An employee mobile mode learning method applied to coal mine underground work type identification

Abstract

The application relates to a staff mobile mode learning method applied to coal mine underground work type identification, and belongs to the technical field of coal mine underground data analysis. A work type identification model is designed by using coal mine underground staff track data, the model realizes identification of which work type the input track data belongs to, and comprises a multi-semantics embedding module, a global feature extraction module and a work type identification module. Advantages: graph embedding and word embedding are combined to learn the embedding representation of staff, external factors are considered to affect the staff mobile track, richer semantic information can be obtained, the transformer technology is used to process variable long track sequences, global semantic features of the track are extracted, a contrast learning network for self-supervised track classification is involved, potential values of unlabeled data are fully tapped, the accuracy of work type identification is improved, technical support is provided for a coal mine underground scheduling platform, the problem of one person with multiple cards is avoided, intelligent scheduling of coal mine underground staff is realized, and underground production safety is maintained.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of coal mine underground data analysis, in particular to a staff mobile mode learning method applied to coal mine underground work type identification. BACKGROUND

[0002] In recent years, with the rapid development of wireless communication technology and global positioning technology, the position information of mobile objects (people, animals, vehicles, etc.) can be easily obtained and the mobile trajectory can be stored and managed. The spatio-temporal trajectory data is a sequence of records of the position and time of the mobile object, which not only contains rich spatio-temporal position information, but also reflects the activity rules and mobile mode of human beings. In addition, the trajectory data provides rich information and value, and trajectory data mining is also applied to many fields, such as location service, traffic prediction, traffic mode identification, and trajectory user linking.

[0003] Underground work type identification can be essentially regarded as an extension of the traditional trajectory classification task. By mining the mobile trajectory of the staff, the basic mobile mode of the staff is revealed, and the trajectory data is linked to the work type according to the mobile mode of the staff. Traditional trajectory measurement methods, such as the longest common subsequence (LCSS) and dynamic time warping (DTW), can identify the most likely user by measuring the similarity between unknown trajectories and known trajectories. Recently, deep trajectory representation learning is used to solve such problems, and most of the work relies on recurrent neural networks (RNN) or its variants, such as long short-term memory (LSTM) and gated recurrent unit (GRU), to capture the mobile mode of human beings, because the RNN-based model can capture the time and semantic information in the variable-length trajectory data. Moreover, most of them are supervised learning, which requires a large amount of labeled data for model training, which not only has challenges in model training but also has high time cost.

[0004] Therefore, the existing method still has three key limitations.

[0005] 1. Due to the sparsity of the trajectory, different staffs usually visit a fixed few of the hundreds of work points in the underground roadway, in addition, the data may involve noise and outliers;

[0006] 2. The existing method only focuses on spatial and temporal features, while ignoring the rich context features in the trajectory data.

[0007] 3. It has a very serious dependence on labels, weak generalization ability, and weak robustness. Work type identification is different from many traditional mobile mode identification problems, which need to extract and analyze various features from the trajectory, which will not only bring the curse of dimensionality, but also infringe the privacy of the staff.

[0008] With the wide application of deep learning in the field of trajectory data, trajectory recognition technology based on artificial intelligence method has become a research hotspot. SUMMARY

[0009] In order to overcome the above-mentioned deficiencies of the prior art, the present application provides a staff movement pattern learning method applied to coal mine underground job identification, which solves the problem that the traditional machine learning method cannot fully utilize the context information of trajectory data, the trajectory data lack labels, and the trajectory data are sparse, resulting in the inability to accurately and quickly identify the job of coal mine underground staff.

[0010] The purpose of the present application is achieved in that the method introduces UWB ultra-wideband positioning data of underground staff, underground roadway map data and staff surrounding environment monitoring data; a job identification model is designed using coal mine underground staff trajectory data; the job identification model includes a multi-semantics embedding module, a global feature extraction module and a job identification module; the UWB is represented as Ultra Wide Band.

[0011] After the trajectory data is preprocessed, it is subjected to the multi-semantics embedding module to obtain multi-aspect semantic information embedding representation; then it is subjected to the global feature extraction module to learn feature representation and obtain global space-time dependency; finally, it is subjected to the job identification module to calculate the probability that the input trajectory data belongs to each job, thereby improving the accuracy of job identification and providing technical support for the coal mine underground scheduling platform, that is, information can be pushed according to the classification results of the job, different messages can be pushed to different staff, the problem of one person with multiple cards can be avoided, the intelligent scheduling of coal mine underground staff can be realized, and the safety of underground production can be maintained.

[0012] The specific steps are as follows:

[0013] Step 1, collect the movement trajectory information of coal mine underground staff, and pre-process the collected data to obtain a plurality of sub-trajectory sequences of each staff; the pre-processing is sequentially grouping processing, cleaning processing, interpolation processing and sub-trajectory division processing;

[0014] Step 2, performing trajectory enhancement processing on the sub-trajectory sequence pre-processed in step 1; the trajectory enhancement processing is strong enhancement processing and weak enhancement processing, respectively, to obtain strong enhancement, weak enhancement and non-enhancement trajectory sequences; the non-enhancement trajectory sequence is the current input sub-trajectory sequence;

[0015] Step 3, the strong enhancement, weak enhancement and non-enhanced trajectory sequences obtained in step 2 are respectively input into the mobile preference embedding unit, the trajectory sequence embedding unit, the attribute information embedding unit and the embedding connection layer of the multi-semantics embedding module to generate multi-semantics embedding representations of the strong enhancement, weak enhancement and non-enhanced trajectory sequences; the mobile preference embedding unit generates a spatial graph using road network data, generates corresponding access graphs using the strong enhancement, weak enhancement and non-enhanced trajectory sequences obtained in step 2 respectively, and merges the spatial graph and the corresponding access graphs to construct corresponding mobile preference graphs, and obtains embedding representations of the mobile preferences of the employees by using a graph embedding method; the trajectory sequence embedding unit obtains trajectory sequence embedding representations by performing position point embedding, timestamp embedding and splicing processing on the strong enhancement, weak enhancement and non-enhanced trajectory sequences obtained in step 2 respectively; the attribute information embedding unit calculates mobile speeds and extracts sensor environmental monitoring data, i.e., temperature, humidity, gas concentration and CO data, based on the strong enhancement, weak enhancement and non-enhanced trajectory sequences obtained in step 2, and obtains embedding representations of these attribute information by using a word embedding method; the embedding connection layer is a connection layer that connects the mobile preference embedding representations, the trajectory sequence embedding representations and the attribute information embedding representations to obtain multi-semantics embedding representations;

[0016] Step 4, the multi-semantics embedding representations of the strong enhancement, weak enhancement and non-enhanced trajectory sequences obtained in step 3 are processed by using a transformer encoder and a contrast learning method in self-supervised learning, and a global feature extraction module is used to process the unlabeled and variable long trajectory sequences to obtain global semantic features of the trajectories;

[0017] Step 5, a fully connected neural network is used to calculate probabilities of the input trajectory data belonging to each type of work by using a work type identification module to correctly classify positive samples and negative samples, so as to better train the model;

[0018] Step 6, the employee trajectory data is divided into training samples and test samples, a designed loss function is used to train the work type identification model, and the parameters are gradually trained and optimized to realize identification of the work type to which the employee trajectory belongs.

