Dtw-based multivariate time series feature prediction method and system

Abstract

The application provides a kind of based on DTW's multivariate time sequence feature prediction method and system, comprising: satellite telemetry data is segmented and handled, and original time sequence is obtained;Respectively extract features, combine to generate multivariate feature vector as initial sample dataset for preprocessing;Time scale alignment equal interval interpolation processing is carried out to the processed feature sequence, and sample time sequence feature set is generated;The initial sample dataset and corresponding classification label set are used as the input of LSTM network, and model training is carried out;According to the feature sequence of sample data and corresponding classification label, the corresponding type feature template is generated;Input trained LSTM model, and obtain classification label;The similarity of each segment feature template and mobile feature sequence is calculated in turn;The biggest template sequence of similarity is selected to obtain the predicted value of specified time and the final predicted value is output by inverse normalization processing.The application can simultaneously parallelly calculate the similarity value between each segment sequence, thereby significantly improving the matching speed.

Description

Technical Field

[0001] This invention relates to the field of time series data mining technology, specifically to a multivariate time series feature prediction method and system based on DTW. Background Technology

[0002] DTW (Time Warping) is a typical optimization problem that uses a time warping function that satisfies certain conditions to describe the temporal correspondence between a test template and a reference template. The goal is to find the warping function that minimizes the cumulative distance when the two templates match. Its basic principle is to find the optimal warping path, matching the coordinates of data points in one time series with the coordinates of data points in another time series that have the highest feature similarity.

[0003] LSTM is an improved RNN model. Its improvement lies in replacing traditional hidden nodes with memory blocks, preventing gradients from vanishing or exploding over long training periods, thus solving the problem of RNNs' inability to handle excessively long sequences. The basic unit of the LSTM model is the memory module, which contains memory cells and three gate structures that control the state of the memory cells: the forget gate, the input gate, and the output gate. The forget gate determines whether useless historical information is forgotten from the memory cell state, the input gate determines the impact of the current input data on the memory cell state, and the output gate determines the output information.

[0004] Patent document CN104915568A (application number: CN201510351199.7) discloses a method for anomaly detection in satellite telemetry data based on DTW. Its key feature is that it includes the following steps: Step 1: Segmenting the data using abrupt change points as identifiers to obtain time series containing category labels; Step 2: Calculating the average DTW distance within each category of the time series containing category labels.

[0005] However, patent document CN104915568A addresses the problems of existing detection methods, such as missed detection of satellite component anomalies due to abnormal parameters not exceeding alarm thresholds, significant deviations in segmenting periodic satellite telemetry data with fixed point counts, and inaccurate measurement results due to slight time series shifts, thus leading to inaccurate anomaly detection results.

[0006] Patent document CN104915434A (application number: CN201510351181.7) discloses a multidimensional time series classification method based on Mahalanobis distance DTW, characterized by the following steps: Step 1: Segmenting historical satellite telemetry data Y under normal satellite operation status using amplitude abrupt change points as identifiers to obtain normal multidimensional time series; Step 2: Clustering the segmented multidimensional time series using a hierarchical clustering method with a set number of clustering target categories to obtain category labels for the multidimensional time series; Step 3: Extracting test data within the time points corresponding to adjacent m+1 amplitude abrupt change points in the latest satellite telemetry data; Step 4: Calculating the DTW distance sequence between the multidimensional time series to be classified and the multidimensional time series containing category labels; Step 5: Using the KNN classification method based on Mahalanobis distance DTW distance, classifying the multidimensional time series to be classified according to the set number of K nearest neighbors to determine the category to which the multidimensional time series to be classified belongs.

[0007] However, patent document CN104915434A aims to address the problems of unsatisfactory fixed-point segmentation of satellite telemetry data, inaccurate measurement results due to correlations between multidimensional time series and slight shifts in time series, which in turn lead to inaccurate classification results. It proposes a multidimensional time series classification method based on Mahalanobis distance (DTW).

