Method for identifying sewage in lake-adjacent strip mine

Abstract

The invention discloses a method for identifying sewage in a lake-near strip mine. The method comprises the following steps: establishing a three-dimensional digital model based on an initial database; establishing a clear sewage migration image database; and constructing a clean water and sewage identification model. Compared with the prior art, in the aspect of model construction, multi-source data are fused, and a high-precision three-dimensional terrain and geologic model is constructed through investigation data and on-site unmanned aerial vehicle laser radar scanning data; according to the invention, clean sewage is identified from the source, so that corresponding clean sewage diversion engineering measures can be built according to the identification result; according to the method, a numerical simulation result and an on-site sewage electrochemical monitoring result are fused to construct clear sewage migration images of different pollution source types; the sewage identification model is constructed by using the deep learning algorithm, and the method has the advantages of low learning cost, high prediction precision and the like.

Description

Technical Field

[0001] The present invention belongs to the technical field of environmental protection, and particularly relates to the identification and zoning of clean and sewage water in the open-pit mine area by the lake. Background Art

[0002] During the open-pit or underground mining process in China, the broken ore is exposed to the air, and the broken ore contains various metal minerals. During the rainy season, the broken ore is oxidized under the combined action of air and rainwater, the mine soil is acidified, and sewage (acidic water) will be generated in the overlapping part of the rainfall runoff area and the acidified soil body. The heavy metal ion concentration in the sewage is high and has certain toxicity, resulting in crop yield reduction or death of animals and plants in the flowing area. There are few relevant studies at home and abroad on how to reduce the generation of acidic water in the open-pit mine area by the lake from the source. At present, the traditional method for disposing of clean and sewage water will collect all the clean water and sewage. After being purified by the sewage treatment station, the water resources are reused or discharged into the outer lake again. This method converts clean water into sewage, thus increasing the water purification cost. If the distribution areas of clean and sewage water can be accurately distinguished and differential treatment is carried out, the water treatment cost can be reduced. In view of this, it is crucial to design a method for identifying the clean and sewage water areas. Summary of the Invention

[0003] The present invention aims to solve the technical problem that the distribution areas of clean water and sewage water cannot be clearly distinguished during the rainy season and flood season. A method for identifying clean and sewage water in the open-pit mine by the lake is provided, which can accurately and real-time monitor the flow conditions of clean water and sewage water in the entire mining area, and can predict in advance the flow conditions of clean and sewage water in the whole area according to the rainfall situation.

[0004] The present invention is achieved through the following technical solutions.

[0005] A method for identifying clean and sewage water in the open-pit mine by the lake according to the present invention is characterized by including the establishment of a three-dimensional digital model based on the initial database, the establishment of a clean and sewage water migration image database, and the construction of a clean and sewage water identification model. The research area is an open-pit mine by the lake, so the sewage storage capacity is limited. In the face of strong rainfall, the flood control and flood resistance pressure is extremely high.

[0006] Step S1: Establishment of a three-dimensional digital model based on the initial database

[0007] (1) Data collection: including: topography, geomorphology, geology, etc.; collecting data on existing clean and sewage diversion measures; collecting historical rainfall data and hydrological and other environmental data.

[0008] (2) UAV field scanning: Use a UAV equipped with a lidar to scan the ground surface, collect complex information such as the undulation changes of the ground surface, vegetation coverage, and water body distribution, and synchronously record the accurate geographical location information of each scanning point.

[0009] (3) Terrain surface modeling: Using the rich surface modeling tools in ArcGIS software, through precise interpolation calculations and reasonable meshing operations on terrain data, a high-precision digital elevation model (DEM) is generated.

[0010] (4) Geological body modeling: The collected underground stratigraphic data and geological structure data are transformed into a continuous and intuitive three-dimensional geological body model through complex triangulation construction, volume meshing, etc., so as to clearly and accurately display the spatial distribution characteristics of the underground geological structure.

