Underground water tank water level image monitoring system and water level monitoring control method based on improved convolutional neural network VGGNet10

Abstract

The invention relates to the technical field of intelligent water level monitoring, in particular to a water level image monitoring system for an underground water tank and a water level monitoring control method based on an improved convolutional neural network VGGNet10, and the water level monitoring control method comprises the steps: obtaining a real-time image of the position relation between the water surface position of the underground water tank and a water level gauge; constructing an improved convolutional neural network model to monitor the water level, wherein the improved convolutional neural network model is an improved convolutional neural network based on a multi-layer improved VGGNet16 convolutional neural network model and an improved VGGNet10 model; and classifying the obtained real-time images by using the improved convolutional neural network model, and controlling the water pump to work. According to the method, the algorithm of the convolutional neural network (CNN) is improved, the improved multi-layer VGGNet16 convolutional neural network model and the VGGNet10 model with higher recognition accuracy are combined, and a shared pooling mode is provided, so that the loss of data information amount is reduced, and the water level image recognition accuracy is remarkably improved.

Description

Technical Field

[0001] The present application relates to the technical field of intelligent water level monitoring, and in particular to an underground water tank water level image monitoring system and a water level monitoring and control method based on an improved convolutional neural network VGGNet10. Background Art

[0002] During the mining process of the working face, water-bearing and water-conducting geological structures that would not cause water inrush accidents under static conditions will gradually develop into water-conducting channels between the coal seam and the aquifer under the influence of mining activities, leading to water inrush accidents. Therefore, it is necessary to precisely monitor the water level in the underground water tank.

[0003] At present, the intelligent control technology of underground water tank water level mainly relies on sensors to collect water level. The installation of water level sensors requires special installation conditions or precise installation positions to ensure accurate measurement, which may increase installation costs and complexity. In addition, the solution of sensor water level collection has limitations in accuracy, especially when the water level changes drastically or the water surface is unstable, the sensor may not be able to track the water level changes in time or give accurate readings. Summary of the invention

[0004] The technical problem to be solved by the present invention is that the existing underground water tank water level monitoring has poor accuracy and real-time performance.

[0005] To this end, the present invention provides an underground water tank water level image monitoring system and a water level monitoring and control method based on an improved convolutional neural network VGGNet10.

[0006] The technical solution adopted by the present invention to solve its technical problem is:

[0007] A water level monitoring and control method based on an improved convolutional neural network VGGNet10 comprises the following steps:

[0008] Step 1, obtaining a real-time image of the positional relationship between the water surface position of the underground water tank and the water level scale;

[0009] Step 2: construct an improved convolutional neural network model to monitor the water level. The improved convolutional neural network model is an improved convolutional neural network based on a multi-layer improved VGGNet16 convolutional neural network model and an improved VGGNet10 model.

[0010] Step three, use the improved convolutional neural network model to classify the acquired real-time images and control the operation of the water pump.

[0011] Furthermore, the improved convolutional neural network model includes an improved pooling layer, which is a contribution pooling layer, and the output of the contribution pooling layer is jointly determined by the contribution pooling factor and the convolution layer output, wherein the contribution pooling factor is jointly determined by the adjustment factor and the convolution layer output according to the interpolation principle. The adjustment factor is determined by the current contribution pooling step size and the number of iterations.

[0012] Furthermore, the calculation method of the contribution pooling layer output is: Among them, ɑ is the contribution pooling factor, s is the step size of the pooling area, and F ij is the output of the convolution layer (contributing to the pooling layer input), and pvalue is the output of the contributing pooling layer.

[0013] Furthermore, the calculation method of the contribution pooling factor is: Where δ is the adjustment factor, s is the step size of the pooling area, and F ij is the output of the convolutional layer.

[0014] Furthermore, the calculation method of the adjustment factor is: Among them, s is the contribution pooling step size and m is the number of iterations.

