Industrial internet abnormal data treatment method and system based on space-time coupling characteristics

Abstract

The invention discloses an industrial internet abnormal data treatment method and system based on space-time coupling characteristics, and belongs to the technical field of industrial internet. The invention discloses a treatment method, which comprises the following steps of: constructing a GCN and LSTM (Long Short Term Memory) collaborative model, effectively representing a space structure of data by using the GCN, mining space correlation information among data point locations, modeling dynamic evolution of the data in a time dimension by using the LSTM, and predicting the data by using a double-layer GCN model and a double-layer LSTM model respectively. And finally, carrying out weighted fitting on prediction results of the two, thereby realizing more comprehensive and accurate analysis and prediction on complex spatio-temporal data. According to the invention, data prediction is carried out by using the GCN and LSTM collaborative model in the industrial data prediction task, and the prediction result is more accurate.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial Internet, and specifically relates to an industrial Internet abnormal data management method and system based on spatiotemporal coupling characteristics. Background Art

[0002] With the rapid development of the Industrial Internet, the data generated in the industrial production process has exploded. These data come from a wide range of sources, including industries on the production line, smart devices, monitoring systems, etc. However, among these massive amounts of data, there are a large number of abnormal data, which brings many challenges to the application of the Industrial Internet.

[0003] Industrial data often has obvious spatiotemporal characteristics. In the temporal dimension, data changes may be related to factors such as the periodicity of the production process and the aging of equipment; in the spatial dimension, different data points of the same equipment and data points of different equipment may affect each other. Therefore, the spatiotemporal characteristics of data are usually considered when managing abnormal data in industry.

[0004] Traditional data governance methods face many challenges when dealing with such data with complex structures and dynamic characteristics. For models based solely on graph convolutional networks (GCN), although they can effectively extract the spatial features of data and capture the topological structure information between nodes, they have limitations when processing time series information and it is difficult to accurately model the dynamic changes of data in the time dimension. Long short-term memory networks (LSTM) are good at processing time series data and can learn long-term dependencies in time series, but their ability to extract spatial structural features of data is relatively weak. Summary of the invention

[0005] The purpose of the present invention is to provide an industrial Internet abnormal data governance method and system based on spatiotemporal coupling characteristics, which is used to solve the technical problem of low analysis and prediction accuracy when existing data governance methods are used to manage complex spatiotemporal data.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] The present invention discloses an industrial Internet abnormal data governance method based on spatiotemporal coupling characteristics, comprising the following steps:

[0008] Collect and preprocess data on the industrial Internet to obtain industrial data sets;

[0009] Construct a topological graph G based on the industrial data set;

[0010] Construct a GCN and LSTM collaborative model, convert the industrial data set into the form of a topological graph G, and divide it into a training set and a test set. Train and test the constructed GCN and LSTM collaborative model to obtain a trained GCN and LSTM collaborative model.

[0011] Obtain historical data before the time period required for prediction, input it into the trained GCN and LSTM collaborative model, and output the prediction results.

[0012] Furthermore, the specific steps of data collection and data preprocessing to obtain industrial data sets are as follows:

[0013] Based on the database management system, industrial data is collected and converted into csv format; then data cleaning is used to pre-process the industrial data to obtain an industrial data set;

[0014] The industrial data set includes the data point name, data collection time and industrial detection value.

[0015] Furthermore, the specific steps of constructing the topology graph G according to the industrial data set are as follows:

[0016] Define a non-directed weighted graph G = (V, A, X);

[0017] Where: V = {v1, v2, ... v n} represents the node set of the graph; A∈Rn*n represents the adjacency matrix, a ij is the data in the i-th row and j-th column of matrix A, whose value is between 0 and 1, indicating that node v i and node v j The level of correlation between

[0018] X∈Rn*p represents the feature matrix, p is the length of the historical time series, x ij Represents the industrial data of the i-th node at the j-th moment.

[0019] Furthermore, the specific steps of converting the industrial data set into the form of a topological graph G are as follows:

[0020] The collected industrial data sets are classified according to the point names, and the historical time series of each data point is used as the feature matrix X. The adjacency matrix A is obtained using the multi-head attention mechanism.

