A method for open-pit regional function classification based on deep regional embedding

Abstract

The application discloses a kind of open pit regional function classification method based on depth area embedding, belong to open pit management field, first to open pit uniform unit area division, then the feature information of different unit area is extracted by trajectory data, and different unit area is constructed into a sparse unit area graph;Second, the spatial embedding representation of unit area node is learned using structured depth network embedding model, and the spatial area embedding of unit area is realized;Then, the unit area time sequence feature is obtained by taking the unit area feature time sequence as the input of bidirectional long short-term memory neural network, the unit area time sequence embedding is realized, to make up the problem of insufficient area time characteristics;Second, the embedding representation of unit area node obtained and unit area sequence feature are fused, and the final feature fusion result is input into attention layer to solve the problem of uneven distribution of different feature weights;Finally, more accurate classification results are obtained by training.

Description

Technical Field

[0001] This invention relates to the field of open-pit mine management, and in particular to a method for functional classification of open-pit mine areas based on deep region embedding. Background Technology

[0002] In today's society, various geographic information acquisition technologies, especially "3S" technologies—namely, Remote Sensing Satellite (RS), Global Positioning System (GPS), and Geographic Information System (GIS)—are developing rapidly, making the collection and storage of massive amounts of geographic and traffic data increasingly simple. Open-pit mines are large-scale production systems integrating data collection and transportation. Trucks and excavators within these mines are equipped with GPS devices, generating a large amount of trajectory data daily. With the development of technology, more and more open-pit mines are equipped with production scheduling systems. These systems rationally allocate truck and excavator pairings within the mine through internal calculations, dispatching orders to trucks in an optimal manner, thus reducing operating costs and improving mining efficiency. By accurately identifying functional areas, the scheduling system can quickly and rationally adjust truck status. Since open-pit mines have multiple mining areas and multiple spoil heaps, and these areas are constantly changing over time, the current method of manually marking mining areas and spoil heaps suffers from problems such as high manpower consumption, untimely updates, and inaccurate data. Therefore, using trajectory data from open-pit mines to classify functional areas is crucial.

[0003] Open-pit mines have multiple functional areas, such as mining areas and unloading areas. Different functional areas perform different duties. Previous research has largely focused on the temporal characteristics of GPS trajectories, which are determined by travel patterns. Different functional areas exhibit different temporal characteristics at different times; for example, on weekday mornings, people travel from residential areas to office areas. However, in the open-pit mine scenario, since trucks work according to dispatch times, their temporal information is insufficient to clearly identify the functional areas. But trucks pass through multiple functional areas during a single dispatch, and these areas are spatially interconnected based on truck trajectories. Therefore, accurately classifying the functional areas of open-pit mines using truck trajectory data is a key research focus for optimizing open-pit mine truck dispatching systems.

[0004] GPS trajectories are among the most important data for regional function identification. They primarily include the latitude and longitude of passengers getting on and off the vehicle, as well as the time of arrival and departure. Previous studies have largely identified urban functions by examining the origin and departure points of taxis. However, since these trajectories are typically continuous, taxi trajectories only consider the starting point and ignore the regional information contained in the intermediate journey, thus neglecting global features. Furthermore, urban taxi trajectories have a wider coverage than open-pit mine truck trajectories, while the coverage of mining area truck trajectories is relatively sparse, posing a challenge to the identification of mining area functions. Summary of the Invention

[0005] The purpose of this invention is to provide a method for functional classification of open-pit mine areas based on deep region embedding, which solves the problems of labor-intensive, untimely updates, and inaccurate data in the existing technology of manually marking open-pit mine areas for identifying mining areas and spoil heaps.

[0006] The technical solution adopted in this invention is as follows: a functional classification method for open-pit mine areas based on deep region embedding. First, the open-pit mine is divided into uniform unit regions. Then, feature information of different unit regions is extracted through trajectory data, and different unit regions are constructed into a sparse unit region map. Next, a structured deep network embedding model is used to learn the spatial embedding representation of unit region nodes, realizing spatial region embedding of unit regions. Then, the temporal sequence of unit region features is used as input to a bidirectional long short-term memory neural network to obtain the temporal features of the unit regions, realizing temporal region embedding of unit regions to compensate for the lack of regional temporal features. Next, the obtained embedding representation of unit region nodes and the unit region sequence features are fused, and the final feature fusion result is input into an attention layer to solve the problem of uneven weight distribution of different features. Finally, training is performed to obtain more accurate classification results.

