Time sequence classification method based on improved discrete wavelet transform and trend similarity measurement

Abstract

The invention discloses a time sequence classification method based on improved discrete wavelet transform and trend similarity measurement. The method comprises the following steps: standardizing the length of a time sequence; time sequence data dimension reduction and feature extraction; normalizing the trend characteristics; the method is constructed based on a classification method of trend similarity measurement, under the condition that data and labels in a training set are known, test data are input, trend features of the test data are compared with corresponding trend features in the training set, the first K trend features which are most similar to the trend features in the training set are found out, and the test data are obtained. The category corresponding to the test data is the category with the highest occurrence frequency in the K trend characteristics, and the trend similarity measurement is to perform similarity matching on the trend characteristics of the two time sequences from the relative deviation of the fluctuation degree and the consistency of the fluctuation direction. According to the method, the accuracy and efficiency of high-dimension and strong-noise time sequence classification can be effectively improved, so that the frontier and accuracy of the time sequence classification method are ensured.

Description

Technical Field

[0001] The invention relates to the fields of time series classification and artificial intelligence technology, and in particular to a time series classification method based on improved discrete wavelet transform and trend similarity measurement. Background Art

[0002] In the field of artificial intelligence technology, time series classification technology is the core technology for processing time series data. Its performance has an important impact on the efficient operation and high-precision guarantee of tasks such as pattern recognition, trend prediction, automation and control. Time series data is a sequence of values ​​arranged in chronological order, which is widely used in various industries, such as the Internet of Things, environment, medical care and social media. However, with the advancement of big data technology, time series data has highlighted characteristics such as high dimensionality and strong noise, and is affected by many complex factors, including but not limited to multi-sensor deployment, multi-modal data collection, high-frequency data collection technology, multi-level feature extraction, sensor data collection errors, external random events, etc. These factors will increase the dimensionality and noise of the data, thereby reducing the quality of the results of data mining tasks such as classification.

[0003] Traditionally, the classification of time series data mainly relies on point-to-point metric distance and dynamic time warping methods. Although these technologies are widely used, they often show problems such as data dimensionality disaster, increased model complexity, low classification accuracy, poor ability to capture time-varying noise, poor robustness and high computational cost when faced with high-dimensional, strong noisy and complex nonlinear time series data.

[0004] In recent years, with the rapid development of dimensionality reduction, time-varying noise processing technology, and similarity matching technology, it has been widely used in various fields with its high efficiency and high precision, and has provided new ideas for the classification of high-dimensional and strong noise time series. In recent years, some patents that have been published have been used to improve the classification effect of time series, such as China Patent (publication number CN118656774A) a time series classification method based on improved genetic algorithm and Fourier transform, China Patent (publication number CN118606844A) a time series classification method based on frequency domain feature extraction, China Patent (publication number CN118427708A) a classification method for time series data, a training method and device for classification model, and China Patent (publication number CN118820905A) a time series classification method, device, equipment and storage medium, compared with traditional methods, can more effectively improve the accuracy and efficiency of time series classification. Based on the above patents, it can be seen that time series classification can be effectively performed based on wavelet transform and time-frequency method. Therefore, it is of great significance to study time series classification based on wavelet transform, but there are still certain limitations. It is still unable to capture the time-varying noise of time series well, and it is unable to accurately measure the similarity of the intrinsic change patterns of time series. At the same time, the high dimensionality of the data brings computational costs to the model, resulting in low accuracy and high cost of mining tasks.

[0005] Therefore, how to provide a time series classification method that combines dimensionality reduction and denoising with trend similarity measurement for high-dimensional, strong-noise time series data has become a technical problem that technicians in this field urgently need to solve. Summary of the invention

[0006] In view of this, the purpose of the present invention is to provide a time series classification method based on improved discrete wavelet transform and trend similarity measurement, which can effectively improve the classification accuracy and efficiency of high-dimensional and strong noise time series, thereby ensuring the cutting-edge and accuracy of the time series classification method and the reliability of related mining task results.