[0021] Furthermore, the self-attention calculation formula and the parallel attention calculation formula in the multi-head attention mechanism are as follows:

[0022] Self-attention calculation formula:

[0023]

[0024] Among them, Q, K, and V represent query, key, and value matrices respectively, and d k is the dimension of the key vector, used to scale clicks to stabilize the softmax function;

[0025] Parallel attention calculation formula:

[0026]

[0027] in, Represent the transformation matrices of query, key, and value of the i-th head respectively;

[0028] The specific steps of using the multi-head attention mechanism to obtain the adjacency matrix A are as follows: Head i The matrix is ​​concatenated on the last dimension, and the concatenated multi-attention head fusion matrix is ​​averaged on the head dimension, and the final matrix is ​​used as the adjacency matrix A.

[0029] Furthermore, the GCN and LSTM collaborative model includes a double-layer GCN model, a double-layer LSTM model, and an MLP model that weightedly fits the output data of the double-layer GCN model and the output data of the double-layer LSTM model;

[0030] The double-layer GCN model and the double-layer LSTM model are used to predict the data of each point in the industrial Internet respectively. Finally, the connection layer is used to weighted fit the prediction results of the two models, and the final prediction results are output.

[0031] Furthermore, the GCN model is:

[0032]

[0033] in, I is the identity matrix; yes degree matrix; H is the feature of each layer. For the input layer, H is X; σ is the nonlinear activation function;

[0034] The two-layer GCN model is:

[0035] f(A,X)=A·(ReLU(A·(XW0))W1);

[0036] Among them, f(A,X) represents the final output result; ReLU() represents the linear rectification function; W0 represents the parameter matrix of the first layer; W1 represents the weight matrix mapping the hidden layer to the output layer;

[0037] The LSTM model includes:

[0038] Forget Gate:

[0039] f ι =σ(W f [h ι-l ,x ι ]+b f );

[0040] Among them, W f Represents the weight coefficient matrix of the forget gate; h ι-l represents the memory information of the hidden layer at the moment before time t, and the sigmoid activation function is used to discard the information; x ι represents the input information at time t; b f Represents the bias condition of the forget gate;

[0041] Input Gate:

[0042] i t =σ(W t [h t-1 ,x t ]+b t );

[0043]

[0044] Among them, the sigmoid activation function is first used to decide which state to update; W t , W c Represent the weight coefficient matrices of the input gate and memory unit respectively; then this state and vector Addition, vector It is created using the tanh function; finally combined with the current memory and long-term memory C t-1 Determine the new cell state C t ; b t and b c They represent the bias conditions of the input gate and memory unit respectively;

[0045] Output Gate:

[0046] o t =σ(W o [h t-l ,x t ]+b o );

[0047] h t =o t ×tanh(C t );

[0048] Among them, h t Represents the output of the first layer LSTM model; Wo Represents the weight coefficient matrix of the output gate; b o Represents the bias condition of the output gate.

[0049] Furthermore, the weighted fitting MLP model is:

[0050]

[0051] Among them, y j represents the jth component of the output layer; represents the input of the i-th neuron in the k-1-th hidden layer; w i,j,k represents the connection weight between the i-th component on the k-1-th hidden layer and the j-th neuron on the k-th hidden layer; f(x) represents the activation function, which is a nonlinear function.

[0052] The present invention also discloses an industrial Internet abnormal data governance system based on time-space coupling characteristics for implementing the above-mentioned governance method.

[0053] Furthermore, the industrial Internet abnormal data governance system based on spatiotemporal coupling characteristics includes:

[0054] Data collection and preprocessing module, used for data collection and data preprocessing to obtain industrial data sets;

[0055] A topology graph construction module, used to construct a topology graph G;

[0056] The data set conversion module is used to convert the industrial data set into the form of a topological graph G as the input data of the GCN model;

[0057] Model building module, used to build GCN and LSTM collaborative models;

[0058] The model training module is used to divide the data set into a training set and a test set, and train and test the GCN and LSTM collaborative model to obtain a trained GCN and LSTM collaborative model;