[0015] Furthermore, in the improved VGGNet10 model, the input data undergoes two convolutions and ReLU operations with multiple 3*3 convolution kernels and then passes through a contribution pooling layer. Finally, the data passing through the last contribution pooling layer is fully connected and SoftMax is used to output two prediction results.

[0016] Furthermore, the improved VGGNet10 model uses the Dropout method to temporarily remove neurons from the network with a probability of 50%, and adopts a small batch gradient descent algorithm to update the network weight parameters.

[0017] Furthermore, the improved VGGNet10 algorithm is divided into a forward propagation process and a back propagation process. In the forward propagation process, the input is a picture sample, and the output result is the CNN model output (a L ), in the back propagation process, the input is m picture samples, and the output result is the hidden layer and output layer parameters W, b.

[0018] An underground water tank water level image monitoring system, comprising:

[0019] A camera, which is mounted on the wall of the underground water tank through a bracket, and captures an image of the position relationship between the water surface and the water level scale;

[0020] A host computer is loaded with a soft PLC having an improved convolutional neural network model, the camera is connected to the host computer, and the host computer can control the operation of a water pump in an underground water tank.

[0021] Furthermore, the camera is located directly in front of the water level gauge, and there is an angle between the vertical reference line of the camera and the length direction of the water level gauge, and the angle is -30° to 30°.

[0022] The beneficial effect of the present invention is that the present application combines the improved multi-layer VGGNet16 convolutional neural network model and the VGGNet10 model with higher recognition accuracy by improving the algorithm of the convolutional neural network (CNN), and proposes a shared pooling method to reduce the loss of data information and significantly improve the accuracy of water level image recognition. This technology can accurately detect the water level line from the complex and changeable underground water tank images and calculate the accurate water level value. The improved CNN can better adapt to the complex environment of the underground water tank, such as changes in lighting, turbid water quality, water surface fluctuations, etc., thereby ensuring stable and reliable water level monitoring under different conditions.

[0023] In this application, by introducing advanced technologies such as deep learning in water level monitoring, the improved CNN not only has powerful image recognition capabilities, but can also automatically make decisions and controls based on monitoring results. For example, when an abnormal water level is found, an alarm mechanism can be automatically triggered or the operating status of equipment such as a water pump can be adjusted. In addition, CNN has the ability to continuously learn, and can continuously optimize the algorithm model as monitoring data accumulates, thereby improving recognition accuracy and intelligence.

[0024] Compared with traditional methods, this method has a high degree of automation, realizing a fully automated process from image acquisition, processing to water level calculation, greatly reducing manual participation and intervention, and improving work efficiency. It can monitor the water level of the underground water tank in real time and promptly feed back the monitoring results to the control system, providing timely and accurate data support for subsequent water level control. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0026] Figure 1 It is a structural schematic diagram of the underground water tank water level image monitoring system of the present invention.

[0027] Figure 2 It is a schematic diagram of the construction of the convolutional neural network model and image processing process in the present invention.

[0028] Figure 3 It is a schematic diagram of contribution pooling in the present invention.

[0029] Figure 4 It is a schematic diagram of the structure of the improved VGGNet10 model in the present invention.

[0030] Figure 5It is a schematic diagram of the structural parameters of the improved VGGNet10 model in the present invention.

[0031] In the figure: 1. Underground water tank; 2. Water level gauge; 3. Camera; 4. Host computer; 5. Water pump. DETAILED DESCRIPTION

[0032] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0033] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In addition, features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0034] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0035] A water level image monitoring system for a downhole water tank includes a camera 3 and a host computer 4. A water level scale 2 is preset in the downhole water tank 1. A water pump 5 is arranged in the downhole water tank 1. The camera 3 is installed in the downhole water tank 1 through an adjustable bracket. The bracket should be installed on the downhole wall. The adjustable bracket can adjust the position and angle of the camera 3 so that the camera 3 is located about 5 meters in front of the water level scale 2. The camera 3 is horizontally facing the water level scale 2. In other embodiments, there is an angle of -30° to 30° between the vertical reference line of the camera 3 and the length direction of the water level scale 2, that is, the camera 3 can have a deflection angle of about 30 degrees. The camera 3 captures the position relationship image between the water surface and the water level scale 2. The camera 3 uses the RTSP (Real-Time Stream Protocol) protocol to transmit the image in real time to the host computer 4 that has been loaded with a soft PLC with an improved convolutional neural network model, and classifies and processes the image. The host computer 4 is connected to the water pump 5.

