CN-115937612-B - Strip mine region function classification method based on depth region embedding

Abstract

The invention discloses a strip mine area function classification method based on deep area embedding, which belongs to the field of strip mine management, and comprises the steps of firstly carrying out uniform unit area division on strip mines, extracting characteristic information of different unit areas through track data, constructing the different unit areas into a sparse unit area graph, then learning space embedding representation of unit area nodes by using a structured deep network embedding model, realizing unit area space area embedding, taking a unit area characteristic time sequence as input of a two-way long-short period memory neural network so as to obtain unit area time sequence characteristics, realizing unit area time sequence area embedding, so as to solve the problem of insufficient area time characteristics, carrying out characteristic fusion on the obtained unit area node embedding representation and unit area sequence characteristics, inputting a final characteristic fusion result into an attention layer, solving the problem of uneven distribution of different characteristic weights, and finally training to obtain more accurate classification results.

Inventors

• Zhao Benzhuang
• ZHANG LEI
• Liu bailong
• LIANG ZHIZHEN

Assignees

• 中国矿业大学

Dates

Publication Date

20260508

Application Date

20230109

Claims (5)

1. 1. A surface mine area function classification method based on depth area embedding is characterized by comprising the steps of S1, preprocessing truck GPS track data and surface mine area areas in the surface mine area, carrying out data cleaning on original surface mine track data, carrying out unit area division on the surface mine areas, extracting unit area characteristics, taking the area functions after the existing surface mine manual division as tag data, namely, each unit area corresponds to one mine area function, wherein the mine area functions comprise a mining area, a mining area and a slag field, S2, constructing a track unit area graph through the track and the unit area data preprocessed in the S1, S3, inputting the track unit area graph obtained in the S2 into a structured depth network embedding model, then utilizing an unsupervised structured depth network embedding model, generating a trained unit area embedding model after training is finished, further obtaining a low-dimensional representation vector of a unit area, S4, constructing a unit area time sequence, extracting unit area characteristics from the time sequence by using a bidirectional length-phase memory neural network, obtaining the unit area characteristics from the time sequence, obtaining a full-sequence layer characteristic, and finally inputting the time sequence vectors into a final classification model, namely, carrying out step 5, obtaining a full-scale feature classification model, and finally inputting the time sequence vectors into the step 5, and finally inputting the final classification model to the step 5, and finally obtaining the final classification model, the specific steps are that the track is mapped to the unit area to obtain the unit area characteristics, and the unit area structure is modeled as a graph Wherein the set of nodes Representing cell regions, edge sets Representing the connection relationship between two unit areas, a unit area track diagram Adjacent matrix of (a) Wherein Representing a cell region And Whether or not to connect; the step S3 is specifically that the step S31 is performed by using the unit area track map obtained in the step S2 Input into a structured depth network embedded model, wherein the structured depth network embedded model consists of an encoder and a decoder, the encoder is responsible for mapping input data to low-dimensional vectors, the decoder is responsible for mapping the low-dimensional vectors to an original representation space, and a network structure of the deep layer of the structured depth network embedded model is better for sensing a track unit area graph Simultaneously, the network senses the local and the whole structure in the graph by jointly optimizing the first-order similarity and the second-order similarity constraint and relieves the track unit area graph The negative influence of sparsity in the structure, step S32, the first order similarity assumes that two unit areas are similar, if two unit areas are connected in the unit area track network, the second order similarity assumes that two unit areas are similar, if the neighborhood structures of two unit areas are similar, even if they are not connected in the unit area track network, the second order proximity enriches the area relation in the unit area track network, so that the structured depth network embedded model can capture the global structure of the unit area track network, and the encoder of the structured depth network embedded model gives the neighborhood structure of the unit area The process of the encoder is expressed as: Wherein the method comprises the steps of Is a matrix of parameters that can be learned, Is a bias term that is used to determine, Is a Sigmoid activation function, step S33, the decoder is the inverse of the encoder, thus obtaining the cell area Output of (2) The goal of self-encoder construction is to minimize the reconstruction loss of the input data and output data, since the input is a neighborhood structure for each cell region, the reconstruction process will tend to have cell regions with similar neighborhood structures to similar low-dimensional representations, and therefore the loss function of second-order similarity is defined as: Wherein the method comprises the steps of Representing multiplication by element, in which the road network of the strip mine is sparse, so that the matrix is contiguous In order to make the model more prone to reconstruct non-zero elements, vectors are referenced In particular, if Then 0, Otherwise The loss function of the structured depth network embedded model is obtained by connecting the loss functions of the first-order similarity and the second-order similarity, and is defined as: Wherein the method comprises the steps of Is a super-parameter which is used for the processing of the data, The method comprises the steps of obtaining a low-dimensional representation vector set of a unit area after minimizing a loss function, wherein the low-dimensional representation vector set is a parameter matrix in the decoding process, the last term in an equation is a regularization term for preventing overfitting, and therefore the local and global structures of a unit area track graph are maintained.
2. 2. The surface mine area function classification method based on depth area embedding is characterized in that the step S1 specifically comprises the steps of S11 cleaning point data which are not in an area to be identified and are invalid in the GPS track data of the mine area, S12 matching the GPS track data of the mine area subjected to data cleaning operation into corresponding surface mine areas by adopting a map matching algorithm to generate unit area characteristic data in the areas, and S13 carrying out area segmentation through the surface mine areas to obtain a plurality of unit areas.
3. 3. The method for classifying strip mine area functions based on depth area embedding of claim 2, wherein the step S13 of dividing the strip mine area into areas comprises the specific steps of uniformly dividing the strip mine area into a plurality of unit area data The areas of the cells that are not to be overlapped, Wherein Representing the area of the strip mine, Representing the i-th cell region after uniform division, constructing the region by dividing the latitude and longitude space equally, namely , Wherein The number of unit areas in latitude and longitude are respectively shown, The maximum latitude is indicated as such, The minimum latitude is indicated as being the one that, The maximum longitude is indicated as being the maximum longitude, Representing the minimum longitude, so I.e. the number of unit areas is Each unit area The four boundary points are defined, and the variables have the same meaning as the longitude and latitude.
4. 4. A method for classifying functions of a strip mine area based on depth area embedding as set forth in claim 3, wherein said specific training step of step S5 is as follows, step S51, embedding the unit area in a low-dimensional manner to represent Final hidden vector representation in cell region timing characteristics Feature stitching is performed, the matrix inputs the attention layer, the feature matrix is distributed by using an attention mechanism, in order to distribute different weights without losing feature information, the input dimension and the output dimension are required to be consistent, and the self-attention mechanism is selected in the attention layer: Wherein the method comprises the steps of Is three different vectors for each feature by combining the feature vectors Multiplied by 、 Vector quantity Weight matrix of (2) Obtained; Is the fraction, gradient stability is divided by Normalizing by Function as activation function, multiplied by Obtaining the final result Step S52, inputting the obtained attention weighted feature matrix into a multi-layer perceptron for further learning, then inputting the obtained attention weighted feature matrix into a full-connection layer for dimension adjustment, and finally using the obtained attention weighted feature matrix Classifying the functions; In the training of the model, a cross entropy loss function is used, for the number of samples The loss function is as follows: Wherein the method comprises the steps of Representing all parameters involved in the training of the model, The method comprises the steps of regularization, parameter learning through an Adam random gradient descent optimizer and back propagation, and step S53, wherein a deep neural network classification model for classifying functions of open-air mining area unit areas is finally obtained after the steps.
5. 5. The method for classifying surface mine areas based on depth area embedding of claim 4, wherein the depth neural network classification model in the step S6 is trained in the step S5, and the surface mine unit area feature vector to be identified is input into the model to obtain the unit area function with higher accuracy and is applied to the surface mine area.

