Similarity measurement and classification method of time series signals based on multi-feature fusion

Abstract

The present invention discloses a time series signal similarity measurement and classification method based on multi-feature fusion. The classification method first takes the input time series signal as a benchmark, screens out signals with extremely large differences through hierarchical filtering, and selects a measurement index for measuring the similarity of the time series signals. The time series signal similarity measurement method based on multi-feature fusion is used to calculate the importance of a statistical data set on various similarity measurement indicators, while improving the scaling method in the hierarchical analysis method and constructing a more accurate judgment matrix to obtain the weight coefficients of the data set on various similarity measurement indicators. Finally, the method is used to perform similarity measurement on a measured data set, implement classification testing, and conduct comparative tests of different similarity measurement methods. The test results show that the proposed method can accurately reflect the similarity between time series signals in different data sets, and performs well in both classification accuracy and calculation speed.

Description

Technical Field

[0001] The present invention belongs to the field of data mining technology and relates to data similarity analysis, and specifically to a time series signal similarity measurement and classification method based on multi-feature fusion. Background Art

[0002] Time series signals, a set of data points arranged in chronological order, are ubiquitous in fields such as medicine, economics, meteorology, and engineering. With the development of big data and large models, and the widespread adoption of sensor technology, extracting useful information from massive amounts of time series signals using data mining techniques has become increasingly important and has broad application prospects.

[0003] To date, data mining of time series signals has primarily focused on tasks such as prediction, classification, clustering, and anomaly detection. A key component of these tasks is measuring the similarity of time series signals. Time series similarity measures can help quickly locate and analyze new time series signals, providing a basis for decision-making and improving the efficiency and quality of data mining.

[0004] While existing technologies have improved the efficiency and effectiveness of metrics at various levels through methods such as feature extraction, dimensionality reduction, and algorithm optimization, they still have some shortcomings. First, they lack adaptability to datasets in different application scenarios, and the stability of results and robustness of algorithms need to be improved. Second, processing long time series, especially when real-time analysis is required, still requires a significant amount of computing time. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this paper proposes a method for similarity measurement and classification of time series signals based on multi-feature fusion. This method introduces the concepts of multi-feature fusion and hierarchical filtering, utilizes multiple similarity feature indices to measure the similarity between time series signals, and provides a method for calculating feature weight coefficients to accurately characterize the importance of selected similarity feature indices across different datasets. This method not only enhances its versatility and generalizability, but also significantly improves computational speed, accuracy, and stability of measurement results.

[0006] The time series signal similarity measurement method based on multi-feature fusion specifically includes the following steps:

[0007] Step 1: Select N similarity feature indicators to measure the similarity between different time series signals.

[0008] Step 2: Set the similarity interval threshold for each signal in the data set. Construct a similarity interval. Calculate the similarity of each signal in the data set under different similarity feature indicators, sort them from large to small according to the similarity, and take the top signal, construct signal In high similarity intervals under different similarity feature indicators, compare the elements in the similar intervals and the high similarity intervals to determine the signal Contribution value to the D-th similarity feature index:

[0009]

[0010]

[0011] in Indicates signal The similarity interval of Indicates signal The high similarity interval of Represents similar interval The first A signal.

[0012] Normalize the contribution value of each signal to obtain the percentage of the Dth similarity feature index in the data set :

[0013]

[0014]

[0015] Step 3: Based on the hierarchical analysis method, the difference between the maximum and minimum percentages of the similarity feature index in a data set is calculated. It is divided into 9 distribution intervals of importance, forming a scaling method based on interval constraints. According to the importance between each similarity feature index, a judgment matrix M is constructed to distribute the weight distribution results of each similarity feature index in the data set.

[0016] Step 4: When measuring the similarity between time series signals X and Y, first calculate the similarity between X and Y under multiple similarity feature indicators in step 1. Then, according to the methods of steps 2 and 3, calculate the weight values ​​of the data set where the time series signal Y is located for different similarity feature indicators Finally, the results of multiple similarity feature indicators are weighted and summed according to the weight value as the similarity measurement result of time series signals X and Y :

[0017]

[0018] The classification method of time series signal similarity measurement based on multi-feature fusion first determines the weight value of the similarity feature index under different known data sets. Then obtain the length of the time series signal T to be classified , remove the known data sets whose length range is in the interval Signals other than Is the set length threshold parameter. Further obtain the total amplitude of the time series signal T to be classified , remove the remaining signals with a total amplitude of Signals other than is the total amplitude threshold parameter, is the total signal amplitude of the time series signal T.