[0007] The present invention solves the technical problem by adopting the following technical solution:

[0008] A time series classification method based on improved discrete wavelet transform and trend similarity measurement includes the following steps:

[0009] S1, time series length normalization;

[0010] S2, dimensionality reduction and feature extraction of time series data, an improved time series dimensionality reduction and denoising method is proposed based on discrete wavelet transform. This method is used to simultaneously achieve dimensionality reduction and denoising of time series data, and finally obtain the trend characteristics of the time series;

[0011] S3, normalization processing of trend characteristics, normalization processing of trend characteristics, and unification of the value range of trend characteristics into the interval [1, 2] through normalization processing;

[0012] S4, a classification method based on trend similarity measurement is constructed. When the data and labels in the training set are known, the test data is input, and the trend characteristics of the test data are compared with the corresponding trend characteristics in the training set. The top K trend characteristics in the training set that are most similar to them are found. The category corresponding to the test data is the category that appears most frequently among the K trend characteristics. Among them, the trend similarity measurement is to perform similarity matching on the trend characteristics of the two time series based on the relative deviation of the fluctuation degree and the consistency of the fluctuation direction.

[0013] Furthermore, in step S1, the method for normalizing the length of the time series is as follows:

[0014] For time series, if the length of the time series is a multiple of 2, it is retained directly; if not, it is processed by shrinking, that is, deleting unstable data or irrelevant data at the beginning of the sequence to make its length a multiple of 2.

[0015] Furthermore, in step S2, the specific method for obtaining the trend characteristics of the time series is as follows: first, the time series is decomposed by discrete wavelet transform, and then the wavelet decomposition coefficients are filtered using the threshold function of the improved compromise method of the following formula, and finally, the high-frequency information output in the first step is not considered when reconstructing the wavelet coefficients, so as to obtain the trend characteristics of the time series;

[0016]

[0017] Among them, λ is the threshold of the j-layer decomposition scale; σ is the noise intensity, j is the decomposition scale, and N is the signal length.

[0018] Furthermore, in step S3, the following formula is used for normalization processing:

[0019] in,

[0020] x i and x i ' is the time series t i The original data value and the normalized data value corresponding to the moment; x_min is the minimum value in the time series, and x_max is the maximum value in the time series.

[0021] Furthermore, in step S4, the classification method based on trend similarity measurement is constructed as follows:

[0022] Step (1): Calculate the degree of dissimilarity between the morphological trends and change patterns of the test data and each training data using the trend similarity measurement method;

[0023] Step (2): sorting according to the increasing relationship of dissimilarity degree;

[0024] Step (3): Select K trend features with the smallest degree of dissimilarity;

[0025] Step (4): Determine the frequency of occurrence of the categories of the top K trend features;

[0026] Step (5): Return the category with the highest frequency among the first K trend features as the predicted category of the test data.

[0027] The present invention discloses a time series classification method based on improved discrete wavelet transform and trend similarity measurement, which has the following beneficial effects:

[0028] (1) Improving classification accuracy: By performing length standardization, dimensionality reduction and denoising, trend feature extraction, and normalization processing on time series data, and using the K-nearest neighbor classification method based on the trend similarity measurement method for training and prediction, the prediction accuracy of the time series classification method of the present invention reaches 89.3%. Compared with the K-nearest neighbor classification method based on the original time series data and the K-nearest neighbor classification based on the Euclidean distance, the accuracy and efficiency of time series classification are effectively improved.

[0029] (2) Improve real-time response capabilities: Process more time series data in a shorter time, ensure that the system responds quickly and takes timely actions in a rapidly changing environment, and reduce system maintenance costs.

[0030] (3) Improve system stability: High-accuracy and fast time series classification helps improve the overall stability of the monitoring system and enhance the reliability of the system.