[0036] The soft PLC hardware is RV1126, which is a high-performance visual processor SoC for IPC / CVR. It is based on the quad-core ARM Cortex-A7 core, has a built-in 2T computing power NPU, and supports 4K30FPS H.264 / H.265 video encoding and decoding. It also has multi-level noise reduction, 3-frame HDR, and black light full color technology features1. The computing power of RV1126 is up to 2.0TOPs23.

[0037] A water level monitoring and control method based on an improved convolutional neural network VGGNet10 comprises the following steps:

[0038] Step 1: The position relationship image between the water surface and the water level gauge 2 is obtained in real time through the camera 3 and transmitted to the host computer 4.

[0039] Step 2: Build an improved convolutional neural network model to monitor water level

[0040] In order to solve the problems of low accuracy in water level recognition and long training time, the 10-layer improved convolutional neural network model VGGNet10 based on VGGNet16 is used for water level monitoring, with the number of layers as the naming rule. The data set shot by network camera 3 is sent to VGGNet10 after pooling improvement and overall model improvement for training and saving the training model for foreign body classification. The overall construction process includes:

[0041] S2.1 Improved pooling layer

[0042] The pooling operation is mainly used to reduce feature dimensionality, compress data, reduce the number of parameters, reduce overfitting, and improve the fault tolerance of the model.

[0043] Generally speaking, pooling layers are periodically inserted between each convolutional layer of a convolutional neural network. There are mainly two pooling methods: maximum pooling and average pooling.

[0044] If the maximum pooling process is a 2*2 area, it can be understood as calculating the maximum value of four numbers in a 2*2 matrix. If the average pooling process is a 2*2 area, it can be understood as calculating the average value of four numbers in a 2*2 matrix.

[0045] The maximum pooling calculation method is to only retain the largest value among the four pixels in the area, and the values ​​of the other three pixels are completely discarded. The final value is only the value of the largest pixel, which sharpens the image in the current area and loses some detail information. Average pooling retains the average value of the four pixels in the pooling area. Once the value difference in this area is too large, the calculated value will be too small, resulting in the passivation of the current area and information loss.

[0046] For images with low resolution, there are not many pixels, and only a small number of pixels are used to characterize the image features. The pooling layer will continuously reduce the amount of data, so the number of parameters and the overall amount of calculation will be greatly reduced. Although overfitting is controlled to a certain extent, this will reduce the amount of data information, and the amount of data information retained is limited. Therefore, this patent proposes an improved pooling method to improve the loss of data information, which is called contribution pooling. The calculation method of contribution pooling is shown in formulas (1) to (3).

[0047]

[0048] In formula (1), ɑ is the contribution pooling factor, s is the step size of the pooling area, and F ij is the output of the convolution layer (contributing to the pooling layer input), and pvalue is the output of the contributing pooling layer.

[0049]

[0050] In formula (2), δ is the adjustment factor, s is the step size of the pooling area, and F ij is the output of the convolutional layer.

[0051]

[0052] In formula (3), s is the contribution pooling step size and m is the number of iterations.