[0011] (5) Multi-source information fusion and model generation: Deeply fuse the high-precision surface model generated by terrain surface modeling with the underground geological model generated by geological body modeling. By establishing a special data interface between GIS software and geological modeling software, efficient interaction and seamless integration of different model data are realized. On this basis, combined with the relevant information in the annual rainfall data, hydrographic environment data, and sewage and rainwater diversion measures data collected and sorted out before, the fused three-dimensional terrain and geological model is further corrected and improved.

[0012] Step S2: Establishment of a sewage and rainwater migration image database

[0013] (1) Simulate real environmental factors: In the hydrodynamic simulation software, according to the actual topographic and geomorphic parameters of the research area, such as slope, aspect, terrain undulation, etc., the model is precisely set to ensure that the simulation environment highly matches the actual terrain.

[0014] (2) Simulate the water flow under different rainfall conditions: Determine the range of rainfall intensities and the variation of durations that may occur in the research area during the rainy season, input these different combinations of rainfall intensities and durations, and obtain the rainfall-runoff generation and concentration paths under various rainfall scenarios.

[0015] (3) Collect sewage flow data during the rainfall period: According to the actual situation on-site in the open-pit mine area, reasonably arrange multiple representative electrochemistry detection points, and obtain information such as the sewage flow position and concentration in the open-pit mine area during the rainfall period in real time at the arranged detection points to ensure obtaining the dynamic change information of sewage flow.

[0016] (4) Construct a sewage and rainwater migration image database: After collecting the sewage flow data, organically combine it with the rainfall-runoff generation and concentration path data previously simulated by the fluid motion analysis software. Using image processing software, generate a series of images sorted by different rainfall intensities, durations, and time sequences, etc., and construct a rainy season sewage and rainwater migration image database.

[0017] Step S3: Construction of a sewage and rainwater identification model

[0018] (1) Image data preprocessing: The pixel values of the rainy-season clean and sewage migration images are uniformly mapped to a specific range. After denoising, enhancement, and normalization, the resulting images form a high-quality training sample library.

[0019] (2) Construction of clean and sewage migration prediction model: By writing a loop to traverse the images, each small-region image is extracted according to the set size and stored as a new data set. The hierarchical multi-region convolutional neural network method HMCNN is used to construct a rainy-season clean and sewage recognition model.

[0020] (3) Optimization of clean and sewage migration prediction model: Use real-time monitored water quality data to generate new images. The water quality parameters are converted into pixel values of the images according to a certain mapping rule to generate images reflecting the real-time water quality situation. Then, use the newly generated images to determine the error. The newly generated images are input into the trained model for prediction to obtain the predicted clean and sewage category results. Compare the prediction results with the actual water quality situation (i.e., the true label) to calculate the error. If the error is unacceptable, use the real-time monitored water quality data for continuous model training and parameter adjustment. Add the images generated from the real-time monitored water quality data to the training sample library and restart the training process. In this process, add multi-head attention for supervised retraining of the model. The specific principle is that after converting the data into vector form, the key vectors and the corresponding value vectors can be represented in pairs as [(K1, V1), (K2, V2), …, (Kn, Vn)]. The attention mechanism is expressed as:

[0021]

[0022] In the formula, Atten is the attention value, V is the value vector of the (Kn, Vn) pair. (Q, K) are the feedback vector and the key vector respectively, U represents the weight corresponding to V, and f represents the weight conversion function.

[0023] The self-attention mechanism uses three trainable parameters for measurement, namely W K 、W V and W Q , and V, K, and Q can be obtained from the input vector. For the weight transformation, the Softmax function is used to obtain the corresponding weights, which can be expressed in the following form:

[0024]

[0025] In the formula, d k represents the dimension number k.

[0026] The multi-head attention mechanism divides the self-attention mechanism into h parallel parts, each part is calculated using formula (3) to generate different weight matrices. The different weight matrices can be divided into h groups according to different heads, and then the outputs are connected through the corresponding weights to form a new output matrix.