[0008] Patent document CN108764460A (application number: CN201810469767.7) discloses a time series prediction method based on temporal convolution and LSTM, comprising: acquiring sample data for a time series prediction problem, including a time series and a target sequence; establishing multiple candidate models based on the TC#LSTM model using different hyperparameter combinations; preprocessing the time series to obtain a window sequence and a target set, and dividing them into a training set and a test set; training the candidate models using the training set; using the trained candidate models to predict a subset of the window sequence in the test set, and calculating the root mean square error between the prediction result of each candidate model and the target set subset in the test set; selecting the candidate model with the smallest root mean square error as the prediction model; preprocessing the time series to be predicted to obtain the window sequence to be predicted, and then using the prediction model to predict the window sequence to obtain the target value.

[0009] However, there is still room for improvement in the accuracy of time series prediction obtained from patent document CN108764460A.

[0010] Therefore, given the limited availability of actual satellite telemetry data, intelligent feature prediction algorithms require a large amount of sample data for training. There is a market need for a multivariate time-series feature prediction method and system based on DTW that can improve matching speed and feature prediction accuracy. Summary of the Invention

[0011] In view of the shortcomings of the prior art, the purpose of this invention is to provide a multivariate time series feature prediction method and system based on DTW.

[0012] A multivariate time-series feature prediction method based on DTW provided by the present invention includes:

[0013] Step S1: Segment the satellite telemetry data according to its periodic characteristics to obtain the original time series;

[0014] Step S2: Extract features based on each dimension of variables and combine them to generate a time-based multivariate feature vector as the initial sample dataset;

[0015] Step S3: Preprocess the set of original time series;

[0016] Step S4: Set a time interval and perform time-scale aligned equal-interval interpolation on the processed feature sequence to generate a sample time series feature set;

[0017] Step S5: Use the initial sample dataset and the corresponding classification label set as input to the LSTM network to train the model and store the trained model parameters.

[0018] Step S6: Generate feature templates of the corresponding type based on the feature sequences of the sample data and the corresponding classification labels;

[0019] Step S7: Normalize the feature sequence to be predicted, input it into the trained LSTM model, and obtain the classification label;

[0020] Step S8: Compare the feature sequence to be predicted with the feature template of the matching type, slide along the feature template according to the set step size, and calculate the similarity between each feature template segment and the moving feature sequence in turn.

[0021] Step S9: Select the template sequence with the highest similarity, perform time-stamp alignment, obtain the predicted value at the specified time, and perform inverse normalization on the predicted value to output the final predicted value.

[0022] Preferably, the feature extraction process in step S2 includes the following sub-steps:

[0023] Step S2.1: Extract feature elements based on the time-varying effects on sample characteristics and classification from the sample data;

[0024] Step S2.2: Generate a multivariate feature vector based on the extracted feature elements.

[0025] Preferably, the data preprocessing in step S3 includes the following sub-steps:

[0026] Step S3.1: Calculate the expected value and standard deviation of the feature sequence dataset;

[0027] Step S3.2: Remove abnormal data, including data larger than μ. k +3σ k and less than μ k -3σ k The data, in which μ k σ represents the mathematical expectation of the k-dimensional sample dataset. k Represents the standard deviation of the k-dimensional sample dataset;

[0028] Step S3.3: Use the mean imputation method to complete the data at the rejection locations, as shown in the following formula:

[0029] x′ k =(x k-1 +x k+1 )σ k / 2

[0030] Where, x k-1 This indicates the data preceding the location data to be removed, x k+1 σ represents the data after removing the location data. k This represents the standard deviation of the sample data in k dimensions;

[0031] Step S3.4: Normalize the cleaned data using the deviation standardization method, as shown in the following formula:

[0032] x * =( x -x min ) / (x max -x min )

[0033] Where, x i Represents the original data, x min x represents the minimum value of the sample. max This represents the maximum value of the sample.

