Generative adversarial network method for electronic nose univariate time sequence data enhancement

Abstract

The invention discloses a generative adversarial network method for electronic nose univariate time series data enhancement, which comprises the following steps: independently fitting a Gaussian mixture model for each category of data in an electronic nose data set, and carrying out adversarial training on a generative adversarial network which comprises a generator and a discriminator; enhanced data is generated using the trained generator. According to the method, the distribution consistency constraint is applied between the noise potential space and the generated data, the common mode collapse problem of the generative adversarial network in the time sequence data generation is relieved, the generated data better conforms to the real data distribution, the adopted total loss LR3 is combined with the gradient penalty strategy, and the time sequence data generation efficiency is improved. The stability of model training is remarkably improved, the common instability problem in the generative adversarial network training process is relieved, and the quality of generated data is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of sensor data processing, and in particular to a method for enhancing lung cancer electronic nose data. Background Art

[0002] The diagnosis of lung cancer is crucial to improving patient survival rates, and electronic nose-based lung cancer detection provides a new, non-invasive, rapid, and economical approach for detecting lung cancer markers. However, similar to other medical datasets, existing lung cancer time series datasets suffer from limited sample size and class imbalance, which limits the generalization performance of advanced deep learning techniques on these datasets. To address this challenge, generative models such as Generative Adversarial Networks (GANs) have been widely used in time series data generation tasks due to their powerful data generation capabilities. By using data augmentation synthesized by generative models, the performance of deep learning models on small datasets can be effectively improved.

[0003] Traditional generative adversarial networks generally use standard Gaussian noise or uniformly distributed noise as the input noise of the generator G, that is: Or z~U(-1,1). Here, z represents the input noise. However, noise that follows a single distribution has a relatively limited ability to fit multimodal data. Mode collapse occurs when the generator's ability to generate data during training is limited to a few modes or categories, failing to cover the full distribution of real data.

[0004] Because traditional GAN models face not only mode collapse but also training instability when generating multi-modal time series, there is an urgent need to address these technical challenges to improve the quality of generated data. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a generative adversarial network method for electronic nose univariate time series data enhancement, so as to solve the technical problems of mode collapse and unstable training often faced by GAN models when generating multi-modal time series.

[0006] The generative adversarial network method for enhancing time series data of a lung cancer electronic nose according to the present invention comprises the following steps:

[0007] I) Fit a Gaussian mixture model for each category of data in the electronic nose dataset. Each Gaussian mixture model is expressed as:

[0008]

[0009] Among them, V iis the Gaussian mixture model corresponding to the i-th category, i∈{1,2,…,C} represents the category number, k i Indicates V i kth i Gaussian components, K i The Gaussian mixture model representing category i has a total of K i Gaussian components, is the kth Gaussian mixture model of category i i The weights of the Gaussian components, The representative mean is The covariance is Gaussian distribution;

[0010] Each sample in the electronic nose dataset is assigned an index pair (i, k i ), used for noise generation of different patterns within a class.

[0011] II) performing adversarial training on a generative adversarial network, which includes a generator and a discriminator. The adversarial training process is divided into two stages, and each stage of adversarial training includes:

[0012] 1) Gaussian mixture model is based on the label and index (i, k) of the sample i ) Find the corresponding Gaussian component k i , for Gaussian component k i Sampling Gaussian mixture noise:

[0013]

[0014] in, is a Gaussian mixture noise.

[0015] 2) Gaussian mixture noise is used as the input noise of the generator, and the generator generates fake samples.

[0016] 3) The fake samples generated by the generator and the real samples from the real data are input into the discriminator, which judges the fake samples and the real samples to obtain the discrimination results;

[0017] In the first stage of the adversarial training process, the total loss L is introduced latent Acting on the generator and introducing the total loss L R3 Acts on both the generator and the discriminator; in the second stage of the anti-training training process, only the total loss L is introduced R3 Acts on both the generator and the discriminator;

[0018] Total loss L latent The expression is as follows:

[0019] L latent =L cst +λLspec (3)

[0020] Among them, L cst is the consistency loss, L spec is spectrum loss; is a weight hyperparameter used to balance the contribution of consistency and spectral loss;

[0021] Consistency loss L cst Used to constrain the distribution between the data generated by the generator and the real data, L cst The expression is as follows:

[0022]

[0023] in, For the generator Fake samples generated for input, is the category i in the real data and belongs to the kth i True samples of clusters;

[0024] Spectrum loss L spec It is used to constrain the spectrum similarity between the data generated by the generator and the real data, L spec The expression is as follows:

