Time sequence data generation method based on space-time diffusion model and application

Abstract

The invention discloses a time sequence data generation method based on a space-time diffusion model and application. According to the method, firstly, a U-shaped network (SE-U-Net) with a compression-and-Excitation (SE) module is provided, the SE module is added on the basis of an original U-shaped network to enhance the spatial feature extraction capability of the U-shaped network, and the U-shaped network is a trunk model for noise prediction. Besides, a time sequence embedding module is designed, when a new sample block is generated, the position and trend information of a previous time sequence is embedded in a decreasing mode, and the time sequence relation between the generated sample blocks is guaranteed. According to the method, testing is carried out in the blast furnace process and the TE process, high-quality time sequence data can be generated, and the time sequence and space relation of a real sequence can be effectively restored. Finally, the invention also provides an application method of the data generation method in a fault diagnosis task.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of time series data generation, and particularly relates to a time series fault diagnosis data generation method based on a Temporal-Spatial Diffusion (TS-Diffusion) model, which is suitable for scenarios where the collected data is time series data and there is a sample scarcity problem in the fault diagnosis task. BACKGROUND

[0002] As a data-driven method, the performance of a fault diagnosis model depends largely on the quality of process data. However, in time series data fault diagnosis, fault samples are not only scarce but also severely imbalanced, because workers will actively intervene to prevent faults from occurring or developing. This characteristic of sample sparsity and imbalance makes it difficult for the model to determine accurate classification boundaries, often leading to model overfitting and reduced robustness. To address these challenges, various methods have been proposed, including sampling techniques, generative models, and loss function methods. Among them, generative models effectively enhance the diversity of samples by learning the distribution characteristics of real data and introducing dynamic features in generated samples, essentially solving the problem of a small number of samples. Therefore, generating high-quality samples to improve the stability and classification accuracy of diagnosis models is an important technical means at present.

[0003] Generative models aim to generate new samples by learning the distribution of existing data. The most direct method is to define an explicit distribution function p model (x, θ) to approximate the real data distribution p dataThen samples are drawn from the approximated distribution function. However, it is very difficult to accurately describe the complex distribution of high-dimensional data, and the emergence of deep learning makes it possible to achieve this task. Among them, the generative adversarial network (GAN) and the variational autoencoder (VAE) are the most typical and famous two generation methods. GANs generate data samples that are difficult to distinguish from real samples from noise using an adversarial training mechanism, but have shortcomings such as poor sample quality, mode collapse, and unstable training. In contrast, VAE uses an end-to-end training method and does not need to use two models for adversarial training, which is more convenient from the perspective of training. However, the loss function of VAE contains two parts, on the one hand, the reconstruction error guarantees the generation quality, and on the other hand, the distribution constraint guarantees the diversity of the generated samples, and the two parts of the loss function are in conflict during the training process, which limits the quality of the generated samples and reduces the practicability of VAE, especially in the case where high-quality samples must be generated. To this day, the diffusion model has become the darling of deep learning-based generation methods, and the generation performance based on the diffusion model (Diffusion) has broken the long-term dominance of GAN in the field of image generation. In the field of time series, Diffusion is also widely used for interpolation and completion of time series data, but it has not been widely applied to time series fault diagnosis. From the perspective of the method, the basic diffusion model still has shortcomings in considering the spatio-temporal relationship of time series data. When the diffusion model is applied to complex time series data, targeted improvements need to be made to the method itself to adapt to the characteristics of complex time series data. SUMMARY

[0004] The present application is aimed at the problem of the scarcity of effective fault samples in time series process fault diagnosis, and proposes a time series data generation method and application based on a spatio-temporal diffusion model. The method uses the lightweight network architecture of SE-U-Net and combines a time series embedding module to generate high-quality time series data and effectively restore the temporal and spatial relationships of real sequences.

[0005] A time series data generation method based on a spatio-temporal diffusion model adds a Squeeze-and-Excitation (SE) module to enhance the spatial feature extraction capability of the original U-shaped network, constructs an SE-U-Net as a noise prediction network in the diffusion model, and designs a new time series embedding module in the process of predicting noise in the original DDPM model. In the denoising process of generating new samples, the position and trend information of the previous time series is embedded in a decreasing form, thereby ensuring the temporal relationship between the generated sample blocks.