[0007] The specific steps are as follows:

[0008] Step S1: Preprocess the truck GPS trajectory data and open-pit mining area in the open-pit mine area. Clean the original open-pit mine trajectory data and divide the open-pit mine area into unit areas. Extract the features of the unit areas and use the existing open-pit mine area functions as label data. That is, each unit area corresponds to a mining area function, which includes mining area, unloading area and slag yard.

[0009] Step S2: Construct a trajectory cell region map using the preprocessed trajectory and cell region data from step S1;

[0010] Step S3: Input the trajectory unit region map obtained in step S2 into the structured deep network embedding model, and then use unsupervised training to train the structured deep network embedding model. After training, generate the trained unit region embedding model, and then obtain the low-dimensional representation vector of the unit region.

[0011] Step S4: Based on the unit region features obtained in step S1, construct a unit region time series sequence, and use a bidirectional long short-term memory neural network to extract the time series features of the unit region units from the time series sequence.

[0012] Step S5: The low-dimensional representation vector of the unit region obtained in step S3 and the temporal features of the unit region extracted in step S4 are fused together. The final feature vector is then input into the attention layer to adjust the feature weights. Finally, it is input into the fully connected layer and the existing classifier softmax is trained to obtain the trained classification model.

[0013] Step S6: Input the final feature vector corresponding to the region in step S5 into the final classification model described in step S5 to identify the mining area function corresponding to the study area.

[0014] Furthermore, step S1 specifically includes:

[0015] Step S11: Clean the GPS trajectory data of the mining area of ​​points that are not in the area to be identified and invalid points;

[0016] Step S12: After data cleaning, the GPS trajectory data of the mining area is matched to the corresponding open-pit mining area using a map matching algorithm to generate unit area feature data within the area.

[0017] Step S13: Divide the open-pit mine area into several unit areas.

[0018] Furthermore, the specific steps for dividing the open-pit mine area into several unit area data as described in step S13 are as follows:

[0019] The open-pit mine area was evenly divided into Non-overlapping unit regions, ,in Indicates an open-pit mine area. This represents the i-th unit region after uniform division, constructed by equally dividing the latitude and longitude space. , ,in The table lists the number of unit regions at latitude and longitude, respectively. Indicates the maximum latitude. Indicates the minimum latitude. Indicates the maximum longitude. This indicates the minimum longitude, so That is, the number of unit regions is Each unit area Defined by four boundary points, its variables have the same meaning as the latitude and longitude mentioned above.

[0020] Furthermore, in step S2, the specific steps are: mapping the trajectory to the unit region to obtain the unit region features, and modeling the unit region structure as a graph. , where the set of nodes Represents a unit region, edge set Indicates the connection relationship between two unit regions; unit region trajectory diagram adjacency matrix ,in Represents a cell region and Is it connected?

[0021]

[0022] Furthermore, step S3 specifically includes:

[0023] Step S31: Obtain the cell region trajectory map obtained in step S2. The input is fed into a structured deep network embedding model; the structured deep network embedding model consists of an encoder and a decoder; the encoder is responsible for mapping the input data to a low-dimensional vector, and the decoder is responsible for mapping the low-dimensional vector to the original representation space; the deep network structure of the structured deep network embedding model better perceives the trajectory unit region map. The network simultaneously perceives the local and global structure of the graph by jointly optimizing first-order similarity and second-order similarity constraints, thus mitigating the impact on the trajectory unit region graph. The negative impact of sparsity in the structure;

[0024] Step S32: First-order similarity assumes two unit regions are similar if they are connected in the unit region trajectory network; second-order approximation assumes two unit regions are similar if their neighborhood structures are similar, even if they are not connected in the unit region trajectory network. Second-order proximity enriches the regional relationships in the unit region trajectory network, enabling the structured deep network embedding model to capture the global structure of the unit region trajectory network; the encoder of the structured deep network embedding model is given the neighborhood structure of the unit region. The encoder process is represented as follows:

[0025]

[0026] in It is a learnable parameter matrix. It is a bias term. It is the Sigmoid activation function;