[0019] Finally, the similarity measurement result between the time series signal T and the remaining signals is calculated, and the similarity measurement result is used as the distance between the time series signal T and the remaining signals. Based on the idea of ​​KNN, Z signals closest to the time series signal T are selected from the remaining signals, and the categories of these signals are counted. The category with the highest number of occurrences is taken as the category of the time series signal T to complete the classification.

[0020] The present invention has the following beneficial effects:

[0021] (1) It eliminates the limitations of traditional methods on single features and provides multiple similarity measurement indicators and their corresponding similarity measurement methods. In particular, a hierarchical analysis method based on similarity interval constraints is proposed to solve the problem of allocating weight coefficients to each similarity measurement indicator. This method optimizes the scaling method in the traditional hierarchical analysis method by quantifying the importance of multiple similarity measurement indicators, and then obtains a judgment matrix suitable for time series signal data sets, thereby calculating accurate weight coefficients.

[0022] (2) The concept of hierarchical filtering is introduced to reduce unnecessary computation. Experimental results on real datasets show that this method achieves a classification accuracy of over 90% in classification tasks, and the computation time is reduced by an order of magnitude compared to other methods. The overall solution not only takes into account excellent usability and strong scalability, but also ensures high generalization. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a flow chart of the time series signal similarity measurement method based on multi-feature fusion;

[0024] Figure 2 This is an example of the similarity feature index percentage result used in Example 1;

[0025] Figure 3 This is a schematic diagram of the target placement in the test site obtained from the data set in Example 2;

[0026] Figure 4 is the original collected signal of the data set DS1 in Example 2 of the method of the present invention;

[0027] Figure 5 is the original collected signal of the data set DS2 in Example 2 of the method of the present invention;

[0028] Figure 6 is the original collected signal of the data set DS3 in Example 2 of the method of the present invention;

[0029] Figure 7 is the original collected signal of the data set DS4 in Example 2 of the method of the present invention;

[0030] Figure 8 The data set DS1 in Example 2 of the method of the present invention is a signal after data preprocessing;

[0031] Figure 9 is the signal after data preprocessing of the data set DS2 in Example 2 of the method of the present invention;

[0032] Figure 10 The data set DS3 in Example 2 of the method of the present invention is a signal after data preprocessing;

[0033] Figure 11 The data set DS4 in Example 2 of the method of the present invention is a signal after data preprocessing;

[0034] Figure 12 This is a flow chart of the classification method based on the similarity measurement of time series signals based on multi-feature fusion. DETAILED DESCRIPTION

[0035] The present invention will be further explained below with reference to the accompanying drawings;

[0036] Example 1

[0037] This embodiment introduces a time series signal similarity measurement method based on multi-feature fusion, such as Figure 1 As shown, it includes the following steps:

[0038] Step 1: First, preprocess the data, including standardization, data format conversion, and noise removal, to improve the quality, accuracy, and usability of the data and lay the foundation for subsequent processing steps. Then, select signal length, total signal amplitude, signal dispersion, maximum signal cross-correlation coefficient, signal shape, and signal distance as similarity feature indicators to measure the similarity between time series signals X and Y:

[0039] s1.1. Similarity of signal length Pay attention to the length difference between the two signals. The smaller the length difference, the more similar the two signals are:

[0040] (1)

[0041] Where n and m represent the lengths of the time series signals X and Y, respectively.

[0042] s1.2, signal amplitude covers a lot of information such as signal energy, trend information and noise level. The similarity of the total signal amplitude Focus on the amplitude difference between the two signals. The smaller the total amplitude of the signals, the more similar the two signals are:

[0043] (2)

[0044] in, represents the i-th element of the time series signal X, Represents the jth element of the time series signal Y; Indicates the absolute value. Indicates finding the maximum value;

[0045] s1.3, signal dispersion represents the degree of change or volatility of the signal, and the similarity of signal dispersion The smaller it is, the more similar the two signals are:

[0046] (3)

[0047] in, and Represents the standard deviation of time series signals X and Y respectively.

[0048] s1.4, the maximum cross-correlation coefficient measures the similarity of two signals under a certain time delay. The similarity of the maximum cross-correlation coefficient is The larger it is, the more similar the shapes and changing trends of the two signals are at the same time:

[0049] (4)

[0050] in, , K is the number of time series signals from the same dataset as the time series signal Y, represents the maximum mutual correlation coefficient between the time series signal X and the kth time series signal in the data set, k=1,2,…,K, Indicates the maximum cross-correlation coefficient of time series signals X and Y.