[0053] From formula (1) to formula (3), it can be seen that the contribution pooling layer output is determined by the contribution pooling factor and the convolution layer output. The contribution pooling factor is determined by the adjustment factor and the convolution layer output according to the interpolation principle. The adjustment factor is determined by the current contribution pooling step size and the number of iterations. The proposed contribution pooling method can reduce the loss of data information and enable the convolutional neural network to accurately extract data information features in different pooling areas and different numbers of iterations. The contribution pooling diagram is shown in the attached figure. Figure 3 As shown in the figure, X, Y represent the horizontal and vertical coordinates with the lower left corner of the image as the origin, F11, F12, F13, F14, F21... represent the value of the current pixel. Figure 3 The pixels in fall into four pooling areas. After the pooling operation, four contribution pooling layer outputs are generated. Contribution pooling performs pooling operations in each pooling area according to formulas (1), (2), and (3), solving the data loss problem of maximum pooling and average pooling.

[0054] S2.2 Improved VGGNet10 model

[0055] Since VGGNet16 was proposed to participate in the ILSVRC2014 competition for 1000 classifications of deep convolutional neural network structures, VGGNet16 increased the model depth and the number of network channels. The number of channels in the first layer of the network is 64, and each subsequent layer doubles the number of channels, up to 512. The increase in the number of channels allows more information to be extracted.

[0056] The VGGNet16 model is targeted at hundreds of categories, and for the water level monitoring system, it is a binary classification, that is, normal state and over-limit state. With the increase of the number of convolutional layers, the recognition accuracy is improved, but too many network structure models are complex and prone to data overfitting. Therefore, the VGGNet10 structure model is selected by comparing the network structure, the number of parameters, and the recognition accuracy. The schematic diagram of the VGGNet10 structure is shown in the attached figure. Figure 4 shown.

[0057] The specific structure of VGGNet10 is as follows: The input of VGGNet10 is 224*224*1 data, and the two 224s represent the length and height of the image respectively. Since all data were converted into grayscale images during the previous data set processing, 1 represents a single grayscale value channel. The input of VGGNet10 undergoes two convolutions and ReLU operations with 32 3*3 convolution kernels, and the size becomes 224*224*32. After the contribution pooling layer, the pooling unit size is 2*2, and the size after pooling becomes 112*112*32. After 64 3*3 convolution kernels undergo two convolutions and ReLU operations, the size becomes 112*112*64. After the 2*2 contribution pooling layer, the size becomes 56*56*64. After 128 3*3 convolution kernels undergo two convolutions and ReLU operations, the size becomes 56*56*128. After the 2x2 contribution pooling layer, the size becomes 28*28*128. After 256 3*3 convolution kernels are used for two convolutions and ReLU operations, the size becomes 28*28*256. After a 2x2 contribution pooling layer, the size becomes 14*14*256. It is fully connected with a layer of 1*1*4096 and a layer of 1*1*2, and finally SoftMax is used to output two prediction results. The detailed structure and parameters are shown in the attached figure. Figure 5 shown.

[0058] The activation function of VGGNet10 uses ReLu, the pooling layer uses contribution pooling, and the Dropout method is used to temporarily remove neurons from the network with a probability of 50%. The small batch gradient descent algorithm is used to update the network weight parameters.

[0059] The VGGNet10 algorithm is divided into the forward propagation process, the back propagation process and some function selection. The specific process of parameter selection in the algorithm is given in order below:

[0060] S2.2.1, forward propagation process, during the forward propagation process, the input is an image sample, and the output result is the CNN model output (a L ), the forward propagation process specifically includes:

[0061] S2.2.1.1 Initialize all hidden layer parameters W, b

[0062] S2.2.1.2ForL=2toL-1

[0063] (a) If layer L is a convolutional layer, the output is:

[0064] a L =RELU(z L )=RELU(a L-1 *W L +b L ) (4)

[0065] (b) If layer L is a pooling layer, the output is:

[0066] a L =Contribution pooling(a L-1 ) (5)

[0067] (c) If layer L is a fully connected layer, the output is:

[0068] a L =δ(z L )=δ(W L a L-1 +b L ) (6)

[0069] S2.2.1.3 The output is:

[0070] a L =softmax(z L )=softmax(W L a L-1 +b L ) (7)

[0071] S2.2.2, Back propagation process, in the back propagation process, the input is m image samples, and the output result is the hidden layer and output layer parameters W, b. The back propagation process specifically includes:

[0072] S2.2.2.1, initialize the parameters W, b of each hidden layer and output layer to random values.