[0059] The data prediction module is used to take historical data as input data and use the trained GCN and LSTM collaborative model to obtain predicted data.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] The present invention discloses an abnormal data governance method for industrial Internet based on spatiotemporal coupling characteristics. By constructing a GCN and LSTM collaborative model, GCN is used to effectively characterize the spatial structure of data, spatial correlation information between data points is mined, and the dynamic evolution of data in the time dimension is modeled with the help of LSTM. A double-layer GCN model and a double-layer LSTM model are used to predict the data respectively, and finally the prediction results of the two are weightedly fitted, so as to achieve more comprehensive and accurate analysis and prediction of complex spatiotemporal data. The present invention realizes data prediction using the collaborative model of GCN and LSTM in industrial data prediction tasks. The prediction results include both the spatial correlation information between nodes and the dynamic evolution characteristics of data in the time dimension. The prediction results are more accurate than those of a single GCN model and LSTM model. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a flow chart of the abnormal data governance method of the industrial Internet based on the spatiotemporal coupling characteristics of the present invention;

[0063] Figure 2 This is a schematic diagram of the structure of the GCN and LSTM collaborative model of the present invention.

[0064] Figure 3 Schematic diagram of the double-layer LSTM model of the present invention. DETAILED DESCRIPTION

[0065] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0066] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0067] The present invention discloses an industrial Internet abnormal data governance method based on spatiotemporal coupling characteristics, comprising the following steps:

[0068] S1: data collection and data preprocessing to obtain industrial data sets;

[0069] S2: construct a topological graph G based on the industrial data set;

[0070] S3: Convert the industrial dataset into a topological graph G as input data for the GCN model;

[0071] S4: Construct a GCN and LSTM collaborative model to analyze the spatiotemporal coupling characteristics of data;

[0072] S5: Divide the industrial data set into a training set and a test set, train and test the constructed GCN and LSTM collaborative model, and obtain a trained GCN and LSTM collaborative model;

[0073] S6: Obtain some historical data before the time period that needs to be predicted, input it into the trained GCN and LSTM collaborative model, and output the prediction result.

[0074] Preferably, the data collection and data preprocessing methods are as follows:

[0075] Based on MySQL and other database management systems, collect industrial data and convert the data into csv format;

[0076] Preprocess the data through operations such as data cleaning;

[0077] The industrial data includes: data point name, data collection time, and industrial detection value.

[0078] Preferably, the method for constructing the industrial data topology graph G is as follows:

[0079] Define an undirected weighted graph G = (V, A, X), where: V = {v1, v2, ... vn} represents the node set of the graph; A ∈ Rn*n represents the adjacency matrix, a ij is the data in the i-th row and j-th column of matrix A, whose value is between 0 and 1, indicating that node v i and node v j The level of correlation between them; X∈Rn*p represents the feature matrix, p is the length of the historical time series, x ij Represents the industrial data of the i-th node at the j-th moment.

[0080] Preferably, the data set conversion method is as follows:

[0081] The collected industrial data are classified according to the point names, and the historical time series of each data point is used as the feature matrix X. The adjacency matrix A is obtained using the multi-head attention mechanism.

[0082] Preferably, the method of obtaining the adjacency matrix using the multi-head attention mechanism includes:

[0083] Self-attention calculation formula:

[0084]

[0085] Among them, Q, K, and V represent query, key, and value matrices respectively, and d k is the dimension of the key vector, used to scale clicks to stabilize the softmax function;

[0086] Parallel attention calculation formula:

[0087]

[0088] in, Represent the transformation matrices of query, key, and value of the i-th head respectively;

[0089] For all Heads calculated i The matrix is ​​concatenated on the last dimension, and the concatenated multi-attention head fusion matrix is ​​averaged on the head dimension, and the final matrix is ​​used as the adjacency matrix.

[0090] Preferably, the GCN and LSTM combination model mentioned includes:

[0091] A two-layer GCN model, a two-layer LSTM model, and an MLP model that weightedly fits the output data of the GCN model and the output data of the LSTM model. The GCN model and the LSTM model are used to predict the data of each point respectively. Finally, the prediction results of the two models are weightedly fitted using the connection layer, and the final prediction result is output.

[0092] Preferably, the GCN model mentioned includes:

[0093]

[0094] in I is the identity matrix; yes degree matrix; H is the feature of each layer. For the input layer, H is X; σ is the nonlinear activation function;

[0095] The two-layer GCN model includes:

[0096] f(A,X)=A·(ReLU(A·(XW0))W1)(4);

[0097] Where f(A,X) represents the final output result; ReLU() represents the linear rectification function; W0 represents the parameter matrix of the first layer; W1 represents the weight matrix mapping the hidden layer to the output layer.

