Univariate time sequence enhancement method applying TEW-TCGAN

Abstract

The invention discloses a univariate time sequence enhancement method applying TEW-TCGAN. The method comprises the following steps: acquiring an original data set of a univariate time sequence; constructing a TEW-TCGAN enhancement model, wherein the TEW-TCGAN enhancement model comprises a generator and a discriminator structure; a TEW-TCGAN enhancement model generator and discriminator loss function is set; inputting the original time sequence data into a TEW-TCGAN enhancement model for training, and generating synthetic time sequence data; and evaluating the performance of the TEW-TCGAN enhancement model through the evaluation index. The method can generate synthetic time sequence data with higher distribution consistency with real samples, is remarkable in enhancement effect, is suitable for various single-variable time sequence enhancement tasks, and has a wide application prospect.

Description

Technical Field

[0001] The present invention relates to the fields of deep learning, artificial intelligence technology, and univariate time series enhancement, and in particular to a univariate time series enhancement method using TEW-TCGAN. Background Art

[0002] As a common and important data carrier and a collection of data points arranged in chronological order, time series is widely used in downstream tasks such as classification, interpolation, and prediction. In practical applications, time series data often faces problems such as insufficient sample size and unbalanced category distribution, which affect the model training effect. To alleviate the above problems, the time series data augmentation (TSDA) method is often used to expand the training data by generating representative synthetic samples, thereby improving the generalization ability and classification effect of the model. In the univariate TSDA task, the TCGAN model has the characteristics of high parallel computing efficiency and strong local feature extraction capabilities in the TSDA task. However, it still has problems such as difficulty in capturing long-term temporal dependency features, failure to dynamically model channel importance, and prone to gradient vanishing.

[0003] In the prior art:

[0004] 1) Patent application CN119066379A, "Time Series Generation Method, Device, and Storage Medium Based on the MDTimeGAN Model," proposes an MDTimeGAN model. This model adds data processing and a multi-discriminator module to the TimeGAN model. This model extracts multidimensional features from the raw data and discriminates each extracted feature separately, allowing the MDTimeGAN model to quickly generate high-quality time series data. This invention improves on the Time Series Convolutional Generative Adversarial Network (TCGAN), differing from the aforementioned invention's approach of using the TimeGAN model for improvement.

[0005] 2) The patent application "Industrial load data enhancement method based on cyclic convolutional generative adversarial network" with application publication number CN119598195A discloses a wind-solar-load coupling scenario generation method and system based on LSTM-GAN, which includes: obtaining provincial wind-solar-load data, which includes wind power output data, photovoltaic output data and load data; according to the provincial wind-solar-load data, using the generator in the improved generative adversarial network model to obtain the generated scenario; the improved generative adversarial network model is formed by introducing a fully connected layer and a long short-term memory network LSTM into the generator of the generative adversarial network model; Gaussian random noise and real scenes are input into the discriminator of the improved generative adversarial network model, and the generator and discriminator are cross-trained through adversarial training; the trained generator is extracted from the improved generative adversarial network model, and Gaussian random noise is input into the trained generator to generate random daily wind-solar-load scenarios. The generator of the present invention consists of a fully connected layer FC, a reshape layer Reshape, a one-dimensional transposed convolution layer Conv1DTranspose, a batch normalization layer BN, a ReLU activation function and an efficient channel attention ECA module, which is different from the method of introducing a fully connected layer and a long short-term memory network LSTM into the generator in the above invention.