[0019] In step 1, the data is preprocessed as follows:

[0020] The underground employee mobile trajectory information of the coal mine includes: employee ID, position coordinates of multiple time points based on UWB positioning, and environmental monitoring data around the employee during movement;

[0021] The grouping processing is to group the corresponding trajectory points according to the unique identifier (employee ID) of the employee;

[0022] The cleaning processing is to remove the trajectory points with frequent long-distance fluctuations in each time slice, and the specific process is as follows:

[0023] Step 1-1, the trajectory data obtained after grouping according to the employee ID is defined with appropriate time slices according to the sampling interval of the trajectory points, and the trajectory points are corresponded to the time slices one by one; the appropriate time slice is between 3s-7s;

[0024] Step 1-2, the center position of all data in each time slice is calculated;

[0025] Step 1-3, the distance between all position information of the group and the center position is calculated, and the nearest data to the center position is retained, and the distance calculation formula is as follows:

[0026]

[0027] Wherein, represents the distance between points A and B, R represents the radius of the earth, A lat , A lot , B lat , B lot respectively represent the longitude and latitude of point A and point B;

[0028] Step 1-4, the center position of all retained data is recalculated; each group represents the trajectory data of an employee, and only one position data of each employee is retained at the same time;

[0029] The interpolation processing is that for the missing trajectory point of a time slice, if the adjacent time before and after the time slice exists trajectory point, the interpolation is carried out according to the adjacent time trajectory point, the interpolation position is the center of the trajectory points of the previous and next time slices, and the missing trajectory point is obtained;

[0030] The sub-trajectory division processing is to divide the trajectory data of each employee after cleaning processing and interpolation processing into a plurality of sub-trajectories according to a fixed time interval of 6 hours.

[0031] In step 2, the trajectory enhancement includes strong enhancement and weak enhancement, and the process is as follows:

[0032] The trajectory enhancement processing includes strong enhancement processing and weak enhancement processing, and the specific process is that the sub-trajectory sequence tr after step 1 preprocessing is respectively subjected to strong enhancement processing to obtain a strong enhancement trajectory sequence and weak enhancement processing to obtain a weak enhancement trajectory sequence The unenhanced trajectory sequence, i.e. the current input sub-trajectory sequence tr, the strong enhancement trajectory sequence and the weak enhancement trajectory sequence Meanwhile input to the multi semantic embedding module, the unenhanced trajectory sequence tr after passing through the multi semantic embedding module and the transformer encoder serves as anchor data for contrastive learning in the global feature extraction module, the strongly enhanced trajectory sequence After passing through the multi semantic embedding module and the transformer encoder, it serves as positive sample data for contrastive learning in the global feature extraction module, the weakly enhanced trajectory sequence After passing through the multi semantic embedding module and the transformer encoder, it serves as negative sample data for contrastive learning in the global feature extraction module;

[0033] The strong enhancement processing adopts a neighborhood enhancement method, for the sub trajectory sequence tr obtained after preprocessing, when the input time interval is t i , the k+1 sub trajectories in the time interval [t i-k / 2 :t i+k / 2 ] are spliced with tr to form a long-term enhanced trajectory

[0034] The weak enhancement processing adopts a random enhancement method, for the sub trajectory sequence tr obtained after preprocessing, when the input time interval is t i , k sub trajectories are randomly sampled from all trajectories of the corresponding employee, and then spliced according to the time information and the input trajectory sequence to form a long-term trajectory

[0035] In step 3, the multi semantic embedding module includes mobile preference embedding, trajectory sequence embedding and attribute information embedding, specifically including:

[0036] The road network data is underground roadway map data of a coal mine, including underground roadway information, underground working area and position coordinates, name and type of each sensor, used to generate a static spatial graph according to the road network data;

[0037] The attribute information includes the moving speed of the employee and the sensor environment monitoring data in the moving process; the sensor environment monitoring data includes temperature, humidity, gas concentration and CO;

[0038] The mobile preference embedding unit is: generating a spatial graph using the road network data, generating an access graph using the trajectory sequence obtained in step 2, merging the two graphs to construct a mobile preference graph, using graph embedding to project the employee synthesized trajectory in the mobile preference graph into a low-dimensional space to obtain the mobile preference embedding representation of the employee;

[0039] The spatial graph generation method is to take the position coordinates of the work points or sensors extracted from the underground roadway map as a spatial graph G spatial (V, E snodes V, the feature vector of each node is initialized as the information of corresponding work point or sensor, such as position, sensor type, work point name; edges E in the space graph s represent a group of edges connected by work points or sensors of the same type;

[0040] The access graph generation method is to take the work points and sensor positions accessed by the employee track sequence in a time period as an access graph G visit (V, E) v nodes V, the feature vector of each node is initialized as the position point information of the employee at the work point or sensor, including employee ID, timestamp, position coordinates, E v represents all access connections, and each e ij ∈E v represents the access order of the employee from site i to site j;

[0041] The mobile preference graph generation method is to combine the space graph G spatial and the access graph G visit to generate a mobile preference graph G full , G full is initially the same as G spatial , in the combination process, the edges appearing in G visit but not in G spatial are added to G full , and the initial feature vectors of the nodes are connected, after the combination, the initial feature vector of each node includes employee ID, timestamp, position coordinates, sensor type or work point name;

[0042] The graph embedding method adopts a GraphSAGE model to learn the embedding representation of each node, the input is a mobile preference graph G full , the initial feature vector of each node is x v , v∈V, after sampling, aggregation, splicing and normalization operations, the final output of the mobile preference embedding representation is The GraphSAGE is Graph Sample and Aggregate;

[0043] The track sequence embedding unit includes position point embedding, timestamp embedding and splicing processing, a transformation matrix is used to project the one-hot vectors of the timestamps and position points in the track sequence to a low-dimensional space and perform splicing processing, to obtain the track sequence embedding representation of the employee;

[0044] The position point embedding process is: for each position point l in an unenhanced track sequence tr represented by a |l|-dimensional one-hot vector, a transformation matrix E lThe l is changed into a low-dimensional vector to capture accurate semantic information; the strong enhancement and weak enhancement trajectory sequence also undergoes the same processing; when the position point embedding of an unenhanced trajectory sequence tr at t time can be expressed as:

[0045]

[0046] Wherein, the matrix E l In the model training process, the learning is carried out constantly, wherein The learning representation matrix D l Is the dimension of the position embedding;

[0047] The timestamp embedding process is: for the time sequence in the unenhanced trajectory sequence tr, the timestamp information is composed, first, the time value is discretized, that is, each timestamp value is mapped to different time intervals in each day in hours, at this time, the number of time intervals is 24, a learnable vector representation is associated with each interval, then a time representation matrix Wherein D t Is the dimension of the timestamp embedding; the strong enhancement and weak enhancement trajectory sequence also undergoes the same processing; when the timestamp embedding of an unenhanced trajectory sequence tr at t time can be expressed as:

[0048]

[0049] Wherein, map(t) is obtained by mapping a one-hot vector, the matrix E t In the model training process, the learning is carried out constantly;

[0050] The splicing processing splices the position point embedding and the timestamp embedding to generate a global representation Wherein D g =D l +D t ; the strong enhancement and weak enhancement trajectory sequence also undergoes the same processing;

[0051] The attribute information embedding unit is: the word vector embedding method is used to convert each attribute into a low-dimensional real vector to obtain the attribute information embedding representation of the employee trajectory, and each attribute is the moving speed, the sensor environment monitoring data; the specific process is as follows:

[0052] Step 3-1, the employee moving speed v uses the distance and time change between the i point and the (i+1) point to describe different speed characteristics, which can be expressed as:

[0053] A spe =△d·(t i+1 -t i ) -1

[0054] where d represents the distance between the i-th point and the (i+1)-th point, t i represents the timestamp of the i-th point;

[0055] Step 3-2, calculate the attribute information of each timestamp of the strong enhancement, weak enhancement and non-enhancement trajectory sequence, and embed these attribute information into a low-dimensional real vector; the attribute information feature of a non-enhancement trajectory sequence tr at time t can be expressed as:

[0056]

[0057]

[0058]

[0059] where W e , W s represent the weights of the fully connected layer, A env represents the sensor environment monitoring data at time t, A spe represents the speed at time t, W e , W s , b e , b s are learnable parameters, and CONCAT represents the connection between attribute vectors;

[0060] The connection operation is that a connection layer is used to connect the movement preference embedding representation, trajectory sequence embedding representation and attribute information embedding representation for the strong enhancement, weak enhancement and non-enhancement trajectory sequence respectively, to obtain a multi-semantics embedding representation:

[0061]

[0062] where H tr ∈R m*n , m represents the dimension of the matrix row, and n represents the dimension of the column.

