A multi-dimensional asynchronous time series data classification method based on spatial structure graph learning

Abstract

The application belongs to the field of graph computing and time series data prediction, and discloses a multi-dimensional asynchronous time series data classification method based on spatial structure graph learning, which considers the interaction between sensors, constructs a graph for multi-dimensional data at each time point respectively, uses a graph neural network to realize information update of each sensor data, fills in the missing observation values in the process of information propagation, and learns the mutual relationship between different time intervals by using a time interval-based Transformer mechanism. In the process of classifying asynchronous time series data, the application realizes mutual propagation of information between different sensors, fills in the missing observation values, improves the robustness of the algorithm under a large amount of missing data, considers the correlation of time series data in time series, and improves the classification effect of multi-dimensional asynchronous time series data.

Description

Technical Field

[0001] This invention belongs to the field of graph computing and time series data prediction, and particularly relates to a multidimensional asynchronous time series data classification method based on spatial structure graph learning. Background Technology

[0002] Multidimensional asynchronous time series data are characterized by multiple dimensions, asynchrony, and time sequence. In real-life scenarios (such as medical, meteorological, financial, and industrial fields), the sampling and observation time intervals are irregular, which poses a great challenge to the classification and analysis of multidimensional asynchronous time series data.

[0003] For multidimensional asynchronous time-series data, the current mainstream approach is to select the smallest asynchronous sampling time interval and perform data prediction and imputation on the missing data in each dimension to form regular multidimensional time-series data; then, existing deep learning algorithms are used to complete the classification task. However, in terms of data prediction and imputation, existing methods only use missing value imputation techniques to fill in missing data in single-dimensional data, without deeply considering the data correlation between multiple dimensions, resulting in low data feasibility; in terms of deep learning algorithms, current classification algorithms do not consider the temporal correlation of time-series data, resulting in poor classification performance. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a multidimensional asynchronous time-series data classification method based on spatial structure graph learning.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A multidimensional asynchronous time-series data classification method based on spatial structure graph learning includes the following steps:

[0007] Step 1: Based on the individual multidimensional asynchronous time series sample data, assign the missing observation value to 0, align the time series data on the individual sample, and align the duration of all samples according to the longest duration of all samples.

[0008] Step 2: For each sample, initialize the relationship graph structure between all sensors as a directed fully connected graph, where the weight of each directed edge is 1. This relationship graph structure will be updated and iterated during the message passing process.

[0009] Step 3: For each sample, update all sensor data at each time point of the sample through a message passing mechanism, and update the graph structure according to the attention weights generated during message propagation;

[0010] Step 4: Consider the correlation between each sensor before and after different times in the sample, and use the time-interval-based Transformer method to integrate the data from each time point of the sensor to generate a single sensor embedding representation;

[0011] Step 5: Combine the data from each sensor of the sample with the static features of the sample to obtain the final features of the sample, and finally use a linear classification method to classify and train the sample.

[0012] Furthermore, step one includes the following specific steps:

[0013] Given N multidimensional asynchronous time-series sample data Where y i ∈{1,2,...,C} is a sample S i The labels are first applied to each individual sample. in This indicates that the sample is located in F sensors {u1, u2, ..., u...} F The time when the data was observed on} Indicates sample S i The sensor u takes observations at time τ; if an observation is missing, it is filled with 0; all time series are then processed. The duration is uniformly converted to T, where T is the longest duration of all observed samples. The specific operation is as follows: for all sequences with a duration greater than T, take the T nearest durations; for sequences with a duration less than T, pad them with 0s to make the duration T.

[0014] Furthermore, step two includes the following specific steps:

[0015] In the initialization phase of the sample sensor relationship graph, for each sample S i Construct a directed weighted graph G i ={V,E i}, where V represents the set of nodes consisting of all sensors, and E i Let represent the set of directed edges between any two sensors. This represents the magnitude of the influence of sensor u on sensor v, and when there are observations on sensor u, they will be transmitted through the directed edge. Transmit information to sensor v, when At this time, there is no information transmission between sensor u and sensor v. According to the definition of a directed edge, in general, when u ≠ v,

[0016] Before information propagation, the directed graph is defined as a directed fully connected graph, and the weight of each directed edge is initialized to 1, that is, for sample S i For any two sensors u and v,

[0017] Furthermore, step three is implemented through the following sub-steps:

[0018] Step 3.1: Map the observation values ​​of the sample on the sensor to a high-dimensional space, and realize the propagation and updating of observation data information between sensors according to the relationship diagram between sensors;

[0019] Step 3.2: Update the relationship graph structure and weights between sensors based on the intermediate results calculated during the information propagation process.