[0027] MAtten(Q,K,V)=Concat(h1,...,h h )W o ,h i =Atten(Q i ,K i ,v i ) (4)

[0028]

[0029] In the formula, W i Q 、W i K and W i V are projected from Q, K, and V. W o is a parameter matrix that remaps the concatenation result to the d model dimension.

[0030] During the training process, the model will further adjust the weight parameters according to the newly added data to better adapt to the real and changeable rainy season environment. Repeat the above training and evaluation process, continuously optimize the model until the error reaches an acceptable range. If the error is acceptable, the model construction ends. At this time, the model can be used for the identification of clean and sewage water in the open-pit mine area by the lake, providing accurate identification results for the actual management and treatment of clean and sewage water.

[0031] The present invention is an innovative method. Compared with the prior art, in terms of model construction, the present invention integrates multi-source data, constructs a high-precision three-dimensional terrain and geological model through exploration data and on-site UAV lidar scanning data; the present invention facilitates the construction of corresponding clean and sewage diversion engineering measures based on the identification results by identifying clean and sewage water from the source; the present invention integrates the numerical simulation results and on-site sewage electrochemistry monitoring results to construct the clean and sewage water migration images of different pollution source types; the present invention uses a deep learning algorithm to construct a clean and sewage water identification model, which has the advantages of low learning cost and high prediction accuracy. Brief Description of the Drawings

[0032] Figure 1 is the overall operation flow chart of the present invention.

[0033] Figure 2 is the schematic diagram of the multi-head supervision mechanism during model retraining.

[0034] Figure 3It is a natural terrain plan view.

[0035] Figure 4 It is a schematic diagram of 3D terrain modeling.

[0036] Figure 5 It is a high-precision 3D model diagram integrating multi-source information.

[0037] Figure 6 It is a simulation diagram of surface runoff generation and concentration for the first rainfall.

[0038] Figure 7 It is a diagram of the migration of clean and sewage water during the first rainy season.

[0039] Figure 8 It is a schematic diagram of HMCNN modeling.

[0040] Figure 9 It is a diagram of the migration of clean and sewage water predicted by HMCNN.

[0041] Figure 10 It is an image of the migration of sewage water after retraining. Specific implementation manners

[0042] In order to comprehensively and clearly demonstrate the purpose, technical solution and its significant advantages of the present invention, the following will combine detailed drawings and specific practical cases to conduct a detailed analysis of the present invention. It should be emphasized that these practical cases are intended to assist in understanding the core idea of the present invention, rather than constituting a limitation on its scope of application or boundaries.

[0043] A method for identifying clean and sewage water in a lakeside open-pit mine described in this embodiment is divided into three main processes, including the process of establishing a 3D digital model based on an initial database, the process of establishing a database of clean and sewage water migration images, and the process of constructing a clean and sewage water identification model. The research area is a certain lakeside open-pit mine. Therefore, the ability to store sewage water is limited, and in the face of strong rainfall, the flood control and flood resistance pressure is extremely high.

[0044] In the process of establishing a 3D digital model based on the initial database, topographic and geomorphic maps, geological structure maps, existing data on sewage and rainwater diversion measures, historical rainfall data, hydrological data, etc. of the research area are obtained. The unmanned aerial vehicle and lidar equipment are used to conduct scanning operations during a period with clear weather and low wind. The TerraSolid software suite for point cloud processing is used to conduct meticulous preprocessing on the massive point cloud data collected by lidar, and key information closely related to the construction of the 3D topographic and geological model is extracted therefrom. The ArcGIS software is selected to import the preprocessed topographic data, including the digital elevation model (DEM) data generated after processing the lidar point cloud data, and detailed geomorphic information into the software to generate a high-precision digital elevation model (DEM). The GOCAD geological modeling software is used to orderly import the collected underground stratigraphic data and geological structure data into the software, and convert the discrete geological data into a continuous and intuitive 3D geological body model, so as to clearly and accurately display the spatial distribution characteristics of the underground geological structure.