[0063] In step 4, the global feature extraction module includes a transformer encoder and contrastive learning, and specifically includes:

[0064] The transformer encoder includes a plurality of encoder blocks, and the specific process is as follows:

[0065] Based on the feature matrix H tr ∈R k*d , H is decomposed into m n-dimensional vectors x t as the input of the transformer encoder, where x tis the input vector for the t-th time step, i.e., the t-th component of the input sequence X; the transformer encoder is composed of a plurality of encoder blocks, and the vector x t After sequentially passing through the multi-head self-attention layer in the encoder block, the Add&Norm operation, the feedforward neural network and the Add&Norm operation, a vector with the same dimension is finally output.

[0066] The contrastive learning is a self-supervised learning method. For unlabeled data, the similarity between the anchor data and the positive sample is much greater than that between the anchor data and the negative sample by comparing the similarity between the anchor data and the positive and negative samples, so as to learn the general features of the data. The specific process is as follows:

[0067] Step 4-1, positive and negative sample selection, in the transformer encoder module, the transformer-based sequence-to-sequence neural network is used to convert the employee's unenhanced trajectory sequence tr ≤t ={c1,c2,…,c t} into a vector h t with the same dimension as the anchor data; the strong enhanced trajectory sequence in the trajectory enhancement module is also passed through the multi-semantics embedding module and the transformer encoder to obtain the positive sample of the input trajectory sequence for contrastive learning The weak enhanced trajectory sequence in the trajectory enhancement module is also passed through the multi-semantics embedding module and the transformer encoder to obtain the negative sample of the input trajectory sequence for contrastive learning

[0068] Step 4-2, use the self-supervised learning method to train the model, and learn the encoder f so that:

[0069]

[0070] Wherein, score(·,·) is a metric function to measure the similarity between samples. In contrastive learning, the representation is learned by comparing positive and negative samples, so that the left part is as large as possible than the right part.

[0071] During the training process, the InfoNCE loss function is used to calculate the loss, and the final feature representation of the trajectory is obtained after the training. The formula of the InfoNCE loss function is as follows:

[0072]

[0073] Each encoder block in the transformer encoder specifically includes:

[0074] The multi-head self-attention layer specifically includes the following process:

[0075] Step 4-3-1, calculate input vector x t Q, K and V matrix of each value; the Q matrix is the full name of Query matrix in English, the K matrix is the full name of Key matrix in English, and the V matrix is the full name of Value matrix in English;

[0076] Q = W q x t

[0077] K = W k x t

[0078] V = W v x t

[0079] wherein, W q , W k and W v are learnable matrices;

[0080] Step 4-3-2, calculate the dot product between Q and K, in order to prevent the result from being too large, divide by wherein d k is the dimension of the Key vector, then normalize the result to a probability distribution by using the normalization index Softmax activation function, and multiply by the matrix V to obtain the weighted sum representation, that is, the output of the self-attention layer of this encoder block, the calculation process is as follows:

[0081]

[0082] Step 4-3-3, the multi-head self-attention layer contains multiple self-attention layers, and the input x t is respectively passed into h different self-attention layers to obtain h output matrices Attention(Q,K,V), and different Attention are spliced; then perform a linear transformation, the calculation process is as follows:

[0083] MultiHead(Q,K,V) = Concat(head1,…,head h )W O

[0084]

[0085] wherein, and W O are learnable matrices;

[0086] The feedforward neural network is: input the weight obtained by the self-attention layer into the feedforward neural network, the feedforward neural network is a two-layer fully connected layer, the activation function of the first layer is Relu, and the second layer does not use the activation function, and the corresponding formula is as follows:

[0087] FFN(x) = max(0, xW1 + b1)W2 + b2

[0088] The Add&Norm operation is: after the input sequence respectively passes through the multi-head self-attention layer and the feedforward neural network, there is a residual connection Add operation, and then a layer normalization operation is performed, the layer normalization operation is denoted as Norm in English, and the calculation process is as follows:

[0089] sub_layer_output = LayerNorm(x + SubLayer(x))

[0090] Wherein, x represents the input of the multi-head self-attention layer and the feedforward neural network, SubLayer(x) represents the output of the multi-head self-attention layer and the feedforward neural network, and the dimensions are the same.

[0091] The step 5 specifically comprises:

[0092] The full connection neural network is: a full connection neural network using a normalized exponential Softmax function is used for classification probability statistics, all neurons in each network are controlled between 0-1, and the calculation formula is as follows:

[0093]

[0094] Wherein, x i is the output value of the i th node, and n is the number of output nodes, that is, the number of classification categories.

[0095] The step 6, model training specifically comprises:

[0096] The training sample and test sample division mode is: the collected employee trajectory data is mixed according to the employee ID, and is divided into training samples and test samples according to the proportion of 70%, 30%;

[0097] The model loss function adopts a cross-entropy loss function to measure the difference between the recognized work type category probability and the real work type category:

[0098]

[0099] Wherein, M is the number of coal mine underground employee work types; y ic Is a symbol function, if the real category of sample i is equal to c, then 1, otherwise 0; p icis the prediction probability that the input sample i belongs to the class c.

[0100] The beneficial effect is that, due to the adoption of the above scheme, the employee movement mode learning is performed by combining GNN and transformer, the data sparsity problem is solved by contrast learning in self-supervised learning, and the rich information of unlabeled data can be fully mined.

[0101] The method comprises three modules, namely a multi-semantics embedding module, a global feature extraction module and a job type recognition module. The multi-semantics embedding module adopts the latest progress of graph representation learning technology, learns the embedding representation of the employee by using context information, and enriches the data information by considering the influence of external factors on the employee movement trajectory; the global feature extraction module adopts transformer technology to process variable long trajectory sequences, capture time correlation, and extract global semantic features of the trajectory, and the module involves a contrast learning network for self-supervised trajectory classification, solves the problem of unlabeled data, and can understand higher-order features of the data before performing the classification task. Compared with other job type recognition models, the method has obvious advantages in application in the job type recognition of employees in the coal mine underground.

[0102] The coal mine underground target accurate positioning system adopts a UWB positioning method, realizes 0.3-meter positioning accuracy, can realize accurate positioning of underground personnel, vehicles and equipment, and can vividly and visually show the distribution and movement of underground personnel, vehicles and equipment. Using deep learning to learn the movement mode of employees and identifying the job type according to the trajectory data is an important part of the management task. It has great significance for intelligent scheduling of coal mine underground employees, solving the one-person multi-card problem, and maintaining the safety of underground production.

[0103] The method solves the problem that the existing classification method cannot fully utilize the context information of trajectory data, the trajectory data lack labels, and the trajectory data are sparse, so that the job type of the coal mine underground employee cannot be accurately and quickly recognized, and the purpose of the present application is achieved.

[0104] Advantages: 1. The contrast learning method in self-supervised learning fully mines the potential value of unlabeled data and solves the data sparsity problem; 2. When learning the employee movement preference embedding representation, the special nature of the coal mine underground roadway is considered, and the rich road network information of the coal mine underground roadway is fully utilized to mine the movement preference of the employee; 3. Not simply relying on historical observation and space-time factors, the synthetic trajectory with real movement constraints is explicitly generated to enhance the trajectory representation and thus enhance the downstream task; 4. The attribute information of the trajectory, the rich space-time context and the synthetic trajectory are used to enhance the trajectory representation learning; 5. The transformer encoder is adopted to process the variable long trajectory sequence after trajectory enhancement, and the long-term time dependency and global semantic features of the trajectory sequence can be captured. BRIEF DESCRIPTION OF DRAWINGS

[0105] Fig. 1 Flow chart of the present application.

