CN111767815A - A method for identifying water leakage in tunnels - Google Patents

Abstract

The invention relates to the field of monitoring, in particular to a method for identifying water leakage in a tunnel, comprising the following steps: (1) accessing a video stream, and performing single-frame image preprocessing; (2) using a trained LVQ model to process the processed image (3) Judging whether there is water leakage; (4) Once water leakage is found, save the current frame video image and use the Canny algorithm to perform edge calculation on the leakage, determine the degree of leakage and output; (5) ) The operator arranges personnel testing according to the severity. After the above scheme is adopted, the efficiency of water leakage detection and prevention is greatly improved, which is beneficial to the tunnel operator's maintenance work.

Description

A method for identifying water leakage in tunnels

technical field

The invention relates to the field of monitoring, in particular to a method for identifying leakage water in a tunnel.

Background technique

Statistics from the highway department show that about 30% of my country's highway tunnels have serious water leakage. Statistics from the railway department show that 28.4% of my country's railway tunnels have serious water leakage. With the rapid development of tunnel engineering in my country, the problem of tunnel engineering diseases has become increasingly prominent. Leakage water is one of the main diseases of tunnels, and it is also the root cause of other diseases in tunnels. The health problems of tunnels have become increasingly prominent.

The current method of detecting water leakage in tunnels is mainly the method of manual inspection. Judging based on the results of visual observation is greatly affected by human factors, and there are problems of low efficiency and poor accuracy.

Therefore, the present inventor has conducted further research on this, and developed a method for identifying water leakage in a tunnel, which is how this case came into being.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method for identifying water leakage in a tunnel, which greatly improves the efficiency and prevention of water leakage detection, and is beneficial to the maintenance of the tunnel by the tunnel operator.

In order to achieve the above object, technical scheme of the present invention is as follows:

A method for identifying water leakage in a tunnel, comprising the following steps:

(1) Access the video stream and perform single-frame image preprocessing;

(2) Use the trained LVQ model to identify the processed image;

(3) Judging whether there is water leakage;

(4) Once water leakage is found, save the current frame video image and use the Canny algorithm to perform edge calculation on the leakage, judge the degree of leakage and output;

(5) The operator arranges personnel testing according to the severity.

By using a series of leaking water images collected in the tunnel as the identification data set, and using the data set to train and adjust the LVQ neural network model, adjust the optimal decision threshold to improve its detection accuracy, and use the Canny algorithm to detect The degree of water leakage shall be judged, and the tunnel operation and maintenance personnel shall be rectified according to the severity of the leakage.

Further, in step (1), the video stream adopts the surveillance camera group in the tunnel.

It can make full use of the original cameras in the tunnel for reconstruction and save resources.

Further, the LVQ model consists of three layers of neurons, namely the input layer, the competition layer and the linear output layer. The LVQ neural network algorithm is a learning algorithm for training the competition layer in a supervised state. The specific calculation steps are as follows:

Step 1: Initialize the weight ω _ij between the input layer and the competitive layer and the learning rate η (η>0);

Step 2: Use the LVQ1 algorithm to learn all input patterns;

Step 3: Send the input vector x=(x ₁ , x ₂ ,...,x _R ) ^T to the input layer, and calculate the distance between the competition layer and the input vector;

Step 4: Select the two competing layer neurons i, j with the smallest distance from the input vector;

Step 5: If neuron i and neuron j satisfy the following two conditions:

① Neuron i and neuron j correspond to different categories;

②The distances d _i and d _j between neuron i and neuron j and the current input vector satisfy

Among them, ρ is the window width that the input vector may fall into which is close to the plane of the mid-section of the two vectors;

Then there are:

①If the category C _i corresponding to neuron i is consistent with the category C _x corresponding to the input vector, that is, C _i =C _x , the weights of neuron i and neuron j are modified according to the following formula;

②If the category C _j corresponding to neuron j is consistent with the category C _x corresponding to the input vector, that is, C _j =C _x , the weights of neuron i and neuron j are modified according to the following formula;

Step 6: If neuron i and neuron j do not meet the conditions of step 5, only the neuron weights closest to the input vector are updated;

Step 7: Finally evaluate the performance of the model using the test dataset.

The typical learning vector quantization algorithms are LVQ1, LVQ2 and LVQ3, among which the first two algorithms are widely used, especially LVQ2 is the most widely used and effective. It has been successfully applied to many fields such as statistics, pattern recognition, and machine learning. In the LVQ1 algorithm, only one neuron can win, that is, only one neuron's weight can be updated and adjusted. In order to improve the classification accuracy, the LVQ2 algorithm is obtained by improving the LVQ1 algorithm, which approaches the Bayesian limit based on the smooth moving decision boundary. The LVQ2 version was then revised to LVQ2.1 and eventually to LVQ3. These later versions of LVQ all introduced secondary winning neurons, and both the winning neuron's weight vector and the secondary winning neuron's weight vector were updated.