[0045] In the establishment of the sewage and rainwater migration image database, based on the high-precision 3D topographic and geological model of the research area that has been constructed, the MIKE FLOOD fluid motion analysis software is selected to conduct simulation settings for different rainfall intensities and durations during the rainy season. By consulting the local meteorological data and historical rainfall data, the range of rainfall intensities and the variation of durations that may occur in the research area during the rainy season are determined. Different combinations of rainfall intensity and duration conditions are sequentially input into the software to obtain the rainwater runoff generation and concentration paths under various rainfall scenarios. Multiple representative electrochemical detection points are reasonably arranged to obtain representative sewage flow information. An electrochemical analyzer is used to measure the conductivity, pH value, and specific ion concentration indicators in the sewage to ensure the acquisition of dynamic change information on the sewage flow. Professional data processing algorithms and software tools are used to superimpose information such as the flow position and concentration of the sewage onto the rainwater runoff generation and concentration path data. The image processing toolbox of MATLAB is used to generate the sewage and rainwater migration images during the rainy season.

[0046] For the construction of the clear and sewage water identification model, preprocess the images of the clear and sewage water migration during the rainy season, including denoising, image enhancement, and normalization, to form a high-quality training sample library with the obtained images. Divide all the clear and sewage water migration images into multiple small regional images and store them as a new dataset. Use the hierarchical multi-region convolutional neural network method (HMCNN) to construct the clear and sewage water identification model during the rainy season. Generate new images using real-time monitored water quality data, and determine whether the error is acceptable by comparing the accuracy with a pre-set threshold. If the error is unacceptable, use the real-time monitored water quality data for continuous model training and parameter adjustment. Add the images generated from the real-time monitored water quality data to the training sample library and restart the training process. Continuously optimize the model until the error reaches an acceptable range. If the error is acceptable, the model construction is completed, and at this time, the model can be used for the identification of clear and sewage water in the open-pit mine area by the lake.

[0047] Specifically, in this embodiment, an area of an open-pit mine to be exploited is selected. The entire area is approximately 5 km 2 , and there are approximately 700,000 tons of copper ore underground. It is planned to extract the minerals in the form of open-pit mining. This area is surrounded by water on all sides. Due to environmental protection requirements, the heavy metal ion water during the mining process cannot be discharged into the surrounding public waters. Therefore, it is necessary to accurately identify the flow areas of clear water and sewage water to reduce the sewage treatment cost.

[0048] First, obtain the topographic and geomorphic maps, geological structure maps, rainfall, rainfall intensity, rainfall duration, rainfall frequency, etc. of the study area, and perform visualization processing based on the obtained data. As Figure 3 shown, the darker areas represent large rainfall runoff. Obtain comprehensive hydrological data of the study area, including dynamic data such as river water level, flow rate, velocity, and water quality, as well as static information such as groundwater level and aquifer characteristics. After completing the data collection, use a drone equipped with lidar to scan the terrain. Select a suitable model of lidar according to the size of the study area, the complexity of the terrain, and the required scanning accuracy. Select a time period with clear weather and low wind for scanning operations. After completing the data acquisition, use the point cloud processing software TerraSolid software suite to carefully preprocess the massive point cloud data collected by the lidar. Apply data filtering algorithms to remove the noise points generated due to measurement errors, instantaneous jitters of ground objects, etc. Through multiple filtering processes, improve the quality of the data and retain the effective data points that can truly reflect the surface characteristics. Finally, integrate and model the multi-source data to obtain Figure 5 .