[0106] Fig. 2 Coal mine underground job identification structure diagram of the present application. DETAILED DESCRIPTION

[0107] The method introduces UWB ultra-wideband positioning data of underground employees, underground roadway map data and employee surrounding environment monitoring data; a job identification model is designed using coal mine underground employee trajectory data; the job identification model includes a multi-semantics embedding module, a global feature extraction module and a job identification module; the UWB is represented as Ultra Wide Band in English;

[0108] After the trajectory data is preprocessed, multi-aspect semantic information embedding representation is obtained through the multi-semantics embedding module; then the global feature extraction module is used to learn feature representation to obtain global space-time dependency; finally, the job identification module is used to calculate the probability of the input trajectory data belonging to each job, thereby improving the accuracy of job identification and providing technical support for the coal mine underground scheduling platform, that is, information can be pushed according to the classification results of the job to push different messages to different employees, avoiding the problem of one person with multiple cards, realizing intelligent scheduling of coal mine underground employees and maintaining underground production safety;

[0109] The specific steps are as follows:

[0110] Step 1, collect coal mine underground employee movement trajectory information, and preprocess the collected data to obtain a plurality of sub-trajectory sequences of each employee; the preprocessing is sequentially grouping processing, cleaning processing, interpolation processing and sub-trajectory division processing;

[0111] Step 2, perform trajectory enhancement processing on the sub-trajectory sequence preprocessed in step 1; the trajectory enhancement processing is strong enhancement processing and weak enhancement processing, respectively, to obtain strong enhancement, weak enhancement and non-enhancement trajectory sequences; the non-enhancement trajectory sequence is the current input sub-trajectory sequence;

[0112] Step 3, the strong enhancement, weak enhancement and non-enhanced trajectory sequences obtained in step 2 are respectively input into the mobile preference embedding unit, the trajectory sequence embedding unit, the attribute information embedding unit and the embedding connection layer of the multi-semantics embedding module to generate multi-semantics embedding representations of the strong enhancement, weak enhancement and non-enhanced trajectory sequences; the mobile preference embedding unit generates a spatial graph using road network data, generates corresponding access graphs using the strong enhancement, weak enhancement and non-enhanced trajectory sequences obtained in step 2 respectively, and merges the spatial graph and the corresponding access graphs to construct corresponding mobile preference graphs, and obtains embedding representations of the mobile preferences of the employees by using a graph embedding method; the trajectory sequence embedding unit obtains trajectory sequence embedding representations by performing position point embedding, timestamp embedding and splicing processing on the strong enhancement, weak enhancement and non-enhanced trajectory sequences obtained in step 2 respectively; the attribute information embedding unit calculates mobile speeds and extracts sensor environmental monitoring data, i.e., temperature, humidity, gas concentration and CO data, based on the strong enhancement, weak enhancement and non-enhanced trajectory sequences obtained in step 2, and obtains embedding representations of these attribute information by using a word embedding method; the embedding connection layer is a connection layer that connects the mobile preference embedding representations, the trajectory sequence embedding representations and the attribute information embedding representations to obtain multi-semantics embedding representations;

[0113] Step 4, the multi-semantics embedding representations of the strong enhancement, weak enhancement and non-enhanced trajectory sequences obtained in step 3 are processed by using a transformer encoder and a contrast learning method in self-supervised learning, and a global feature extraction module is used to process the unlabeled and variable long trajectory sequences to obtain global semantic features of the trajectories;

[0114] Step 5, a fully connected neural network is used to calculate probabilities of the input trajectory data belonging to each type of work by using a work type identification module to correctly classify positive samples and negative samples, so as to better train the model;

[0115] Step 6, the employee trajectory data is divided into training samples and test samples, a designed loss function is used to train the work type identification model, and the parameters are gradually trained and optimized to realize identification of the work type to which the employee trajectory belongs.

[0116] In step 1, the data is preprocessed as follows:

[0117] The underground employee mobile trajectory information of the coal mine includes: employee ID, position coordinates of multiple time points based on UWB positioning, and environmental monitoring data around the employee during movement;

[0118] The grouping processing is to group the corresponding trajectory points according to the unique identifier (employee ID) of the employee;

[0119] The cleaning processing is to remove the trajectory points with frequent long-distance fluctuations in each time slice, and the specific process is as follows:

[0120] Step 1-1, the trajectory data obtained after grouping according to the employee ID is defined with appropriate time slices according to the sampling interval of the trajectory points, and the trajectory points are corresponded to the time slices one by one; the appropriate time slice is between 3s-7s;

[0121] Step 1-2, the center position of all data in each time slice is calculated;

[0122] Step 1-3, the distance between all position information of the group and the center position is calculated, and the nearest data to the center position is retained, and the distance calculation formula is as follows:

[0123]

[0124] Wherein, The distance between points A and B, R represents the radius of the earth, A lat , A lot , B lat , B lot Respectively represent the longitude and latitude of point A and point B;

[0125] Step 1-4, the center position of all retained data is recalculated; each group represents the trajectory data of an employee, and only one position data of each employee is retained at the same time;

[0126] The interpolation processing is that for the missing trajectory points in a time slice, if there are trajectory points in the adjacent time before and after the time slice, the interpolation is carried out according to the adjacent time trajectory points, the interpolation position is the center of the trajectory points of the previous and next time slices, and the missing trajectory points are obtained;

[0127] The sub-trajectory division processing is to divide the trajectory data of each employee after cleaning processing and interpolation processing into a plurality of sub-trajectories according to a fixed time interval of 6 hours.

[0128] In step 2, the trajectory enhancement includes strong enhancement and weak enhancement, and the process is as follows:

[0129] The trajectory enhancement processing includes strong enhancement processing and weak enhancement processing, and the specific process is as follows: the sub-trajectory sequence tr after step 1 preprocessing is subjected to strong enhancement processing to obtain a strong enhancement trajectory sequence And weak enhancement processing to obtain a weak enhancement trajectory sequence The unenhanced trajectory sequence, i.e. the current input sub-trajectory sequence tr, the strong enhancement trajectory sequence The weak enhancement trajectory sequence Meanwhile input to the multi semantic embedding module, the unenhanced trajectory sequence tr after passing through the multi semantic embedding module and the transformer encoder serves as anchor data for contrastive learning in the global feature extraction module, the strongly enhanced trajectory sequence After passing through the multi semantic embedding module and the transformer encoder, it serves as positive sample data for contrastive learning in the global feature extraction module, the weakly enhanced trajectory sequence After passing through the multi semantic embedding module and the transformer encoder, it serves as negative sample data for contrastive learning in the global feature extraction module;

[0130] The strong enhancement processing adopts a neighborhood enhancement method, for the sub trajectory sequence tr obtained after preprocessing, when the input time interval is t i , the k+1 sub trajectories in the time interval [t i-k / 2 :t i+k / 2 ] are spliced with tr to form a long-term enhanced trajectory

[0131] The weak enhancement processing adopts a random enhancement method, for the sub trajectory sequence tr obtained after preprocessing, when the input time interval is t i , k sub trajectories are randomly sampled from all trajectories of the corresponding employee, and then spliced according to the time information and the input trajectory sequence to form a long-term trajectory

[0132] In step 3, the multi semantic embedding module includes mobile preference embedding, trajectory sequence embedding and attribute information embedding, specifically including:

[0133] The road network data is underground roadway map data of a coal mine, including underground roadway information, underground working area and position coordinates, name and type of each sensor, used to generate a static spatial graph according to the road network data;

[0134] The attribute information includes the moving speed of the employee and the sensor environment monitoring data in the moving process; the sensor environment monitoring data includes temperature, humidity, gas concentration and CO;