[0006] 3) The patent application for “Ocean Data Augmentation Method Based on Global Attention Generative Adversarial Network” with application publication number CN118643867A discloses an ocean data augmentation method based on global attention generative adversarial network, which uses a new global attention mechanism to establish closer feature associations between different convolutional layers to improve the overall expression ability of the model; by analyzing the gradient vanishing and training instability in the generative adversarial network process, the network is optimized and improved in a targeted manner; the network structure and loss function are improved to solve the training instability problem, and residual modules are added to the generator and discriminator to avoid gradient vanishing caused by increasing the network depth. The TEW-TCGAN enhanced model disclosed in the present invention embeds an ECA module in the generator, introduces a TCN structure in the discriminator, and applies the Wasserstein distance in the objective function, which is different from the method of using a global attention mechanism to optimize the generator and discriminator in the above invention. Summary of the Invention

[0007] To solve the above technical problems, the purpose of the present invention is to provide a univariate time series enhancement method using TEW-TCGAN (Temporal Convolution Network, Efficient Channel Attention module, and Wasserstein distance-based Time-series Convolutional Generative Adversarial Network, based on time series convolutional generative adversarial network with temporal convolutional network, ECA module, and Wasserstein distance). Based on the TCN structure, ECA module, and Wasserstein distance, this method optimizes TCGAN into a TEW-TCGAN model with outstanding feature expression ability, strong time series modeling ability, and high training process stability, effectively improving the quality and distribution consistency of generated synthetic time series data, in the field of univariate time series enhancement methods.

[0008] The purpose of the present invention is achieved through the following technical solutions:

[0009] A univariate time series enhancement method using TEW-TCGAN, comprising:

[0010] A. Obtain the original univariate time series dataset;

[0011] B. Build the TEW-TCGAN enhanced model, including the generator and discriminator structures;

[0012] C. Set the TEW-TCGAN enhanced model generator and discriminator loss functions;

[0013] D. Input the original time series data into the TEW-TCGAN enhanced model for training to generate synthetic time series data;

[0014] E. Evaluate the performance of the TEW-TCGAN enhanced model through evaluation metrics.

[0015] Compared with the prior art, one or more embodiments of the present invention may have the following advantages:

[0016] By embedding an efficient channel attention module ECA in the generator, the modeling ability of inter-channel dependencies can be enhanced and the feature expression quality of generated samples can be improved; combined with the introduction of the temporal convolutional network TCN structure in the discriminator, the model's ability to model long-range dependent features in time series can be enhanced, and the ability to distinguish the differences between generated samples and real samples can be improved; by adopting an adversarial training strategy that combines Wasserstein distance and gradient penalty, the stability and convergence efficiency of the model training process are significantly improved. The constructed TEW-TCGAN model can generate synthetic time series data with higher consistency with the real sample distribution, and the enhancement effect is significant; the method of the present invention can be applied to various univariate time series enhancement tasks and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of the univariate time series enhancement method using TEW-TCGAN;

[0018] Figure 2 This is the structure diagram of the TEW-TCGAN model;

[0019] Figure 3 This is the structure diagram of efficient channel attention ECA;

[0020] Figure 4 It is the structure diagram of the temporal convolutional network residual block TCN Resblock;

[0021] Figure 5 This is the experimental result of the TEW-TCGAN model on the UCR dataset. DETAILED DESCRIPTION

[0022] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be described in further detail below with reference to embodiments and accompanying drawings.

[0023] like Figure 1 As shown in Figure 2, the univariate time series enhancement method process using TEW-TCGAN includes the following steps:

[0024] Step 10: Obtain the original univariate time series data set;

[0025] Step 20: Build the TEW-TCGAN enhanced model, including the generator and discriminator structures;

[0026] Step 30 sets the TEW-TCGAN enhanced model generator and discriminator loss functions;

[0027] Step 40: Input the original time series data into the TEW-TCGAN enhanced model for training to generate synthetic time series data;

[0028] Step 50 evaluates the performance of the TEW-TCGAN enhanced model through evaluation indicators.

[0029] The above step 10 specifically includes: obtaining a univariate time series original data set, selecting 6 UCR univariate time series original data sets with smaller sample sizes, specifically including Beef, CBF, DiatomSizeR, ECGFiveDays, FaceFour, and GunPoint, and dividing the data sets into a training set and a test set.