[0027] Step S33: The decoder is the reverse process of the encoder, thus obtaining the cell region. Output The goal of autoencoder construction is to minimize the reconstruction loss between the input and output data. Since the input is the neighborhood structure of each unit region, the reconstruction process will cause unit regions with similar neighborhood structures to tend towards similar low-dimensional representations. Therefore, the loss function for second-order similarity is defined as:

[0028]

[0029] in This represents element-wise multiplication. Because the road network in an open-pit mine is sparse, the adjacency matrix... There are many zero elements; to make the model more inclined to reconstruct non-zero elements, vectors are used. Specifically, if ,but 0, otherwise ;

[0030] The loss function for the structured deep network embedding model is obtained by concatenating the loss functions of first-order similarity and second-order similarity, and is defined as follows:

[0031]

[0032] in It's a hyperparameter. It is the parameter matrix in the decoding process; the last term in the equation is the regularization term to prevent overfitting; after minimizing the loss function, a low-dimensional representation vector set of the unit region is obtained, thus preserving the local and global structure of the unit region trajectory map.

[0033] Furthermore, in step S4, the specific steps are as follows:

[0034] In step S4, the ordinary long short-term memory neural network uses units and hidden states to store long-term and short-term memories of past sequence information, respectively, based on the input sequence. Each element in It will be processed by the recurrent long short-term memory neural network unit, along with the previous unit state. and hidden state The information will be controlled by three specially designed gates, namely the input gate. Forgotten Gate and output gate To obtain the current unit state and hidden output state The specific calculation formula is as follows:

[0035]

[0036] in It is the sigmoid activation function. Represents element-wise product. and These represent the corresponding weights and biases to be learned, respectively. In the proposed framework, the bidirectional long short-term memory neural network consists of two independent long short-term memory neural networks that can summarize the forward and backward information of the temporal sequence of the unit region, and then merge the information from both directions. For each time... The feedforward long short-term memory neural network is based on the previous hidden vector. and the feature vector of the current time input in the input cell region sequence. To calculate the current hidden vector Backward Long Short-Term Memory Neural Networks are based on the reverse of the last hidden vector. and the current input feature vector To calculate the hidden vector Then, the forward hidden vector... and backward hidden vector The final hidden vector merged into a bidirectional long short-term memory neural network model In a bidirectional long short-term memory neural network model, the parameters in two opposite directions are independent, but they share the input of the same unit region's temporal sequence. (Open-pit mine unit region) The time series samples are ,in Each time step has four normalized features: outflow, inflow, point density, and average velocity. This is the sequence length. Therefore, recurrent neural network layers need to be forward and backward... A long short-term memory neural network unit. As shown in the figure. . and These represent the forward hidden vector and the backward hidden vector, respectively. Indicates by and The final hidden vector of the connection.

[0037]

[0038] Among them, hidden state values , , This represents the number of hidden neurons in a Long Short-Term Memory (LSTM) neural network unit. (Function) It is the transition function of a recurrent network. Internally, it uses input gate, forget gate and output gate mechanisms to obtain the current cell state and hidden state, so as to maintain long-term dependencies in the time sequence.

[0039] Furthermore, the specific training steps described in step S5 are as follows:

[0040] Step S51: Embed the low-dimensional representation of the cell region. and the final hidden vector representation in the temporal features of the unit region Feature concatenation is performed, and the matrix is ​​input into the attention layer. An attention mechanism is used to assign weights to the feature matrix. To ensure consistent input and output dimensions when assigning different weights without losing feature information, a self-attention mechanism is selected in the attention layer.

[0041]

[0042] in These are three distinct vectors for each feature, obtained by transforming the feature vectors... Multiply , ,vector weight matrix Obtained; It is a fraction; gradient stability is achieved by dividing by... Perform normalization; use The function is used as an activation function, multiplied by To obtain the final result ;

[0043] Step S52: Input the obtained attention-weighted feature matrix into a multilayer perceptron for further learning, then input it into a fully connected layer to adjust the dimensions, and finally use it. Classify functions;

[0044]

[0045] In model training, the cross-entropy loss function is used; for a sample size of... The training set and the loss function are as follows:

[0046]

[0047] in This represents all the parameters involved in training the model. It is a regularization term; parameters are learned using the Adam stochastic gradient descent optimizer and backpropagation.

[0048] Step S53: After the above steps, a deep neural network classification model for functional classification of open-pit mining unit areas is finally obtained.