[0135] The mobile preference embedding unit is: generating a spatial graph using the road network data, generating an access graph using the trajectory sequence obtained in step 2, merging the two graphs to construct a mobile preference graph, using graph embedding to project the employee synthesized trajectory in the mobile preference graph into a low-dimensional space to obtain the mobile preference embedding representation of the employee;

[0136] The spatial graph generation method is to take the position coordinates of the work points or sensors extracted from the underground roadway map as a spatial graph G spatial (V, E snodes V, the feature vector of each node is initialized as the information of corresponding work point or sensor, such as position, sensor type, work point name; edges E in the space graph s a group of edges representing the connection of work points or sensors of the same type;

[0137] The access graph generation method is to take the work points and sensor positions accessed by the employee in a time period on the employee trajectory sequence as an access graph G visit (V, E) v nodes V, the feature vector of each node is initialized as the position point information of the employee at the work point or sensor, including employee ID, timestamp, position coordinates, E v representing all access connections, each e ij ∈E v representing the access order of the employee from site i to site j;

[0138] The mobile preference graph generation method is to combine the space graph G spatial and the access graph G visit to generate a mobile preference graph G full , G full is the same as G spatial at the beginning, in the combination process, the edges appearing in G visit but not in G spatial are added to G full , and the initial feature vectors of the nodes are connected and combined, after the combination, the initial feature vector of each node finally includes employee ID, timestamp, position coordinates, sensor type or work point name;

[0139] The graph embedding method adopts a GraphSAGE model to learn the embedding representation of each node, the input is a mobile preference graph G full , the initial feature vector of each node is x v v∈V, after sampling, aggregation, splicing and normalization operations, the finally output mobile preference embedding representation is The GraphSAGE is Graph Sample and Aggregate;

[0140] The trajectory sequence embedding unit includes position point embedding, timestamp embedding and splicing processing, a transformation matrix is used to project the one-hot vectors of the timestamps and position points in the trajectory sequence to a low-dimensional space and perform splicing processing, to obtain the trajectory sequence embedding representation of the employee;

[0141] The position point embedding process is: for each position point l in an unenhanced trajectory sequence tr represented by a |l| -dimensional one-hot vector, a transformation matrix E lThe l is converted into a low-dimensional vector to capture accurate semantic information; the strong enhancement and weak enhancement trajectory sequences are also processed in the same way; when the position point embedding of an unenhanced trajectory sequence tr at time t can be expressed as:

[0142]

[0143] Wherein, the matrix E l In the model training process, wherein is a learnable representation matrix composed of all position vector representations, D l is the dimension of the position embedding;

[0144] The timestamp embedding process is as follows: for the time sequence in the unenhanced trajectory sequence tr composed of timestamp information, first, the time value is discretized, that is, each timestamp value is mapped to different time intervals in each day in units of hours, at this time, the number of time intervals is 24, and a learnable vector representation is associated with each interval, then a time representation matrix Wherein D t is the dimension of the timestamp embedding; the strong enhancement and weak enhancement trajectory sequences are also processed in the same way; when the timestamp embedding of an unenhanced trajectory sequence tr at time t can be expressed as:

[0145]

[0146] Wherein, map(t) is obtained by mapping a one-hot vector, the matrix E t In the model training process, learning is carried out continuously;

[0147] The splicing processing splices the position point embedding and the timestamp embedding to generate a global representation Wherein D g =D l +D t ; the strong enhancement and weak enhancement trajectory sequences are also processed in the same way;

[0148] The attribute information embedding unit is: each attribute is converted into a low-dimensional real vector by using a word vector embedding method to obtain an attribute information embedding representation of the employee trajectory, and the each attribute is respectively a moving speed, sensor environment monitoring data; the specific process is as follows:

[0149] Step 3-1, the employee moving speed v uses the distance and time change between the i-th point and the (i+1)-th point to describe different speed characteristics, which can be expressed as:

[0150] A spe =△d·(t i+1 -t i ) -1

[0151] wherein d represents the distance between the i th point and the (i+1) th point, and t i represents the timestamp of the i th point;

[0152] Step 3-2, calculate the attribute information of each timestamp of the strong enhancement, weak enhancement and non-enhancement trajectory sequence, and embed the attribute information into a low-dimensional real vector; the attribute information feature of a non-enhancement trajectory sequence tr at time t can be expressed as:

[0153]

[0154]

[0155]

[0156] wherein W e , W s represent the weights of the full connection layer, A env represents the sensor environment monitoring data at time t, A spe represents the speed at time t, W e , W s , b e and b s are learnable parameters, and CONCAT represents the connection between attribute vectors;

[0157] The connection operation is that a connection layer is used to connect the moving preference embedding representation, trajectory sequence embedding representation and attribute information embedding representation of the strong enhancement, weak enhancement and non-enhancement trajectory sequence respectively, to obtain a multi-semantics embedding representation:

[0158]

[0159] wherein H tr ∈R m*n , m represents the dimension of the matrix row, and n represents the dimension of the column.

[0160] In step 4, the global feature extraction module includes a transformer encoder and contrastive learning, and specifically includes:

[0161] The transformer encoder includes a plurality of encoder blocks, and the specific process is as follows:

[0162] Based on the feature matrix H tr ∈R k*d obtained in the multi-semantics embedding module, H is decomposed into m n-dimensional vectors x t as the input of the transformer encoder, wherein x tThe vector x is the input vector at time step t, which is the t-th component of the input sequence X; the transformer encoder consists of multiple encoder blocks, and the vector x... t The vectors sequentially pass through the multi-head self-attention layer, Add&Norm operation, feedforward neural network, and Add&Norm operation in the encoder block, and finally output a vector of the same dimension.

[0163] The contrastive learning described is a self-supervised learning method. For unlabeled data, it learns the general features of the data by comparing the similarity between anchor data and positive and negative samples, ensuring that the similarity between anchor data and positive samples is much greater than its similarity with negative samples. The specific process is as follows:

[0164] Step 4-1, positive and negative sample selection: In the transformer encoder module, a transformer-based sequence-to-sequence neural network is used to process the employee's unenhanced trajectory sequence tr. ≤t ={c1,c2,…,c t} Convert to a vector h of the same dimensions t As anchor data; the strongly enhanced trajectory sequence in the trajectory enhancement module. Similarly, after passing through a multi-semantic embedding module and a transformer encoder, positive samples for the comparative learning of this input trajectory sequence are obtained. The weakly enhanced trajectory sequence in the trajectory enhancement module Similarly, after passing through a multi-semantic embedding module and a transformer encoder, negative samples for the comparative learning of the input trajectory sequence are obtained.

[0165] Step 4-2: Train the model using self-supervised learning. By learning the encoder f, the following is achieved:

[0166]

[0167] Here, score(·,·) is a metric function to measure the similarity between samples. In contrastive learning, the representation is learned by comparing positive and negative samples, so that the left part is as large as possible as the right part.

