Multi-dimensional time series data anomaly discrimination and restoration method based on convolutional neural network

Abstract

The invention discloses a convolutional neural network-based multi-dimensional time series data anomaly discrimination and restoration method, which comprises the steps of reading historical multi-dimensional time series data, calculating the spatial similarity of the time series data, and forming a correlation matrix; inputting the correlation matrix into a convolutional neural network to obtain a multi-dimensional feature map, and reconstructing the correlation matrix; training a neural network model by minimizing a reconstruction error of the correlation matrix, and storing model parameters as a file; model parameters are loaded in a real-time system, online data are compared with model reconstruction data, whether abnormity exists or not is judged, and if yes, repairing is carried out. According to the method, the problem that multi-channel sensor data of the same equipment is high in linkage but the incidence relation cannot be effectively utilized is effectively avoided, hidden and special anomalies can be effectively detected, the anomalies are repaired through neural network reconstruction data, the accuracy and timeliness of data anomaly judgment and repair are improved, and the method is suitable for popularization and application. And online application in engineering practice is promoted.

Description

Technical Field

[0001] The present invention belongs to the technical field of real-time monitoring of flue gas emissions, and in particular relates to a method for distinguishing and repairing abnormalities in multidimensional time series data based on a convolutional neural network. Background Art

[0002] Whether the flue gas emission data is accurate and reliable is related to whether the enterprise can grasp its own emission footprint and rules, and whether the regulatory authorities can form emission reports and policies in a timely manner. It plays an important role in guiding the improvement of the ecological environment and the optimization and development of the industry. Large thermal power station units are arranged with 5,000 to 10,000 on-site measurement points to monitor the state quantities of steam, flue gas, drainage, steam extraction, exhaust, water supply, and over-air in each equipment channel in real time, such as flow rate, temperature, pressure, flow, concentration, density, pressure difference, etc., and also monitor the overall coal consumption, load, plant transformer power, speed, thermal efficiency, etc. of the unit in real time, and calculate and analyze the online emissions of various types of flue gas through the above large amount of measurement point data.

[0003] The above-mentioned measurement point data involves multiple disciplines such as power plant boilers, steam turbines, environmental protection, chemistry, thermal engineering, and electrical engineering, which makes it very difficult to timely discover and correct hidden abnormal data. In the existing technology, there are four types of abnormality detection methods for time series data: one is statistical methods; the second is machine learning methods; the third is deep learning methods; and the fourth is manual methods.

[0004] The statistical learning method first calculates the mean and standard deviation of the time series data, and then determines whether the new data has anomalies according to the three sigma criterion. This method is simple to calculate, but the accuracy of the judgment is low; the machine learning method uses the differences in the numerical distribution or distance density of the data at each measuring point to screen out outliers, but this method is very sensitive to the choice of kernel density function, and the judgment result is unstable; deep learning methods mostly use LSTM long short-term memory network or its variant model to consider the time trend characteristics of the data, or use GCN graph convolution network to consider the overall dependency of multi-dimensional time series data, but LSTM network has a certain time lag effect, and the number of nodes and calculation of GCN network are too large, which may affect the generalization and application ability of the model; the manual method is to arrange experts to make joint judgments on the time series data of multiple measuring points, use the equipment operation rules and engineering experience to screen out hidden and special anomalies, and repair them through manual reasoning. However, this method is only suitable for static data analysis research, and it is difficult to meet the timeliness and intelligence requirements under online working conditions. In addition, the existing technologies mostly provide methods for anomaly detection, and rarely provide scientific and practical methods for anomaly repair. Summary of the invention

[0005] The technical problem solved by the present invention is: to provide a method for distinguishing and repairing anomalies in multi-dimensional time series data based on convolutional neural networks, which can effectively detect hidden and specific anomalies, and repair the anomalies by reconstructing data through neural networks, which helps to improve the accuracy and timeliness of data anomaly distinction and repair, and promote online applications in engineering practice.

[0006] Technical solution: In order to solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0007] A method for identifying and repairing abnormalities in multidimensional time series data based on a convolutional neural network comprises the following steps:

[0008] Step S1, reading historical multi-dimensional time series data, calculating the spatial similarity of the time series data, and forming a correlation matrix;

[0009] Step S2, input the correlation matrix into the convolutional neural network to obtain a multi-dimensional feature map, and reconstruct the correlation matrix by combining the LSTM network, the Attention mechanism and the deconvolution operation;

[0010] Step S3, training the neural network model by minimizing the reconstruction error of the correlation matrix, and saving the model parameters as a file;

[0011] Step S4, load the model parameters in the real-time system, compare the online data with the model reconstruction data, determine whether there is an abnormality, and if so, repair it.