[0098] Preferably, the LSTM model mentioned includes:

[0099] Forget Gate:

[0100] f ι =σ(W f [h ι-l ,x ι ]+b f )(5);

[0101] Among them, W f Represents the weight coefficient matrix of the forget gate; h ι-l represents the memory information of the hidden layer at the moment before time t, and the sigmoid activation function is used to discard the information; x ι represents the input information at time t; b f Represents the bias condition of the forget gate;

[0102] Input Gate:

[0103] i t =σ(W t [h t-1 ,x t ]+b t )(6);

[0104]

[0105] Among them, the sigmoid activation function is first used to decide which state to update; W t , W c Represent the weight coefficient matrices of the input gate and memory unit respectively; then this state and vector Addition, vector It is created using the tanh function; finally combined with the current memory and long-term memory C t-1 Determine the new cell state C t ; b t and b c They represent the bias conditions of the input gate and memory unit respectively;

[0106] Output Gate:

[0107] o t =σ(W o [h t-l,x t ]+b o )(9);

[0108] h t =o t ×tanh(C t )(10);

[0109] Among them, h t Represents the output of the first layer LSTM model; W o Represents the weight coefficient matrix of the output gate; b o Represents the bias condition of the output gate.

[0110] Preferably, the MLP model mentioned for weighted fitting includes:

[0111]

[0112] Among them, y j represents the jth component of the output layer; represents the input of the i-th neuron in the k-1-th hidden layer; w i,j,k represents the connection weight between the i-th component on the k-1-th hidden layer and the j-th neuron on the k-th hidden layer; f(x) represents the activation function, which is a nonlinear function.

[0113] The present invention also discloses an industrial Internet abnormal data governance system based on spatiotemporal coupling characteristics, including:

[0114] Data collection and preprocessing module, used to collect industrial data and perform simple preprocessing on the data set;

[0115] Topology map construction module, used to construct multi-point industrial data topology map G;

[0116] The dataset conversion module is used to convert the dataset into a topological graph G as input data for the GCN model;

[0117] Model building module, used to build GCN and LSTM collaborative models;

[0118] The model training module is used to divide the data set into a training set and a test set, and train and test the GCN and LSTM collaborative model to obtain a trained GCN and LSTM collaborative model;

[0119] The data prediction module is used to take some industrial historical data as input data and use the trained GCN and LSTM collaborative model to obtain predicted data.

[0120] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0121] like Figure 1 As shown, an industrial Internet abnormal data governance method based on spatiotemporal coupling characteristics provided by an embodiment of the present invention can predict industrial data, which mainly includes the following steps:

[0122] S1: Data acquisition and data preprocessing.

[0123] The collected industrial data mainly include: data point name, data collection time, and industrial detection value.

[0124] In this embodiment, industrial data is collected based on database management systems such as MySQL and converted into csv format; the data is preprocessed through operations such as data cleaning, abnormal data in the data set is deleted, and missing values ​​are supplemented.

[0125] S2: Construct a topological graph G of multi-point industrial data.

[0126] Define an undirected weighted graph G = (V, A, X), where: V = {v1, v2, ... vn} represents the node set of the graph; A∈Rn*n represents the adjacency matrix, aij is the data in the i-th row and j-th column of the matrix A, and its value is between 0 and 1, indicating that the node v i and node v j The level of correlation between them; X∈Rn*p represents the feature matrix, p is the length of the historical time series, x ij Represents the industrial data of the i-th node at the j-th moment.

[0127] Specifically, to construct a multi-point industrial data topology map, first, a node set V in the topology map is constructed according to the industrial data points, and each node in the set corresponds to a data point.

[0128] In order to express the correlation between data points, the adjacency matrix A is used to represent it, A∈Rn*n, n is the number of nodes in the topology graph, and the element aij in the matrix represents the correlation between node vi and node vj, and its size is between 0 and 1.

[0129] The characteristic attributes of each node on the topological graph G are expressed using the characteristic matrix X, where X∈Rn*p, n is the number of nodes in the graph, p is the length of the historical time series, and x ij Represents the industrial data of the i-th node at the j-th moment.