Description

Strip mine region function classification method based on depth region embedding Technical Field The invention relates to the field of strip mine management, in particular to a strip mine regional function classification method based on deep regional embedding. Background Various geographic information acquisition technologies, particularly "3S" technologies, namely remote sensing satellites (RSs), global Positioning Systems (GPS), geographic Information Systems (GIS) are rapidly developing in today' S society, so that it is becoming increasingly simple to collect and store massive geographic traffic data. The strip mine is a large-scale production system integrating collection and transportation, vehicles such as trucks, diggers and the like are also provided with GPS equipment, and a large amount of track data is generated every day. With the development of the era, more and more strip mines are equipped with a production scheduling system, and the system distributes the truck in the mine and pairs the diggers reasonably through internal calculation, so that the truck is dispatched in an optimal mode, the operation cost is reduced, and the mining efficiency is improved. By accurately judging the functional area, the dispatching system can quickly and reasonably adjust the state of the truck. Since the strip mine has a plurality of stopes and a plurality of muck sites, and both the stopes and the muck sites are constantly changing over time. At present, the manual marking mode is adopted to realize the identification of stopes and dregs fields, which has the problems of labor consumption, untimely updating, inaccurate data and the like, so that the classification of functional areas through track data on strip mines is very important. Open-pit mining areas have a variety of functional areas, such as mining areas, unloading areas, and the like. Different functional areas assume different responsibilities. In the past, most of researches focus on the time sequence characteristics of the GPS track, which are determined by the travel rule, and the functional areas show different time characteristics in different time periods, for example, people go from a living area to an office area in the morning of work day. In the open pit scene, the trucks work according to the dispatching time, the information characteristics of the trucks in the time dimension are insufficient to clearly identify the regional functions, but the trucks pass through a plurality of functional areas in one dispatching task, and the different functional areas have spatial interrelations depending on truck tracks, so how to accurately classify the open pit functional areas by utilizing truck track data is an important study for optimizing the open pit truck dispatching system. The GPS track is one of the most important data for identifying regional functions, the content of the GPS track mainly comprises the longitude and latitude and time of getting on and off, most of previous researches identify urban functions by researching the starting point of a taxi, and as the track is usually continuous, the taxi track only considers the regional information contained in the middle journey ignored by the starting position, and the global feature is ignored. Meanwhile, the coverage rate of urban taxi tracks is wider than that of open pit truck tracks, and the coverage rate of mining area truck track areas is sparse, so that a certain challenge is brought to mining area function identification. Disclosure of Invention The invention aims to provide a strip mine area function classification method based on depth area embedding, which solves the problems that in the prior art, the identification of stopes and dregs fields is labor-consuming, untimely in updating and inaccurate in data due to the fact that the strip mine areas are marked manually. The technical scheme includes that the open pit area function classification method based on deep area embedding is characterized in that firstly uniform unit area division is conducted on open pit, characteristic information of different unit areas is extracted through track data, the different unit areas are built into a sparse unit area diagram, then space embedding representation of unit area nodes is learned through a structured deep network embedding model, unit area space area embedding is achieved, unit area characteristic time sequence sequences are used as input of a two-way long-short term memory neural network, unit area time sequence characteristics are obtained, unit area time sequence area embedding is achieved, the problem of insufficient area time characteristics is solved, feature fusion is conducted on the obtained unit area node embedding representation and unit area sequence characteristics, a final feature fusion result is input into an attention layer, the problem of uneven distribution of different feature weights is solved, and finally training is conducted, and more accurate classifica