Further, the LVQ model leakage identification training steps are as follows:

Step 1: Manually label all images in the dataset;

Step 2: Use the training dataset to train the LVQ model until the model converges;

Step 3: Use the validation data set to evaluate the generalization ability of the model, adjust the hyperparameters, and take the model after 1500 iterations as the initial recognition model;

Step 4: Repeat steps 2 and 3 until the generalization ability of the model on the validation dataset meets the task requirements;

Step 5: Fix the trained network structure;

Step 6: Use the validation dataset to assess whether the model is overfitting;

The performance of the leakage water identification model will determine the identification result. This system trains the model to an accuracy of more than 95%, and the average inference time is less, which can meet the real-time requirements.

Further, in step (1), the preprocessing of the water leakage image includes the following steps:

Step 1: Grayscale binary processing;

Step 2: Gaussian filtering; weighted average of the entire image, the value of each pixel is obtained by itself and other pixel values in the neighborhood after weighted average; after smoothing, the influence of noise is reduced;

Step 3: Corrosion; remove the boundary points of the object, make the target smaller, and eliminate noise points smaller than structural elements;

Step 4: Expansion; merge all the background points of the object into the object, make the target larger, and fill the holes in the target;

Step 5: Edge Detection.

By eroding and then dilating, the outline of the object is generally smoothed, narrow discontinuities are broken and thin protrusions are eliminated.

After adopting the above scheme, the present invention has the following advantages compared with the prior art:

1. Using non-contact monitoring technology, the system has high reliability and durability;

2. The camera can monitor multiple diseased parts at the same time, and it is an existing equipment, saving equipment investment;

3. Have the ability to judge the degree of water leakage disease;

4. Information access tunnel maintenance management cloud platform for storage and analysis.

Description of drawings

Figure 1 is a flow chart of water leakage identification and detection;

Figure 2 Learning vector quantization LVQ algorithm neural network diagram

Figure 3 is a framework for judging the severity of water leakage

Detailed ways

The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

A method for identifying water leakage in tunnels. Based on the method of deep learning, the technology of identifying and judging the degree of water leakage in tunnels is studied. This technology can effectively perform real-time water leakage detection on the tunnel monitoring screen. The LVQ algorithm is used in the identification field. The advantages of the system are to identify the leakage water, and its recognition accuracy is higher than that of the traditional neural network algorithm, and the Canny algorithm is used to perform edge calculation on the leakage water, and the leakage water area is used to judge its degree, so that operators can arrange according to the severity. Personnel detection.

(1) Leakage water identification and detection process

The system's identification and detection of water leakage is mainly divided into two parts, one is the identification of water leakage, and the other is the judgment of the degree of water leakage. Through the trained LVQ model, real-time water leakage identification is performed on the tunnel situation in the video stream. After identifying the water leakage picture, the edge processing of the Canny algorithm is performed on it, and the severity of the water leakage is judged by the water leakage area. Combined with the tunnel maintenance cloud platform for data analysis and linkage of governance plans. See Figure 1. The specific identification and detection process is as follows:

Access the video stream for single-frame image preprocessing;

Use the trained LVQ model for recognition;

Determine whether there is water leakage;

Once water leakage is found, save the video image of the current frame and use edge computing to determine the degree of water leakage;

Use filters to reduce image noise;

The enhancement algorithm is used to highlight the points with significant changes in gray level in the neighborhood;

Use gradient amplitude threshold judgment as edge calculation judgment standard;

Use the leakage water area as the leakage water level and output.

(2) LVQ neural network model

Learning Vector Quantization (LVQ) neural network is a kind of neural network with supervised learning method for training competitive layers, which has a wide range of applications in the field of recognition and optimization. The LVQ neural network consists of 3 layers of neurons, namely the input layer, the competition layer and the linear output layer. The input layer and the competition layer are fully connected, and the competition layer and the linear output layer are partially connected. The number of neurons in the competition layer is always greater than the number of neurons in the linear output layer, and each linear output layer neuron can be connected with multiple competition layer neurons. See Figure 2, the advantages are as follows:

1. The network structure is simple.

2. Accurate classification of input vector patterns can be accomplished through supervised learning.

3. There is no need to normalize and orthogonalize the input vector.

4. The distance between the input vector and the competition layer can be directly calculated to realize the recognition.