[0130] S3: Convert the dataset into a topological graph G as input data for the GCN model.

[0131] The collected industrial data are classified according to the point names, and the historical time series of each data point is used as the feature matrix X. The adjacency matrix A is obtained using the multi-head attention mechanism.

[0132] Specifically, a time series of length p is intercepted for each data point to construct a feature matrix X, X∈Rn*p, where n is the number of nodes in the graph and p is the length of the historical time series.

[0133] Self-attention calculation formula:

[0134]

[0135] Among them, Q, K, and V represent query, key, and value matrices respectively, and d k is the dimension of the key vector, used to scale clicks to stabilize the softmax function;

[0136] The parallel attention calculation formula based on this is:

[0137]

[0138] in, Represent the transformation matrices of query, key, and value of the i-th head respectively;

[0139] On the basis of obtaining the feature matrix X, multi-head attention is used to calculate the adjacency matrix, and the corresponding parallel attention is calculated for each point. i The matrices are concatenated in the last dimension to obtain a multi-attention head fusion matrix. The multi-attention head fusion matrix is ​​averaged in the head dimension, and the resulting matrix is ​​the adjacency matrix of this embodiment, which represents the correlation between the data points in the topology diagram of this embodiment.

[0140] S4: Construct a GCN and LSTM collaborative model to analyze the spatiotemporal coupling characteristics of data.

[0141] Figure 2 It is a structural schematic diagram of the GCN and LSTM collaborative model of the present invention, which includes a double-layer GCN model, a double-layer LSTM model, and an MLP model that weightedly fits the output data of the GCN model and the output data of the LSTM model; the collaborative model uses the GCN model and the LSTM model to predict the data of each point respectively, and finally uses the MLP model to weightedly fit the prediction results of the two models to output the final prediction result.

[0142] Specifically, the GCN model includes:

[0143]

[0144] in I is the identity matrix; yes degree matrix; H is the feature of each layer. For the input layer, H is X; σ is the nonlinear activation function;

[0145] The two-layer GCN model includes:

[0146] f(A,X)=A·(ReLU(A·(XW0))W1)(4);

[0147] Where f(A,X)∈Rn*T represents the output result of the prediction length T; ReLU(·) represents the linear rectification function; W0 represents the parameter matrix of the first layer; W1 represents the weight matrix mapping the hidden layer to the output layer.

[0148] Figure 3 The figure shows a schematic diagram of the structure of a two-layer LSTM model in this experiment. The LSTM model includes:

[0149] Forget Gate:

[0150] f ι =σ(W f [h ι-l ,x ι ]+b f )(5);

[0151] Among them, W f Represents the weight coefficient matrix of the forget gate; h ι-l represents the memory information of the hidden layer at the moment before time t, and the sigmoid activation function is used to discard the information; x ι represents the input information at time t; b f Represents the bias condition of the forget gate;

[0152] Input Gate:

[0153] i t =σ(W t [h t-1 ,x t ]+b t )(6);

[0154]

[0155] Among them, the sigmoid activation function is first used to decide which state to update; W t , W c Represent the weight coefficient matrices of the input gate and memory unit respectively; then this state and vector Addition, vector It is created using the tanh function; finally combined with the current memory and long-term memory C t-1 Determine the new cell state C t ; bt and b c They represent the bias conditions of the input gate and memory unit respectively;

[0156] Output Gate:

[0157] o t =σ(W o [h t-l ,x t ]+b o )(9);

[0158] h t =o t ×tanh(C t )(10);

[0159] Among them, h t Represents the output of the first layer LSTM model; W o Represents the weight coefficient matrix of the output gate; b o Represents the bias condition of the output gate.

[0160] The second layer of LSTM repeatedly converts the output h of the first layer t Treat as x ι As the input of the second layer, repeat equations (5) to (10) to output the final prediction result H t Similar to the output of the GCN model, the final output of the LSTM model is H t ∈Rn*T, which represents the output result of the predicted length T.

[0161] Finally, a new connection layer is used to perform weighted fitting on the output results of the GCN model and the output results of the LSTM model. In this embodiment, the MLP model is used to perform weighted fitting on the output results of the GCN and LSTM models, and its working principle is shown in formula (11):

[0162]

[0163] Among them, y j represents the jth component of the output layer; represents the input of the i-th neuron in the k-1-th hidden layer; w i,j,k represents the connection weight between the i-th component on the k-1-th hidden layer and the j-th neuron on the k-th hidden layer; f(x) represents the activation function, which is a nonlinear function.