[0034] Preferably, in step S4, the interpolation process needs to be performed by setting a time interval according to the temporal characteristics of each dataset, and the time interval shall not be greater than the actual interval of the feature sequence set.

[0035] Preferably, step S5 includes:

[0036] Step S5.1: Normalize the sample feature dataset;

[0037] Step S5.2: Match the sample feature data with the classification labels one by one and combine them as a new training feature dataset;

[0038] Step S5.3: Train the LSTM neural network using the normalized data to obtain the prediction model.

[0039] Preferably, the method for generating the feature template in step S6 includes the following sub-steps:

[0040] Step S6.1: Select two feature sequences from sample data of the same type;

[0041] Step S6.2: Use the longer feature sequence as the fixed comparison feature template and the shorter feature sequence as the moving comparison feature template;

[0042] Step S6.3: Fill the fixed comparison template sequence with elements of a fixed length and a value of 0 before and after;

[0043] Step S6.4: Slide the moving feature sequence along the filled fixed template sequence with a set step size, and calculate the similarity between each feature sequence and the moving feature sequence in turn;

[0044] Step S6.5: Select the two sequences with the highest similarity. If the similarity value is not less than the threshold, fit the two sequences together. Otherwise, the feature sequence fails to match and is treated as a new category.

[0045] Step S6.6: After matching multiple curves, the resulting fitted curve segments are spliced ​​together to form a complete curve, which is the template for the time feature sequence.

[0046] Preferably, step S6.4 includes the following sub-steps:

[0047] Step S6.4.1: Calculate the distance matrix between the two sequences;

[0048] Step S6.4.2: Calculate the dynamic normalization matrix of the two sequences;

[0049] Step S6.4.3: Backtrack to find the matching relationship between each feature point;

[0050] Step S6.4.4: Calculate the feature matching matrix of the two sequences;

[0051] Step S6.4.5: Remove feature point pairs that are repeatedly matched in the feature matching matrices of the two sequences;

[0052] Step S6.4.6: Calculate the similarity between the two sequences using the following formula:

[0053]

[0054] Where, x i Let y represent the i-th feature of the time series x. i Let N represent the i-th feature of the time series y, and N represent the feature dimension.

[0055] Preferably, step S7 includes the following steps:

[0056] Step S7.1: Normalize the feature sequence to be predicted;

[0057] Step S7.2: Input the normalized data into the trained prediction model to obtain the classification labels.

[0058] Preferably, step S8 includes the following steps:

[0059] Step S8.1: Perform time-series filling on the fixed template sequence;

[0060] Step S8.2: Slide the moving feature sequence along the filled fixed template sequence with a set step size to obtain two feature sequences with equal time spans;

[0061] Step S8.3: Calculate the similarity between each feature sequence and the moving feature sequence in turn.

[0062] According to the present invention, a multivariate time series feature prediction system based on DTW is provided, comprising:

[0063] Module M1: Segments the satellite telemetry data according to its periodic characteristics to obtain the original time series;

[0064] Module M2: Extracts features based on variables of each dimension and combines them to generate a time-based multivariate feature vector as the initial sample dataset;

[0065] Module M3: Preprocesses the set of original time series;

[0066] Module M4: Sets a time interval to perform time-stamped, equally spaced interpolation on the processed feature sequence to generate a sample time series feature set;

[0067] Module M5: Takes the initial sample dataset and the corresponding classification label set as input to the LSTM network, trains the model, and stores the trained model parameters;

[0068] Module M6: Generates feature templates of the corresponding type based on the feature sequences of the sample data and the corresponding classification labels;

[0069] Module M7: Normalizes the feature sequence to be predicted, inputs it into the trained LSTM model, and obtains the classification label;

[0070] Module M8: Compares the feature sequence to be predicted with the feature template of the matching type, slides along the feature template at the set step size, and calculates the similarity between each feature template segment and the moving feature sequence in turn;

[0071] Module M9: Selects the template sequence with the highest similarity, performs time-stamp alignment, obtains the predicted value at the specified time, and performs inverse normalization on the predicted value to output the final predicted value.