[0006] In the SE-U-Net module, the SE module is added to each layer of the U-net network to assign variable weights to different channels; the attention weight is constructed by the input feature, and then the attention weight is multiplied by the input feature to distinguish the importance of the feature; when it is located in the lth layer of the SE-U-Net network, the input feature is Among them C l is the number of channels, H l and W l is the spatial dimension of the variable; SE module is used to calculate X l Perform the following operations:

[0007] (1.1) Global average pooling: by formula Calculate the global feature z for each channel c c ,get where x c (i,j) is X l The element at position (i, j) in channel c;

[0008] (1.2) Compression and excitation: Input Z into a network containing two fully connected layers. The number of neurons in the first fully connected layer is r is the compression factor, the activation function is ReLU; the number of neurons in the second fully connected layer is C l , the activation function is Sigmoid, and the channel attention weight vector is obtained

[0009] (1.3) Feature weighting: by formula For the original input feature X l Weighted to obtain enhanced feature map It can distinguish the importance of different channel features and enhance the ability of spatial feature extraction.

[0010] 3. The method according to claim 2 is characterized in that a new time series embedding module is designed to embed the position and trend information of the previous time series in a decreasing manner during the denoising process of generating new samples: let the kth generated sample block be S k ∈R T×m , where T is the sample block length, i.e., the sequence length, and m is the variable dimension; when generating the k+1th sample block, select S k The sequence of the last p moments As an embedding basis, then do something like this:

[0011] (2.1) Embedding weight setting: Build n embedding segments based on Then the weight of the embedding vector W={w1,w2,…,w n},in i = 1, 2, …, n; if T cannot be divided by p, let the remainder be r = T % p, then the last remaining part constructs an embedding segment with a length of r, and the weight is

[0012] (2.2) Weighted embedding: multiply the embedding segment with the corresponding weight and splice according to the order to obtain the time sequence embedding vector E ∈ R T×m In the denoising process of generating the k+1th sample block, E is added to other input information to form a new input, which is input into the model together to ensure the time sequence relationship between the generated sample blocks in a conditional generation manner.

[0013] The method obtains data, and the generated data has a space-time relationship, so that the generated samples are spliced according to the generation order to obtain the final long sequence data.

[0014] The application of the generated data in time sequence fault diagnosis assumes that the number of fault categories is C, the number of samples of the category with the largest number of samples is N, and the time sequence data generation method based on the space-time diffusion model is applied to the category with a small number of samples to generate samples, so that the number of samples of each category is enhanced to N, so that the data is balanced, and then a classification method is applied.

[0015] The application of the generated data in time sequence fault diagnosis is applied to blast furnace data generation.

[0016] The beneficial effects of the present application are:

[0017] 1. The SE module improves the feature extraction capability of the original U-Net network and maintains the lightweight feature.

[0018] 2. The time sequence embedding module is proposed to ensure the time sequence consistency between the generated sample blocks.

[0019] 3. The model fully considers the characteristics of the scarcity of fault samples in time sequence fault diagnosis, and proposes a TS-Diffusion model for data enhancement of fault samples, which can effectively improve the accuracy of fault diagnosis. DETAILED DESCRIPTION

[0020] Figure 1 It is a schematic diagram of a U-shaped network (SE-U-Net) with a compression excitation module.

[0021] Figure 2 It is a schematic diagram of a compression excitation model.

[0022] Figure 3 It is a schematic diagram of time sequence embedding in the sample generation process.

[0023] Figure 4The construction of the data set in the training process.

[0024] Figure 5 The variable generation process schematic diagram.

[0025] Figure 6 The input and output schematic diagram of the SE-U-Net, and the position of the diffusion step embedding and the time series embedding added.

[0026] Figure 7-1 The original data curve, the LSTM-GAN generated data curve and the LSTM-VAE generated data curve of the 16th, 17th and 18th variables in the TE process.

[0027] Figure 7-2 The Time-GAN generated data curve and the TS-Diffusion generated data curve of the 16th, 17th and 18th variables in the TE process.

[0028] Figure 8 The t-SNE distribution diagram of the results generated by different models in the TE process.