[0168] During training, the Noise Contrast Estimation (InfoNCE) loss function is used to calculate the loss. After training, the final feature representation of the trajectory is obtained. The formula for calculating the Noise Contrast Estimation loss function is as follows:

[0169]

[0170] Each encoder block in the transformer encoder specifically includes:

[0171] The multi-head self-attention layer described above is as follows:

[0172] Step 4-3-1, calculate input vector x t Q, K and V matrix of each value; the Q matrix is the full name of Query matrix, the K matrix is the full name of Key matrix, and the V matrix is the full name of Value matrix;

[0173] Q = W q x t

[0174] K = W k x t

[0175] V = W v x t

[0176] Wherein, W q , W k and W v are learnable matrices;

[0177] Step 4-3-2, calculate the dot product between Q and K, in order to prevent the result from being too large, divide by Where d k is the dimension of the Key vector, then use the normalized index Softmax activation function to normalize the result to a probability distribution, and multiply by the matrix V to get the weighted sum representation, that is, the output of the self-attention layer of this encoder block, the calculation process is as follows:

[0178]

[0179] Step 4-3-3, multi-head self-attention layer contains multiple self-attention layers, input x t is respectively passed into h different self-attention layers, h output matrices Attention(Q,K,V) are calculated, different Attention is spliced; then perform a linear transformation, the calculation process is as follows:

[0180] MultiHead(Q,K,V) = Concat(head1,…,head h )W O

[0181]

[0182] Wherein, and W O are learnable matrices;

[0183] The feedforward neural network is: input the weight obtained by the self-attention layer into the feedforward neural network, the feedforward neural network is a two-layer fully connected layer, the activation function of the first layer is Relu, and the second layer does not use the activation function, and the corresponding formula is as follows:

[0184] FFN(x) = max(0, xW1 + b1)W2 + b2

[0185] The Add&Norm operation is: after the input sequence respectively passes through the multi-head self-attention layer and the feedforward neural network, there is a residual connection Add operation, and then a layer normalization operation is performed, the layer normalization operation is denoted as Norm in English, and the calculation process is as follows:

[0186] sub_layer_output = LayerNorm(x + SubLayer(x))

[0187] Wherein, x represents the input of the multi-head self-attention layer and the feedforward neural network, SubLayer(x) represents the output of the multi-head self-attention layer and the feedforward neural network, and the dimensions are the same.

[0188] The step 5 specifically comprises:

[0189] The full connection neural network is: a full connection neural network using a normalized exponential Softmax activation function is used for classification probability statistics, all neurons in each network are controlled between 0-1, and the calculation formula is as follows:

[0190]

[0191] Wherein, x i is the output value of the i-th node, and n is the number of output nodes, that is, the number of classification categories.

[0192] The step 6, the model training specifically comprises:

[0193] The training sample and the test sample division mode are: the collected employee trajectory data is mixed according to the employee ID, and is divided into training samples and test samples according to the proportion of 70%, 30%;

[0194] The model loss function adopts a cross-entropy loss function to measure the difference between the recognized work type category probability and the real work type category:

[0195]

[0196] Wherein, M is the number of coal mine underground employee work types; y ic Is a symbol function, if the real category of sample i is equal to c, then 1, otherwise 0; p icis the predicted probability that the input sample i belongs to class c.

[0197] The technical solutions of the present application will be described in further detail below in combination with the drawings and specific embodiments:

[0198] Embodiment 1:

[0199] The specific embodiments of the present application are described below in combination with the drawings and specific embodiments: Figs. 1-2 The specific embodiments of the present application are described below in combination with the drawings and specific embodiments:

[0200] An employee movement pattern learning method applied to coal mine underground job identification, in an embodiment of a coal mine underground employee movement trajectory data UWB data set, includes the following steps:

[0201] Step 1, based on the obtained employee UWB trajectory data, the UWB trajectory data is preprocessed.

[0202] Based on the UWB trajectory data of the employees working in a coal mine underground for a day shift and a night shift, a total of 922560 UWB trajectory data are included, each trajectory data records the employee ID, time, latitude and longitude, work point and sensor environmental monitoring data in the employee movement process, such as temperature, humidity, gas concentration, CO and other information.

[0203] Based on the obtained employee unique identification ID and time sequence chaotic employee UWB trajectory data, for each time slice, the corresponding trajectory points are grouped according to the employee ID, the trajectory data of each employee is cleaned and interpolated, the data set is mixed according to the employee ID, and is divided into training samples and test samples according to the proportion of 70%, 30%.

[0204] Step 2, according to the fixed time interval (6 hours), the trajectory data of each employee is divided into multiple sub-trajectory.

[0205] Step 3, the current input sub-trajectory sequence is subjected to strong enhancement processing and weak enhancement processing, and the strong enhancement, weak enhancement and unenhanced trajectory sequence is obtained, wherein the unenhanced trajectory sequence is the current input sub-trajectory sequence.

[0206] Step 4, the strong enhancement, weak enhancement and unenhanced trajectory sequence is input into the multi semantic embedding module, which first extracts the position coordinates and types of underground work points and sensors based on the underground tunnel map data, generates a space graph; extracts the position coordinates of the work points and sensors passed by the employees based on the trajectory sequence of the employees, and generates an access graph; merge the two graphs to construct a movement preference graph, input the movement preference graph into the GraphSAGE model, project the synthesized trajectory of the employees into a low-dimensional space, and learn the movement preference embedding representation of the employees, the processed embedding representation is At the same time, the sparse trajectory sequence is processed and spliced using the space transformation matrix and the time transformation matrix, and the processed trajectory sequence embedding representation is At the same time, the attribute information features are extracted using the word embedding method, and the processed attribute information embedding representation is Finally, the movement preference embedding, trajectory sequence embedding and attribute information embedding are connected through a connection layer to obtain a multi semantic embedding representation.

[0207] Step 5, input the multi semantic embedding representation obtained in step 3 into the transformer encoder of the global feature extraction module, the transformer encoder is composed of multiple encoder blocks, each encoder block is composed of a multi-head self-attention layer and a feedforward neural network, to process variable trajectory sequences, capture time correlation, and extract global semantic features of the trajectory; after transformer coding, input into the contrast learning module, learn the feature representation by comparing positive and negative samples, that is, make the similarity between anchor data and positive samples as large as possible greater than the similarity between anchor data and negative samples, wherein, the noise contrast estimation loss function is used to calculate the loss and train, and the final feature representation of the trajectory is obtained after training.

[0208] Step 6, input the feature representation processed by the global feature extraction module into the type of work recognition module, which is a single-layer fully connected network with a Softmax activation function, to calculate the probability of the input trajectory data belonging to each type of work.

[0209] Step 7, use the cross entropy loss function to train the type of work recognition model, and gradually train and optimize the parameters to realize the recognition of the type of work to which the employee trajectory belongs.

[0210] Step 8, experimental environment and hyperparameter setting:

[0211] All experiments are implemented on a server with Nvidia GTX 3080Ti GPU using Python, the deep learning framework is PyTorch 1.1.0, the programming language is Python 3.6, and Adam is used to optimize the training during the training. The initial learning rate is 0.001, which is reduced by 10% every 10 cycles, the batch size is 16, the total number of iterations for training is 160K, and the number of K layers in the GNN module is 2; the output size of the GNN module is 128.

[0212] The above only describes the preferred embodiments of the present application, and it should be noted that those skilled in the art can make several improvements and modifications without departing from the technical principles of the present application, and these improvements and modifications should also be considered as the protection scope of the present application.