[0072] Compared with the prior art, the present invention has the following beneficial effects:

[0073] 1. This invention segments the original time series and can simultaneously calculate the similarity values ​​between each segment, thereby significantly improving the matching speed.

[0074] 2. This invention improves the accuracy of feature templates by using interpolation, sliding window iteration, and piecewise fitting processes, thereby improving the accuracy of feature prediction. Attached Figure Description

[0075] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0076] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0077] Figure 2 This is a schematic diagram of the satellite telemetry data preprocessing process in this invention;

[0078] Figure 3 This is a schematic diagram of the LSTM neural network structure in this invention;

[0079] Figure 4 This is a schematic diagram of the LSTM network model training process in this invention;

[0080] Figure 5 This is a schematic diagram illustrating the principle of the DTW algorithm in this invention. Detailed Implementation

[0081] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0082] According to the present invention, a multivariate time-series feature prediction method based on DTW is provided, such as... Figure 1 As shown, it includes the following:

[0083] Step S1: Based on the periodic characteristics of the satellite telemetry data, segment the satellite telemetry data to obtain the original time series.

[0084] Step S2: Extract features based on each dimension of the variable, and combine them to generate a time-based multivariate feature vector as the initial sample dataset. The feature extraction process in step S2 includes the following sub-steps:

[0085] Step S2.1: Extract feature elements based on the time changes that affect the sample characteristics and classification of the sample data.

[0086] Step S2.2: Generate a multivariate feature vector based on the extracted feature elements.

[0087] Step S3: Preprocess the original time series dataset. For example... Figure 2 As shown, the data preprocessing in step S3 includes the following sub-steps:

[0088] Step S3.1: Calculate the expected value and standard deviation of the feature sequence dataset.

[0089] Step S3.2: Remove abnormal data, including data larger than μ. k +3σ k and less than μ k -3σ k Data,

[0090] Where, μ k σ represents the mathematical expectation of the k-dimensional sample dataset. k This represents the standard deviation of the k-dimensional sample dataset.

[0091] Step S3.3: Use the mean imputation method to complete the data at the elimination positions, according to the formula.

[0092] x′ k =(x k-1 +x k+1 )σ k / 2

[0093] Where, x k-1 This indicates the data preceding the location data to be removed, x k+1 σ represents the data after removing the location data. k The standard deviation of the sample data in k dimensions

[0094] Step S3.4: Normalize the cleaned data using the deviation standardization method, according to the formula...

[0095] x * =(x i -x min ) / (x max -x min )

[0096] Where, x i Represents the original data, x min x represents the minimum value of the sample. max This represents the maximum value of the sample.

[0097] Step S4: Set a time interval and perform time-scale aligned equal-interval interpolation on the processed feature sequences to generate a sample time series feature set. The interpolation process in Step S4 requires setting a time interval based on the temporal characteristics of each dataset, and this time interval must not exceed the actual interval of the feature sequence set.

[0098] Step S5: Use the initial sample dataset and the corresponding classification label set as input to the LSTM network to train the model and store the trained model parameters. Step S5 includes:

[0099] Step S5.1: Normalize the sample feature dataset.

[0100] Step S5.2: Match the sample feature data with the classification labels one by one and combine them to form a new training feature dataset.

[0101] Step S5.3: Train an LSTM neural network using the normalized data to obtain a prediction model. The LSTM neural network structure is as follows: Figure 3 As shown, the LSTM network includes input, output, and gate structures. The input includes the cell state, hidden state, and the input vector at time t. The output includes the cell state and hidden state; the gate structures include a forget gate, an input gate, and an output gate. The LSTM network model training process is as follows: Figure 4 As shown, the process includes: first, inputting training samples and initializing grid parameters; then, performing forward computation to calculate the model error function and determining whether the model error function meets the requirements. If yes, LSTM training is completed; otherwise, backward error update is performed, returning to forward computation and recalculating the model error function until the model error function meets the requirements.