[0029] Figure 9-1 The original data curve, the LSTM-GAN generated data curve and the LSTM-VAE generated data curve of the 1st, 2nd and 10th variables in the blast furnace process.

[0030] Figure 9-2 The Time-GAN generated data curve and the TS-Diffusion generated data curve of the 1st, 2nd and 10th variables in the blast furnace process.

[0031] Figure 10 The t-SNE distribution diagram of the results generated by different models in the blast furnace process.

[0032] Figure 11 The t-SNE distribution diagram of the original data and the enhanced data. DETAILED DESCRIPTION

[0033] The application will be further described in conjunction with the accompanying drawings and examples.

[0034] A time series data generation method based on a space-time diffusion model TS-Diffusion, which adds an SE module to enhance the spatial feature extraction capability of the original U-shaped network, constructs an SE-U-Net as a noise prediction network in the Diffusion model; in the process of predicting noise in the original DDPM model, a new time series embedding module is designed, which embeds the position and trend information of the previous time series in a decreasing form in the denoising process of generating new samples, thereby ensuring the time series relationship between the generated sample blocks.

[0035] In the SE-U-Net module, the SE module is added to each layer of the U-net network to assign variable weights to different channels; the attention weight is constructed by the input feature, and then the attention weight is multiplied by the input feature to distinguish the importance of the feature; when it is located in the lth layer of the SE-U-Net network, the input feature is Among them C l is the number of channels, H l and W l is the spatial dimension of the variable; SE module is used to calculate X l Perform the following operations:

[0036] (1.1) Global average pooling: by formula Calculate the global feature z for each channel c c ,get where x c (i,j) is X l The element at position (i, j) in channel c;

[0037] (1.2) Compression and excitation: Input Z into a network containing two fully connected layers. The number of neurons in the first fully connected layer is r is the compression factor, the activation function is ReLU; the number of neurons in the second fully connected layer is C l , the activation function is Sigmoid, and the channel attention weight vector is obtained

[0038] (1.3) Feature weighting: by formula For the original input feature X l Weighted to obtain enhanced feature map It can distinguish the importance of different channel features and enhance the ability of spatial feature extraction.

[0039] The above-mentioned design of a new time series embedding module embeds the position and trend information of the previous time series in a decreasing form during the denoising process of generating new samples: let the kth generated sample block be S k ∈R T×m , where T is the sample block length, i.e., the sequence length, and m is the variable dimension; when generating the k+1th sample block, select S k The sequence of the last p moments As an embedding basis, then do something like this:

[0040] (2.1) Embedding weight setting: Build n embedding segments based on Then the weight of the embedding vector W={w1,w2,…,w n},in i = 1, 2, …, n; if T cannot be divided by p, let the remainder be r = T % p, then the last remaining part constructs an embedded segment with a length of r, and the weight is

[0041] (2.2) Weighted embedding: multiply the embedded segment with the corresponding weight and splice them in order to obtain the time sequence embedding vector E ∈ R T×m In the denoising process of generating the k+1th sample block, E is added to other input information at the first layer and the last layer of the SE-U-Net network to form a new input, which is input into the model together to ensure the time sequence relationship between the generated sample blocks in a conditional generation manner.

[0042] Since the generated data has good space-time relationship, the final long sequence data can be obtained by splicing the generated samples in the generation order.

[0043] When the generated data is applied to time sequence fault diagnosis, assuming that the number of fault categories is C and the number of samples of the category with the largest sample number is N, the TS-Diffusion algorithm is applied to sample generation for the categories with small sample numbers, so that the number of samples of each category is enhanced to N, so that the data is balanced, and then a classification method is applied.

[0044] Embodiment

[0045] The Tennessee-Eastman (TE) process and the blast furnace process are taken as representatives for implementation, and the two processes are typical time sequence processes, wherein the data generation capability is tested on two data sets, and the fault diagnosis after data enhancement is further tested on the blast furnace process.