[0025]

[0026] Where STFT(·) represents the short-time Fourier transform operation;

[0027] During the training process, the total loss L cst Backward pass to the generator to constrain the distribution consistency between the latent space and the generated sample space;

[0028] Total loss L R3 The expression is as follows:

[0029]

[0030] L RpGAN The loss is described as follows:

[0031]

[0032] Among them, p z (z) is the distribution of noise z, p data (x) is the distribution of the real sample x, D(G(z)) is the discriminant output of the discriminator for the generated sample G(z), D(x) is the discriminant output of the discriminator for the real sample x, and function f is the Softplus function;

[0033] Gradient penalty term L R1 The description is as follows:

[0034]

[0035] Where L is the length of the input time series, d is the number of sensors, is the partial derivative of the discriminator with respect to the value of the j-th variable at the l-th time step, that is, the gradient of the data point;

[0036] Gradient penalty term L R2 The description is as follows:

[0037]

[0038] in, Represents the partial derivative of the discriminator with respect to the value of the j-th variable at the l-th time step in the generated data, that is, the gradient of the data point;

[0039] γ is a hyperparameter used to control the gradient penalty term L R1 and L R2 Strength;

[0040] During the training process, the total loss L R3 Backward pass to the generator and discriminator.

[0041] III) The Gaussian mixture model finds the corresponding Gaussian components according to the labels and indices of the samples in the electronic nose dataset, samples the Gaussian components to obtain Gaussian mixture noise, and inputs the Gaussian mixture noise into the generative adversarial network trained in step II), and the generator generates enhanced data.

[0042] Furthermore, the generator in step II is composed of an encoder and a decoder.

[0043] Beneficial effects of the present invention:

[0044] 1. The generative adversarial network method for electronic nose univariate time series data enhancement in the present invention alleviates the common mode collapse problem of generative adversarial networks in time series data generation by imposing distribution consistency constraints between the noise latent space and the generated data, making the generated data more consistent with the real data distribution.

[0045] 2. The generative adversarial network method for enhancing univariate time series data of electronic noses in this invention adopts a total loss L R3 Combined with the gradient penalty strategy, the stability of model training is significantly improved, the instability problem common in the training process of generative adversarial networks is alleviated, and the quality of generated data is improved.

[0046] 3. The generative adversarial network method for enhancing univariate time series data of electronic noses in the present invention transforms the generator of the existing generative adversarial network into an encoder-decoder structure, thereby improving the model's ability to gradually model from the Gaussian mixture noise latent space to the generated data space.

[0047] 4. The present invention's generative adversarial network method for enhancing univariate time series data for electronic noses employs a two-stage training mechanism during model training: The first stage introduces latent space distribution consistency constraints and R3 regularization to strengthen the generator's initial alignment with the real data distribution, improving the model's recognition and coverage of latent patterns. The second stage removes the distribution consistency constraints, retaining only R3 regularization, further stabilizing the model training process and enhancing the diversity and authenticity of generated samples. This two-stage training mechanism enables a gradual transition from generator learning of distributional structure to stable optimization of generation quality, improving overall training efficiency and model robustness, and preventing the negative impact of early pattern collapse on subsequent training. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is the UTS-GMN-GAN framework diagram.

[0049] Figure 2 Direct visual comparison of UTS data synthesized for different models.

[0050] Figure 3 A visual comparison of UTS data synthesized by different models after PCA dimensionality reduction.

[0051] Figure 4 Visual comparison of UTS data synthesized by different models after t-SNE dimensionality reduction. DETAILED DESCRIPTION

[0052] The present invention will be further described below with reference to the accompanying drawings and examples.

[0053] The generative adversarial network method for enhancing univariate time series data of an electronic nose in this embodiment includes the following steps:

[0054] I) Fit a Gaussian mixture model (GMM) for each category of time series data in the electronic nose dataset. Each Gaussian mixture model is expressed as:

[0055]

[0056] Among them, V i is the Gaussian mixture model corresponding to the i-th category, i∈{1,2,…,C} represents the category number, k i Indicates V i kthi Gaussian components, K i The Gaussian mixture model representing category i has a total of K i Gaussian components, is the kth Gaussian mixture model of category i i The weights of the Gaussian components, The representative mean is The covariance is Gaussian distribution;

[0057] Each sample in the electronic nose dataset is assigned an index pair (i, k i ), used for noise generation of different patterns within a class.