[0049] In step S6, the deep neural network classification model is trained in step S5. By inputting the feature vector of the open-pit mine unit area to be identified into the model, a unit area function with high accuracy can be obtained and applied to the open-pit mine area.

[0050] The beneficial effects of this invention are as follows: This invention fully considers the basic details of GPS trajectories of vehicles in mining areas and the topological relationships between unit areas in open-pit mines. Based on the temporal relationships within unit areas, it can significantly improve the classification effect. Using trajectory data from open-pit mines for functional area classification is crucial.

[0051] (1) This method utilizes the advantages of structured deep network embedding models in maintaining local and global features of road unit area maps and robustness to sparse networks to obtain low-dimensional representation vectors of unit area nodes in the map, which improves the recognition results and makes model building simpler.

[0052] (2) It takes into account both the spatial and temporal characteristics of the open-pit mine area and uses an attention mechanism to assign weights to different features, thus overcoming the problem of insufficient regional representation caused by single-dimensional features.

[0053] (3) It can effectively combine historical truck GPS trajectory data and open-pit mine area data for the study of open-pit mine functional classification, fully explore the connection between trajectory and open-pit mine function, and significantly improve classification accuracy.

[0054] Based on the above effects, this invention solves the problems of labor-intensive, untimely updates, and inaccurate data associated with the manual marking of functional zones in existing open-pit mines for identifying mining areas and spoil heaps. Thus, it achieves the objective of this invention. Attached Figure Description

[0055] Figure 1 This is a flowchart of a functional classification method for open-pit mine areas based on deep region embedding.

[0056] Figure 2 A map is constructed for the trajectory unit region. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0058] refer to Figure 1 This is a flowchart of the method of the present invention. The present invention proposes a functional classification method for open-pit mine areas based on deep region embedding, which includes the following steps:

[0059] Step S1: Preprocess the GPS trajectory data of trucks in the open-pit mine and the open-pit mine area. Clean the original open-pit mine trajectory data, divide the open-pit mine area into unit areas, extract the features of the unit areas, and assign different functional labels to different unit areas according to the actual situation of the mine area to facilitate later training.

[0060] The specific implementation steps are as follows:

[0061] Preprocessing was performed on the GPS trajectory data of trucks within the open-pit mine area and the open-pit mine area itself. Data cleaning was performed on the original trajectory data of 45 trucks from the open-pit mine, removing data from the GPS trajectory data that was not in the area to be identified and from invalid points. Figure 2 As shown, the GPS trajectory data of the mining area after data cleaning is matched to the corresponding open-pit mining area using a map matching algorithm to generate unit area feature data. The open-pit mining area is then divided into uniform unit areas, generating several uniform unit area data according to 150m×70m. The open-pit mining area is further divided into unit areas to obtain 400×45 unit areas. Unit area features are extracted, with the time period divided into half-hour segments. The features obtained within half an hour include the number of trajectory points, outflow, inflow, and average speed within the unit area.

[0062] The specific steps for dividing the open-pit mine area into several unit area data are as follows: The open-pit mine area is evenly divided into... Non-overlapping unit regions, ,in Indicates an open-pit mine area. This represents the i-th unit region after uniform division, constructed by equally dividing the latitude and longitude space. , ,in The table lists the number of unit regions at latitude and longitude, respectively. Indicates the maximum latitude. Indicates the minimum latitude. Indicates the maximum longitude. This indicates the minimum longitude, so That is, the number of unit regions is Each unit area Defined by four boundary points, its variables have the same meaning as the latitude and longitude mentioned above.

[0063] Step S2: Construct a trajectory unit region map using the preprocessed trajectory and unit region data from step S1.

[0064] The specific steps are: mapping the trajectory to the unit region to obtain the unit region features, and modeling the unit region structure as a graph. , where the set of nodes Represents a unit region, edge set Indicates the connection relationship between two unit regions; unit region trajectory diagram adjacency matrix ,in Represents a cell region and Is it connected?

[0065]

[0066] Step S3: Input the trajectory unit region map obtained in step S2 into the structured deep network embedding model; then train the embedding model in unsupervised training. After training, generate the trained unit region embedding model and obtain the low-dimensional representation vector of the unit region.