[0046] 1. Algorithm framework

[0047] (1) U-shaped network with compression excitation module

[0048] As mentioned above, the generation model should have the ability to extract spatial relationships. Previous research results show that CNN-based networks have outstanding performance in spatial relationship extraction due to the local receptive field of convolution kernels and the ability of hierarchical feature abstraction. Therefore, the TS diffusion model proposed in this paper takes the SE-U-Net network with an SE module (SE-U-Net) as the main network for noise prediction. The structure of the SE-U-Net is as follows: Figure 1As shown, its function is to predict the noise of each step in the reverse process. In the traditional convolutional neural network, the output is obtained by simply adding the convolution results of different channels. This equal-weight addition method makes it difficult for the network to distinguish the importance of the features extracted by different channels. Considering this drawback, the proposed method adds an SE module to each layer of the U-Net, which uses a channel attention mechanism to assign different weights to different channels. The first triangle in each layer represents a convolution operation, and the second triangle represents a convolution operation with an added SE module. This operation allows the network to assign greater weights to valuable features, improving the model's feature extraction ability and prediction performance. The SE structure diagram is shown in Figure 2 As shown, the channel attention weight is based on the input features, and then multiplied by the input features to distinguish the importance of different channel features. The introduction of the compression factor is to further reduce the complexity of the module. A simple module design can enable the model to perform well on a limited data set.

[0049] (2) Time sequence embedding module

[0050] The above SE-U-Net prediction network provides a good foundation for the temporal and spatial consistency within a single sample block, but to generate real and length-unlimited time series data, it must also overcome the difficulty of temporal consistency between sample blocks. In the field of image generation, the generated images are independent of each other, and there is no temporal relationship between the images before and after. However, for time series data, not only the temporal relationship within the sample block must be guaranteed, but also the temporal relationship between the sample blocks. From this perspective, the position and trend information at the end of the previous generated sequence is crucial for the generation of the next sequence data, as it determines the position and direction of the next sequence generation. Therefore, this subsection proposes a new time sequence embedding module, which will embed the position and trend information at the end of the previous generated sequence in a decreasing manner as additional information for conditional generation when generating a new sample block. The s-length subsequence segment at the end of the previous sample block will be embedded in the generation of the next sequence in a repeatedly decreasing manner as a time sequence embedding, as shown in Figure 3 The reason for setting the decreasing weight is to consider that the subsequence comes from the position and trend information at the end of the generated sequence, which has a greater impact on the closer part of the generated sample in the next segment, and a smaller impact on the farther part. In this way, the model can not only guarantee the consistency of the sample block in time sequence, but also make the sample generation diverse.

[0051] Figure 3 It shows how to implement time sequence embedding during sample generation. Since a new time sequence embedding model is added, a corresponding training set needs to be constructed to enable the model to have this capability. Therefore, Figure 4The process of constructing training samples is demonstrated. First, the sequence under different diffusion steps is constructed by adding noise step by step through the forward process. The training sample of the basic diffusion model includes two inputs: the sequence with noise and the embedding of diffusion steps, while the label of training is the added noise. In the proposed method, the input consists of three parts, in addition to the above two parts, a timing embedding is added. Among them, the label of training, the sequence with noise and the embedding of diffusion steps will change with the change of diffusion steps, but the timing embedding is always composed of the subsequence fragment at the end of the last generated sample. This way can ensure that the generated model always understands the position and trend information of the last generated sequence in the denoising process, and thus guarantees the quality of the final generated sample.

[0052] At this point, with SE-U-Net, the variable features within the sample block can be well extracted, and the timing relationship between the sample blocks can be ensured by the timing embedding. The proposed method can generate more realistic time series samples. In addition, since the timing relationship between the sample blocks is guaranteed, the model can generate longer time series by splicing the sample blocks.

[0053] (3) Time series generation without length limit

[0054] Since the traditional data generation method cannot guarantee the time relationship between the generated sample blocks, it can only generate real samples with limited length, and they need to increase the length of a single sample block to increase the effective sample length, which undoubtedly brings a rapidly growing burden to model training. In contrast, TS-Diffusion can ensure the continuity of sample generation through timing embedding, and can splice sequence data of any length under the condition that the length of a single sample block is very small, and can efficiently generate a large number of real time series samples. Figure 5 The entire process of long time series data generation is demonstrated. When generating the first sample, the last real sample block can be used as the basis for timing embedding, and then random noise is sampled in the prior distribution. The random noise is denoised by TS-Diffusion to generate a sample that maintains a timing relationship with the last sample block. Subsequent samples can then be generated in the same way based on the first sample block. Once the required number of sample blocks is generated, they only need to be concatenated in order to form consistent time series data.