[0030] like Figure 2 As shown, the above step 20 specifically includes: constructing a TEW-TCGAN enhanced model, including a generator and a discriminator structure, wherein the generator G TEW The structure consists of a fully connected layer FC, a shape reshaping layer Reshape, a one-dimensional transposed convolution layer Conv1DTranspose, a batch normalization layer BN, a ReLU activation function, and an efficient channel attention ECA module, which is used to generate synthetic time series data x from the input random noise z g .

[0031] like Figure 3 As shown, the above step 20 specifically includes: the ECA module inputs the feature (C e , L e The global information is extracted by global average pooling GAP, and then k e The size 1D convolution obtains the channel importance weight, and Sigmiod maps the channel importance weight to [0,1] to form the channel attention weight. Multiply the output element-wise with the channel attention weight Complete the pair The weight of each channel is adjusted to highlight key features and suppress unimportant features. The 1D convolution kernel size is k e According to the following formula, the number of channels C e , hyperparameters γ and β are determined adaptively.

[0032]

[0033] The above step 20 specifically includes: the discriminator D TEW The structure consists of a temporal convolutional network residual block TCN Resblock, a flattening layer Flatten, and a fully connected layer FC, which are used to distinguish the original time series data x from the synthetic time series data x g (like Figure 2 As shown); the TCN Resblock is D TEWThe core component consists of a dilated causal convolution layer, a batch normalization layer (BN), a ReLU activation function, and a Dropout layer (such as Figure 4 shown).

[0034] The above step 30 specifically includes: setting the TEW-TCGAN enhanced model generator and discriminator loss function, the loss function is constructed based on Wasserstein distance and combined with gradient penalty strategy, the generator loss function Discriminator loss function Expressed as:

[0035]

[0036] Where the synthetic time series data is G TEW (z), the discriminator output is D TEW (·), from the noise distribution p z The sampling is z~p z (z), sampled from the original time series data distribution as x~p data (x), the mathematical expectation operator is E[·], and the gradient penalty term is The linear interpolation points of the original and synthetic time series data are The gradient penalty coefficient is λ GP , the L2 norm is ||·||2, the discriminator is The gradient is

[0037] The above step 40 specifically includes: inputting the original time series data into the TEW-TCGAN enhanced model for training to generate synthetic time series data, wherein the training process adopts an adversarial optimization strategy, D TEW After training n times, G TEW Update once and use Adam optimizer to update D TEW , G TEW Parameters; the synthetic time series data x g By G TEW Receive random noise z to generate a structure similar to the original time series data; use the Tensorflow network framework equipped with NVIDIA GeForce RTX 3090GPU, and the specific parameter settings include: initial learning rate α Ad =1.0×10 -4 , first-order momentum coefficient β1 Ad =0, second-order momentum coefficient β2 Ad =0.9, gradient penalty coefficient λ GP =10, D TEW With G TEW Iteration ratio n critic=5 Set batch size Batchsize = 16 and training epoch number Epoch = 300.

[0038] The above step 50 specifically includes: evaluating the performance of the TEW-TCGAN enhanced model through evaluation indicators, wherein the evaluation indicators are the maximum mean difference (MMD) and the nearest neighbor distance (NND). Figure 5 The experimental results of this method on the UCR dataset are shown in Figure 2. The calculation method of MMD and NND is expressed as:

[0039]

[0040] In the formula, the number of original samples and synthetic samples is N, and the real sample set is Generate a sample set of The Gaussian radial basis kernel function with L2 norm is K(·,·), and the L2 norm is ||·||.

[0041] Although the embodiments disclosed herein are as described above, the contents described herein are merely embodiments for facilitating understanding of the present invention and are not intended to limit the present invention. Any person skilled in the art may make any modifications and variations in the form and details of the embodiments without departing from the spirit and scope of the present invention. However, the scope of patent protection of the present invention shall remain subject to the scope defined by the appended claims.