[0020] Furthermore, step three includes the following specific steps:

[0021] First, use a nonlinear mapping. Sample S i One-dimensional observation at time τ on sensor u Mapping to a higher dimension Where σ is a nonlinear activation function, O u These are trainable parameters that are only related to the sensor u, d h =4 is a predefined dimension parameter;

[0022] Then, sample S i The observation data at each time point will undergo L-layer information propagation based on the constructed inter-sensor relationship diagram, where L=2 is a predefined parameter. In the l-th layer, sample S i Information at time τ on sensor u It will aggregate the neighbor nodes it observes. The information is updated based on its information from the previous moment, using the following update formula:

[0023]

[0024] Where ⊙ represents the dot product operation, μ = 0.3 is the predetermined weighting parameter for the influence of the previous time step on the information at the current time step, and w u , These are trainable parameters that are only related to the sensor values ​​u and v. Indicates in sample S i The magnitude of the influence of sensor u on sensor v, It is the normalized attention weight of sensor v on u at time τ, that is, in sample S i The magnitude / importance weight of the influence of sensor v information on sensor u at time τ is calculated as follows:

[0025]

[0026]

[0027] in d are trainable vector parameters that are only related to sensor u. l =16 is a predefined parameter. It is From d h Mapping 1D space to d l +d p Trainable parameters in dimensional space, d p =16 is a predefined parameter. The position code at time τ is defined as a series of trigonometric functions with frequency F:

[0028]

[0029] Where 2k and 2k+1 represent feature dimensions, the odd and even columns of position encoding are calculated differently, F = 10000 and d p =16 is a predefined parameter;

[0030] During information dissemination, the weights of the global sensor relationship graph are updated by aggregating the attention weight information learned from different sensors at different times:

[0031]

[0032] The structure of the sensor relationship graph on the sample is updated by pruning the minimum K% edge weights for each sample, where K=50 is a predefined parameter.

[0033] Furthermore, step four is implemented through the following sub-steps:

[0034] Step 4.1: Concatenate the embedded representation information of the sample at each time step on each sensor with the temporal location encoding information to obtain the embedded representation containing temporal location information;

[0035] Step 4.2: Use a time-interval-based Transformer mechanism to learn the information about the mutual influence between each time step;

[0036] Step 4.3: Sum and average the embedding representations of each time point with observations on each sensor for each sample to obtain the final masked sensor embedding representation for each sensor of each sample.

[0037] Furthermore, step four includes the following specific steps:

[0038] Sensor embedding expression generation stage, targeting sample S i First, embed the information at time τ on sensor u. With time and position encoding p τ splicing * T This represents the transpose of a matrix. Indicates sample S i The embedded representation of time and location information at time τ on sensor u, matrix Indicates sample S i Embedded representation on sensor u;

[0039] Then, a time-interval-based self-attention mechanism is used instead of the standard self-attention mechanism in Tansformer. generate and value(V) i u ):

[0040]

[0041]

[0042]

[0043] in These represent the trainable matrix parameters related to generating the query, key, and value, respectively. A fully connected layer is used to obtain the time interval information. The specific calculation process is as follows:

[0044]

[0045]

[0046] in For time-interval-based self-attention weights, For the final output, σ is the Softmax function. For sample S k The time interval matrix, where T k ={t k,1 ,t k,2 ,...,t k,T} represents sample S k The time series, and Represents time t k,i and t k,j The time intervals are then normalized using the time interval matrix. in Representing the time interval matrix The minimum value other than 0;

[0047] Sample S i The data at the moments when observations are recorded on sensor u are aggregated to obtain sample S. i Mask embedding representation on sensor u:

[0048]

[0049] in Representation matrix The j-th column vector, For sample S i The mask vector on sensor u represents the sample S. i Observations on sensor u Which part is the missing value?

[0050]

[0051] Representing vectors The j-th element.

[0052] Furthermore, step five is implemented through the following sub-steps:

[0053] Step 5.1: Concatenate the embedded representations of all sensors for each sample with the static attribute features of the sample to obtain the embedded representation of a single sample;

[0054] Step 5.2: Use a classifier to classify the samples and use the SDG gradient descent method to train the model. During the model training process, both the cross-entropy loss function and the difference loss function between the graph structures of each sample are considered.