[0055] Figure 6Two details about the temporal embedding and the diffusion step number encoding are further demonstrated. First, the position encoding adopted by the diffusion step number is the sinusoidal position embedding, and the diffusion step number is added to the network of each layer through the dimension conversion of the fully connected network. The significance of introducing this embedding is to enable the SE-U-Net to know the current diffusion step number in the reverse process. Second, the temporal embedding and the input have the same size, so the temporal embedding is added to the input and output of the SE-U-Net in the same way as the position embedding, that is, the addition operation with the variable.

[0056] (4) Sample generation enhances the blast furnace few-sample fault diagnosis architecture

[0057] In actual production process, there are many fault categories but they do not occur frequently, and once a fault occurs, the operator will actively intervene to avoid further deterioration. Therefore, the number of fault samples collected on site is usually small, and there is a serious inter-class imbalance phenomenon. Considering that the sample sparsity and imbalance problem will affect the training of the diagnosis model, a high-quality sample generation method TS-Diffusion model is designed in this chapter, which can effectively supplement data for fault classes with few samples. The TS-Diffusion model not only supports multi-variable input, but also effectively maintains the spatial consistency between multi-variable inputs. The fault diagnosis process based on data generation enhancement for inter-class imbalance dataset contains 4 parts. The first part is data collection and analysis, which analyzes the data characteristics of the process and mines the potential characteristics of the data, so as to propose a generation method. The second part generates samples corresponding to each fault category, and performs data enhancement on the few-sample fault categories to achieve the goal of data balancing. The third part evaluates the generated data, and if the quality of the generated data is not high, the model can be adjusted until high-quality data is generated. The fourth part trains the model based on the enhanced dataset to complete the fault diagnosis task.

[0058] 2、Parameter setting

[0059] In the embodiment, the sequence length T of the sample block is selected as 128, the diffusion step number K of the diffusion model is 1000, the diffusion coefficient β of the diffusion model is 0.01, and the sub-sequence length p of the temporal embedding module is 16. In the comparative method, the sequence length T of the sample block is selected as 128; the discriminator and the generator of the LSTM-GAN are composed of LSTM and fully connected layers; the encoder and the decoder of the LSTM-VAE are composed of LSTM models and fully connected layers; and the Time-GAN uses the standard model in the ydata_synthetic library.

[0060] 3、TE process data generation test

[0061] The TE dataset is introduced for experiments, on the one hand, as a simulation dataset, the variable change form and spatio-temporal relationship are representative, on the other hand, it is also an industrial dataset, which can increase the diversity of case analysis. The TE dataset contains 1 group of normal data and 21 groups of fault data, which is composed of 52 variables. The selected comparison methods include LSTM-GAN, LSTM-VAE, Time-GAN and TS-Diffusion, among which variables 16, 17 and 18 of the TE process are visualized, and the original variable graph and the generated sample results are shown in Figure 7-1 、 Figure 7-2 The t-SNE graph of the generated data and the original data is shown in Figure 8 .

[0062] As can be seen from Figure 8 , the distribution of the generated data of all methods has a certain coincidence with the distribution of the original data, and the method with the most similar distribution is the proposed TS-Diffusion. As can be seen from Figure 7-1 、 Figure 7-2 , LSTM-GAN can generate samples with time relationship, but due to the very unstable training of GAN, the generation performance on complex data is not reliable. In contrast, the training stability and generation performance of LSTM-VAE are much better, but due to the strict Gaussian distribution constraint of the latent variable, the data quality generated by VAE will be affected. In addition, LSTM-VAE cannot guarantee the time sequence relationship between the generated sample blocks, so the variables at the junction of the data blocks will change sharply, which is particularly evident in variable 18. This change will destroy the time sequence relationship of the data. In contrast, TS-Diffusion not only generates data with similar distribution to the real data, but also generates data with high consistency with the trend of the real data, because the introduction of temporal embedding maintains the time sequence relationship between the generated sample blocks. In addition, the proposed temporal and spatial consistency indicators are further used here to evaluate the generated data, to objectively evaluate the generation performance of the model.