(3) LVQ neural network model learning algorithm

The LVQ neural network algorithm is a learning algorithm that trains the competitive layer in a supervised state, rather than the traditional learning algorithm, which is a self-organizing feature mapping algorithm. The typical learning vector quantization algorithms are LVQ1, LVQ2 and LVQ3, among which the first two algorithms are widely used, especially LVQ2 is the most widely used and effective. It has been successfully applied to many fields such as statistics, pattern recognition, and machine learning.

In the LVQ1 algorithm, only one neuron can win, that is, only one neuron's weight can be updated and adjusted. In order to improve the classification accuracy, the LVQ2 algorithm is obtained by improving the LVQ1 algorithm, which approaches the Bayesian limit based on the smooth moving decision boundary. The LVQ2 version was then revised to LVQ2.1 and eventually to LVQ3. These later versions of LVQ all introduced secondary winning neurons, and both the winning neuron's weight vector and the secondary winning neuron's weight vector were updated. Specific calculation steps:

Step 1: Initialize the weight ω _ij between the input layer and the competitive layer and the learning rate η (η>0).

Step 2: Use the LVQ1 algorithm to learn all input patterns.

Step 3: Send the input vector x=(x ₁ , x ₂ ,...,x _R ) ^T to the input layer, and calculate the distance between the competition layer and the input vector.

Step 4: Select the two competing layer neurons i, j with the smallest distance from the input vector.

Step 5: If neuron i and neuron j satisfy the following two conditions:

③ Neuron i and neuron j correspond to different categories;

④ The distances d _i and d _j between neuron i and neuron j and the current input vector satisfy

where ρ is the width of the window that the input vector may fall into which is close to the plane of the mid-section of the two vectors.

Then there are:

③ If the category C _i corresponding to neuron i is consistent with the category C _x corresponding to the input vector, that is, C _i =C _x ,

Then the weights of neuron i and neuron j are modified according to the following formula.

④ If the category C _j corresponding to neuron j is consistent with the category C _x corresponding to the input vector, that is, C _j =C _x ,

Then the weights of neuron i and neuron j are modified according to the following formula.

Step 6: If neuron i and neuron j do not meet the conditions of step 5, only the neuron weights closest to the input vector are updated.

Step 7: Finally evaluate the performance of the model using the test dataset.

(4) Leakage identification model training

The main purpose of water leakage identification is to detect whether there is water leakage in the tunnel monitoring screen, and there are only water leakage and no water leakage.

Step 1: Manually label all images in the dataset.

Step 2: Use the training dataset to train the LVQ model until the model converges.

Step 3: Use the validation dataset to evaluate the generalization ability of the model, adjust the hyperparameters, and take the model after 1500 iterations as the initial recognition model.

Step 4: Repeat steps 2 and 3 until the generalization ability of the model on the validation dataset meets the task requirements.

Step 5: Fix the trained network structure.

Step 6: Use the validation dataset to assess whether the model is overfitting.

The performance of the leakage water identification model will determine the identification result. This system trains the model to an accuracy of more than 95%, and the average inference time is less, which can meet the real-time requirements.

(5) Evaluation algorithm of water leakage severity

The tunnel water leakage area is the tunnel wall, which is located around the video image. Therefore, the surrounding part of the picture can be extracted as the area of interest to reduce the interference of vehicles and tunnel lighting on water leakage detection and reduce the processing time of the detection algorithm.

The system first preprocesses the obtained leakage water images, as shown in Figure 3.

Step 1: Grayscale binary processing. Due to the complex background of the tunnel, there will be some noise after processing, and the binary image needs to be denoised. Morphological operations, namely dilation and erosion operations, can solve this problem well.

Step 2: Gaussian filtering. A weighted average is performed on the entire image, and the value of each pixel is obtained by weighted averaging of itself and other pixel values in its neighborhood. Reduces the effect of noise after smoothing.

Step 3: Corrosion. By removing the boundary points of the object and making the target smaller, noise points smaller than structuring elements can be eliminated.

Step 4: Inflate. Merge all the background points of the object into the object, make the target larger, and fill the holes in the target.

Step 5: Edge Detection. Use the finite difference of first-order partial derivatives to calculate the magnitude and direction of the gradient; perform non-maximum suppression on the gradient magnitude; detect and connect edges with a double-threshold algorithm.

Then calculate the size of the leaking water, and further its size, divide the severity into minor leaking water and severe leaking water.

The visual image processing method used in this system is to use the difference feature to complete the detection through the grayscale difference in the water leakage area in the visible light image compared with the non-leakage area. And can make full use of the original camera in the tunnel for transformation, saving resources.