[0055] Furthermore, step five includes the following specific steps:

[0056] For data that lacks static attribute features, sample S i All sensors u1, u2, ..., u F Data Perform a concatenation operation to obtain the final feature embedding representation of the sample.

[0057]

[0058] For data with static attribute features, sample S i All sensors u1, u2, ..., u F Data and static feature data Perform a concatenation operation to obtain the final feature embedding representation of the sample.

[0059]

[0060] After obtaining the sample embeddings, a classifier is used to classify the samples. The classifier is defined as a two-layer MLP, with the final output layer having C neurons, where C is the number of sample classes. The specific calculation formula is as follows:

[0061]

[0062] in For sample S i The predicted output, σ is an activation function, W (0) and W (1) These are the trainable matrix parameters in the model;

[0063] Finally, the model was trained using the stochastic gradient descent (SGD) optimization method; the model's loss function consists of two parts, Loss = L CE +λL r Where λ = 0.001 is a predefined parameter. Represents the cross-entropy loss function. F represents the difference loss function between the sample image structures learned by the model, which is the average Euclidean distance between each pair of samples and each pair of sensors, where F represents the number of sensors and N represents the number of samples.

[0064] The beneficial effects of this invention are as follows: This invention is a multidimensional asynchronous time series data classification method based on the time interval Transformer, oriented towards the fields of graph computation and time series data prediction, and has the following characteristics:

[0065] (1) This invention proposes a multidimensional asynchronous time-series data classification method based on spatial structure graph learning. By modeling the interaction between multiple sensors as a graph structure, a graph neural network is used to realize the mutual propagation of sensor data information. During the information propagation process, missing observations are automatically filled in. The algorithm has high robustness when a large amount of sensor data is missing.

[0066] (2) This invention proposes to use a time-time-based Transformer mechanism, which considers the correlation between time-series data and learns the interrelationship between different times in the sensor, thereby improving the classification effect of the algorithm. Attached Figure Description

[0067] Figure 1 The diagrams provided are example diagrams of multidimensional asynchronous time-series data and sample duration alignment, as presented in this invention.

[0068] Figure 2 This is a diagram illustrating the overall framework of the algorithm of this invention. Detailed Implementation

[0069] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are for illustrative purposes only and are not intended to limit the scope of the invention.

[0070] like Figure 2 As shown, the present invention provides a multidimensional asynchronous time-series data classification method based on spatial structure graph learning, comprising the following steps:

[0071] Step 1: Based on the data of a single multidimensional asynchronous time series sample, assign the missing part to 0, align the time series data on a single sample, and align the duration of all samples according to the longest duration of all samples.

[0072] During the sample duration alignment phase, for the given N multidimensional asynchronous time-series sample data... Where y i ∈{1,2,...,C} is a sample S i The labels are first applied to each individual sample. in This indicates that the sample is located in F sensors {u1, u2, ..., u...} F The time when the data was observed on} Indicates sample S i The sensor u takes observations at time τ, and if an observation is missing, it is filled with 0. All time series are then processed. All durations are uniformly converted to T, where T is the longest duration of all observed samples. Specifically, for all sequences with a duration greater than T, the T nearest durations are taken; for sequences with a duration less than T, they are padded with zeros to reach a duration of T. For example... Figure 1 As shown, (a) is multidimensional asynchronous time series sample data, and (b) is the result after the sample duration is aligned to T=5. Solid dots represent observed values, and hollow dots represent the filled value 0.

[0073] Step 2: For each sample, initialize the relationship graph structure between all sensors as a directed fully connected graph, where the weight of each directed edge is 1. This relationship graph structure will be updated and iterated during the message passing process.