[0058] II) Conduct adversarial training on a Generative Adversarial Network (GAN), which includes a generator and a discriminator. The adversarial training process is divided into two stages. Each stage of adversarial training includes:

[0059] 1) Gaussian mixture model is based on the label and index (i, k) of the sample i ) Find the corresponding Gaussian component k i , for Gaussian component k i Sampling Gaussian Mixture Noise (GMN):

[0060]

[0061] in, is a Gaussian mixture noise.

[0062] Compared to traditional single-distribution noise, the GMN obtained in this example can effectively increase the diversity of generated data during training, making it more closely aligned with the distribution characteristics of real data, thereby significantly alleviating the problem of mode collapse. Furthermore, because the GMM is fitted based on real data, its statistical characteristics are closer to the actual distribution of the target data, further improving the quality of the synthesized data.

[0063] 2) Gaussian mixture noise is used as the input noise of the generator, and the generator generates fake samples; in this embodiment, the generator is composed of an encoder and a decoder.

[0064] Because the dimensionality of a GMN is typically the same as real data, directly performing a significant dimensionality reduction on the GMN can result in the loss of key information. Furthermore, employing multiple layers of equal-dimensional transformations can lead to excessive model complexity or redundant model parameters. Therefore, the generator in this embodiment employs an encoder-decoder structure to more efficiently process the GMN. During the encoding phase, the dimensionality of the input noise is gradually compressed to reduce redundant information and extract key features. During the decoding phase, the data dimensionality is gradually restored, allowing the generated data to better match the target distribution. This hierarchical transformation approach reduces information loss while enhancing the generator's expressiveness and generalization capabilities, making it more suitable for time series data generation tasks.

[0065] In addition, model structure adaptive logic can be introduced to dynamically adjust the number of layers of the generator and discriminator using a logarithmic scale according to the length of the time series. This mechanism can automatically adapt to univariate time series data of different lengths without manually modifying the model structure. This makes the GMN-GAN model in this embodiment more scalable and able to adapt to data generation tasks of univariate time series of different lengths.

[0066] 3) The fake samples generated by the generator and the true samples from the real data are input into the discriminator, and the discriminator judges the authenticity of the fake samples and the true samples to obtain the discrimination results.

[0067] In the first stage of the adversarial training process, the total loss L is introduced latent Acting on the generator and introducing the total loss L R3 Acts on both the generator and the discriminator; in the second stage of the anti-training training process, only the total loss L is introduced R3 It works on both the generator and the discriminator.

[0068] Total loss L latent The expression is as follows:

[0069] L latent =L cst +λL spec (3)

[0070] Among them, L cst is the consistency loss, L spec is spectrum loss; is a weight hyperparameter used to balance the contribution of consistency and spectral loss.

[0071] Consistency loss L cst Used to constrain the distribution between the data generated by the generator and the real data, L cst The expression is as follows:

[0072]

[0073] in, For the generator Fake samples generated for input, is the category i in the real data and belongs to the kth i Based on this approach, the generator can find the corresponding cluster in the real data set according to the category to which the GMN belongs and the components of the corresponding Gaussian variable.

[0074] In model training, although GMN is used as an implicit conditional input for the generator, the weights of the GAN model are randomly initialized, so the generator cannot quickly learn the distribution of real data. i The different components in represent different patterns in the category, i.e., subcategories. However, due to the lack of explicit generation constraints on subcategory-level data in the design of GMN-GAN, the generator cannot correctly learn the correspondence between the latent space (i.e., noise space) and the subcategory pattern. Therefore, the data generated by the generator through the GMN corresponding to the specified subcategory cannot match the corresponding real data, that is, the mapping of the generator's latent space to the sample space at the subcategory level is chaotic, which in turn affects the pattern correspondence of the generated data. To solve the above problem, this embodiment uses the total loss L latent Introducing consistency loss L cst , which is achieved by dividing the output of the generator by the noise in its input latent space. The generator is encouraged to learn the accurate mapping relationship between the GMN latent space and the sample space, so that the subclasses in the latent space can be accurately mapped to the corresponding subclasses in the sample space, thereby maintaining their distribution consistency.

[0075] Due to L cst The data distribution is matched mainly by minimizing the Euclidean distance between the synthetic samples and the real samples. In practical applications, this may cause the generator to generate more noise, affecting the quality of the generated data. Therefore, based on the consistency loss, this embodiment reduces the total loss L latent Spectral loss is further introduced to reduce the noise interference in the generated data. spec It is used to constrain the spectrum similarity between the data generated by the generator and the real data, L spec The expression is as follows:

[0076]

[0077] Here, STFT(·) represents the short-time Fourier transform operation.