[0102] Step S6: Generate a feature template of the corresponding type based on the feature sequence of the sample data and the corresponding classification label. The method for generating the feature template in step S6 includes the following sub-steps:

[0103] Step S6.1: Select two feature sequences from the same type of sample data.

[0104] Step S6.2: Use the longer feature sequence as the fixed comparison feature template and the shorter feature sequence as the moving comparison feature template.

[0105] Step S6.3: Fill the fixed comparison template sequence with elements of a fixed length and a value of 0 before and after it.

[0106] Step S6.4: Slide the moving feature sequence along the padded fixed template sequence with a set step size, and calculate the similarity between each feature sequence segment and the moving feature sequence sequentially. Step S6.4 includes the following sub-steps:

[0107] Step S6.4.1: Calculate the distance matrix between the two sequences.

[0108] Step S6.4.2: Calculate the dynamic normalization matrix of the two sequences.

[0109] Step S6.4.3: Backtrack to find the matching relationship between each feature point.

[0110] Step S6.4.4: Calculate the feature matching matrix of the two sequences.

[0111] Step S6.4.5: Remove feature point pairs that are repeatedly matched in the two sequence feature matching matrices.

[0112] Step S6.4.6: Calculate the similarity between the two sequences using the following formula:

[0113]

[0114] Where, x i Let y represent the i-th feature of the time series x. i Let N represent the i-th feature of the time series y, and N represent the feature dimension.

[0115] Step S6.5: Select the two sequences with the highest similarity. If the similarity value is not less than the threshold, fit the two sequences together. Otherwise, the feature sequence fails to match and is treated as a new classification.

[0116] Step S6.6: After matching multiple curves, the resulting fitted curve segments are spliced ​​together to form a complete curve, which is the template for the time feature sequence.

[0117] Step S7: Normalize the feature sequence to be predicted, input it into the trained LSTM model, and obtain the classification label. Step S7 includes the following steps: Perform model prediction:

[0118] Step S7.1: Normalize the feature sequence to be predicted.

[0119] Step S7.2: Input the normalized data into the trained prediction model to obtain the classification labels.

[0120] Step S8: Compare the feature sequence to be predicted with the feature template of the matching type, slide along the feature template according to the set step size, and calculate the similarity between each segment of the feature template and the moving feature sequence in turn. Step S8 includes the following steps:

[0121] Step S8.1: Perform time-series filling on the fixed template sequence.

[0122] Step S8.2: Slide the moving feature sequence along the filled fixed template sequence with a set step size to obtain two feature sequences with equal time spans.

[0123] Step S8.3: Calculate the similarity between each feature sequence and the moving feature sequence in turn.

[0124] Step S9: Select the template sequence with the highest similarity, perform time-stamp alignment, obtain the predicted value at the specified time, and perform inverse normalization on the predicted value to output the final predicted value.

[0125] Furthermore, such as Figure 5 As shown, both the horizontal and vertical axes represent time. First, the two time series, namely the reference trajectory X and the trajectory to be reconstructed Y, are aligned; then, a continuous matching relationship containing all corresponding points in both time series is found; based on the matching relationship, a time-based regularized path is generated, which is the reconstructed trajectory Y*, satisfying the conditions of boundary property, continuity, and monotonicity.

[0126] The present invention also provides a multivariate time series feature prediction system based on DTW. The multivariate time series feature prediction system based on DTW can be implemented by executing the process steps of the multivariate time series feature prediction method based on DTW. That is, those skilled in the art can understand the multivariate time series feature prediction method based on DTW as a preferred embodiment of the multivariate time series feature prediction system based on DTW.

[0127] According to the present invention, a multivariate time-series feature prediction system based on DTW includes the following:

[0128] Module M1: Based on the periodic characteristics of satellite telemetry data, segment the satellite telemetry data to obtain the original time series.