[0067] Step S3 is as follows:

[0068] Step S31: Obtain the cell region trajectory map obtained in step S2. The input is fed into a structured deep network embedding model; the structured deep network embedding model consists of an encoder and a decoder; the encoder is responsible for mapping the input data to a low-dimensional vector, and the decoder is responsible for mapping the low-dimensional vector to the original representation space; the deep network structure of the structured deep network embedding model better perceives the trajectory unit region map. The network simultaneously perceives the local and global structure of the graph by jointly optimizing first-order similarity and second-order similarity constraints, thus mitigating the impact on the trajectory unit region graph. The negative impact of sparsity in the structure;

[0069] Step S32: First-order similarity assumes two unit regions are similar if they are connected in the unit region trajectory network; second-order approximation assumes two unit regions are similar if their neighborhood structures are similar, even if they are not connected in the unit region trajectory network. Second-order proximity enriches the regional relationships in the unit region trajectory network, enabling the structured deep network embedding model to capture the global structure of the unit region trajectory network; the encoder of the structured deep network embedding model is given the neighborhood structure of the unit region. The encoder process is represented as follows:

[0070]

[0071] in It is a learnable parameter matrix. It is a bias term. It is the Sigmoid activation function;

[0072] Step S33: The decoder is the reverse process of the encoder, thus obtaining the cell region. Output The goal of autoencoder construction is to minimize the reconstruction loss between the input and output data. Since the input is the neighborhood structure of each unit region, the reconstruction process will cause unit regions with similar neighborhood structures to tend towards similar low-dimensional representations. Therefore, the loss function for second-order similarity is defined as:

[0073]

[0074] in This represents element-wise multiplication. Because the road network in an open-pit mine is sparse, the adjacency matrix... There are many zero elements; to make the model more inclined to reconstruct non-zero elements, vectors are used. Specifically, if ,but 0, otherwise ;

[0075] The loss function for the structured deep network embedding model is obtained by concatenating the loss functions of first-order similarity and second-order similarity, and is defined as follows:

[0076]

[0077] in It's a hyperparameter. It is the parameter matrix in the decoding process; the last term in the equation is the regularization term to prevent overfitting; after minimizing the loss function, a low-dimensional representation vector set of the unit region is obtained, thus preserving the local and global structure of the unit region trajectory map.

[0078] Step S4: Based on the unit region features obtained in step S1, construct a unit region time series sequence, and use a bidirectional long short-term memory neural network to extract the time series features of the unit region units from the time series sequence.

[0079] The specific steps are as follows:

[0080] In step S4, the ordinary long short-term memory neural network uses units and hidden states to store long-term and short-term memories of past sequence information, respectively, based on the input sequence. Each element in It will be processed by the recurrent long short-term memory neural network unit, along with the previous unit state. and hidden state The information will be controlled by three specially designed gates, namely the input gate. Forgotten Gate and output gate To obtain the current unit state and hidden output state The specific calculation formula is as follows:

[0081]

[0082] in It is the sigmoid activation function. Represents element-wise product. and These represent the corresponding weights and biases to be learned, respectively. In the proposed framework, the bidirectional long short-term memory neural network consists of two independent long short-term memory neural networks that can summarize the forward and backward information of the temporal sequence of the unit region, and then merge the information from both directions. For each time... The feedforward long short-term memory neural network is based on the previous hidden vector. and the feature vector of the current time input in the input cell region sequence. To calculate the current hidden vector Backward Long Short-Term Memory Neural Networks are based on the reverse of the last hidden vector. and the current input feature vector To calculate the hidden vector Then, the forward hidden vector... and backward hidden vector The final hidden vector merged into a bidirectional long short-term memory neural network model In a bidirectional long short-term memory neural network model, the parameters in two opposite directions are independent, but they share the input of the same unit region's temporal sequence. (Open-pit mine unit region) The time series samples are ,in Each time step has four normalized features: outflow, inflow, point density, and average velocity. This is the sequence length. Therefore, recurrent neural network layers need to be forward and backward... A long short-term memory neural network unit. As shown in the figure. . and These represent the forward hidden vector and the backward hidden vector, respectively. Indicates by and The final hidden vector of the connection.

[0083]

[0084] Among them, hidden state values , , This represents the number of hidden neurons in a Long Short-Term Memory (LSTM) neural network unit. (Function) It is the transition function of a recurrent network. Internally, it uses input gate, forget gate and output gate mechanisms to obtain the current cell state and hidden state, so as to maintain long-term dependencies in the time sequence.