[0078] During the training process, the total loss L cstThe backpropagation is passed to the generator to constrain the distribution consistency between the latent space and the generated sample space.

[0079] The above-mentioned introduction of distribution consistency constraints in the GMN-GAN generator training process improves the mapping consistency between the latent space and the sample space, and effectively reduces the noise interference in the generated data, thereby improving the quality of data generation.

[0080] Traditional GAN training often faces the problems of training instability (such as gradient disappearance and explosion) and mode collapse. To solve this problem, this embodiment uses the total loss L R3 RpGAN loss and zero-center gradient penalty are introduced in

[15] .

[0081] Total loss L R3 The expression is as follows:

[0082]

[0083] Among them, L RpGAN is the RpGAN loss, L R1 and L R2 is the gradient penalty, γ is a hyperparameter used to control the gradient penalty term L R1 and L R2 strength.

[0084] L RpGAN The loss is described as follows:

[0085]

[0086] Among them, p z (z) is the distribution of noise z, p data (x) is the distribution of the real sample x, D(G(z)) is the discriminant output of the discriminator for the generated sample G(z), D(x) is the discriminant output of the discriminator for the real sample x, and function f is the Softplus function. L RpGAN This loss enables the discriminator to judge the authenticity of generated samples based on their relative authenticity to real samples, rather than focusing solely on the degree of classification between generated samples and real samples. Measuring the authenticity of each generated sample relative to real samples, this coupling of real and generated data helps maintain the decision boundary within the distribution of real data, thereby mitigating mode collapse.

[0087] Although RpGAN alleviates the mode collapse problem by introducing relative pairing loss, another major challenge in GAN training is the gradient instability of the discriminator, which may lead to gradient vanishing or gradient explosion, thus affecting the training of the generator. In order to solve the gradient instability problem, this embodiment uses the total loss L R3 Two gradient penalty terms are introduced.

[0088] Gradient penalty term L R1 The description is as follows:

[0089]

[0090] Where L is the length of the input time series, d is the number of sensors (i.e. the total number of variable dimensions of the time series), is the partial derivative of the discriminator with respect to the value of the j-th variable at the l-th time step, that is, the gradient of the data point.

[0091] Gradient penalty term L R2 The description is as follows:

[0092]

[0093] in, Represents the partial derivative of the discriminator with respect to the value of the j-th variable at the l-th time step in the generated data, that is, the gradient of the data point.

[0094] Gradient penalty term L R1 Acting on real data, the gradient penalty term L R2 Acting on the generated data. During the training process, the total loss L R3 Passed back to the generator and discriminator, if the generator G can already generate data with real distribution, the discriminator D will not push the generator G away from the optimal state; by constraining the gradient norm of the discriminator, the oscillation generated during the training process can be effectively reduced.

[0095] III) The Gaussian mixture model finds the corresponding Gaussian components according to the labels and indices of the samples in the electronic nose dataset, samples the Gaussian components to obtain Gaussian mixture noise, and inputs the Gaussian mixture noise into the generative adversarial network trained in step II), and the generator generates enhanced data.

[0096] The following experiments and analysis are conducted on the generative adversarial network method for enhancing the univariate time series data of the electronic nose proposed in this embodiment. The model architecture of the generative adversarial network method for enhancing the univariate time series data of the electronic nose proposed in this embodiment is as follows: Figure 1 As shown, it is named UTS-GMN-GAN model.

[0097] Experimental setup

[0098] The electronic nose dataset used in all the following experiments is the univariate version of the lung cancer dataset, referred to as LC-UTS. According to existing studies, the data of the TGS2602 channel has the highest contribution in the classification task of lung cancer detection. Therefore, LC-UTS is constructed using channel data from the TGS2602 sensor. The detailed information of the sensor is shown in Table 1. LC-UTS contains 169 samples, 107 from lung cancer patients and 62 from healthy controls. Each sample is a univariate time series (i.e., it only contains data collected from a single channel), and each sample contains 360 data points collected during the sampling phase, which can be expressed as S i =[s1,s2,…,s L ] T , where L = 360, i = 1. Table 2 lists the training set of the lung cancer dataset (X tr ), and the test set (X te ). All models are trained only on the training set without using any data from the test set.

[0099] Table 1 Information of gas sensors in the sensor array

[0100]

[0101] Table 2 Lung cancer dataset division

[0102]

[0103] All experiments were implemented using the PyTorch deep learning framework in Python and run on an NVIDIA TITANXp graphics card (3840 cores, 12GB of global memory). The hyperparameter settings used during UTS-GMN-GAN model training are shown in Table 3. Epochs is the total number of epochs trained for the UTS-GMN-GAN model, and Guided Epochs is the number of epochs within the epochs that were guided using the latent space distribution consistency loss.