[0012] Furthermore, step S1 is implemented as follows:

[0013] S11: reading data files collected by multi-channel sensors at different times within a period of time, and organizing them into original multi-dimensional time series data;

[0014] S12: In the s time segment, the dot product similarity between the multidimensional sequences is calculated to form a correlation matrix of the time series data.

[0015] Furthermore, in step S11, a historical data table is constructed in the relational database, and for each data file, each line of measurement point information is read cyclically and saved in the historical data table until the last line of the file;

[0016] The historical data is sorted, and the sorting keywords include the measuring point name and time. Then, SQL query statements are used to return the sorted measuring point data record set, and the returned record set is grouped by the measuring point name. The data in each group corresponds to a measuring point data sequence, and the measuring point values ​​in the group are normalized to form the original multi-dimensional time series data.

[0017] Furthermore, in step S12, the specific implementation method is: for any time t on any measurement point data sequence, trace back s consecutive time points, and calculate the dot product similarity between the multidimensional sequences in s time segments. The formula is as follows:

[0018] ;

[0019] in, , Represent any two measuring points subsequence, s represents the number of backtracking moments, N represents the total number of measurement points, Represents any two measuring points at time t The dot product similarity between A matrix with elements is the original multidimensional sequence correlation matrix, Respectively represent the data values ​​of measuring points i and j at time t-ε.

[0020] Furthermore, the implementation method of step S2 is:

[0021] S21: Input the multi-dimensional sequence correlation matrix into the multi-layer convolutional neural network, calculate and obtain the multi-dimensional feature map of the corresponding layer network, and realize spatial feature extraction;

[0022] S22: Input the multi-dimensional feature map into the LSTM network, calculate and obtain a feature map sequence that integrates long-term and short-term information, and based on the Attention mechanism, perform weighted summation of the feature map sequence with m consecutive moments as the time window to achieve temporal feature extraction;

[0023] S23: Use the Attention mechanism to adaptively select information related to the current moment, determine the time window length m, and form a refined output of the current hidden state by aggregating the information features within the time window ;

[0024] S24: Input the feature maps in steps S21 and S22 into the convolutional neural network, and after multiple layers of deconvolution and splicing operations, obtain a correlation matrix reconstructed by the neural network model.

[0025] Furthermore, the specific implementation process of step S21 is as follows:

[0026] The original multidimensional sequence correlation matrix Feed it to the first convolutional layer for convolution calculation, the calculation output of this layer is a multi-dimensional feature map ,in is the feature map dimension of the first convolutional layer, is the feature map depth of the first convolutional layer; assuming express The feature map of the layer, For the The feature map dimension of the layer, For the The feature map depth of the layer, then The feature map output of a layer is calculated as follows:

[0027] ;

[0028] in, represents the convolution operation, g(.) represents the activation function, express The size is The convolution kernel, is the deviation term, where For the The convolution kernel dimension of the layer, For the The feature map depth of the layer, Indicates that the dimension of the lth layer is , the depth is The output feature map of .

[0029] Furthermore, the specific implementation process of step S23 is as follows:

[0030] ;

[0031] Among them, softmax(.) represents the normalized exponential function, Mul(.) represents the product function of the corresponding elements of the tensor, and Sum(.) represents the sum function of the tensor elements; Represents any moment between tm and t, where m is the number of moments in the Attention mechanism time window. represents the hidden state of the lth convolutional layer at time i (i∈(tm, t)); It represents the importance weight of the previous moment information, and utilizes the time information between each convolutional layer in the time window to realize the time feature extraction on the spatial feature map.

[0032] Furthermore, the specific implementation process of step S24 is as follows:

[0033] Design a convolutional decoder with the following formula:

[0034] ;

[0035] in, represents the deconvolution operation, represents the concatenation operation, g(.) represents the activation function, and They represent the convolution kernel and bias term of the lth deconvolution layer respectively, represents the deconvolution output feature map of the lth layer at time t; from back to front, the lth layer ConvLSTM Feed to the deconvolutional neural network and output feature map It is concatenated in series with the output of the l-1th ConvLSTM layer, and the concatenated matrix is ​​further input into the next deconvolution layer; the final output , represents the reconstructed correlation matrix.