[0074] In the initialization phase of the sample sensor relationship graph, for each sample S i Construct a directed weighted graph G i ={V,E i}, where V represents the set of nodes consisting of all sensors, and E i Let represent the set of directed edges between any two sensors. For any This represents the magnitude of the influence of sensor u on sensor v, and when there are observations on sensor u, they will be transmitted through the directed edge. Transmit information to sensor v, when At this time, there is no information transmission between sensor u and sensor v. According to the definition of a directed edge, in general, when u ≠ v,

[0075] Before information propagation, the directed graph is defined as a directed fully connected graph, and the weight of each directed edge is initialized to 1, that is, for sample S i For any two sensors u and v,

[0076] Step 3: For each sample, update all sensor data at each time point of the sample through a message passing mechanism, and update the graph structure according to the attention weights generated during message propagation;

[0077] Step 3.1: Map the observation values ​​of the sample on the sensor to a high-dimensional space, and realize the propagation and updating of observation data information between sensors according to the relationship diagram between sensors;

[0078] First, use a nonlinear mapping. Sample S i One-dimensional observation at time τ on sensor u Mapping to a higher dimension Where σ is a nonlinear activation function (e.g., a linear mapping + ReLU activation function), O u These are trainable parameters that are only related to the sensor u, d h =4 is a predefined dimension parameter.

[0079] Then, sample S i The observation data at each time point will undergo L-layer information propagation based on the constructed inter-sensor relationship diagram, where L=2 is a predefined parameter. In the l-th layer, sample S i Information at time τ on sensor u It will aggregate the neighbor nodes it observes. The information is updated based on its information from the previous moment, using the following update formula:

[0080]

[0081] Where ⊙ represents the dot product operation, μ = 0.3 is the predetermined weighting parameter for the influence of the previous time step on the information at the current time step, and w u , These are trainable parameters that are only related to the sensor values ​​u and v. Indicates in sample S i The magnitude of the influence of sensor u on sensor v, It is the normalized attention weight of sensor v on u at time τ, that is, in sample S iThe magnitude / importance weight of the influence of sensor v information on sensor u at time τ is calculated as follows:

[0082]

[0083]

[0084] in d are trainable vector parameters that are only related to sensor u. l =16 is a predefined parameter. It is From d h Mapping 1D space to d l +d p Trainable parameters in dimensional space, d p =16 is a predefined parameter. The position code at time τ is defined as a series of trigonometric functions with frequency F:

[0085]

[0086] Where 2k and 2k+1 represent feature dimensions, the odd and even columns of position encoding are calculated differently, F = 10000 and d p =16 is a predefined parameter.

[0087] Step 3.2: Update the relationship graph structure and weights between sensors based on the intermediate results calculated during the information propagation process;

[0088] During information dissemination, the weights of the global sensor relationship graph are updated by aggregating the attention weight information learned from different sensors at different times:

[0089]

[0090] The structure of the sensor relationship graph on the sample is updated by pruning the smallest K% edge weights for each sample. Here, K=50 is a predefined parameter.

[0091] Step 4: Consider the correlation between each sensor before and after different times in the sample, and use the time-interval-based Transformer method to integrate the data from each time point of the sensor to generate a single sensor embedding representation;

[0092] Step 4.1: Concatenate the embedded representation information of the sample at each time step on each sensor with the temporal location encoding information to obtain the embedded representation containing temporal location information;

[0093] Sensor embedding expression generation stage, targeting sample S i First, embed the information at time τ on sensor u. With time and position encoding p τ splicing * T This represents the transpose of a matrix. Indicates sample S i The embedded representation of time and location information at time τ on sensor u, further, matrix Indicates sample S i Embedded representation on sensor u.

[0094] Step 4.2: Use a time-interval-based Transformer mechanism to learn the information about the mutual influence between each time step;

[0095] Then, a time-interval-based self-attention mechanism is used instead of the standard self-attention mechanism in Tansformer. Specifically, using... generate and value(V) i u ):

[0096]

[0097]

[0098]

[0099] in These represent the trainable matrix parameters related to generating the query, key, and value, respectively. A fully connected layer is used to obtain the time interval information; the specific calculation process is as follows:

[0100]

[0101]

[0102] in For time-interval-based self-attention weights (multi-head attention can also be used here), The final output is σ, which is the Softmax function. For sample S k The time interval matrix, where T k ={t k,1 ,t k,2 ,...,t k,T} represents sample S k The time series, and Represents time t k,i and t k,j The time interval between each point. Here, to accelerate the convergence of the model, the time interval matrix is ​​normalized: in Representing the time interval matrix The minimum value other than 0.

[0103] Step 4.3: Sum and average the embedding representations of each time point with observations on each sensor for each sample to obtain the final masked sensor embedding representation for each sensor of each sample;

[0104] Sample S i The data at the moments when observations are recorded on sensor u are aggregated to obtain sample S. i Mask embedding representation on sensor u:

[0105]

[0106] in Representation matrix The j-th column vector, For sample S i The mask vector on sensor u represents the sample S. i Observations on sensor u Which part is the missing value?

