CN-121999371-A - Improved YOLOv-based earth-rock dam ERT image leakage feature identification method and system

CN121999371ACN 121999371 ACN121999371 ACN 121999371ACN-121999371-A

Abstract

The invention discloses an improved YOLOv-based earth-rock dam ERT (resistivity tomography) image leakage feature identification method and system, and belongs to the technical field of disaster prevention and reduction of hydraulic engineering. The method comprises the steps of collecting original resistivity tomography data of an earth and rockfill dam, carrying out inversion imaging and standardization processing to construct an initial resistivity tomography image data set, expanding the resistivity tomography image data set by adopting data augmentation operation, marking a leakage area and dividing the data set, constructing an example segmentation neural network model with an improved YOLOv architecture, training by using the data set, inputting an ERT image to be detected into the trained model, outputting a boundary frame, confidence coefficient and segmentation mask of the leakage area, and finally carrying out post-processing and evaluating leakage characteristics by using the improved linear growth model. The invention effectively improves the recognition accuracy and robustness of the leakage characteristics in the ERT image, and realizes the automation and the intellectualization of the leakage detection.

Inventors

ZHENG WENXIAO
He Beihao
JI HAORAN
GAN JIANJUN
LIU DANZHU
ZHANG WEN
DING MEIQI
ZHOU ZAIYONG
QU ZHIHUI
TIAN TAO
XU PAN

Assignees

江西水利电力大学
中铁水利水电规划设计集团有限公司

Dates

Publication Date: 20260508
Application Date: 20260407

Claims (9)

1. An earth and rockfill dam ERT image leakage feature identification method based on improvement YOLOv is characterized by comprising the following steps: S1, acquiring original resistivity tomography data of a earth-rock dam, performing inversion imaging and standardization processing to construct an initial resistivity tomography image data set; s2, accurately labeling leakage areas in the resistivity tomography image by using a labeling tool, uniformly classifying the leakage areas into single categories, and dividing a training set, a verification set and a test set according to a proportion; S3, constructing an example segmentation neural network model of the improved YOLOv architecture; S4, training the example segmentation neural network model by using the marked data set, and optimizing parameters through forward propagation and backward propagation; s5, inputting the resistivity tomography image to be detected into a trained example segmentation neural network model, outputting a boundary box, a category confidence coefficient and a segmentation mask of a leakage area, and completing automatic identification and positioning of leakage characteristics; S6, carrying out post-processing on the identified leakage area, and evaluating leakage characteristics by using the improved linear growth prediction model.
2. The earth-rock dam ERT image leakage feature recognition system based on the improvement YOLOv of claim 1, wherein the example segmentation neural network model of the improved YOLOv architecture comprises a backbone network, a neck network and a head network which are sequentially connected, wherein the neck network is composed of a multi-level channel attention scale sequence module, the head network is composed of a multi-parameter monte carlo attention architecture, a cooperative two-way attention mechanism and a head network output layer which are sequentially connected, and the cooperative two-way attention mechanism introduces multi-level channel attention and leakage risk attention.
3. The earth-rock dam ERT image leakage characteristic recognition system based on the improvement YOLOv as claimed in claim 2, wherein the multi-level channel attention scale sequence module is formed by sequentially connecting a scaling-splicing sub-module, a multi-scale sequence sub-module, a multi-parameter element-by-element adding sub-module and a multi-level channel attention sub-module, wherein a resistivity tomography geological structure constraint weight map is introduced into the scaling-splicing sub-module, and the resistivity tomography geological structure constraint weight map is specifically as follows: ; Wherein the method comprises the steps of A resistivity value represented as the pixel in the ERT image; Typical resistivity for saturated aqueous media; Typical resistivity for dry media; For the depth of the pixel point, The maximum depth of the pixel point; 、 is a learnable scale factor; Is that A function; And (5) representing a resistivity tomography geological constraint weight map.
4. The improved YOLOv-based earth-rock dam ERT image leakage feature identification system of claim 2, wherein the multi-level channel monte carlo attention architecture includes an improved base component module, a multi-parameter monte carlo attention module, and a parallel ERT physical multi-scale convolution attention module.
5. The improved YOLOv-based earth-rock dam ERT image leak feature identification method of claim 4, wherein the improved base assembly module comprises: The channels are aligned, the channel number of the input feature image is adjusted to be an integer multiple of a preset divisor through 1X 1 convolution, and an aligned feature image is obtained; Tensor scrambling, namely randomly grouping the aligned feature images along a channel dimension in a training stage, and carrying out channel rearrangement in each group to obtain a scrambled feature image; and (3) carrying out normalized convolution, carrying out two-dimensional convolution, batch normalization and SiLU activation function processing on the scrambled feature map, and outputting a normalized feature map.
6. The earth-rock dam ERT image leakage feature recognition system based on improvement YOLOv of claim 1, wherein the example segmented neural network model of the improved YOLOv architecture is trained using a multi-task joint optimization framework during training, the loss function is derived from weighted summation of physical perception modulation classification loss, edge gradient coupling regression loss, distribution focus loss, and physical consistency mask loss, and is trained using a task risk perception task alignment distributor.
7. The improved YOLOv-based earth-rock dam ERT image leakage feature identification method of claim 6, wherein the physical perception modulation classification loss dynamically amplifies positive sample loss by a resistivity modulation factor, wherein the bounding box regression loss is an edge gradient coupling regression loss, and introduces a resistivity gradient edge consistency constraint, wherein the instance segmentation mask loss is a physical consistency mask loss, and comprises a resistivity spatial smoothness constraint term, and wherein the risk perception task alignment distributor defines a task alignment score as: ; Wherein: Aligning the scores for the risk perception tasks to replace the original TAL scores for positive and negative sample distribution; For classification confidence; IoU for the predicted and real frames, r for the risk factor, Is an exponential superparameter.
8. The improved YOLOv-based earth-rock dam ERT image leak feature identification method of claim 1, wherein the improved linear growth prediction model uses the following formula to calculate leak area: ; Wherein, the In order to predict the area of leakage, To obtain an empirical upper limit for leakage area based on dam type statistics, For the leakage development rate coefficient, t is the current time, In order for the leak to be initially detected for a period of time, Is the initial detection area.
9. An earth-rock dam ERT image leakage feature recognition system based on improvement YOLOv, comprising the following modules: the ERT data acquisition and processing module is used for acquiring original data of the earth and rockfill dam ERT acquired in ERT equipment, carrying out resistivity inversion imaging, standardization and data augmentation processing, and outputting a standardized ERT image; an ERT image leakage characteristic recognition module is internally provided with an example segmentation neural network model based on an improved YOLOv framework and is used for processing the input standardized ERT image and outputting a recognition result of a leakage area; An ERT image leak prediction module evaluates the characteristics of the leak region using an improved linear growth prediction model based on the recognition results.