[0036] Furthermore, the specific implementation process of step S3 is as follows:

[0037] The loss function is defined as:

[0038] ;

[0039] in, represents the F norm of the matrix, N represents the total number of measurement points, and the Adam optimizer is used to minimize the above loss; represents the reconstruction error matrix at time t, represents the original multidimensional serial correlation matrix; represents the reconstructed correlation matrix;

[0040] After multiple training cycles, the offline training parameters of the model are obtained, and the trained parameters are saved to the file system for loading and use by the online system.

[0041] Furthermore, in step S4, the neural network model parameters are loaded when the real-time system is running, and the multidimensional data at the current moment is reconstructed through the model. The online data is compared with the reconstructed data. If the difference between the data in a certain dimension exceeds a threshold, an abnormality exists, and the multidimensional reconstructed data is used to repair the abnormality. At the same time, the repair time and content are recorded in the abnormality handling log, and relevant personnel are notified to investigate the source of the abnormality.

[0042] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0043] (1) The present invention can organize and form structured original multi-dimensional time series data from massive original data files through line-by-line reading, full-data sorting, category grouping and normalization processing, which is conducive to improving the data processing efficiency of complex neural networks.

[0044] (2) The present invention extracts subsequences of a set length from long time series data and measures the correlation of multidimensional data by the dot product similarity of the subsequence data. This makes it possible to capture the dependency of multidimensional data by constructing a general convolutional neural network instead of constructing a more complex graph convolutional neural network, which helps to improve the practicality of the algorithm.

[0045] (3) The present invention can effectively combine convolutional neural networks, long short-term memory networks and attention mechanisms, completely covering the temporal and spatial characteristics of multi-dimensional time series data, and can effectively detect hidden and specific anomalies.

[0046] (4) The present invention can repair anomalies through reconstructed data using a neural network based on deconvolution and splicing operations, and send the anomaly information to relevant personnel in a timely manner, which helps to improve the accuracy and timeliness of data anomaly identification and repair, and promotes online applications in engineering practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a structural diagram of a multi-dimensional time series data anomaly identification and repair method based on a convolutional neural network. DETAILED DESCRIPTION

[0048] The present invention is further illustrated below in conjunction with specific examples. The examples are implemented based on the technical solutions of the present invention. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0049] Example 1

[0050] like Figure 1 As shown, the multi-dimensional time series data anomaly identification and repair method based on convolutional neural network of the present application includes the following steps:

[0051] Step S1, reading historical multi-dimensional time series data, calculating the spatial similarity of the time series data, and forming a correlation matrix;

[0052] Step S2, input the correlation matrix into the convolutional neural network to obtain a multi-dimensional feature map, and reconstruct the correlation matrix by combining the LSTM network, the Attention mechanism and the deconvolution operation;

[0053] Step S3, training the neural network model by minimizing the reconstruction error of the correlation matrix, and saving the model parameters as a file;

[0054] Step S4, load the model parameters in the real-time system, compare the online data with the model reconstruction data, determine whether there is an abnormality, and if so, repair it.

[0055] Example 2

[0056] Step S1: read historical multi-dimensional time series data, calculate the spatial similarity of the time series data, and form a correlation matrix. The implementation method includes:

[0057] S11: Read the data files collected by the multi-channel sensor at different times within one year, and organize them into original multi-dimensional time series data. The specific implementation process is as follows:

[0058] A historical data table is constructed in a relational database, which contains three fields: measuring point name, measuring point value, and time. For each data file, each line of measuring point information is read in a loop and saved in the historical data table until the last line of the file.

[0059] The historical data is sorted using database SQL statements, with the sorting keyword 1 being the measuring point name and the keyword 2 being time. Then, the SQL query statement is used to return the sorted measuring point data record set, and the returned record set is grouped by the measuring point name. The data in each group corresponds to a measuring point data sequence, and the measuring point values ​​in the group are normalized to form the original multi-dimensional time series data.

[0060] S12: In the s time segment, calculate the dot product similarity between the multidimensional sequences to form a correlation matrix of the time series data. The specific implementation process is as follows:

[0061] For any time t on any measurement point data sequence, trace back s consecutive moments and calculate the dot product similarity between the multidimensional sequences within the s time segment. The formula is as follows:

[0062] ;

[0063] in, , Represent any two measuring points subsequence, s represents the number of backtracking moments, N represents the total number of measurement points, Represents any two measuring points at time t The dot product similarity between A matrix with elements (Right now ) is the original multidimensional sequence correlation matrix, Respectively represent the data values ​​of measuring points i and j at time t-ε.