[0051] s1.5. Similarity of signal shapes Focus on the difference between the shapes or waveforms of two signals:

[0052] First calculate the upper and lower envelopes of the time series signal X and , and the overlap length of time series signals X and Y Then intercept 、 and the first L elements of Y, to obtain a subsequence of length L 、 、 Using subsequences Falling within the envelope and The number of elements between To reflect the shape similarity between time series signals X and Y, and use the time series signal of the data set to which time series signal Y belongs to normalize, and obtain the signal shape similarity of time series signals X and Y :

[0053] (5)

[0054] in, , Indicates the shape similarity between the time series signal X and the kth time series signal in the dataset.

[0055] s1.6. Use dynamic time warping to measure signal distance between time series , starting from the first data point of the two time series, each alignment operation between data points is regarded as a decision stage, thus The calculation of distance value is converted into the calculation of subsequence distance value:

[0056] (6)

[0057] in, Represents the last element of the time series signal X, Y. Respectively represent the time series signals X and Y removed After that, the sequence of remaining elements.

[0058] After normalization, the similarity of the signal distance of time series signals X and Y is obtained :

[0059] (7)

[0060] in, , Represents the signal distance between the time series signal X and the kth time series signal in the dataset.

[0061] Step 2: For different data sets, the weights of the above N similarity feature indicators in measuring the similarity of time series signals need to be determined based on the data characteristics in the data set:

[0062] s2.1, for a data set , set the similarity interval threshold to , find the location away from the signal Recent signals, which together constitute the similar interval of the signal :

[0063] (8)

[0064] s2.2, Calculation Signal The similarity between each similarity feature indicator and other signals in the dataset S ~ :

[0065] (9)

[0066] s2.3, Since the positions of signals in similar intervals are close, the similarity of signals in similar intervals will be higher in theory. In order to avoid the interference of abnormal signals, the concept of high similarity interval is proposed. Similarity Sort from largest to smallest, take the first signals, forming signals High similarity interval .

[0067] Comparison signal Similar interval With high similarity interval Signal within, confirm signal Contribution value to the D-th similarity feature index:

[0068] (10)

[0069] (11)

[0070] s2.4. Determine the contribution of dataset S to the Dth similarity feature index:

[0071] (12)

[0072] After normalization, the percentage of the Dth similarity feature index is obtained :

[0073] (13)

[0074] Step 3: Based on the hierarchical analysis method, determine the weight coefficients of different similarity feature indicators on the data set S:

[0075] s3.1. Based on the calculation results of step 2, determine the percentage of the most important similarity feature index in the data set S , the percentage of similarity feature indicators with the least importance , calculate the difference between the two Then the difference Divided into 9 distribution intervals of importance, the width of each distribution interval is , forming a scaling method based on interval constraints. The importance scales corresponding to different distribution intervals are shown in Table 1:

[0076] Table 1

[0077]

[0078] s3.2, based on the scaling method given in Table 1, compare the importance of each similarity feature index and construct the judgment matrix M. Figure 2 The similarity feature index percentage results shown are as follows:

[0079]

[0080] s3.3. Calculate the maximum eigenvalue of the judgment matrix M And the order E, then calculate its consistency index CI and consistency ratio CR, and perform consistency test on the judgment matrix M:

[0081] (14)

[0082] (15)

[0083] Among them, RI is the average random consistency index, which is determined by the order of the judgment matrix M. If the consistency ratio , then the judgment matrix M passes the consistency test, otherwise it is necessary to adjust the size of the similarity interval and the high similarity interval in step 2 and reconstruct the judgment matrix M.

[0084] s3.4, according to the judgment matrix M, determine Figure 2 The weight distribution results of each similarity feature index under the similarity feature index percentage shown are shown in Table 2:

[0085] Table 2

[0086]

[0087] Step 4: When measuring the similarity between time series signals X and Y, first calculate the similarity between X and Y under multiple similarity feature indicators in step 1. Then, according to the methods of steps 2 and 3, calculate the weight values ​​of the data set where the time series signal Y is located for different similarity feature indicators Finally, the results of multiple similarity feature indicators are weighted and summed according to the weight value as the similarity measurement result of time series signals X and Y :

[0088] (16)

[0089] Example 2

[0090] This embodiment introduces a classification method based on the similarity measurement of time series signals based on multi-feature fusion. The known data set used in this embodiment is the active sonar target echo signal in the underwater acoustic field. Figure 3 As shown in Figure 2, the experimental scenario for acquiring measured data is as follows: the transmitted signal is a linear frequency modulation (LFM) signal, the transducer, hydrophone, and target are all located 7 meters underwater, the distance between the transducer and the target is 8 meters, the transmitted signal pulse width is 2 milliseconds, the signal frequency range is 40-80 kHz, and the sampling frequency is 1 MHz. The experimental equipment collects the active sonar target echo signal in the form of a time series signal.