According to the characteristics of tunnel water leakage images, based on the learning vector quantization (LVQ) neural network model, a water leakage identification technology is researched, which is used for real-time water leakage judgment on the pictures taken during the tunnel detection process. Firstly, by using a series of leaking water images collected in the tunnel as the identification data set, and using the data set to train and adjust the LVQ neural network model, the optimal decision threshold is adjusted to improve its detection accuracy. And the Canny algorithm is used to judge the degree of water leakage, and the tunnel operation and maintenance personnel are linked to rectify according to the severity of the leakage.

Advantages of this detection system:

1. Using non-contact monitoring technology, the system has high reliability and durability;

2. The camera can monitor multiple diseased parts at the same time, and it is an existing equipment, saving equipment investment;

3. Have the ability to judge the degree of water leakage disease;

4. Information access tunnel maintenance management cloud platform for storage and analysis.

The above are only specific embodiments of the present invention, and all words such as "upper, lower, left, right, middle" involved in the present invention are only for reference and are not absolutely limited. Any modification shall be regarded as an act infringing the protection scope of the present invention.

Claims (5)

1. A method for identifying water leakage in a tunnel, comprising the steps of: (1) Access the video stream and perform single-frame image preprocessing; (2) Use the trained LVQ model to identify the processed image; (3) Judging whether there is water leakage; (4) Once water leakage is found, save the current frame video image and use the Canny algorithm to perform edge calculation on the leakage, judge the degree of leakage and output; (5) The operator arranges personnel testing according to the severity. 2 . The method for identifying water leakage in a tunnel according to claim 1 , wherein in step (1), the video stream adopts a group of surveillance cameras in the tunnel. 3 . 3. a kind of tunnel seepage water identification method according to claim 1 is characterized in that: LVQ model is made up of three layers of neurons, namely input layer, competition layer and linear output layer, LVQ neural network algorithm is supervised A learning algorithm for training the competition layer in the state, the specific calculation steps are as follows: Step 1: Initialize the weight ω _ij between the input layer and the competitive layer and the learning rate η (η>0); Step 2: Use the LVQ1 algorithm to learn all input patterns; Step 3: Send the input vector x=(x ₁ , x ₂ ,...,x _R ) ^T to the input layer, and calculate the distance between the competition layer and the input vector;

Step 4: Select the two competing layer neurons i, j with the smallest distance from the input vector; Step 5: If neuron i and neuron j satisfy the following two conditions: ① Neuron i and neuron j correspond to different categories; ②The distances d _i and d _j between neuron i and neuron j and the current input vector satisfy

Among them, ρ is the window width that the input vector may fall into which is close to the plane of the mid-section of the two vectors; Then there are: ①If the category C _i corresponding to neuron i is consistent with the category C _x corresponding to the input vector, that is, C _i =C _x , the weights of neuron i and neuron j are modified according to the following formula;

②If the category C _j corresponding to neuron j is consistent with the category C _x corresponding to the input vector, that is, C _j =C _x , the weights of neuron i and neuron j are modified according to the following formula;

Step 6: If neuron i and neuron j do not meet the conditions of step 5, only the neuron weights closest to the input vector are updated; Step 7: Finally evaluate the performance of the model using the test dataset. 4. a kind of tunnel seepage water identification method according to claim 1, is characterized in that: LVQ model seepage water identification training step is: Step 1: Manually label all images in the dataset; Step 2: Use the training dataset to train the LVQ model until the model converges; Step 3: Use the validation data set to evaluate the generalization ability of the model, adjust the hyperparameters, and take the model after 1500 iterations as the initial recognition model; Step 4: Repeat steps 2 and 3 until the generalization ability of the model on the validation dataset meets the task requirements; Step 5: Fix the trained network structure; Step 6: Use the validation dataset to evaluate whether the model is overfitting; The performance of the leakage water identification model will determine the identification result. This system trains the model to an accuracy of more than 95%, and the average inference time is less, which can meet the real-time requirements. 5. The method for identifying water leakage in a tunnel according to claim 1, wherein in step (1), the preprocessing of the water leakage image comprises the following steps: Step 1: Grayscale binary processing; Step 2: Gaussian filtering; weighted average of the entire image, the value of each pixel is obtained by weighted average of itself and other pixel values in the neighborhood; after smoothing, the influence of noise is reduced; Step 3: Corrosion; remove the boundary points of the object, make the target smaller, and eliminate noise points smaller than structural elements; Step 4: Expansion; merge all the background points of the object into the object to make the target larger and fill the holes in the target; Step 5: Edge Detection.

Images