[0064] Step S2: Input the multidimensional sequence correlation matrix into the convolutional neural network to obtain the multidimensional feature map, and reconstruct the multidimensional sequence correlation matrix by combining the LSTM network, Attention mechanism and deconvolution operation.

[0065] S21: Input the multi-dimensional sequence correlation matrix into the multi-layer convolutional neural network, calculate and obtain the multi-dimensional feature map of the corresponding layer network, and realize spatial feature extraction. The specific implementation process is as follows:

[0066] The original multidimensional sequence correlation matrix Feed it to the first convolutional layer for convolution calculation, the calculation output of this layer represents a multi-dimensional feature map ,in represents the feature map dimension of the first convolutional layer, represents the feature map depth of the first convolutional layer. Assume express The feature map of the layer, Indicates The feature map dimension of the layer, Indicates The feature map depth of the layer, then The feature map output of a layer is calculated as follows:

[0067] ;

[0068] in, represents the convolution operation, g(.) represents the activation function, express The size is The convolution kernel, represents the deviation term, where Indicates The convolution kernel dimension of the layer, Indicates The feature map depth of the layer, Indicates that the dimension of the lth layer is , the depth is The output feature map of .

[0069] The convolutional neural network is denoted as Conv. According to engineering experience, the scaled exponential linear unit is used as the activation function in Conv, and the number of convolutional layers is set to 4, denoted as Conv1~Conv4, with 32 cores of size 3×3×1, 64 cores of size 3×3×32, 128 cores of size 2×2×64, and 256 cores of size 2×2×128, respectively. In addition, the step size of the convolution operation is set to 1×1, 2×2, 2×2, and 2×2, respectively.

[0070] S22: Input the multi-dimensional feature map into the LSTM network, calculate and obtain a feature map sequence that integrates long-term and short-term information, and based on the Attention mechanism, perform weighted summation of the feature map sequence with m consecutive moments as the time window to achieve temporal feature extraction. The specific implementation process is as follows:

[0071] The LSTM network based on the convolutional feature map input is recorded as ConvLSTM. According to the feature map of the lth convolutional layer and the hidden state at the previous moment , update the current hidden state , where the formula of the ConvLSTM unit is:

[0072] ;

[0073] in, represents the convolution operation, represents the Hadamard product, is the sigmoid activation function, yes The size is The convolution kernel, represents the bias term of the l-th layer ConvLSTM, They represent the input gate, forget gate, memory unit, output gate and hidden state of the lth layer network at time t. When performing convolution operations, the convolution kernel size and convolution step size of each layer are consistent with the parameters of Conv1~Conv4 mentioned above.

[0074] S23: Considering that the previous moment information is not consistent with the current hidden state Equally relevant, the Attention mechanism is used to adaptively select information related to the current moment. According to engineering practice, the time window length m=5 is taken, and the information features within the time window are aggregated to form a refined output of the current hidden state. , the specific formula is as follows:

[0075] ;

[0076] Among them, softmax(.) represents the normalized exponential function, Mul(.) represents the product function of the corresponding elements of the tensor, and Sum(.) represents the summation function of the tensor elements. Represents the importance weight of the previous moment information, Indicates the current state; Represents any moment between tm and t, where m is the number of moments in the Attention mechanism time window. Indicates time The hidden state of the lth convolutional layer.

[0077] Importance weight based on previous moment information , the time information between each convolutional layer in the time window is utilized to realize the temporal feature extraction on the spatial feature map.

[0078] S24: Input the feature graphs in steps S21 and S22 into the convolutional neural network, and after multi-layer deconvolution and concatenation operations, obtain the correlation matrix reconstructed by the neural network model. The specific implementation process is as follows:

[0079] Design a convolutional decoder with the following formula:

[0080] ;

[0081] in, represents the deconvolution operation, represents the concatenation operation, g(.) represents the activation function, and They represent the convolution kernel and bias term of the lth deconvolution layer respectively, represents the deconvolution output feature map of the lth layer at time t; from back to front, the lth layer ConvLSTM Feed to deconvolutional neural network; output feature map It is concatenated with the output of the l-1th ConvLSTM layer in series, and the concatenated matrix is ​​further input into the next deconvolution layer; the final output , represents the reconstructed correlation matrix.