[0091] The target is a diver that rotates slowly with the platform. The data is classified according to the target posture and formed into four data sets: DS1, DS2, DS3, and DS4. The original collected signals in the four data sets are as follows: Figures 4-7 As shown in the figure, after the original collected signal is subjected to data preprocessing operations such as filtering, noise reduction and segmentation, the signal processing results are as follows Figures 8-11 The detailed information in each dataset is shown in Table 3:

[0092] Table 3

[0093]

[0094] like Figure 3 As shown, first, the weight values ​​of the similarity feature index in the four data sets DS1, DS2, DS3, and DS4 are determined according to the method of steps 2 and 3 in Example 1. The calculation results are shown in Table 4:

[0095] Table 4

[0096]

[0097] The threshold values ​​involved in the weight calculation process are shown in Table 5:

[0098] Table 5

[0099]

[0100] Then get the length of the time series signal T to be classified , remove the known data sets whose length range is in the interval Signals other than =0.5, which is the set length threshold parameter. Further obtain the total amplitude of the time series signal T to be classified , remove the remaining signals with a total amplitude of Signals other than =0.5, is the total amplitude threshold parameter set, is the total signal amplitude of the time series signal T.

[0101] Finally, the similarity measurement result between the time series signal T and the remaining signals is calculated according to the method of step 4 in Example 1. The similarity measurement result is used as the distance between the time series signal T and the remaining signals. Based on the idea of ​​KNN, Z signals closest to the time series signal T are selected from the remaining signals, and the categories of these signals are counted. The category with the highest number of occurrences is used as the category of the time series signal T to complete the classification.

[0102] Multiple samples were randomly selected from the data sets DS1, DS2, DS3, and DS4 to test the accuracy of this classification method. The results are shown in Table 6:

[0103] Table 6

[0104]

[0105] As can be seen, the proposed method achieved good results in the classification test of the four types of measured data set sample signals. This reflects the high stability and robustness of the proposed method, which can maintain good separation between different categories and maintain high similarity consistency within the same category.

[0106] In addition, this embodiment also selected several traditional similarity measurement methods for comparative analysis, and the selected classifier was the KNN classifier, and the Z value was set to 1. The final classification accuracy and calculation time results are shown in Table 7:

[0107] Table 7

[0108]

[0109] From the results in Table 7, it can be seen that compared with other time series signal similarity measurement methods, this method has more advantages in terms of classification accuracy and calculation time.

[0110] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A time series signal similarity measurement method based on multi-feature fusion, characterized by: The specific steps include: Step 1: Select N similarity feature indicators to measure the similarity between different time series signals; Step 2: Set the similarity interval threshold for each signal S in the data set. k Construct a similar interval A k Calculate the similarity of each signal in the data set under different similarity feature indicators, sort them from large to small according to the similarity, take the first φ% signals, and construct signal S k High similarity interval HSI under the Dth similarity feature index D (S k ); compare the elements in the similar interval and the high similarity interval to determine the signal S k Contribution value to the D-th similarity feature index, D = 1, 2, ..., N: Among them, S l Represents similarity interval A k The lth signal within; Normalize the contribution value of each signal to obtain the percentage of the Dth similarity feature index in the data set Step 3: Based on the analytic hierarchy process, the difference R(S) between the maximum and minimum percentages of similarity feature indicators in a data set is divided into multiple distribution intervals of importance, forming a scaling method based on interval constraints. According to the importance between each similarity feature indicator, a judgment matrix M is constructed to distribute the weight distribution results of each similarity feature indicator in the data set; Step 4: When measuring the similarity between time series signals X and Y, first calculate the similarity Sim between X and Y under multiple similarity feature indicators in step 1. D (X, Y), and then follow the methods of steps 2 and 3 to calculate the weight value w of the data set where the time series signal Y is located for different similarity feature indicators D Finally, the results of multiple similarity feature indicators are weighted and summed according to the weight value, which is used as the similarity measurement result Sim of time series signals X and Y. total (X,Y):

2. The method for measuring similarity of time series signals based on multi-feature fusion according to claim 1, characterized in that: Signal length, signal total amplitude, signal dispersion, signal maximum cross-correlation coefficient, signal shape and signal distance are selected as similarity feature indicators.