[0073] S2.2.2.2, The iterative process includes:

[0074] A1, Fori = 1tom

[0075] A11 forward propagation

[0076] A12 calculates the gradient of the output layer through the loss function

[0077] A13forL=L-1toL=2, execute the back propagation algorithm to calculate:

[0078] A131 If layer L is a convolutional layer, the output is:

[0079]

[0080] A132 If layer L is a pooling layer, the output is:

[0081]

[0082] A2, For L = 2 to L = L, update the parameters W,b of the first layer according to the following two cases:

[0083] A21 If layer L is a fully connected layer, the output is:

[0084]

[0085] A22 If layer L is a convolutional layer, the output is:

[0086]

[0087] A23 If all changes in W,b are less than the iteration stop threshold ε, then exit the iteration loop and go to step S2.2.2.3.

[0088] S2.2.2.3, output the linear relationship coefficient matrix W and bias vector b of each hidden layer and output layer.

[0089] In summary, the selection of functions in the improved VGGNet10 model is as follows:

[0090] Convolutional layer: a L =RELU(z L )=RELU(a L-1 *W L +b L ) (12)

[0091] Pooling layer:

[0092]

[0093] Fully connected layer:

[0094] a L =δ(z L )=δ(W L a L-1 +b L ) (16)

[0095] RELU=max(0,x) (17)

[0096] Softmax:

[0097] Step 3: Water level control

[0098] The trained model is imported into the soft PLC hardware, and the network camera 3 monitors the image in real time and sends it to the soft PLC hardware for real-time image classification. When the current water level exceeds the limit, the control relay is turned on and the water pump 5 works. When the current water level does not exceed the limit, the control relay is turned off and the water pump 5 does not work to complete the water level image monitoring and control of the underground water tank 1.

[0099] In summary, this application improves the convolutional neural network (CNN) algorithm, combines the improved multi-layer VGGNet16 convolutional neural network model and the VGGNet10 model with higher recognition accuracy, and proposes a shared pooling method to reduce the loss of data information and significantly improve the accuracy of water level image recognition. This technology can accurately detect the water level line from the complex and changeable downhole water tank 1 image and calculate the accurate water level value. The improved CNN can better adapt to the complex environment of the downhole water tank 1, such as lighting changes, turbid water, water surface fluctuations, etc., thereby ensuring stable and reliable water level monitoring under different conditions.

[0100] In the water level monitoring of this application, by introducing advanced technologies such as deep learning, the improved CNN not only has a strong image recognition ability, but also can automatically make decisions and control according to the monitoring results. For example, when an abnormal water level is found, the alarm mechanism can be automatically triggered or the operating status of equipment such as the water pump 5 can be adjusted. In addition, CNN has the ability of continuous learning, and can continuously optimize the algorithm model as the monitoring data accumulates, thereby improving the recognition accuracy and intelligence level.

[0101] Compared with traditional methods, this method has a high degree of automation, realizing a fully automated process from image acquisition, processing to water level calculation, greatly reducing manual participation and intervention, and improving work efficiency. It can monitor the water level of the underground water tank 1 in real time, and promptly feed back the monitoring results to the control system, providing timely and accurate data support for subsequent water level control.

[0102] Compared with traditional water level sensors and other equipment, water level monitoring technology based on image processing and CNN has lower costs. This is mainly due to the popularity of cameras and the optimization of CNN algorithms.

[0103] The technology is flexible to deploy and does not rely on specific hardware equipment. Monitoring can be achieved by simply installing a camera near the underground water tank 1. At the same time, due to the flexibility of the CNN algorithm, the algorithm can be adjusted and optimized according to actual needs.