[0129] Module M2: Extracts features based on variables of each dimension and combines them to generate a time-based multivariate feature vector as the initial sample dataset. The feature extraction process in Module M2 includes the following sub-steps: Module M2.1: Extracts feature elements based on the time-varying effects on sample features and classification from the sample data. Module M2.2: Combines the extracted feature elements to generate a multivariate feature vector.

[0130] Module M3: Preprocesses the original time series dataset. The data preprocessing in Module M3 includes the following sub-steps: Module M3.1: Calculates the expected value and standard deviation of the feature sequence dataset. Module M3.2: Removes outliers, including data larger than μ. k +3σ k and less than μ k -3σ k The data, in which μ k σ represents the mathematical expectation of the k-dimensional sample dataset. k This represents the standard deviation of the k-th dimension of the sample dataset. Module M3.3: Uses mean imputation to complete the data at the removal locations, based on the formula: x′ k =(x k-1 +x k+1 )σ k / 2, where x k-1 This indicates the data preceding the location data to be removed, x k+1 σ represents the data after removing the location data. k This represents the standard deviation of the sample data in dimension k. Module M3.4: Normalizes the cleaned data using the deviation standardization method, based on the formula: x * =(x i -x min ) / (x max -x min Where, x i Represents the original data, x min x represents the minimum value of the sample. max This represents the maximum value of the sample.

[0131] Module M4: Sets a time interval to perform time-scale aligned and equally spaced interpolation on the processed feature sequences, generating a sample time series feature set. The interpolation process in module M4 requires setting a time interval based on the temporal characteristics of each dataset; this time interval must not exceed the actual interval of the feature sequence set.

[0132] Module M5: Uses the initial sample dataset and corresponding classification label set as input to the LSTM network for model training and stores the trained model parameters. Module M5 includes: Module M5.1: Normalizes the sample feature dataset. Module M5.2: Matches the sample feature data with the classification labels one-to-one, combining them to form a new training feature dataset. Module M5.3: Trains the LSTM neural network using the normalized data to obtain the prediction model.

[0133] Module M6: Generates feature templates of corresponding types based on the feature sequences and corresponding classification labels of the sample data. The system for generating feature templates in Module M6 includes the following sub-steps: Module M6.1: Select two feature sequences from sample data of the same type. Module M6.2: Use the longer feature sequence as a fixed comparison feature template and the shorter feature sequence as a moving comparison feature template. Module M6.3: Fill the fixed comparison template sequence with elements of a fixed length and a value of 0 before and after it. Module M6.4: Slide the moving feature sequence along the filled fixed template sequence with a set step size, and calculate the similarity between each feature sequence and the moving feature sequence. Module M6.4 includes the following sub-steps: Module M6.4.1: Calculate the distance matrix between the two sequences. Module M6.4.2: Calculate the dynamic normalization matrix between the two sequences. Module M6.4.3: Backtrack to find the matching relationship between each feature point. Module M6.4.4: Calculate the feature matching matrix between the two sequences. Module M6.4.5: Remove feature point pairs that are repeatedly matched in the feature matching matrix between the two sequences. Module M6.4.6: Calculates the similarity between two sequences using the following formula:

[0134]

[0135] Where, x i Let y represent the i-th feature of the time series x. i Let N represent the i-th feature of the time series y, and N represent the feature dimension. Module M6.5: Select the two sequences with the highest similarity. If the similarity value is not less than a threshold, fit the two sequences together; otherwise, the feature sequence fails to match and is treated as a new category. Module M6.6: Piece together the fitted curve segments generated after matching multiple curves into a complete curve, which is the template for the time feature sequence.

[0136] Module M7: Normalizes the feature sequence to be predicted, inputs it into the trained LSTM model, and obtains the classification label. Module M7 includes the following steps: Model prediction: Module M7.1: Normalizes the feature sequence to be predicted. Module M7.2: Inputs the normalized data into the trained prediction model and obtains the classification label.