3. The method for measuring similarity of time series signals based on multi-feature fusion according to claim 2, characterized in that: For a time series signal X with a length of n and a time series signal Y with a length of m, the calculation method of each similarity feature index is: s1.1, similarity of signal length Sim D＝1 (X,Y): s1.2, Similarity Sim of the total signal amplitude D＝2 (X,Y): Among them, x i Represents the i-th element of the time series signal X, y j Represents the j-th element of the time series signal Y; | | indicates finding the absolute value, and max() indicates finding the maximum value; s1.3, Similarity of signal dispersion Sim D＝3 (X,Y): Where σ(X) and σ(Y) represent the standard deviations of the time series signals X and Y, respectively; s1.4, Similarity of Maximum Correlation Coefficient Sim D＝4 (X,Y): Among them, S maxcorr =[r X1 ,r X2 ,…,r Xk ,…,r XK ], K is the number of time series signals from the same dataset as the time series signal Y, r Xk represents the maximum correlation coefficient between the time series signal X and the kth time series signal in the data set, k = 1, 2, ..., K, r XY Indicates the maximum cross-correlation coefficient of time series signals X and Y; s1.5, signal shape similarity Sim D＝5 (X,Y): Among them, S env =[Count(X,1),…,Count(X,k),…,Count(X,K)], where Count(X,k) represents the shape similarity between the time series signal X and the kth time series signal in the dataset; s1.6, Similarity of signal distance Sim D＝6 (X,Y): Among them, S dis =[DTW(X,1),…,DTW(X,k),…,DTW(X,K)], where DTW(X,k) represents the signal distance between the time series signal X and the kth time series signal in the data set.

4. The method for measuring similarity of time series signals based on multi-feature fusion according to claim 3, characterized in that: The calculation method of the signal shape similarity is: first calculate the upper and lower envelopes X of the time series signal X upper and X lower , and the overlap length L = max(m,n) of the time series signals X and Y; then intercept X upper 、X lower and the first L elements of Y, we get a subsequence X′ of length L upper , X′ lower , Y′; use the subsequence Y′ to fall on the envelope X upper and X lower The number of elements between Count(X,Y) reflects the shape similarity between the time series signals X and Y.

5. The method for measuring similarity of time series signals based on multi-feature fusion according to claim 3, characterized in that: Use the dynamic time warping method to measure the signal distance between time series DTW(X,Y): Among them, x n 、y m Represents the last element of the time series signal X, Y; X n-1 、Y m-1 Represents the time series signals X and Y without x n 、y m The sequence consisting of the remaining elements.

6. The method for measuring similarity of time series signals based on multi-feature fusion according to claim 1, characterized in that: Based on the interval constraint scaling method, the importance scales corresponding to different distribution intervals are shown in Table 1: Table 1 in, 7. The method for measuring similarity of time series signals based on multi-feature fusion according to claim 1, characterized in that: If the constructed judgment matrix M cannot pass the consistency test, adjust the size of the similarity interval and the high similarity interval in step 2 and reconstruct the judgment matrix M.

8. A classification method based on similarity measurement of time series signals using multi-feature fusion, characterized by: First, according to any of claims 1 to 7, the time series signal similarity measurement method based on multi-feature fusion is used to determine the weight values ​​of similarity feature indicators under different known data sets; Then, the length |T| of the time series signal T to be classified is obtained, and the signals whose length range is outside the interval [|T|(1-θ),|T|(1+θ)] in the known data set are eliminated, where θ is the set length threshold parameter; further, the total amplitude A(T) of the time series signal T to be classified is obtained, and the signals whose total amplitude is outside [A(T)(1-ψ),A(T)(1-ψ)] in the remaining signals are eliminated, where ψ is the set total amplitude threshold parameter, and A(T) is the total signal amplitude of the time series signal T; Finally, the similarity measurement result between the time series signal T and the remaining signals is calculated, and the similarity measurement result is used as the distance between the time series signal T and the remaining signals. Based on the idea of ​​KNN, Z signals closest to the time series signal T are selected from the remaining signals, and the categories of these signals are counted. The category with the highest number of occurrences is taken as the category of the time series signal T to complete the classification.

Images