Claims

1. A univariate time series enhancement method using TEW-TCGAN, characterized in that: The method comprises: Step A obtains the original univariate time series data set; Step B builds the TEW-TCGAN enhanced model, including the generator and discriminator structures; Step C sets the loss function of the generator and discriminator of the TEW-TCGAN enhanced model; In step D, the original time series data is input into the TEW-TCGAN enhanced model for training to generate synthetic time series data; Step E evaluates the performance of the TEW-TCGAN enhanced model through evaluation indicators.

2. The univariate time series enhancement method using TEW-TCGAN according to claim 1, characterized in that In the step A, a univariate time series original data set is obtained, and 6 UCR univariate time series original data sets with smaller sample sizes are selected. The data sets are divided into a training set and a test set.

3. The univariate time series enhancement method using TEW-TCGAN according to claim 1, characterized in that In step B, the generator G TEW The structure consists of a fully connected layer FC, a shape reshaping layer Reshape, a one-dimensional transposed convolution layer Conv1DTranspose, a batch normalization layer BN, a ReLU activation function, and an efficient channel attention ECA module, which is used to generate synthetic time series data x from the input random noise z g .

4. The univariate time series enhancement method using TEW-TCGAN according to claim 3, characterized in that: The ECA module input features The global information is extracted by global average pooling GAP, and then k e The channel importance weights are obtained by convolution of size 1D, where C e 、L e are the number of channels and length respectively; Sigmiod maps the channel importance weight to [0,1] to form the channel attention weight, Multiply the output element-wise with the channel attention weight Complete the pair The weight of each channel is adjusted, and the 1D convolution kernel size is k e According to the following formula, the number of channels C e , hyperparameters γ and β are adaptively determined:

5. The univariate time series enhancement method using TEW-TCGAN according to claim 1, characterized in that: In step B, the discriminator D TEW The structure consists of a temporal convolutional network residual block TCN Resblock, a flattening layer Flatten, and a fully connected layer FC, which are used to distinguish the original time series data x from the synthetic time series data x g ; The temporal convolutional network residual block TCNResblock consists of a dilated causal convolution layer, a batch normalization layer BN, a ReLU activation function, and a Dropout layer, which is D TEW The core components of 6. The univariate time series enhancement method using TEW-TCGAN according to claim 1, characterized in that: In step C, the generator and discriminator loss functions are constructed based on the Wasserstein distance and combined with the gradient penalty strategy. The generator loss function is The discriminator loss function is Respectively expressed as: Where G TEW (z) is the synthetic time series data, D TEW (·) is the discriminator output, z~p z (z) is the noise distribution p z Sampling in, x~p data (x) is the distribution of the original time series data p data The sampling in , E[·] is the mathematical expectation operator, is the gradient penalty term, is the linear interpolation point of the original and synthetic time series data, λ GP is the gradient penalty coefficient, ||·||2 is the L2 norm, For the discriminator pair gradient.

7. The univariate time series enhancement method using TEW-TCGAN according to claim 1, characterized in that: In step D, the original time series data is input into the training process of the TEW-TCGAN enhanced model using an adversarial optimization strategy. TEW After training n times, G TEW Update once and use Adam optimizer to update D TEW , G TEW Parameters; the synthetic time series data x g By G TEW The received random noise z is generated and has similar structural characteristics to the original time series data.

8. The univariate time series enhancement method using TEW-TCGAN according to claim 1, characterized in that: In step E, the evaluation indicators are the maximum mean difference MMD and the nearest neighbor distance NND, wherein the calculation method of the maximum mean difference MMD and the nearest neighbor distance NND is expressed as: Where N is the number of original samples and synthetic samples, is the real sample set, To generate the sample set, K(·,·) is the Gaussian radial basis kernel function with L2 norm, and ||·|| is the L2 norm.

Images