[0049] Subsequently, based on the completed three-dimensional model of the research area, the fluid motion analysis software MIKE FLOOD was selected to conduct simulation settings for different rainfall intensities and durations during the rainy season. By referring to the local meteorological data and historical rainfall data: the rainfall intensity was divided into four levels: light rain, moderate rain, heavy rain, and rainstorm, and the corresponding rainfall intensity values were set respectively; light rain was 0.1 - 10 mm / h, and moderate rain was 10.1 - 25 mm / h. For the rainfall duration, 1 hour, 3 hours, and 6 hours were set. These different combinations of rainfall intensity and duration were sequentially input into the software, and the simulation program was started to obtain the rainwater generation and confluence paths under various rainfall scenarios. During the simulation process, the software calculated the flow direction, velocity, and confluence area information of the water flow on the ground according to the set terrain parameters and rainfall conditions, and generated detailed rainwater generation and confluence path data, such as Figure 6 shown. According to the actual situation on-site in the research area, multiple representative electrochemical detection points were reasonably arranged to ensure that the detection points could accurately obtain representative sewage flow information. An electrochemical analyzer was used to measure the conductivity, pH value, and specific ion concentration indicators in the sewage, and the geographical coordinate information of the sewage flow position was accurately recorded through a positioning device connected to the equipment. During the rainfall process, at a certain time interval, every 10 minutes, the data of each detection point was collected to ensure the acquisition of dynamic change information of the sewage flow. Through the Geographic Information System (GIS) software, the sewage concentration information was visually displayed on the rainwater generation and confluence path map in the form of different colors or concentration gradients. Using the image processing toolbox of the image processing software MATLAB, the rainwater and sewage migration images during the rainy season were generated. The results of one simulation were as Figure 7 shown.

[0050] Then, preprocess the rainy-season clean and sewage migration images to obtain a high-quality training sample library. For the collected rainy-season clean and sewage migration images, the first step is to perform denoising. Since noise may be introduced during image acquisition and transmission, affecting subsequent recognition results, the Gaussian filtering algorithm can be selected. This algorithm effectively smooths the image and removes common noise types such as Gaussian white noise by performing weighted averaging on each pixel point and its neighborhood in the image. In Python, with the help of the OpenCV library, the cv2.GaussianBlur() function can be used to implement Gaussian filtering. For example, set the convolution kernel size to (5,5) and the standard deviation to 0. Perform image enhancement to improve the contrast and clarity of the image, making the characteristics of clean and sewage more obvious. Histogram equalization is a commonly used method. It expands the gray range of the image and enhances the overall contrast by redistributing the pixel values of the image. In OpenCV, for grayscale images, the cv2.equalizeHist() function can be used. For color images, they can be first converted to the HSV color space, perform histogram equalization on the V channel, and then convert back to the RGB space. Normalize the image, mapping the pixel values of the image to a specific range, such as [0,1] or [-1,1]. This helps to accelerate the convergence speed of the neural network and improve the stability of the model. For an image with pixel values in the range of 0 to 255, if it is to be normalized to [0,1], each pixel value can be divided by 255 through simple mathematical operations, that is, pixel = pixel / 255.0. After denoising, enhancement, and normalization, the resulting images form a high-quality training sample library. Input the processed clean and sewage migration images with different rainfall intensities, durations, rainfall types, and different pollution source types into the model construction process. Divide all clean and sewage migration images into multiple small-region images. This step helps the neural network better capture the local features in the image. The appropriate small-region size can be selected according to the size and feature complexity of the image. For example, set it to 64×64 pixels. In Python, by writing a loop to traverse the image, extract each small-region image according to the set size, and store it as a new dataset.