[0085] Step S5: The low-dimensional representation vector of the unit region obtained in step S3 and the temporal features of the unit region extracted in step S4 are fused together. The final feature vector is then input into the attention layer to adjust the feature weights. Finally, it is input into the fully connected layer and the existing classifier softmax is trained to obtain the trained classification model.

[0086] The specific training steps are as follows:

[0087] Step S51: Embed the low-dimensional representation of the cell region. and the final hidden vector representation in the temporal features of the unit region Feature concatenation is performed, and the matrix is ​​input into the attention layer. An attention mechanism is used to assign weights to the feature matrix. To ensure consistent input and output dimensions when assigning different weights without losing feature information, a self-attention mechanism is selected in the attention layer.

[0088]

[0089] in These are three distinct vectors for each feature, obtained by transforming the feature vectors... Multiply , ,vector weight matrix Obtained; It is a fraction; gradient stability is achieved by dividing by... Perform normalization; use The function is used as an activation function, multiplied by To obtain the final result ;

[0090] Step S52: Input the obtained attention-weighted feature matrix into a multilayer perceptron for further learning, then input it into a fully connected layer to adjust the dimensions, and finally use it. Classify functions;

[0091]

[0092] In model training, the cross-entropy loss function is used; for a sample size of... The training set and the loss function are as follows:

[0093]

[0094] in This represents all the parameters involved in training the model. It is a regularization term; parameters are learned using the Adam stochastic gradient descent optimizer and backpropagation.

[0095] Step S53: After the above steps, a deep neural network classification model for functional classification of open-pit mining unit areas is finally obtained.

[0096] Step S6: Input the final feature vector corresponding to the region in step S5 into the trained deep neural network classification model to identify the mining area function corresponding to the study area.

[0097] In step S6, the deep neural network classification model is trained in step S5. By inputting the feature vector of the open-pit mine unit area to be identified into the model, a unit area function with high accuracy is obtained and applied to the open-pit mine area.

[0098] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for functional classification of open-pit mine areas based on deep region embedding, characterized by: Step S1: Preprocess the truck GPS trajectory data and open-pit mining area. Clean the original open-pit mine trajectory data and divide the open-pit mining area into unit regions. Extract the features of each unit region and use the existing manually divided regional functions as label data. Each unit region corresponds to a mining area function, including mining area, unloading area, and slag yard. Step S2: Construct a trajectory unit region map using the preprocessed trajectory and unit region data from Step S1. Step S3: Input the trajectory unit region map obtained in Step S2 into a structured deep network embedding model. Then, train the structured deep network embedding model using unsupervised training. After training, generate the trained unit region embedding model, thus obtaining the low-dimensional representation vector of the unit region. Step S4: Based on the unit region features obtained in step S1, a unit region time series is constructed, and a bidirectional long short-term memory neural network is used to extract the time series features of the unit regions from the time series. In step S5, the low-dimensional representation vector of the unit region obtained in step S3 and the unit region time series features extracted in step S4 are fused, and the final feature vector is input to the attention layer to adjust the feature weights. Finally, it is input to the fully connected layer and the existing classifier softmax is trained to obtain a trained classification model. In step S6, the final feature vector corresponding to the region in step S5 is input to the final classification model described in step S5 to identify the mining area function corresponding to the study area. Specifically, in step S2, the trajectory is mapped to the unit region to obtain the unit region features, and the unit region structure is modeled as a graph. , where the set of nodes Represents a unit region, edge set Indicates the connection relationship between two unit regions; unit region trajectory diagram adjacency matrix ,in Represents a cell region and Is it connected? The specific steps of step S3 are: Step S31, taking the unit region trajectory map obtained in step S2... Input into a structured deep network embedding model; The structured deep network embedding model consists of an encoder and a decoder; the encoder is responsible for mapping the input data to a low-dimensional vector, and the decoder is responsible for mapping the low-dimensional vector to the original representation space. Structured deep network embedding models have deeper network structures that better perceive trajectory unit region maps. The network simultaneously perceives the local and global structure of the graph by jointly optimizing first-order similarity and second-order similarity constraints, thus mitigating the impact on the trajectory unit region graph. The negative impact of sparsity in the structure; Step S32: First-order similarity assumes that two unit regions are similar if they are connected in the unit region trajectory network; Second-order approximation assumes that two unit regions are similar if their neighborhood structures are similar, even if they are not connected in the unit region trajectory network; Second-order proximity enriches the regional relationships in the unit region trajectory network, enabling the structured deep network embedding model to capture the global structure of the unit region trajectory network; The encoder of a structured deep network embedding model has a neighborhood structure for a given cell region. The encoder process is represented as follows: in It is a learnable parameter matrix. It is a bias term. It is the Sigmoid activation function; step S33, the decoder is the inverse process of the encoder, thus obtaining the cell region. Output The goal of autoencoder construction is to minimize the reconstruction loss between the input and output data. Since the input is the neighborhood structure of each unit region, the reconstruction process will cause unit regions with similar neighborhood structures to tend towards similar low-dimensional representations. Therefore, the loss function for second-order similarity is defined as: in This represents element-wise multiplication. Because the road network in an open-pit mine is sparse, the adjacency matrix... There are many zero elements; to make the model more inclined to reconstruct non-zero elements, vectors are used. Specifically, if ,but 0, otherwise The loss function for the structured deep network embedding model is obtained by connecting the loss functions of first-order similarity and second-order similarity, and is defined as follows: in It's a hyperparameter. It is the parameter matrix in the decoding process; the last term in the equation is the regularization term to prevent overfitting; after minimizing the loss function, a low-dimensional representation vector set of the unit region is obtained, thus preserving the local and global structure of the unit region trajectory map.