Claims

1. A staff mobile mode learning method applied to coal mine underground job identification, characterized in that: The method introduces the UWB ultra-wideband positioning data of the downhole employee, the downhole roadway map data and the employee surrounding environment monitoring data; a work type recognition model is designed by using the trajectory data of the coal mine downhole employee; the work type recognition model comprises a multi-semantics embedding module, a global feature extraction module and a work type recognition module; the UWB is represented as Ultra Wide Band in English; After the trajectory data is preprocessed, multi-aspect semantic information embedding representation is obtained through the multi-semantics embedding module; then, global space-time dependency is obtained through the global feature extraction module; finally, the probability that the input trajectory data belongs to each work type is calculated through the work type recognition module, so as to improve the accuracy of work type recognition and provide technical support for the coal mine downhole scheduling platform, that is, information is pushed according to the classification results of the work type, different messages are pushed to different employees, the problem of one person with multiple cards is avoided, the intelligent scheduling of the coal mine downhole employee is realized, and the safety of downhole production is maintained; The specific steps are as follows: Step 1: collecting the moving trajectory information of the coal mine downhole employee, and preprocessing the collected data to obtain a plurality of sub-trajectory sequences of each employee; the preprocessing is sequentially grouping processing, cleaning processing, interpolation processing and sub-trajectory division processing; Step 2: performing trajectory enhancement processing on the sub-trajectory sequence preprocessed in step 1; the trajectory enhancement processing is strong enhancement processing and weak enhancement processing, respectively, to obtain strong enhancement, weak enhancement and non-enhancement trajectory sequences; the non-enhancement trajectory sequence is the current input sub-trajectory sequence; Step 3: using the strong enhancement, weak enhancement and non-enhancement trajectory sequences obtained in step 2 to pass through the moving preference embedding unit, the trajectory sequence embedding unit, the attribute information embedding unit and the embedding connection layer of the multi-semantics embedding module respectively, to generate multi-semantics embedding representation of the strong enhancement, weak enhancement and non-enhancement trajectory sequences; the moving preference embedding unit generates a space graph using road network data, generates a corresponding access graph using the strong enhancement, weak enhancement and non-enhancement trajectory sequences obtained in step 2, respectively, and merges the space graph and the corresponding access graph to construct a corresponding moving preference graph, and adopts a graph embedding method to obtain an embedding representation of the employee moving preference; the trajectory sequence embedding unit processes the strong enhancement, weak enhancement and non-enhancement trajectory sequences obtained in step 2 through position point embedding, timestamp embedding and splicing processing to obtain trajectory sequence embedding representation; the attribute information embedding unit calculates the moving speed and extracts sensor environment monitoring data, i.e., temperature, humidity, gas concentration and CO data, based on the strong enhancement, weak enhancement and non-enhancement trajectory sequences obtained in step 2, and adopts a word embedding method to obtain embedding representation of these attribute information; the embedding connection layer is a connection operation of the moving preference embedding representation, the trajectory sequence embedding representation and the attribute information embedding representation through a connection layer to obtain multi-semantics embedding representation; Step 4, the multi-semantics embedding representation of the strong enhanced, weak enhanced and non-enhanced trajectory sequence obtained in step 3 is processed by a transformer encoder and a contrast learning method in self-supervised learning, and a global feature extraction module is used to process the unlabeled and variable long trajectory sequence to obtain the global semantic feature of the trajectory; Step 5, the feature representation of the sample data learned in step 4 is used to calculate the probability of the input trajectory data belonging to each type of work by using a fully connected neural network through a type of work recognition module, so as to correctly classify positive samples and negative samples, and then the model can be better trained; Step 6, the employee trajectory data is divided into training samples and test samples, and the designed loss function is used to train the type of work recognition model, and the optimization parameters are gradually trained to realize the recognition of the type of work to which the employee trajectory belongs.

2. The employee movement pattern learning method for underground coal mine job identification according to claim 1, characterized in that: In step 1, the data preprocessing process is as follows: The underground coal mine employee movement trajectory information comprises: employee ID, position coordinates at multiple time points based on UWB positioning, and surrounding environment monitoring data during employee movement; The grouping processing is: grouping the corresponding trajectory points according to the unique identifier (employee ID) of the employee; The cleaning processing is: removing the trajectory points that frequently fluctuate in a long distance within each time slice, and the specific process is as follows: Step 1-1, the trajectory data obtained after grouping according to the employee ID is defined with a suitable time slice according to the sampling interval of the trajectory points, and the trajectory points are corresponded to the time slice one by one; the suitable time slice is between 3s-7s; Step 1-2, the center position of all data in each time slice is calculated; Step 1-3, the distance between all position information and the center position of the group is calculated, and the nearest data to the center position is retained, and the distance calculation formula is as follows: wherein, denotes the distance between points A and B, R represents the earth radius, A lat , A lot , B lat , B lot denote the longitude and latitude of points A and B, respectively; Step 1-4, the center position of all retained data is recalculated; each group represents the trajectory data of an employee, and only one position data of each employee is retained at the same time; The interpolation processing is: for a missing trajectory point in a time slice, if there are trajectory points in the adjacent time before and after the time slice, the trajectory points are interpolated according to the adjacent time trajectory points, and the interpolation position is the center of the trajectory points before and after the time slice, to obtain the missing trajectory point; The sub-trajectory division processing is: the trajectory data of each employee after cleaning processing and interpolation processing is divided into a plurality of sub-trajectory sequences according to a fixed time interval of 6 hours.

3. The employee movement pattern learning method for underground coal mine job identification of claim 1, characterized in that: Step 2 specifically comprises: The trajectory enhancement processing includes strong enhancement processing and weak enhancement processing, and the specific process is as follows: the sub-trajectory sequence tr after step 1 preprocessing is respectively subjected to strong enhancement processing to obtain a strong enhancement trajectory sequence and weak enhancement processing to obtain a weak enhancement trajectory sequence The unenhanced trajectory sequence, i.e. the current input sub-trajectory sequence tr, the strong enhancement trajectory sequence and the weak enhancement trajectory sequence are simultaneously input into a multi-semantics embedding module, the unenhanced trajectory sequence tr after passing through the multi-semantics embedding module and a transformer encoder serves as anchor data for contrast learning in a global feature extraction module, the strong enhancement trajectory sequence after passing through the multi-semantics embedding module and the transformer encoder serves as positive sample data for contrast learning in the global feature extraction module, and the weak enhancement trajectory sequence after passing through the multi-semantics embedding module and the transformer encoder serves as negative sample data for contrast learning in the global feature extraction module. The strong enhancement processing adopts a neighborhood enhancement method. For the sub-trajectory sequence tr obtained after preprocessing, when the input time interval is t i , k+1 sub-trajectories in the time interval [t i-k / 2 :t i+k / 2 ] are spliced with tr to form a long-term enhanced trajectory tr The weak enhancement processing adopts a random enhancement method, for a sub-trajectory sequence tr obtained after preprocessing, when the input time interval is t i k sub-trajectories are randomly sampled from all trajectories of the corresponding employee, and then spliced according to time information and the input trajectory sequence to form a long-term trajectory 4. The employee movement pattern learning method for underground coal mine job identification of claim 1, characterized in that: Step 3 specifically comprises: The road network data is: underground mine roadway map data, comprising underground roadway information, underground working area and position coordinates, name and type of each sensor, used to generate a static space graph according to the road network data; The attribute information comprises: the moving speed of the employee and the sensor environment monitoring data during the movement; the sensor environment monitoring data comprises: temperature, humidity, gas concentration and CO. The mobile preference embedding unit is: generating a space graph using the road network data, generating an access graph using the trajectory sequence obtained in step 2, merging the two graphs to construct a mobile preference graph, using graph embedding to project the synthesized trajectory of the employee in the mobile preference graph into a low-dimensional space to obtain the mobile preference embedding representation of the employee; The spatial graph generation method is to take the position coordinates of the working points or sensors extracted from the underground roadway map as nodes V of a spatial graph G spatial s The feature vector of each node is initialized as information corresponding to the working point or sensor, such as a position, a sensor type, and a working point name. s The edges E of the spatial graph represent a group of edges connecting working points or sensors of the same type.​ The access graph generation method is to take the work points and sensor positions accessed by the employee in a time period on the employee track sequence as nodes V of an access graph G visit <V,E v , the feature vector of each node is initialized as the position point information of the employee at the work point or sensor, including employee ID, timestamp, position coordinates, E v represents all access connections, and each e ij ∈E v represents the access order of the employee from site i to site j; The mobile preference graph generation method is to merge the space graph G spatial and the access graph G visit to generate a mobile preference graph G full , G full is initially the same as G spatial , in the merging process, the edges appearing in G visit but not in G spatial are added to G full , and the node initial feature vectors are connected, after merging, the initial feature vector of each node finally includes employee ID, timestamp, position coordinate, sensor type or work point name; The graph embedding method adopts a graph sampling aggregation GraphSAGE model to learn an embedding representation of each node, and an input is a mobile preference graph G full An initial feature vector of each node is x v , v e V, after sampling, aggregation, splicing, and normalization operations, a final output mobile preference embedding representation is The GraphSAGE is Graph Sample and Aggregate; The trajectory sequence embedding unit includes position point embedding, timestamp embedding and splicing processing, and a transformation matrix is used to project the one-hot vectors of the timestamps and position points in the trajectory sequence into a low-dimensional space and perform splicing processing to obtain the trajectory sequence embedding representation of the employee; The position point embedding process is: for each position point l in the unenhanced trajectory sequence tr represented by a |l| dimension one-hot vector, using the transformation matrix E l Transforming l into a low-dimensional vector to capture accurate semantic information; the strongly enhanced and weakly enhanced trajectory sequences are also processed in the same way; when a position point embedding of an unenhanced trajectory sequence tr at time t can be represented as: where matrix E l The model is continuously learned during the model training process, wherein is a learnable representation matrix composed of all position vector representations, D l is the dimension of the position embedding; The timestamp embedding process is that the time sequence in the non-enhanced track sequence tr is composed of timestamp information, first, the time value is discretized, that is, each timestamp value is mapped to different time intervals in hours per day, at this time, the number of time intervals is 24, a learnable vectorized representation is associated with each interval, then a time representation matrix can be formed where D t is the dimension of the timestamp embedding; the strong enhanced and weak enhanced track sequences are also processed in the same way; when a non-enhanced track sequence tr at time t is timestamp embedded and can be expressed as: Wherein, map(t) is obtained by mapping one-hot vector, matrix E t Continuously learn in the model training process; The concatenation process concatenates the position point embedding and the timestamp embedding to generate a global representation where D g = D l + D t ; the same is done for the strong and weak augmented trajectory sequences; The attribute information embedding unit is: using a word vector embedding method to convert each attribute into a low-dimensional real vector to obtain the attribute information embedding representation of the employee trajectory, and each attribute is respectively a moving speed and sensor environmental monitoring data; The specific process is as follows: Step 3-1, the employee moving speed v uses the distance and time change between the i-th point and the (i+1)-th point to describe different speed characteristics, which can be expressed as: A spe = Δd · (t i+1 -t i ) -1 where d represents the distance between the ith point and the (i+1)th point, t i denotes the timestamp of the ith point; Step 3-2, calculate the attribute information of each timestamp of the strongly enhanced, weakly enhanced and unenhanced trajectory sequence, and embed the attribute information into a low-dimensional real vector; When the attribute information feature of a unenhanced trajectory sequence tr at time t can be expressed as: where W e , W s represent the weights of the fully connected layer, A env represents the sensor environment monitoring data at time t, A spe represents the velocity at time t, W e , W s , b e , b s are learnable parameters, and CONCAT represents the concatenation between the attribute vectors; The connection operation is: using a connection layer to connect the mobile preference embedding representation, the trajectory sequence embedding representation and the attribute information embedding representation of the strongly enhanced, weakly enhanced and unenhanced trajectory sequence respectively to obtain the multi-semantics embedding representation: where H tr ∈ R m*n m denotes the dimension of the matrix rows and n denotes the dimension of the columns.