[0031] (4) Enhanced decision support capabilities: With high-accuracy classification and rapid response capabilities, the monitoring system can detect problems in real time and provide corresponding decision support based on the classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 is a flow chart of the method of the present invention;

[0033] Figure 2 A comparison chart of the original data and trend characteristics of a time series of the present invention;

[0034] Figure 3 It is a schematic diagram of the decomposition process of the three-layer local high-frequency discrete wavelet transform of the time series of the present invention;

[0035] Figure 4 This is a comparison chart of the original time series and trend characteristics of the sensor in Example 1;

[0036] Figure 5 This is a comparison chart of the original time series and trend characteristics of the movement in Example 2. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] The present invention provides a time series classification method based on improved discrete wavelet transform and trend similarity measurement. The method determines the category of the current time series data change pattern and trend by using the improved discrete wavelet transform method and trend similarity measurement to reduce the dimension, denoise, extract trend features and match trend pattern similarity of time series data. This method can effectively improve the classification accuracy and efficiency of high-dimensional and strong noise time series, thereby ensuring the cutting-edge and accuracy of the time series classification method and the reliability of the results of related mining tasks.

[0039] refer to Figure 1 As shown, the present invention discloses a time series classification method based on improved discrete wavelet transform and trend similarity measurement, comprising the following steps:

[0040] S1, time series length normalization;

[0041] For time series, if the length of the time series is a multiple of 2, it is directly retained; if not, it is processed by shrinking, that is, deleting unstable data or irrelevant data at the beginning of the sequence, so that its length becomes a multiple of 2. This can meet the calculation requirements of subsequent algorithms for data.

[0042] S2, time series data dimensionality reduction and feature extraction, based on discrete wavelet transform, an improved time series dimensionality reduction and denoising method is proposed, namely the local high-frequency discrete wavelet transform method, which is used to simultaneously achieve dimensionality reduction and denoising of time series data, and finally obtain the trend characteristics of the time series; trend characteristics refer to the fluctuation patterns that change over time after dimensionality reduction and denoising of the time series. For example, the original data and trend characteristics of a time series are compared. Figure 2 shown.

[0043] In the first step of the discrete wavelet transform decomposition, the high-frequency information output contains most of the noise of the original time series signal, and the wavelet coefficients become almost zero after threshold filtering. Based on this feature of the discrete wavelet transform, a local high-frequency discrete wavelet transform method is proposed. The basic principle is: in step S2, the specific method for obtaining the trend characteristics of the time series is as follows: first, the time series is decomposed by discrete wavelet transform, and then the wavelet decomposition coefficients are filtered using the threshold function of the improved compromise method of the following formula, and finally, the high-frequency information output in the first step is not considered when reconstructing the wavelet coefficients, so as to obtain the trend characteristics of the time series; the specific process is as follows Figure 3 shown.

[0044]

[0045] Where λ is the threshold of the j-layer decomposition scale; σ is the noise intensity, and the value of the present invention is σ=mean(W j,k ) / 0.6745, j is the decomposition scale, and N is the signal length. The calculation basis of the threshold function is: first, the signal and noise can be separated, and the noise is usually manifested as the high-frequency part of the wavelet coefficients; second, the threshold is related to the noise intensity and the signal length; third, the noise intensity is estimated by the median of the high-frequency wavelet coefficients; fourth, the improved compromise method (i.e., combining soft threshold and hard threshold) achieves a better balance between noise removal and signal retention, avoiding the loss of useful signal information due to excessive denoising.

[0046] The local high-frequency discrete wavelet transform method can simultaneously achieve dimensionality reduction and denoising of time series data, which can provide a data basis for subsequent classification and reduce computational costs.

[0047] S3, normalization processing of trend characteristics, normalization processing of trend characteristics, through normalization processing, the value range of trend characteristics is unified into the interval [1, 2]; the calculation process is shown in the following formula:

[0048] in,

[0049] x i and x i ' is the time series t i The original data value and the normalized data value corresponding to the moment; x_min is the minimum value in the time series, and x_max is the maximum value in the time series.