[0107]

[0108] Representing vectors The j-th element.

[0109] Step 5: Combine the data from each sensor of the sample with the static features of the sample to obtain the final features of the sample, and finally use a linear classification method to classify and train the sample.

[0110] Step 5.1: Concatenate the embedded representations of all sensors for each sample with the static attribute features of the sample (if they exist) to obtain the embedded representation of a single sample;

[0111] In the sample embedding, representation, and classification stage, for data lacking static attribute features, sample S is... i All sensors u1, u2, ..., u F Data Perform a concatenation operation to obtain the final feature embedding representation of the sample.

[0112]

[0113] For data with static attribute features, sample S i All sensors u1, u2, ..., u F Data and static feature data Perform a concatenation operation to obtain the final feature embedding representation of the sample.

[0114]

[0115] Step 5.2: Use a classifier to classify the samples and use the SDG gradient descent method to train the model. During the model training process, both the cross-entropy loss function and the difference loss function between the graph structures of each sample are considered.

[0116] After obtaining the sample embeddings, a classifier is used to classify the samples. Here, the classifier is defined as a two-layer MLP, with the final output layer having C neurons, where C is the number of sample classes. The specific calculation formula is as follows:

[0117]

[0118] in For sample S i The predicted output, σ is an activation function (e.g., the sigmoid activation function), W (0) and W (1) These are the trainable matrix parameters in the model.

[0119] Finally, the model is trained using the stochastic gradient descent (SGD) optimization method. The model's loss function consists of two parts: Loss = L CE +λL r Where λ = 0.001 is a predefined parameter. Represents the cross-entropy loss function. F represents the difference loss function between the sample image structures learned by the model (the average Euclidean distance between each pair of samples and each pair of sensors), where F represents the number of sensors and N represents the number of samples.

Claims

1. A method for classifying multi-dimensional asynchronous time series data based on spatial structure graph learning, characterized in that, The method comprises the following steps: Step 1: Collect medical data, assign missing observations as 0 according to single multi-dimensional asynchronous time series sample data, align time series data on a single sample, and align all sample lengths according to the longest time length of all samples; For the given N multi-dimensional asynchronous time series sample data where y i ∈{1,2,...,C} is the label of sample S i , first for each single sample where denotes the time when the sample observes data on F sensors {u1,u2,...,u F}, denotes the observation value of sensor u of sample S i at time τ, if the observation value is missing, it is filled with 0; all time series lengths are uniformly converted to T, where T is the longest length of all observation samples, and the specific operation is: for all sequences with a length greater than T, take the latest T length; for sequences with a length less than T, fill them up to length T by filling 0 after them. Step 2: For each sample, initialize the relationship graph structure between all sensors as a directed fully connected graph, where the weight of each directed edge is 1, and the relationship graph structure is updated iteratively during the message passing process; Step 3: For each sample, update all sensor data information at each time point of the sample through the message passing mechanism, and update the graph structure according to the attention weight generated during the message propagation process; Step 4: Considering the correlation of each sensor in the sample before and after different times, using the time interval based Transformer method, integrating the data of each time point of the sensor, generating a single sensor embedding expression; The sensor embedding representation generation stage, for a sample S i first embeds the information of the sample S at a time instant τ on a sensor u τ is concatenated with a time position encoding p * T denotes the transpose of a matrix, in this case denotes the embedding representation of the sample S i containing time position information at a time instant τ on a sensor u, the matrix denotes the embedding representation of the sample S i on a sensor u; Then, a time interval based self-attention mechanism is used instead of the standard self-attention mechanism in Tansformer, using generating and wherein respectively represent trainable matrix parameters related to generating query, key, value, and use a fully connected layer to obtain time interval information, and the specific calculation process is: wherein is the time interval based self-attention weight, is the final output, and σ is the Softmax function, is the sample S k 's time interval matrix, wherein T k = {t k,1 , t k,2 ,...,t k,T} represents the time sequence of the sample S k , and represents the time interval between time t k,i and t k,j , and the time interval matrix is normalized as follows: wherein represents the minimum value of the time interval matrix except 0; The sample S i The data of the time instants with observations on the sensor u are summarized to obtain the sample S i The mask embedding expression on the sensor u is where the jth column vector of the matrix , is the sample S i the mask vector at sensor u, indicating the sample S i the observation at sensor u which part of the vector the jth element of the vector denotes the jth element of the vector Step 5: Combine the data of each sensor of the sample with the static features of the sample to obtain the final features of the sample, and finally use a linear classification method to train the sample.