[0051] Use the hierarchical multi-region convolutional neural network method (HMCNN) to construct a rainy-season clean and sewage recognition model. The overall process is as Figure 8As shown below. First, determine the hierarchical structure of the network, which generally includes multiple convolutional layers, pooling layers, fully connected layers, etc. In the convolutional layer, set convolutional kernels of different sizes, such as 3×3 or 5×5, and extract different levels of features of the image through convolutional operations. For example, the first convolutional layer can be set with 32 3×3 convolutional kernels to capture the primary features of the image. The pooling layer usually uses max pooling or average pooling to reduce the data dimension while retaining important features, such as setting the pooling kernel size to 2×2 and the stride to 2. The fully connected layer synthesizes the features extracted by the previous layers and finally outputs the classification result. By stacking these layers in sequence, a complete HMCNN model framework is constructed. During the training process of HMCNN, it is crucial to select appropriate optimization algorithms and loss functions. The optimization algorithm can choose the Adam optimizer, which combines the advantages of Adagrad and RMSProp, can adaptively adjust the learning rate, and accelerate the model convergence speed. In the deep learning framework PyTorch, the Adam optimizer can be easily called using the corresponding function. The loss function can choose the cross-entropy loss function. For multi-classification problems, the cross-entropy loss function can effectively measure the difference between the model prediction result and the true label. During the training process, the model adjusts the weight parameters of the network continuously according to the calculation result of the loss function to minimize the loss value.

[0052] After the model training is completed, use the real-time monitored water quality data to generate new images. This step requires converting the real-time monitored water quality parameters, such as pH value, pollutant concentration, etc., into pixel values of the image according to a certain mapping rule to generate an image reflecting the real-time water quality situation, such as Figure 9 As shown below. Then, use the newly generated image to determine the error. Input the newly generated image into the trained model for prediction to obtain the predicted result of the clean and sewage water categories. Compare the prediction result with the actual water quality situation (i.e., the true label), calculate the ratio of the number of correctly predicted samples to the total number of samples to obtain the accuracy rate, and judge whether the error is acceptable by comparing the accuracy rate with the pre-set threshold. If the error is not acceptable, use the real-time monitored water quality data for continuous model training and parameter adjustment. Add the image generated from the real-time monitored water quality data to the training sample library and restart the training process. During the training process, the model will further adjust the weight parameters according to the newly added data to better adapt to the real and changeable rainy season environment. Repeat the above training and evaluation process, and add multi-head attention for supervised retraining of the model during this process, continuously optimize the model until the error reaches an acceptable range. If the error is acceptable, the model construction ends. At this time, this model can be used for the identification of clean and sewage water in the open-pit mine area by the lake, providing accurate identification results for the actual management and treatment of clean and sewage water. After obtaining the HMCNN model that meets the accuracy, set the rainfall amount to 25 mm / h and the rainfall time to 2 h. The prediction result integrating all information is shown in Figure 10。