[0063] 4. Blast furnace process data generation test

[0064] This experiment is aimed at a 2650 cubic meter blast furnace of a certain steel group. The data of the furnace body is sampled once every 10s on average, containing 13 variables, as shown in Table 1.

[0065] Table 1 Variable list of dataset

[0066] 1 Oxygen enrichment rate (%) 8 Total pressure difference (MPa) 2 BF permeability index 9 Resistance coefficient [3 Oxygen enriched flow (m 3 / h)] 10 BF top temperature (℃) 4 Blast kinetic energy (KJ) 11 Hot blast pressure (MPa) 5 Furnace bustle gas index 12 Blast humidity (g / m) 6 Theoretical combustion temperature (℃) 13 Oxygen enrichment pressure (MPa) 7 BF top pressure (kPa)

[0067] The hot blast stove switching phenomenon exists in the blast furnace, and the peak-shaped change brought by this switching operation has high consistency among multiple variables, which can be used to verify the performance of the model. The generation performance of variables 1, 2 and 9 is shown in Figure 9-1 、 Figure 9-2 The t-SNE plot of the generated data and the original data is shown in Figure 10 .

[0068] As can be seen from Figure 10 , the distribution of the generated data and the distribution of the original data have certain coincidence, and TSDiffusion is the best in distribution fitting. But as can be seen from Figure 9-1 、 Figure 9-2 , the data curves generated by different models have significant differences. This part of the result has two aspects worth paying attention to, one is the spatial relationship of the generated sample, and the other is the time sequence relationship between the sample blocks. The training of the LSTM-GAN model is affected by the hot blast stove switching, although it tries to reproduce the hot blast stove switching phenomenon in the generated sample, but this reproduction destroys the generation effect of other normal part of the sample, leading to the final generation result is not ideal. While the LSTM-VAE performs better in the generation of normal working condition samples, but it cannot reproduce the hot blast stove switching phenomenon, leading to the distortion of the generated result. The generation result of Time-GAN is similar to that of LSTM-GAN. In addition, since the above methods do not have the ability to keep the time sequence relationship between the sample blocks, there will be mutations at the junction of the data generated by them. In contrast, TS-Diffusion not only has the time sequence consistency between the data blocks, but also reproduces the hot blast stove switching phenomenon well.

[0069] 5. Fault diagnosis test of blast furnace process data enhancement

[0070] The data set contains four categories: normal, suspension, pipeline travel and sensor failure, and the details of the data set are shown in Table 2.

[0071] Table 2

[0072]

[0073] It should be noted that, in addition to sensor failure, the data in the test set and the training set do not come from the same failure. For example, the pipeline travel failure in the training set and the pipeline travel failure in the test set are actually two failures, with a time interval of tens of days between the two failures. Since the duration of the failure cannot be controlled artificially, there will be a sample imbalance problem. Pipeline travel refers to the phenomenon that the gas flow in a local area of the furnace develops excessively in a particular cross-sectional area. This phenomenon can cause the gas flow to develop abnormally in a certain area, forming a channel similar to a pipeline, usually appearing at the edge or center of the blast furnace. Suspended material refers to the situation where the furnace charge stops descending for more than a certain time, and sensor failure refers to the situation where the sensor of the blast furnace suddenly fails, causing the sensor to return an abnormal value or 0.

[0074] As can be seen from Table 2, the number of samples of normal, pipeline travel, suspended material and sensor failure in the training set is 3000, 798, 1635 and 1100 respectively. In order to achieve a balanced training set, each type of failure is supplemented by 3000 samples generated by the TS-Diffusion model, and the t-SNE distribution of the original data and the enhanced data is shown in Figure 11 .

[0075] As can be seen from Figure 11 , the fault types in the original data distribution are very discrete and discontinuous in space distribution, which is a typical performance of a small sample. However, after sample enhancement, the distribution of fault samples becomes more concentrated and continuous, and from the right part of Figure 11 , it can be seen that after the intra-class samples become more continuous and concentrated, the classification boundary between classes also becomes clearer. In this case, the training of the diagnosis model can become more stable and less prone to overfitting, and the classification boundary can also be more accurate.

[0076] After augmenting the data of each class, different baseline diagnosis methods are used for diagnosis testing to more fully prove the effectiveness of the model sample enhancement, and the results are shown in Table 3.

