the refinement degree of a leakage detection result are low, and the leakage concentration cannot be accurately obtained in real time in the prior art. The technical scheme adopted by the invention is as follows: the invention provides a uranium hexafluoride leakage monitoring method based on machine vision, which comprises the following steps: obtaining an image to be detected containing leakage smoke in a uranium hexafluoride leakage environment; Identifying a smoke area and a background area in the image to be detected based on an image segmentation model, so as to divide the smoke area and the background area into a smoke image and a background image; adopting a twin neural network model to carry out concentration estimation on the smoke concentration reflected by the smoke map; the twin neural network model comprises an input module, a feature extraction module and an output module; the input module is used for receiving input data composed of the smoke graph and a plurality of reference concentration sample graphs with different concentrations; The characteristic extraction module comprises two subnetworks which are identical in structure and share weight values and are used for respectively extracting characteristic vectors of the smoke graph and the reference concentration sample graph; The output module selects a sample graph with the highest similarity with the smoke graph to be detected from the reference concentration sample graphs by calculating Euclidean distance between the smoke graph feature vector and each reference concentration sample graph feature vector, so as to further determine the concentration value of the smoke graph. The preferable technical scheme is as follows: the obtaining of the plurality of reference concentration sample graphs with different concentrations comprises the following steps: the method comprises the steps of collecting a large number of clear smoke images and corresponding parameter data through experiments or simulation, wherein the parameter data comprise gray scale, atmospheric illumination intensity and distance parameters from a camera to smoke; The concentration category labeling of the smoke graph sample comprises the following steps: And calculating a concentration value corresponding to the smoke graph sample by using an aerosolization equation, wherein the calculation formula is as follows: , wherein P, B, A, S respectively represents the gray value of a pixel point in a smoke region of the smoke image, the gray value of a pixel point corresponding to a background region, the atmospheric illumination intensity and the distance from a camera to smoke, and m is a dimming coefficient for quantifying the smoke concentration; Mapping the concentration value of each smoke graph sample to a plurality of divided concentration class labels to obtain a data set; And selecting a plurality of images which have different concentrations and can cover all concentration categories from the data set as the reference concentration sample graph. Training of the twin neural network model, comprising: According to the concentration category labels of the reference concentration sample graphs, respectively matc