[0137] Module M8: Compares the feature sequence to be predicted with a matching feature template, slides along the feature template at a set step size, and calculates the similarity between each feature template segment and the moving feature sequence. Module M8 includes the following steps: Module M8.1: Temporally fills the fixed template sequence. Module M8.2: Slides the moving feature sequence along the filled fixed template sequence at a set step size to obtain two feature sequences with equal time spans. Module M8.3: Calculates the similarity between each feature sequence segment and the moving feature sequence.

[0138] Module M9: Selects the template sequence with the highest similarity, performs time-stamp alignment, obtains the predicted value at the specified time, and performs inverse normalization on the predicted value to output the final predicted value.

[0139] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0140] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A multivariate temporal feature prediction method based on DTW, characterized in that, include: Step S1: Segment the satellite telemetry data according to its periodic characteristics to obtain the original time series; Step S2: Extract features based on each dimension of variables and combine them to generate a time-based multivariate feature vector as the initial sample dataset; Step S3: Preprocess the set of original time series; Step S4: Set a time interval and perform time-scale aligned equal-interval interpolation on the processed feature sequence to generate a sample time series feature set; Step S5: Use the initial sample dataset and the corresponding classification label set as input to the LSTM network to train the model and store the trained model parameters; Step S6: Generate feature templates of the corresponding type based on the feature sequences of the sample data and the corresponding classification labels; Step S7: Normalize the feature sequence to be predicted, input it into the trained LSTM model, and obtain the classification label; Step S8: Compare the feature sequence to be predicted with the feature template of the matching type, slide along the feature template according to the set step size, and calculate the similarity between each feature template segment and the moving feature sequence in turn. Step S9: Select the template sequence with the highest similarity, perform time-stamp alignment, obtain the predicted value at the specified time, and perform inverse normalization on the predicted value to output the final predicted value. The method for generating the feature template in step S6 includes the following sub-steps: Step S6.1: Select two feature sequences from sample data of the same type; Step S6.2: Use the longer feature sequence as the fixed comparison feature template and the shorter feature sequence as the moving comparison feature template; Step S6.3: Fill the fixed comparison template sequence with elements of a fixed length and a value of 0 before and after; Step S6.4: Slide the moving feature sequence along the filled fixed template sequence with a set step size, and calculate the similarity between each feature sequence and the moving feature sequence in turn; Step S6.5: Select the two sequences with the highest similarity. If the similarity value is not less than the threshold, fit the two sequences together. Otherwise, the feature sequence fails to match and is treated as a new category. Step S6.6: After matching multiple curves, the resulting fitted curve segments are spliced ​​together to form a complete curve, which is the template for the time feature sequence.

2. The multivariate temporal feature prediction method based on DTW according to claim 1, characterized in that, The feature extraction process in step S2 includes the following sub-steps: Step S2.1: Extract feature elements based on the time-varying effects on sample characteristics and classification from the sample data; Step S2.2: Generate a multivariate feature vector based on the extracted feature elements.

3. The multivariate time-series feature prediction method based on DTW according to claim 1, characterized in that, The data preprocessing in step S3 includes the following sub-steps: Step S3.1: Calculate the expected value and standard deviation of the feature sequence dataset; Step S3.2: Remove abnormal data, the abnormal data including data larger than... and less than The data, among which, Represents the sample dataset The mathematical expectation of the dimension Represents the sample dataset Standard deviation of dimension; Step S3.3: Use the mean imputation method to complete the data at the rejection locations, as shown in the following formula: in, This indicates removing the data preceding the location data. This indicates the data after the location data has been removed. Representing sample data Standard deviation of dimension; Step S3.4: Normalize the cleaned data using the deviation standardization method, as shown in the following formula: in, Represents the original data. This represents the minimum value of the sample. This represents the maximum value of the sample.