[0164] The predicted data of the GCN model and the LSTM model are used as the input data of the MLP model. The data is fitted through a fully connected layer, and the output result is the final predicted data.

[0165] S5: Divide the industrial data set into a training set and a test set, use the GCN and LSTM collaborative model constructed by the training set for training, and then use the test set to test the trained model; analyze the prediction results and the true values ​​to detect whether the accuracy of the prediction data reaches the expected goal. If the prediction target is reached, the prediction result is output. If the prediction target is not reached, the structural parameter indicators of the model are reset until the accuracy of the prediction result reaches the expected prediction target.

[0166] KL divergence and mean square error are used as loss functions during model training.

[0167] KL divergence, also known as relative entropy, can be used to measure the similarity or difference between two probability distributions. Formula (12) is the definition formula of KL divergence:

[0168]

[0169] Among them, P(x) represents the real data of the training set; Q(x) represents the model prediction result; x represents all possible values, which here represents the index of the training set data. The smaller the KL divergence, the closer the model prediction result is to the actual situation, the better the performance of the model, and the more accurate it can predict future data.

[0170] The mean square error (MSE) is also a commonly used indicator to measure the difference between the model prediction value and the actual observation value, which is used to evaluate the degree of fit of the model on the given data. MSE is obtained by calculating the average of the square of the difference between the prediction value and the actual observation value. Its definition formula is shown in formula (13):

[0171]

[0172] Where n represents the number of samples in the training set; represents the predicted value of the i-th sample; y i Represents the true value of the i-th sample; the smaller the MSE, the closer the model prediction result is to the actual situation, the better the performance of the model, and the more accurate it can predict future data.

[0173] S6: In a specific application scenario, obtain historical data before the time period that needs to be predicted, use this data as input data, input it into the trained GCN and LSTM collaborative model, and output the prediction results. When performing outlier data governance, the data in the original data that differs greatly from the predicted data is determined as abnormal data, deleted and replaced with the predicted data, thereby achieving the purpose of data governance.

[0174] The above contents are only for explaining the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A method for managing abnormal data in the industrial Internet based on spatiotemporal coupling characteristics, characterized in that: The following steps are involved: Collect and preprocess data on the industrial Internet to obtain industrial data sets; Construct a topological graph G based on the industrial data set; Construct a GCN and LSTM collaborative model, convert the industrial data set into the form of a topological graph G, and divide it into a training set and a test set. Train and test the constructed GCN and LSTM collaborative model to obtain a trained GCN and LSTM collaborative model. Obtain historical data before the time period required for prediction, input it into the trained GCN and LSTM collaborative model, and output the prediction results.

2. According to the method of industrial Internet abnormal data management based on spatiotemporal coupling characteristics according to claim 1, it is characterized in that: The specific steps of data collection and data preprocessing to obtain industrial data sets are as follows: Based on the database management system, industrial data is collected and converted into csv format; then data cleaning is used to pre-process the industrial data to obtain an industrial data set; The industrial data set includes the data point name, data collection time and industrial detection value.

3. According to the method of industrial Internet abnormal data management based on spatiotemporal coupling characteristics according to claim 2, it is characterized in that: The specific steps of constructing the topology graph G based on the industrial data set are as follows: Define a non-directed weighted graph G = (V, A, X); Where: V = {v1, v2, ... v n } represents the node set of the graph; A∈Rn*n represents the adjacency matrix, a ij is the data in the i-th row and j-th column of matrix A, whose value is between 0 and 1, indicating that node v i and node v j The level of correlation between X∈Rn*p represents the feature matrix, p is the length of the historical time series, x ij Represents the industrial data of the i-th node at the j-th moment.

4. According to the method of industrial Internet abnormal data management based on spatiotemporal coupling characteristics according to claim 3, it is characterized in that: The specific steps to convert the industrial data set into the form of topological graph G are as follows: The collected industrial data sets are classified according to the point names, and the historical time series of each data point is used as the feature matrix X. The adjacency matrix A is obtained using the multi-head attention mechanism.