[0082] The deconvolution neural network is denoted as ReConv, and there are four deconvolution layers ReConv4~ReConv1, which have 128 cores of size 2×2×256, 64 cores of size 2×2×128, 32 cores of size 3×3×64, and 1 core of size 3×3×64. The step sizes of the deconvolution operation are 2×2, 2×2, 2×2, and 1×1, respectively. In this way, the feature map matrix is ​​merged in different deconvolution layers and ConvLSTM layers to improve the performance of anomaly detection.

[0083] Step S3: Train the neural network model by minimizing the reconstruction error of the correlation matrix, and save the model parameters as a file.

[0084] Calculate the reconstructed residual of the correlation matrix at all times, construct the loss function with the F norm of the residual, optimize the model parameters through the Adam optimizer, and save the trained parameters to the file system. The specific implementation process is:

[0085] The loss function is defined as:

[0086] ;

[0087] in, represents the F norm of the matrix, N represents the total number of measurement points, represents the reconstruction error matrix at time t, and the Adam optimizer is used to minimize the above loss; represents the original multidimensional serial correlation matrix; represents the reconstructed correlation matrix.

[0088] After a sufficient number of training times, the offline training parameters of the model are obtained, and the trained parameters are saved to the file system for loading and use by the online system.

[0089] Based on the trained model, calculate the final error matrix at time t ,in is the final optimized reconstructed correlation matrix, judging whether each column exceeds a given threshold The number of elements of , and form the number vector of the number of elements exceeding the threshold at the current moment ,in The value is related to the specific data set. Take 0.005. Based on the number vector of values ​​exceeding the threshold at all times, obtain the abnormality discrimination parameter:

[0090] ,

[0091] Among them, θ represents the abnormal discrimination parameter, represents the magnification factor. In this embodiment, the data set is set to 1.6. Med{.} is used to obtain the median. Represents any j-column value of the vector of the number of super-threshold values ​​at any time t.

[0092] Step S4: Load the model parameters in the real-time system, compare the online data with the model reconstruction data, determine whether there is an abnormality, and repair it if so.

[0093] When the real-time system is running, the neural network model parameters are loaded, and the multi-dimensional data at the current moment is reconstructed through the model. The online data is compared with the reconstructed data. If the difference in a certain dimension exceeds the threshold, there is an abnormality, and the multi-dimensional reconstructed data is used to repair the abnormality. At the same time, the repair time and content are recorded in the abnormality handling log, and sent to the designated mailbox to notify relevant personnel to investigate the source of the abnormality. The specific implementation process is:

[0094] Calculate the reconstruction error matrix of multidimensional time series data at the current moment

[0095] ;

[0096] in, Represents the reconstruction error matrix of the multidimensional time series data at the current moment, represents the original multidimensional serial correlation matrix; Represents the reconstructed correlation matrix at the latest moment; the threshold is , by judging whether each column of the reconstruction error matrix exceeds The number of elements of , calculate the number vector of super-threshold , for any j-column element in the vector, if , then it is judged that the original data corresponding to this column is abnormal and needs to be corrected.

[0097] For the j columns of data with anomalies, use the reconstructed correlation matrix elements Replace the original correlation matrix elements , and correct the measurement point data corresponding to the jth column according to the following formula:

[0098] ;

[0099] Among them, s represents the number of backtracking moments, N represents the total number of measurement points, Indicates going back to the current moment Subsequence data at each moment, Indicates the correction value.

[0100] The current measurement point number j, current time t, and original data , correct the data The data is grouped, recorded in the exception handling log, and sent to the designated mailbox to notify relevant personnel to investigate the source of the exception.

[0101] The following data from a real power plant are verified using the method of the present invention, and the main data include: superheater outlet steam temperature (°C), economizer front feed water temperature (°C), No. 1 high pressure heater outlet water temperature (°C), No. 2 high pressure heater drain temperature (°C), reheater outlet steam temperature (°C), superheater cooling water flow rate (t / h), economizer front feed water flow rate (t / h), reheater emergency water spray flow rate (t / h), soot blowing steam flow rate (t / h), unit coal combustion amount (t / h), superheater outlet steam pressure (MPa), superheater cooling water pressure (MPa), economizer front feed water pressure (MPa), reheater outlet steam pressure (MPa), reheater inlet steam pressure (MPa), incoming coal received basis carbon Car (%), incoming coal received basis hydrogen Har (%), incoming coal received basis nitrogen Nar (%), incoming coal received basis sulfur Sar (%), incoming coal received basis oxygen Oar (%).