[0077] As shown in Table 3, the tested methods can be divided into four types, A: the most commonly used basic diagnosis method, B: the diagnosis method considering time relationship, C: the diagnosis method considering space relationship, and D: the time-space diagnosis method. These four types are commonly used diagnosis methods in the field of time series. As can be seen from the table, the accuracy of different types of methods has improved to varying degrees after data augmentation, which fully demonstrates the effectiveness of the TS-Diffusion model from the perspective of downstream tasks.

[0078] Table 3 Diagnosis performance and improvement of different types of methods before and after fault augmentation

[0079]

[0080] The embodiments in the above description can be further combined or replaced, and the embodiments are merely used to describe the preferred embodiments of the present application, and are not intended to limit the concept and scope of the present application. Without departing from the design idea of the present application, various changes and improvements made by those skilled in the art to the technical solutions of the present application shall fall within the protection scope of the present application. The protection scope of the present application is given by the appended claims and any equivalents thereof.

Claims

1. A method for generating time series data based on a spatiotemporal diffusion model, characterized in that: Based on the original U-net, a Squeeze-and-Excitation (SE) module is added to enhance its spatial feature extraction capability, and an SE-U-Net is constructed as the noise prediction network in the diffusion model. In the noise prediction process of the original DDPM model, a new time series embedding module is designed. During the denoising process of generating new samples, the position and trend information of the previous time series is embedded in a decreasing manner, thereby ensuring the temporal relationship between the generated sample blocks.

2. The method according to claim 1, characterized in that In the SE-U-Net module, the SE module is added to each layer of the U-net network to assign variable weights to different channels; the attention weight is constructed by the input feature, and then the attention weight is multiplied by the input feature to distinguish the importance of the feature; when it is located in the lth layer of the SE-U-Net network, the input feature is Among them C l is the number of channels, H l and W l is the spatial dimension of the variable; SE module is used to calculate X l Perform the following operations: (1.1) Global average pooling: by formula Calculate the global feature z for each channel c c ,get where x c (i,j) is X l The element at position (i, j) in channel c; (1.2) Compression and excitation: Input Z into a network containing two fully connected layers. The number of neurons in the first fully connected layer is r is the compression factor, the activation function is ReLU; the number of neurons in the second fully connected layer is C l , the activation function is Sigmoid, and the channel attention weight vector is obtained (1.3) Feature weighting: by formula For the original input feature X l Weighted to obtain enhanced feature map It can distinguish the importance of different channel features and enhance the ability of spatial feature extraction.

3. The method according to claim 2, characterized in that The above-mentioned design of a new time series embedding module embeds the position and trend information of the previous time series in a decreasing form during the denoising process of generating new samples: let the kth generated sample block be S k ∈R T×m , where T is the sample block length, i.e., the sequence length, and m is the variable dimension; when generating the k+1th sample block, select S k The sequence of the last p moments As an embedding basis, then do something like this: (2.1) Embedding weight setting: Build n embedding segments based on Then the weight of the embedding vector W={w1,w2,…,w n },in If T is not divisible by p, let the remainder be r = T% p, then the remaining part constructs an embedding segment of length r, with a weight of (2.2) Weighted embedding: Multiply the embedding segments with the corresponding weights and concatenate them in order to obtain the temporal embedding vector E∈R T ×m In the denoising process of generating the k+1th sample block, E is added to other input information in the first and last layers of the SE-U-Net network to form a new input and then input into the model together, ensuring the temporal relationship between the generated sample blocks in a conditional generation manner.

4. Data obtained by the method according to any one of claims 1 to 3, characterized in that: The generated data already has a spatiotemporal relationship, and the final long sequence data can be obtained by splicing the generated samples in the generation order.

5. The application of generated data in timing fault diagnosis according to claim 4, characterized in that: Assume that the number of fault categories is C, and the category with the largest number of samples is N. For categories with a small number of samples, the time series data generation method based on the spatiotemporal diffusion model is applied to generate samples. The number of samples in each category is increased to N to balance the data, and then the classification method is applied.

6. The application of generated data in timing fault diagnosis according to claim 4, characterized in that: Applied to blast furnace data generation.

Images