5. According to the method of industrial Internet abnormal data management based on spatiotemporal coupling characteristics according to claim 4, it is characterized in that: The self-attention calculation formula and parallel attention calculation formula in the multi-head attention mechanism are as follows: Self-attention calculation formula: Among them, Q, K, and V represent query, key, and value matrices respectively, and d k is the dimension of the key vector, used to scale clicks to stabilize the softmax function; Parallel attention calculation formula: in, Represent the transformation matrices of query, key, and value of the i-th head respectively; The specific steps of using the multi-head attention mechanism to obtain the adjacency matrix A are as follows: Head i The matrix is ​​concatenated on the last dimension, and the concatenated multi-attention head fusion matrix is ​​averaged on the head dimension, and the final matrix is ​​used as the adjacency matrix A.

6. According to the method of industrial Internet abnormal data management based on spatiotemporal coupling characteristics according to claim 1, it is characterized in that: The GCN and LSTM collaborative model includes a double-layer GCN model, a double-layer LSTM model, and an MLP model that weights and fits the output data of the double-layer GCN model and the output data of the double-layer LSTM model; The double-layer GCN model and the double-layer LSTM model are used to predict the data of each point in the industrial Internet respectively. Finally, the connection layer is used to weighted fit the prediction results of the two models, and the final prediction results are output.

7. The method for managing abnormal data in the industrial Internet based on spatiotemporal coupling characteristics according to claim 6 is characterized in that: The GCN model is: in, I is the identity matrix; yes degree matrix; H is the feature of each layer. For the input layer, H is X; σ is the nonlinear activation function; The two-layer GCN model is: f(A,X)=A·(ReLU(A·(XW0))W1); Among them, f(A,X) represents the final output result; ReLU() represents the linear rectification function; W0 represents the parameter matrix of the first layer; W1 represents the weight matrix mapping the hidden layer to the output layer; The LSTM model includes: Forget Gate: f ι =σ(W f [h ι-l ,x ι ]+b f ); Among them, W f Represents the weight coefficient matrix of the forget gate; h ι-l represents the memory information of the hidden layer at the moment before time t, and the sigmoid activation function is used to discard the information; x ι represents the input information at time t; b f Represents the bias condition of the forget gate; Input Gate: i t =σ(W t [h t-1 ,x t ]+b t ); Among them, the sigmoid activation function is first used to decide which state to update; W t , W c Represent the weight coefficient matrices of the input gate and memory unit respectively; then this state and vector Addition, vector It is created using the tanh function; finally combined with the current memory and long-term memory C t-1 Determine the new cell state C t ; b t and b c They represent the bias conditions of the input gate and memory unit respectively; Output Gate: the t =σ(W o [h t-l ,x t ]+b o ); h t =o t ×tanh(C t ); Among them, h t Represents the output of the first layer LSTM model; W o Represents the weight coefficient matrix of the output gate; b o Represents the bias condition of the output gate.

8. The method for managing abnormal data in the industrial Internet based on spatiotemporal coupling characteristics according to claim 6 is characterized in that: The weighted fitting MLP model is: Among them, y j represents the jth component of the output layer; represents the input of the i-th neuron in the k-1-th hidden layer; w i,j,k represents the connection weight between the i-th component on the k-1-th hidden layer and the j-th neuron on the k-th hidden layer; f(x) represents the activation function, which is a nonlinear function.

9. An industrial Internet abnormal data management system based on spatiotemporal coupling characteristics, characterized in that: The industrial Internet abnormal data governance system based on time-space coupling characteristics is used to implement the governance method described in any one of claims 1 to 8.

10. The industrial Internet abnormal data management system based on spatiotemporal coupling characteristics according to claim 9 is characterized in that: include: Data collection and preprocessing module, used for data collection and data preprocessing to obtain industrial data sets; A topology graph construction module, used to construct a topology graph G; The data set conversion module is used to convert the industrial data set into the form of a topological graph G as the input data of the GCN model; Model building module, used to build GCN and LSTM collaborative models; The model training module is used to divide the data set into a training set and a test set, and train and test the GCN and LSTM collaborative model to obtain a trained GCN and LSTM collaborative model; The data prediction module is used to take historical data as input data and use the trained GCN and LSTM collaborative model to obtain predicted data.

Images