[0102] Using the above data, the dataset is divided into a simulation dataset and a real power plant dataset. The detailed information of the two datasets is shown in Table 1.

[0103] Table 1 Verification data

[0104]

[0105] The method of the present invention is compared with four types of methods, namely, classification models, density estimation models, time prediction models and graph network models.

[0106] The classification model classifies the test data as abnormal or non-abnormal by learning a decision function, and is compared with the single-class classification model of support vector machine (SVDD).

[0107] The density estimation model detects outliers by data density and is compared with the K nearest neighbor density estimation model (KNN).

[0108] The temporal prediction model models the temporal dependency of the training data and is compared using a vector autoregression model (VAR) and a long short-term memory network (LSTM).

[0109] The graph network model models sequence data and data relationships through nodes and adjacent edges, and is compared using a graph convolutional neural network (GCN).

[0110] This example uses precision, recall, and F1 score to evaluate the effectiveness and advantages of the proposed explanation method, where:

[0111] Precision refers to the proportion of samples that are actually positive among those predicted by the model.

[0112] Recall refers to the proportion of samples that are actually positive that are correctly predicted as positive by the model.

[0113] The F1 score represents the harmonic mean of precision and recall, and is used to comprehensively evaluate the performance of the model.

[0114] Table 2 Experimental results

[0115]

[0116] From Table 2 we can conclude that:

[0117] (1) The temporal prediction model and graph network model perform better than the SVDD and KNN models, indicating that the two datasets contain spatiotemporal correlations.

[0118] (2) The performance of LSTM and GCN is better than VAR, indicating that deep learning models can capture more complex data relationships than traditional methods.

[0119] (3) Compared with other methods, this method performs best in all three indicators, with improvements ranging from 4.9% to 10.1%, indicating that the method of the present invention can effectively model the correlation and time dependence between sensor measurement points of multidimensional time series.

[0120] The present invention establishes a complete method for distinguishing and repairing anomalies in multi-dimensional time series data, which effectively avoids the problem that the data of multi-channel sensors on the same device are highly linked but the correlation relationship cannot be effectively utilized. A feature map extraction method based on convolutional neural networks and LSTM networks is proposed, which completely covers the time and space characteristics of multi-dimensional time series data, can effectively detect hidden and specific anomalies, and repair anomalies by reconstructing data through neural networks, which helps to improve the accuracy and timeliness of data anomaly distinction and repair, and promote online applications in engineering practice.

[0121] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for identifying and repairing abnormalities in multidimensional time series data based on convolutional neural networks, characterized in that: The following steps are involved: Step S1, read the historical multi-dimensional time series data, calculate the spatial similarity of the time series data, and form a correlation matrix; the implementation method of step S1 is: S11: Read the data files collected by the multi-channel sensor at different times within a period of time, and organize them into original multi-dimensional time series data; build a historical data table in the relational database, and for each data file, cyclically read each line of measurement point information and save it in the historical data table until the last line of the file; Sorting the historical data, with the sorting keywords including the measuring point name and time, and then using SQL query statements to return the sorted measuring point data record set, and grouping the returned record set by the measuring point name. The data in each group corresponds to a measuring point data sequence, and the measuring point values ​​in the group are normalized to form the original multi-dimensional time series data; S12: In the time segment s, the dot product similarity between the multidimensional sequences is calculated to form a correlation matrix of the time series data; Step S2, input the correlation matrix into the convolutional neural network to obtain a multi-dimensional feature map, and reconstruct the correlation matrix by combining the LSTM network, the Attention mechanism and the deconvolution operation; Step S3, training the neural network model by minimizing the reconstruction error of the correlation matrix, and saving the model parameters as a file; Step S4, load the model parameters in the real-time system, compare the online data with the model reconstruction data, determine whether there is an abnormality, and if so, repair it.

2. According to the multi-dimensional time series data anomaly identification and repair method based on convolutional neural network according to claim 1, it is characterized in that: In step S12, the specific implementation method is: for any time t on any measurement point data sequence, trace back s consecutive time points, and calculate the dot product similarity between the multidimensional sequences in s time segments. The formula is as follows: ； in, , Represent any two measuring points subsequence, s represents the number of backtracking moments, N represents the total number of measurement points, Represents any two measuring points at time t The dot product similarity between A matrix with elements is the original multidimensional sequence correlation matrix, Respectively represent the data values ​​of measuring points i and j at time t-ε.