[0050] S4, Construction of classification method based on trend similarity measurement:

[0051] The K-nearest neighbor classification method usually uses a metric distance, such as Euclidean distance or Manhattan distance, to calculate the distance between objects as a non-similarity index between each object. Although the metric distance realizes the classification through the non-similarity between data points, it ignores the similarity matching of the change pattern between time series data. Therefore, the present invention proposes an improved trend similarity measurement method, and uses the similarity measurement method as the measurement function of the K-nearest neighbor classification method.

[0052] The basic idea of ​​the classification method based on trend similarity measurement is to input test data when the data and labels in the training set are known, compare the trend features of the test data with the corresponding trend features in the training set, and find the top K trend features in the training set that are most similar to them. Then the category corresponding to the test data is the category that appears most frequently among the K trend features, as follows Figure 1 As shown, the steps of the algorithm are as follows:

[0053] Step 1: Use the trend similarity measurement method to calculate the degree of dissimilarity between the morphological trends and change patterns of the test data and each training data;

[0054] Among them, the trend similarity measurement method is: to perform similarity matching on the trend characteristics of two time series based on the relative deviation of the fluctuation degree and the consistency of the fluctuation direction.

[0055] For any two trend features X of length n t and Y t The calculation process of the non-similarity and trend similarity measurement method of (t=1,2,…,n) is as follows:

[0056]

[0057] (1) Represents X t and Y t First, for X t and Y t For example, the relative offset at time t is expressed as like If the value is <1, the relative deviation at time t is expressed as The larger the ratio is, the larger the relative deviation is at time t. Then, the average value of the relative deviations at the obtained n times is calculated.

[0058] (2) The fluctuation amplitude sequence X representing the trend characteristics t ' and Y t ' Relative deviation. X t and Y t The fluctuations at time t are expressed as Indicates that the corresponding fluctuation amplitude sequence X can be obtained by sequential calculation t ' and Y t '. Use the Euclidean distance ED between the wave sequences to represent X t ' and Y t ' Relative deviation. By taking the logarithm to convert the product form into the sum form, the mathematical calculation is simplified, and the influence of the value of the trend feature on the similarity measurement is reduced.

[0059] (3)r(X t ,Y t ) represents X t and Y t The consistency of the fluctuation trend of is in the range of [-1,1]. The rank correlation coefficient is calculated to show the consistency of the two trend characteristics in terms of fluctuation direction and degree. The calculation process is shown in the formula below.

[0060]

[0061] (4) Finally, add the relative average deviation and the relative deviation of the amplitude, and subtract the consistency in the fluctuation direction to obtain X t and Y t The degree of MFCDD non-similarity in the patterns is then calculated by performing an antilogarithmic calculation.

[0062] Step 2: Sort by increasing degree of dissimilarity;

[0063] Step 3: Select K trend features with the smallest degree of dissimilarity;

[0064] Step 4: Determine the frequency of occurrence of the categories of the top K trend features;

[0065] Step 5: Return the category with the highest frequency among the top K trend features as the predicted category of the test data.

[0066] The category of the test data is predicted based on the first K trend features with the smallest degree of dissimilarity. The category with the highest frequency among the first K trend features is used as the category of the test data.

[0067] The selection of the prediction category is based on the majority voting principle. Specifically, the KNN algorithm finds the K neighbors closest to the target point and then determines the prediction category of the target point based on the categories to which these neighbors belong.

[0068] Here, the class with the highest frequency is selected as the predicted class for the following reasons:

[0069] 1. Majority voting principle

[0070] Assume that the class distribution is uniform: if there are more samples of a certain class in the training data, then the neighboring points of this class will naturally account for the majority when predicting. Therefore, this class is more likely to be predicted as the class of the target point.

[0071] Smoothness of category division: KNN assumes that similar samples in the feature space tend to belong to the same category. Therefore, if most of the neighboring K samples belong to the same category, this category can be considered as the representative category of the area, thereby predicting that the target point belongs to this category.

[0072] 2. Category Frequency and Similarity Hypothesis

[0073] In KNN, the category of each neighbor represents the "local" structure of the point in the feature space. By selecting the K closest neighbors, we can use their category information to make predictions. For points with high similarity (i.e., points with low dissimilarity), we believe that they are similar to some extent, so their categories should also be similar. This means that in a local area, the category with the highest frequency is more likely to be the actual category of the target point.