[0104] Furthermore, this technology is not only applicable to water level monitoring of underground water tanks 1, but can also be extended to other occasions where water level monitoring is required, such as rivers, lakes, reservoirs, etc. With the continuous development and improvement of technology, the underground water tank 1 water level image monitoring and control technology based on improved convolutional neural network is expected to play a greater role in water resource management, flood prevention and disaster reduction, and other fields, and promote the progress and development of related industries.

[0105] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A water level monitoring and control method based on an improved convolutional neural network VGGNet10, characterized in that: The following steps are included: Step 1, obtaining a real-time image of the positional relationship between the water surface position of the underground water tank (1) and the water level scale (2); Step 2: construct an improved convolutional neural network model to monitor the water level. The improved convolutional neural network model is an improved convolutional neural network based on a multi-layer improved VGGNet16 convolutional neural network model and an improved VGGNet10 model. Step three, using the improved convolutional neural network model to classify the acquired real-time image and control the operation of the water pump (5).

2. The water level monitoring and control method based on the improved convolutional neural network VGGNet10 according to claim 1 is characterized in that: The improved convolutional neural network model includes an improved pooling layer, which is a contribution pooling layer. The output of the contribution pooling layer is jointly determined by the contribution pooling factor and the convolution layer output, wherein the contribution pooling factor is jointly determined by the adjustment factor and the convolution layer output according to the interpolation principle. The adjustment factor is determined by the current contribution pooling step size and the number of iterations.

3. The water level monitoring and control method based on the improved convolutional neural network VGGNet10 according to claim 2 is characterized in that: The calculation method of the contribution pooling layer output is: Among them, ɑ is the contribution pooling factor, s is the step size of the pooling area, and F ij is the output of the convolution layer (contributing to the pooling layer input), and pvalue is the output of the contributing pooling layer.

4. The water level monitoring and control method based on the improved convolutional neural network VGGNet10 according to claim 3 is characterized in that: The calculation method of the contribution pooling factor is: Where δ is the adjustment factor, s is the step size of the pooling area, and F ij is the output of the convolutional layer.

5. The water level monitoring and control method based on the improved convolutional neural network VGGNet10 according to claim 4 is characterized in that: The calculation method of the adjustment factor is Among them, s is the contribution pooling step size and m is the number of iterations.

6. The water level monitoring and control method based on the improved convolutional neural network VGGNet10 according to claim 1 is characterized in that: In the improved VGGNet10 model, the input data undergoes two convolutions and ReLU operations with multiple 3*3 convolution kernels and then passes through a contribution pooling layer. Finally, the data passing through the last contribution pooling layer is fully connected and SoftMax is used to output two prediction results.

7. The water level monitoring and control method based on the improved convolutional neural network VGGNet10 according to claim 6 is characterized in that: The improved VGGNet10 model uses the Dropout method to temporarily remove neurons from the network with a probability of 50%, and adopts a small batch gradient descent algorithm to update the network weight parameters.

8. The water level monitoring and control method based on the improved convolutional neural network VGGNet10 according to claim 7 is characterized in that: The improved VGGNet10 algorithm is divided into a forward propagation process and a back propagation process. In the forward propagation process, the input is a picture sample, and the output result is the CNN model output (a L ), in the back propagation process, the input is m picture samples, and the output result is the hidden layer and output layer parameters W, b.

9. An underground water tank water level image monitoring system, characterized in that: include, A camera (3), the camera (3) being mounted on the wall of the underground water tank (1) via a bracket, the camera (3) capturing an image of the positional relationship between the water surface and the water level gauge (2); A host computer (4) is loaded with a soft PLC having an improved convolutional neural network model, the camera (3) is connected to the host computer (4), and the host computer (4) is capable of controlling the operation of a water pump (5) in an underground water tank (1).

10. The underground water tank water level image monitoring system according to claim 9, characterized in that: The camera (3) is located directly in front of the water level gauge (2), and there is an angle between the vertical reference line of the camera (3) and the length direction of the water level gauge (2), and the angle is -30° to 30°.

Images