2. The method for functional classification of open-pit mine areas based on deep region embedding as described in claim 1, characterized in that: The specific steps of step S1 are as follows: Step S11, cleaning up point data that is not in the area to be identified and invalid points in the GPS trajectory data of the mining area; Step S12, using a map matching algorithm to match the GPS trajectory data of the mining area after data cleaning to the corresponding open-pit mining area, generating unit area feature data within the area; Step S13, dividing the open-pit mining area into several unit areas.

3. The method for functional classification of open-pit mine areas based on deep region embedding as described in claim 2, characterized in that: The specific steps for dividing the open-pit mine area into several unit area data in step S13 are as follows: uniformly dividing the open-pit mine area into... Non-overlapping unit regions, ,in Indicates an open-pit mine area. This represents the i-th unit region after uniform division, constructed by equally dividing the latitude and longitude space. , ,in The table lists the number of unit regions at latitude and longitude, respectively. Indicates the maximum latitude. Indicates the minimum latitude. Indicates the maximum longitude. This indicates the minimum longitude, so That is, the number of unit regions is Each unit area Defined by four boundary points, its variables have the same meaning as the latitude and longitude mentioned above.

4. The method for functional classification of open-pit mine areas based on deep region embedding according to claim 3, characterized in that: The specific training steps in step S5 are as follows: Step S51, embed the low-dimensional representation of the unit region. and the final hidden vector representation in the temporal features of the unit region Feature concatenation is performed, and the matrix is ​​input into the attention layer. An attention mechanism is used to assign weights to the feature matrix. To ensure consistent input and output dimensions when assigning different weights without losing feature information, a self-attention mechanism is selected in the attention layer. in These are three distinct vectors for each feature, obtained by transforming the feature vectors... Multiply , ,vector weight matrix Obtained; It is a fraction; gradient stability is achieved by dividing by... Perform normalization; use The function is used as an activation function, multiplied by To obtain the final result Step S52: Input the obtained attention-weighted feature matrix into a multilayer perceptron for further learning, then input it into a fully connected layer to adjust the dimensions, and finally use it. Classify functions; In model training, the cross-entropy loss function is used; for a sample size of... The training set and the loss function are as follows: in This represents all the parameters involved in training the model. It is a regularization term; parameters are learned through Adam stochastic gradient descent optimizer and backpropagation; step S53: after the above steps, a deep neural network classification model for functional classification of open-pit mining unit areas is finally obtained.

5. The method for functional classification of open-pit mine areas based on deep region embedding according to claim 4, characterized in that: In step S6, the deep neural network classification model is trained in step S5. By inputting the feature vector of the open-pit mine unit area to be identified into the model, a unit area function with high accuracy can be obtained and applied to the open-pit mine area.

Images