Description

Improved YOLOv-based earth-rock dam ERT image leakage feature identification method and system Technical Field The invention belongs to the technical field of disaster prevention and reduction of hydraulic engineering, and particularly relates to an earth and rockfill dam ERT (resistivity tomography) image leakage characteristic identification method and system based on improvement YOLOv. Background The leakage phenomenon widely exists in earth and rockfill dams, so that not only can the water loss of the reservoir be caused, but also piping, bubble springs and dam deformation damage can be caused, and even serious dangerous situations such as dam burst, flood inundation and the like can be caused. Therefore, it is important to develop efficient, accurate and intelligent leakage dangerous case detection technology. At present, the traditional leakage detection method mainly comprises drilling, geophysical prospecting, manual inspection and the like. The drilling method has high time and economic cost, is difficult to adapt to the rapid inspection requirement of dangerous situations, has strong interpretation subjectivity due to high expert experience, has low manual inspection efficiency, and is difficult to realize early warning and accurate positioning. With the development of high-density electrical method technology, the resistivity Tomography (ELECTRICAL RESISTANCE tomograph, ERT) technology has shown great potential in the field of earth and rockfill dam leakage detection due to the advantages of non-invasiveness, visualization, relatively low cost and the like. Although ERT technology can generate images reflecting conditions inside a dam, image recognition links still face a plurality of challenges, such as image blurring and artifact interference, difficulty in extracting leakage features, insufficient automation and instantaneity, and the like, which bring great difficulty to rapid detection of leakage dangerous cases of earth and rockfill dams. In recent years, deep learning techniques, particularly target detection algorithms, have been revolutionarily successful in the field of computer vision. The powerful end-to-end characteristic learning and expression capability provides a brand new solution for overcoming the bottleneck. The deep learning model can directly automatically learn the deep abstract mode of the leakage characteristic from the original image, has stronger robustness to blurring, noise and artifacts, and indicates the direction for automatic and intelligent recognition of the ERT leakage image. Among a plurality of target detection algorithms, YOLOv algorithm adopts an Anchor-Free frame (Anchor-Free) design, and introduces a new backbone network and a feature fusion network, thereby providing an advanced technical foundation for ERT leakage image detection. However, the direct application of YOLOv to the ERT image of the earth-rock dam still has the problems of poor field adaptability, insufficient small target detection capability, complicated geological background interference and the like, and the targeted improvement and innovation are needed. Disclosure of Invention In order to solve the technical problems, the invention provides an earth and rockfill dam ERT image leakage characteristic identification method and system based on improvement YOLOv. The invention enhances the detection capability of YOLOv algorithm to small target and low contrast leakage characteristics in the ERT image of the earth and rockfill dam through a series of targeted improvements, effectively overcomes the interference of image blurring and artifact under complex geological conditions, remarkably improves the accuracy and robustness of automatic identification and positioning of leakage positions, and does not need to additionally increase dangerous case inspection cost. In order to achieve the above purpose, the present invention adopts the following technical scheme. An earth-rock dam ERT image leakage characteristic identification method based on improvement YOLOv comprises the following steps: S1, acquiring original resistivity tomography data of a earth-rock dam, performing inversion imaging and standardization processing to construct an initial resistivity tomography image data set; s2, accurately labeling leakage areas in the resistivity tomography image by using a labeling tool, uniformly classifying the leakage areas into single categories, and dividing a training set, a verification set and a test set according to a proportion; S3, constructing an example segmentation neural network model of the improved YOLOv architecture; S4, training the example segmentation neural network model by using the marked data set, and optimizing parameters through forward propagation and backward propagation; s5, inputting the resistivity tomography image to be detected into a trained example segmentation neural network model, outputting a boundary box, a category confidence coefficient and a segmentation m