2. The method of claim 1, wherein, The step 2 comprises the following specific steps: The sample sensor inter-relationship graph initialization phase, for each sample S i A directed weighted graph G is constructed i = {V, E i}, where V represents a node set composed of all sensors, E i represents a directed edge set composed of two sensors, for any represents the influence size of sensor u on sensor v, and when there is an observation value on sensor u, the information will be propagated to sensor v through the directed edge When , there is no information propagation between sensor u and sensor v, according to the definition of the directed edge, generally, when u≠v, Before the information propagation is performed, the directed graph is defined as a directed fully connected graph, and the weight of each directed edge is initialized as 1, that is, for sample S i For any two sensors u, v, 3. The method of claim 1, wherein, The step 3 is implemented through the following sub-steps: Step 3.1: Map the observations of the sample on the sensors to a high-dimensional space, and realize the propagation and update of the observation data information between the sensors according to the relationship graph between the sensors; Step 3.2: Update the relationship graph structure and weight between the sensors according to the intermediate results calculated during the information propagation process.

4. The method of claim 2, wherein, The step 3 comprises the following specific steps: First, a non-linear mapping is used The samples S i The 1-dimensional observation at time t on sensor u is mapped to a higher dimensional space where σ is a non-linear activation function, O u is a trainable parameter only related to sensor u, d h = 4 is a predefined dimensionality parameter; Then, the sample S i The observation data at each time will be propagated according to the constructed inter-sensor relationship graph with L layers of information propagation, where L = 2 is a predefined parameter, at the l-th layer, the sample S i The information at time t on sensor u Will aggregate its observed neighbor nodes And its previous time information, the update formula is as follows: where ⊙ denotes the dot product operation, μ = 0.3 is a predetermined influence weight parameter of the previous time on the current time information, is a trainable parameter only related to sensors u, v, denotes the influence of sensor v on sensor u at time τ in sample S i , is the normalized attention weight of sensor v on u at time τ, i.e., the importance weight of the sensor v information at time τ on sensor u in sample S i , which is calculated as follows: where is a trainable vector parameter related to sensor u only, d l = 16 is a predefined parameter, is a trainable parameter mapping from a d h dimensional space to a d l +d p dimensional space, p = 16 is a predefined parameter, is the position encoding at time step t, defined as a series of trigonometric functions with frequency F: where 2k, 2k+1 represent the feature dimension, the odd and even columns of the position encoding are calculated differently, F = 10000 and d p = 16 are predefined parameters; During the information propagation process, the weight of the global sensor relationship graph is updated by aggregating the learned attention weight information between different sensors at different times: And by pruning the smallest K% edge weight of each sample, the structure of the sensor relationship graph on the sample is updated, where K=50 is a predefined parameter.

5. The method of claim 1, wherein, The step 5 is implemented through the following sub-steps: Step 5.1: Concatenate the embedding expression of all sensors of each sample with the static attribute features of the sample to obtain the embedding expression of a single sample; Step 5.2: Use a classifier to classify the sample, and use the SDG gradient descent method to train the model, considering the cross-entropy loss function and the difference loss function between the graph structures of each sample during the model training process.

6. The method of claim 4, wherein, The step 5 comprises the following specific steps: For data without static attribute features, the sample S i All sensor data u1, u2,..., u F are spliced to obtain the final feature embedding expression of the sample​ For data with static attribute features, the sample S i The data of all sensors u1, u2,..., u F and static feature data are spliced to obtain the final feature embedding expression of the sample​ After obtaining the sample embedding expression, use a classifier to classify the sample, and the classifier is defined as a 2-layer MLP, and the final output layer has C neurons, where C is the number of sample categories, and the specific calculation formula is: wherein is the predicted output result of the sample S i , σ is an activation function, W (0) and W (1) are matrix parameters trainable in the model; The model is trained by using an optimization method of stochastic gradient descent (SGD); the loss function of the model includes two parts, i.e., Loss=L CE +λL r , wherein λ=0.001 is a predefined parameter, represents a cross-entropy loss function, represents a difference loss function between learned graph structures of each sample by the model, i.e., an average Euclidean distance between two sensors of each sample, wherein F represents the number of sensors, and N represents the number of samples.

Images