5. The employee movement pattern learning method for underground coal mine job identification of claim 1, characterized in that: The step 4 specifically includes: The transformer encoder includes a plurality of encoder blocks, and the specific process is as follows: H tr ∈R k*d H is decomposed into m n-dimensional vectors x t as input to a transformer encoder, where x t is the input vector at the t-th time step, i.e. the t-th component of the input sequence X; the transformer encoder consists of multiple encoder blocks, and the vector x t is sequentially passed through a multi-head self-attention layer, an Add&Norm operation, a feed-forward neural network and an Add&Norm operation in the encoder block, and finally outputs a vector of the same dimension. The contrastive learning is a self-supervised learning method, for unlabeled data, by comparing the similarity of anchor data and positive and negative samples, the similarity of anchor data and positive samples is much larger than that of negative samples, to learn the general features of the data, the specific process is as follows: Step 4-1, positive and negative sample selection, in the transformer encoder module, using a transformer-based sequence-to-sequence neural network to convert the employee's unenhanced trajectory sequence tr ≤t = {c1, c2, …, c t} into a vector h t of the same dimension as the anchor data; the strong enhanced trajectory sequence in the trajectory enhancement module is also subjected to the multi-semantics embedding module and the transformer encoder to obtain the positive sample for the contrastive learning of the input trajectory sequence The weak enhanced trajectory sequence in the trajectory enhancement module is also subjected to the multi-semantics embedding module and the transformer encoder to obtain the negative sample for the contrastive learning of the input trajectory sequence Step 4-2, use the method of self-supervised learning to train the model, and learn the encoder f to make: Wherein, score(·,·) is a measure function to measure the similarity between samples, in contrastive learning, by comparing positive and negative samples to learn the representation, let the left part as much as possible greater than the right part; In the training process, the InfoNCE loss function is calculated by using noise contrast estimation, and the final feature representation of the trajectory is obtained after training, and the formula of the noise contrast estimation loss function is as follows:

6. The employee movement pattern learning method for underground coal mine job identification of claim 5, characterized in that: Each encoder block in the transformer encoder specifically includes: The multi-head self-attention layer specifically includes: Step 4-3-1, calculating input vector x t Q, K and V matrices for each value; the Q matrix is the Query matrix, the K matrix is the Key matrix, and the V matrix is the Value matrix; Q = W q x t K = W k x t V = W v x t where W q , W k , and W v are learnable matrices; Step 4 - 3 - 2, the dot product between Q and K is computed, which is divided by where d k is the dimension of the Key vector, and the result is normalized to a probability distribution using the normalized exponential Softmax activation function, and multiplied by the matrix V to get the weighted sum representation, which is the output of the self-attention layer of this encoder block, computed as follows: Step 4-3-3, the multi-head self-attention layer includes multiple self-attention layers, and the input x t is respectively transmitted into h different self-attention layers, h output matrices Attention(Q, K, V) are calculated, different Attention are spliced, and a linear transformation is performed again, and the calculation process is as follows: MultiHead(Q, K, V) = Concat(head1,..., head h ) W O wherein W O is a learnable matrix; The feedforward neural network is: inputting the weight obtained by the self-attention layer into the feedforward neural network, and the feedforward neural network is a two-layer fully connected layer, the activation function of the first layer is Relu, and the second layer does not use the activation function, and the corresponding formula is as follows: FFN(x)=max(0,xW1+b1)W2+b2 The Add&Norm operation is: after the input sequence respectively passes through the multi-head self-attention layer and the feedforward neural network, there is a residual connection Add operation, and then a layer normalization operation is performed, the layer normalization operation is denoted as Norm in English, and the calculation process is as follows: sub_layer_output=LayerNorm(x+SubLayer(x)) Wherein, x represents the input of the multi-head self-attention layer and the feedforward neural network, and SubLayer(x) represents the output of the multi-head self-attention layer and the feedforward neural network, and the dimensions are the same.

7. The employee movement pattern learning method for underground coal mine job identification of claim 1, characterized in that: The step 5 specifically comprises: The full connection neural network is: a full connection neural network using a normalized exponential Softmax activation function is used for classification probability statistics, all neurons in each layer of the network are controlled between 0-1, and the calculation formula is as follows: where x i is the output value of the i-th node, and n is the number of output nodes, i.e., the number of classification categories.

8. The employee movement pattern learning method for underground coal mine job identification of claim 1, characterized in that: The step 6 specifically comprises: The training sample and test sample division mode is: the collected employee trajectory data is mixed according to the employee ID, and is divided into training samples and test samples according to the proportion of 70%, 30%; The model loss function adopts a cross-entropy loss function to measure the difference between the recognized work type category probability and the real work type category: where M is the number of coal mine underground employee job types; y ic is the indicator function that takes the value 1 if the true class of sample i equals c and 0 otherwise; p ic is the predicted probability that input sample i belongs to class c.

Images