Claims

1. A method for identifying clear and sewage water in a lakeside open-pit mine, characterized in that, Including the following steps: Step S1: Establishment of the three-dimensional digital model based on the initial database: (1) Data collection: including topography, geomorphology, and geology; collection of existing sewage and rainwater diversion measures data; collection of annual rainfall data and hydro-environmental data; (2) UAV field scanning: Use a UAV equipped with lidar for surface scanning, collect information on surface undulations, vegetation cover, and water body distribution, and synchronously record the precise geographical location information of each scanning point; (3) Topographic surface modeling: Use ArcGIS to perform precise interpolation calculations and reasonable meshing operations on the topographic data to generate a high-precision digital elevation model; (4) Geological body modeling: The collected underground stratum data and geological structure data are transformed into a continuous and intuitive three-dimensional geological body model through complex triangular mesh construction and volume meshing operations, so as to clearly and accurately display the spatial distribution characteristics of the underground geological structure; (5) Multi-source information fusion and model generation: Deeply fuse the high-precision surface model generated by topographic surface modeling with the underground geological model generated by geological body modeling; by establishing a special data interface between GIS software and geological modeling software, realize the efficient interaction and seamless integration of different model data; on this basis, combine the relevant information in the annual rainfall data, hydro-environmental data, and sewage and rainwater diversion measures data collected and sorted out before, and further correct and improve the fused three-dimensional topographic and geological model; Step S2: Establishment of the sewage and rainwater migration image database: (1) Simulate real environmental factors: In the hydrodynamic simulation software, accurately set the model according to the actual topographic and geomorphic parameters of the study area: slope, aspect, and terrain undulation, to ensure that the simulation environment highly coincides with the actual terrain; (2) Simulate the water flow under different rainfall conditions: Determine the range of rainfall intensity and duration variation that may occur in the study area during the rainy season, input these different combinations of rainfall intensity and duration conditions, and obtain the rainfall-runoff generation and concentration paths under various rainfall scenarios; (3) Collect sewage flow data during the rainfall period: According to the actual situation on the site of the open-pit mine area, reasonably arrange multiple representative electrochemistry detection points, and obtain the sewage flow position and concentration information in the open-pit mine area during the rainfall period in real time at the arranged detection points to ensure the acquisition of dynamic change information of sewage flow; (4) Construct the sewage and rainwater migration image database: After collecting the sewage flow data, organically combine it with the rainfall-runoff generation and concentration path data simulated by the fluid motion analysis software before; use image processing software to generate a series of images sorted by different rainfall intensities, durations, and time sequences, and construct a sewage and rainwater migration image database during the rainy season; Step S3: Construction of the sewage and rainwater identification model: (1) Image data preprocessing: Uniformly map the pixel values of the sewage and rainwater migration images during the rainy season to a specific range, and after denoising, enhancement, and normalization processing, the obtained images form a high-quality training sample library; (2) Construction of the prediction model for the migration of clean and sewage water: By writing a loop to traverse the image, extract the image of each small area according to the set size, and store it as a new data set; Use the hierarchical multi-region convolutional neural network method HMCNN to construct a rainy-season clean and sewage water recognition model; (3) Optimization of the prediction model for the migration of clean and sewage water: Use real-time monitored water quality data to generate new images; Convert water quality parameters into pixel values of the image according to a certain mapping rule to generate an image reflecting the real-time water quality situation; Then, use the newly generated image to determine the error; Input the newly generated image into the trained model for prediction to obtain the predicted results of the clean and sewage water categories; Compare the prediction results with the actual water quality situation and calculate the error; If the error is unacceptable, use the real-time monitored water quality data for continuous model training and parameter adjustment; Add the image generated from the real-time monitored water quality data to the training sample library and restart the training process; Add multi-head attention for supervised re-training of the model during this process: After converting the data into vector form, the key vector and the corresponding value vector are represented in pairs as [(K1, V1), (K2, V2), …, (Kn, Vn)], and the attention mechanism is expressed as: (In the formula, Atten is the attention value, V is the value vector of the (Kn, Vn) pair; (Q, K) are the feedback vector and the key vector respectively, U represents the weight corresponding to V, and f represents the weight conversion function; The self-attention mechanism utilizes three trainable parameter metrics: W K , W V and W Q , and obtains V, K, and Q through the input vector; for weight transformation, the Softmax function is used to obtain the corresponding weights, which can be expressed in the following form: where d k represents the number of dimensions k; (The multi-head attention mechanism divides the self-attention mechanism into h parallel parts, and each part is calculated using formula (3) to generate different weight matrices. The different weight matrices can be divided into h groups according to different heads, and then the outputs are connected through the corresponding weights to form a new output matrix; MAtten(Q,K,V)=Concat(h1,...,h h )W o ,h i =Atten(Q i ,K i ,v i ) (4) where, W i Q , W i K and W i V are projected from Q, K, and V; W o is a parameter matrix that remaps the concatenation result to the d model dimension; (During the training process, the model further adjusts the weight parameters according to the newly added data to better adapt to the changing environment in the real rainy season; Repeat the above training and evaluation process to continuously optimize the model until the error reaches an acceptable range; If the error is acceptable, the model construction ends, and this model can be used for the identification of clean and sewage water in the open-pit mine area by the lake, providing accurate identification results for the actual management and treatment of clean and sewage water.

Images