4. The multivariate time-series feature prediction method based on DTW according to claim 1, characterized in that, In step S4, the interpolation process needs to be performed by setting a time interval according to the temporal characteristics of each dataset. The time interval must not be greater than the actual interval of the feature sequence set.

5. The multivariate time-series feature prediction method based on DTW according to claim 1, characterized in that, Step S5 includes: Step S5.1: Normalize the sample feature dataset; Step S5.2: Match the sample feature data with the classification labels one by one and combine them as a new training feature dataset; Step S5.3: Train the LSTM neural network using the normalized data to obtain the prediction model.

6. The multivariate time-series feature prediction method based on DTW according to claim 1, characterized in that, Step S6.4 includes the following sub-steps: Step S6.4.1: Calculate the distance matrix between the two sequences; Step S6.4.2: Calculate the dynamic normalization matrix of the two sequences; Step S6.4.3: Backtrack to find the matching relationship between each feature point; Step S6.4.4: Calculate the feature matching matrix of the two sequences; Step S6.4.5: Remove feature point pairs that are repeatedly matched in the feature matching matrices of the two sequences; Step S6.4.6: Calculate the similarity between the two sequences using the following formula: in, Representing time series The dimensional features, Representing time series The dimensional features, This represents the feature dimension.

7. The multivariate time-series feature prediction method based on DTW according to claim 1, characterized in that, Step S7 includes the following steps: Step S7.1: Normalize the feature sequence to be predicted; Step S7.2: Input the normalized data into the trained prediction model to obtain the classification labels.

8. The multivariate time-series feature prediction method based on DTW according to claim 1, characterized in that, Step S8 includes the following steps: Step S8.1: Perform time-series filling on the fixed template sequence; Step S8.2: Slide the moving feature sequence along the filled fixed template sequence with a set step size to obtain two feature sequences with equal time spans; Step S8.3: Calculate the similarity between each feature sequence and the moving feature sequence in turn.

9. A multivariate time-series feature prediction system based on DTW, characterized in that, include: Module M1: Segments the satellite telemetry data according to its periodic characteristics to obtain the original time series; Module M2: Extracts features based on variables of each dimension and combines them to generate a time-based multivariate feature vector as the initial sample dataset; Module M3: Preprocesses the set of original time series; Module M4: Sets a time interval to perform time-stamped, equally spaced interpolation on the processed feature sequence to generate a sample time series feature set; Module M5: Takes the initial sample dataset and the corresponding classification label set as input to the LSTM network, trains the model, and stores the trained model parameters; Module M6: Generates feature templates of the corresponding type based on the feature sequences of the sample data and the corresponding classification labels; Module M7: Normalizes the feature sequence to be predicted, inputs it into the trained LSTM model, and obtains the classification label; Module M8: Compares the feature sequence to be predicted with the feature template of the matching type, slides along the feature template at the set step size, and calculates the similarity between each feature template segment and the moving feature sequence in turn; Module M9: Selects the template sequence with the highest similarity, performs time-stamp alignment, obtains the predicted value at the specified time, and performs inverse normalization on the predicted value to output the final predicted value; The method for generating feature templates in module M6 includes the following sub-steps: Module M6.1: Select two feature sequences from sample data of the same type; Module M6.2: Uses longer feature sequences as fixed comparison feature templates and shorter feature sequences as moving comparison feature templates; Module M6.3: Pads the fixed comparison template sequence with elements of a fixed length that are 0 before and after it; Module M6.4: Slide the moving feature sequence along the filled fixed template sequence with a set step size, and calculate the similarity between each feature sequence and the moving feature sequence in turn; Module M6.5: Select the two sequences with the highest similarity. If the similarity value is not less than the threshold, the two sequences are fitted. Otherwise, the feature sequence fails to match and is treated as a new category. Module M6.6: After matching multiple curves, the resulting fitted curve segments are spliced ​​into a complete curve, which is the template for the time feature sequence.

Images