3. According to the multi-dimensional time series data anomaly identification and repair method based on convolutional neural network according to claim 1, it is characterized in that: The implementation method of step S2 is: S21: Input the multi-dimensional sequence correlation matrix into the multi-layer convolutional neural network, calculate and obtain the multi-dimensional feature map of the corresponding layer network, and realize spatial feature extraction; S22: Input the multi-dimensional feature map into the LSTM network, calculate and obtain a feature map sequence that integrates long-term and short-term information, and based on the Attention mechanism, perform weighted summation of the feature map sequence with m consecutive moments as the time window to achieve temporal feature extraction; S23: Use the Attention mechanism to adaptively select information related to the current moment, determine the time window length m, and form a refined output of the current hidden state by aggregating the information features within the time window ; S24: Input the feature maps in steps S21 and S22 into the convolutional neural network, and after multiple layers of deconvolution and splicing operations, obtain a correlation matrix reconstructed by the neural network model.

4. The method for abnormal identification and repair of multi-dimensional time series data based on convolutional neural network according to claim 3 is characterized in that: The specific implementation process of step S21 is as follows: The original multidimensional sequence correlation matrix Feed it to the first convolutional layer for convolution calculation, the calculation output of this layer is a multi-dimensional feature map ,in is the feature map dimension of the first convolutional layer, is the feature map depth of the first convolutional layer; assuming express The feature map of the layer, For the The feature map dimension of the layer, For the The feature map depth of the layer, then The feature map output of a layer is calculated as follows: ； in, represents the convolution operation, g(.) represents the activation function, express The size is The convolution kernel, is the deviation term, where For the The convolution kernel dimension of the layer, For the The feature map depth of the layer, Indicates that the dimension of the lth layer is , the depth is The output feature map of .

5. The method for abnormal identification and repair of multi-dimensional time series data based on convolutional neural network according to claim 3 is characterized in that: The specific implementation process of step S23 is as follows: ； Among them, softmax(.) represents the normalized exponential function, Mul(.) represents the product function of the corresponding elements of the tensor, and Sum(.) represents the sum function of the tensor elements; Represents any moment between tm and t, where m is the number of moments in the Attention mechanism time window. represents the hidden state of the lth convolutional layer at time i (i∈(tm, t)); It represents the importance weight of the previous moment information, and utilizes the time information between each convolutional layer in the time window to realize the time feature extraction on the spatial feature map.

6. The method for abnormal identification and repair of multi-dimensional time series data based on convolutional neural network according to claim 3 is characterized in that: The specific implementation process of step S24 is as follows: Design a convolutional decoder with the following formula: ； in, represents the deconvolution operation, represents the concatenation operation, g(.) represents the activation function, and They represent the convolution kernel and bias term of the lth deconvolution layer respectively, represents the deconvolution output feature map of the lth layer at time t; from back to front, the lth layer ConvLSTM Feed to the deconvolutional neural network and output feature map It is concatenated in series with the output of the l-1th ConvLSTM layer, and the concatenated matrix is ​​further input into the next deconvolution layer; the final output , represents the reconstructed correlation matrix.

7. The method for abnormal identification and repair of multi-dimensional time series data based on convolutional neural network according to claim 1 is characterized in that: The specific implementation process of step S3 is: The loss function is defined as: ； in, represents the F norm of the matrix, N represents the total number of measurement points, and the Adam optimizer is used to minimize the above loss; represents the reconstruction error matrix at time t, represents the original multidimensional serial correlation matrix; represents the reconstructed correlation matrix; After multiple training cycles, the offline training parameters of the model are obtained, and the trained parameters are saved to the file system for loading and use by the online system.

8. The method for abnormal identification and repair of multi-dimensional time series data based on convolutional neural network according to claim 1 is characterized in that: In step S4, the neural network model parameters are loaded when the real-time system is running, and the multidimensional data at the current moment is reconstructed through the model. The online data is compared with the reconstructed data. If the difference between the data in a certain dimension exceeds the threshold, there is an abnormality, and the multidimensional reconstructed data is used to repair the abnormality. At the same time, the repair time and content are recorded in the exception handling log, and the relevant personnel are notified to investigate the source of the abnormality.

Images