[0074] Categories with high frequency are more representative: When a category appears the most times among the K neighbors, we believe that the distribution of this category in this local area is more representative, and thus reasonably infer that the target point should belong to this category.

[0075] The classification method based on trend similarity measurement can better calculate the similarity of trend patterns of time series, thereby improving the classification accuracy of time series.

[0076] The present invention can improve classification accuracy: by standardizing the length of time series data, reducing dimensionality for denoising, extracting trend features, and normalizing the data, and using the K-nearest neighbor classification method based on the trend similarity measurement method for training and prediction, the prediction accuracy of the time series classification method of the present invention reaches 89.3%. Compared with the K-nearest neighbor classification method based on the original time series data and the K-nearest neighbor classification based on the Euclidean distance, the accuracy and efficiency of time series classification are effectively improved. Improve real-time response capabilities: process more time series data in a shorter time, ensure that the system responds quickly and takes timely actions in a rapidly changing environment, and reduce system maintenance costs. Improve system stability: high-accuracy and fast time series classification helps to improve the overall stability of the monitoring system and can enhance the reliability of the system. Enhance decision-making support capabilities: with high-accuracy classification and fast response capabilities, the monitoring system can discover problems in real time and provide corresponding decision support based on the classification results.

[0077] Embodiment 1:

[0078] Classification of sensor time series data

[0079] 1. Data preparation

[0080] Collecting data sets: We use the public sensor time series data set MoteStrain, which has been divided into training data sets and experimental data sets. The classes of each time series are specified in both subsets. A sensor sequence refers to a series of data generated by a sensor and arranged in time or space order. The MoteStrain data set has 2 classes, the length of each time series is 84, and the number of time series in the training data set and experimental data set is 20 and 1252 respectively.

[0081] Length normalization: There is no need to shrink the length of the time series in the MoteStrain dataset.

[0082] 2. Time series dimensionality reduction, denoising and trend feature extraction

[0083] The training data set and the experimental data set are decomposed according to the local high-frequency discrete wavelet transform method, and the Daubechies6 wavelet basis function is used for two-layer decomposition to obtain the trend characteristics. Figure 4 As shown in the figure, the dimension of the trend feature is reduced from 84 to 48, and the noise is also reasonably eliminated.

[0084] 3. Standardization of trend characteristics

[0085] The trend features are normalized, and the value range of the trend features is unified into the interval [1, 2] through normalization.

[0086] 4. Construction of classification method based on trend similarity measurement

[0087] A K-nearest neighbor classification model based on the trend similarity measurement method is constructed, with K=1, and the MoteStrain training data set is trained.

[0088] 5. Time Series Classification Model Evaluation

[0089] The model was evaluated using the test set, and the classification accuracy CA performance index and running time T were calculated. The classification accuracy refers to the ratio of the number of experimental data set sequences with correct category recognition to the total number of data set sequences; the running time measures the efficiency of time series classification. The classification accuracy of the invention reached 89.3% and the running time was 31.03s.

[0090] Embodiment 2:

[0091] Classification of motion time series data

[0092] 1. Data preparation

[0093] Collecting datasets: We use the public motion time series dataset GunPoint, which has been divided into training datasets and experimental datasets. The classes of each time series are specified in both subsets. The GunPoint dataset has 2 classes, the length of each time series is 150, and the number of time series in the training dataset and experimental dataset is 50 and 150 respectively.

[0094] Motion sequences refer to continuous data sequences that describe the motion changes of an object or individual over a certain period of time. These sequences are usually used to represent motion trajectories, posture changes, speed changes, etc. Motion sequences are widely used to describe the actions and behaviors of human bodies, objects, or robots.

[0095] Length normalization: There is no need to shrink the length of the time series of the GunPoint dataset.

[0096] 2. Time series dimensionality reduction, denoising and trend feature extraction

[0097] The training data set and the experimental data set are decomposed according to the local high-frequency discrete wavelet transform method, and the Daubechies6 wavelet basis function is used for three-layer decomposition to obtain the trend characteristics. Figure 5 As shown in the figure, the dimension of the trend feature is reduced from 150 to 80, and the noise is also reasonably eliminated.

[0098] 3. Standardization of trend characteristics

[0099] The trend features are normalized, and the value range of the trend features is unified into the interval [1, 2] through normalization.

[0100] 4. Construction of classification method based on trend similarity measurement

[0101] A K-nearest neighbor classification model based on the trend similarity measurement method is constructed, with K=1, and the GunPoint training dataset is trained.

[0102] 5. Time Series Classification Model Evaluation

[0103] The model was evaluated using the test set, and the classification accuracy CA performance index and running time T were calculated. The classification accuracy refers to the ratio of the number of experimental data set sequences with correct category recognition to the total number of data set sequences; the running time measures the efficiency of time series classification. The classification accuracy of the invention reached 94.67% and the running time was 20.89s.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A time series classification method based on improved discrete wavelet transform and trend similarity measurement, characterized in that: The steps include: S1, time series length normalization; S2, dimensionality reduction and feature extraction of time series data, an improved time series dimensionality reduction and denoising method is proposed based on discrete wavelet transform. This method is used to simultaneously achieve dimensionality reduction and denoising of time series data, and finally obtain the trend characteristics of the time series; S3, normalization processing of trend characteristics, normalization processing of trend characteristics, and unification of the value range of trend characteristics into the interval [1, 2] through normalization processing; S4, a classification method based on trend similarity measurement is constructed. When the data and labels in the training set are known, the test data is input, and the trend characteristics of the test data are compared with the corresponding trend characteristics in the training set. The top K trend characteristics in the training set that are most similar to them are found. The category corresponding to the test data is the category that appears most frequently among the K trend characteristics. Among them, the trend similarity measurement is to perform similarity matching on the trend characteristics of the two time series based on the relative deviation of the fluctuation degree and the consistency of the fluctuation direction.

2. The time series classification method based on improved discrete wavelet transform and trend similarity measurement according to claim 1 is characterized in that: In step S1, the method for normalizing the length of the time series is as follows: For time series, if the length of the time series is a multiple of 2, it is retained directly; if not, it is processed by shrinking, that is, deleting unstable data or irrelevant data at the beginning of the sequence to make its length a multiple of 2.

3. A time series classification method based on improved discrete wavelet transform and trend similarity measurement according to claim 1 or 2, characterized in that: In step S2, the specific method for obtaining the trend characteristics of the time series is as follows: first, the time series is decomposed by discrete wavelet transform, and then the wavelet decomposition coefficients are filtered using the threshold function of the improved compromise method of the following formula, and finally, the high-frequency information output in the first step is not considered when reconstructing the wavelet coefficients, so as to obtain the trend characteristics of the time series; Among them, λ is the threshold of the j-layer decomposition scale; σ is the noise intensity, j is the decomposition scale, and N is the signal length.

4. The time series classification method based on improved discrete wavelet transform and trend similarity measurement according to claim 3 is characterized in that: In step S3, the following formula is used for normalization processing: in, x i and x i ' is the time series t i The original data value and the normalized data value corresponding to the moment; x_min is the minimum value in the time series, and x_max is the maximum value in the time series.

5. The time series classification method based on improved discrete wavelet transform and trend similarity measurement according to claim 4 is characterized in that: In step S4, the classification method based on trend similarity measurement is constructed as follows: Step (1): Calculate the degree of dissimilarity between the morphological trends and change patterns of the test data and each training data using the trend similarity measurement method; Step (2): sorting according to the increasing relationship of dissimilarity degree; Step (3): Select K trend features with the smallest degree of dissimilarity; Step (4): Determine the frequency of occurrence of the categories of the top K trend features; Step (5): Return the category with the highest frequency among the first K trend features as the predicted category of the test data.

Images