[0104] Table 3 Hyperparameter settings of the UTS-GMN-GAN model

[0105]

[0106] Model ablation experiment

[0107] In order to evaluate L latent and L R3 To investigate the impact of the model performance, an ablation experiment was conducted on the UTS-GMN-GAN structure. The experimental results are shown in Table 4.

[0108] Table 4 Ablation experiment of UTS-GMN-GAN

[0109]

[0110] From the experimental results, it can be seen that without introducing L latent and L R3 When L is introduced alone, the FID score of the model is 8.547, the discrimination score and the prediction score are 0.461 and 0.070 respectively. latent After the test, all three metrics improved significantly, with the FID in particular dropping to 0.059, demonstrating that the loss function effectively optimizes the distribution of the latent space, making the generated data more consistent with the statistical characteristics of real data, and improving data quality and diversity. Furthermore, the discriminant score and prediction score also decreased to 0.074 and 0.025, respectively, indicating that the generated data is closer to the distribution of real data in the classification task.

[0111] However, using L alone R3 Time (Baseline+L R3 ), the model performance degraded significantly, the FID increased to 18.467, the discrimination score further increased to 0.500, and the prediction score also increased significantly to 0.861. This shows that in the absence of potential distribution constraints, L R3 It will excessively suppress the learning ability of the discriminator, making it difficult for the generator to obtain effective gradient feedback, resulting in a decrease in the quality of generated data.

[0112] When L is introduced at the same time latent and L R3 Time (Baseline+L latent +L R3 ), the model performance reached the best, FID was further reduced to 0.008, and the discrimination score and prediction score were also optimized to 0.025 and 0.026 respectively. This result shows that L latent On the basis of optimizing the generated data, a more stable data distribution is provided for the discriminator, so that L R3 It can improve the generalization ability of the discriminator while avoiding excessive suppression of the generator learning. In summary, L latent is the key factor to improve the quality of generated data, and L R3 On this basis, the stability and generalization ability of the model are further enhanced, so that the overall performance of the final model is improved.

[0113] Model comparison experiment

[0114] The UTS-GMN-GAN model is comprehensively compared with other mainstream time series generation models (TimeGAN, Sig-WGAN, TimeVAE, and BioDiffusion) to evaluate its advantages in data quality, pattern coverage, computational overhead, etc. The following methods are used to evaluate the experimental results:

[0115] (1) Visual analysis: including direct comparison of samples of real and fake data, observation of pattern coverage of synthetic data from different channels, and visual analysis after dimensionality reduction.

[0116] (2) Synthetic data quality assessment (objective indicators): including discriminative score, predictive score, and Fréchet inception distance (FID). When evaluating the quality of generated data, referring to the common settings in existing studies, the default method of using a 1:1 ratio of generated data to real data is used for evaluation, that is, using synthetic samples with the same number of real data to ensure consistency and fairness of the evaluation. In order to reduce the randomness of model weight initialization and training process, all objective evaluation indicators are calculated based on the average accuracy of 5 runs.

[0117] The number of samples synthesized by each model is equivalent to the training set, and the quality is evaluated by combining subjective and objective evaluation methods. In addition, the training time, inference time, parameter scale and computational complexity of different models are compared and analyzed. There are two types of subjective indicators. One is to directly draw the time series curve of real data and synthetic data to intuitively evaluate its generation quality (such as Figure 2 The other is to reduce the dimension of the real data and the synthetic data to a 2D plane for visualization through dimensionality reduction methods, including PCA (such as Figure 3 ) and t-SNE methods (as shown Figure 4 ) as shown.

[0118] Figure 2 In the results shown, the red curve represents the curve drawn from real data, and the blue curve represents the curve drawn from the synthetic dataset. The results show that the data synthesized by the UTS-GMN-GAN model proposed in this example can effectively cover the different patterns in real data and has the best performance in pattern coverage.

[0119] Figure 3 and Figure 4In the results shown, red dots represent sample points after dimensionality reduction of real data, and blue dots represent sample points after dimensionality reduction of synthetic data. Experimental results show that the UTS-GMN-GAN model's PCA dimensionality reduction results are most similar to the distribution of real data. Similarly, the t-SNE dimensionality reduction results, similar to those of BioDiffusion, are consistent with the distribution of real data.

[0120] Table 5 shows the quality evaluation results of the data synthesized by the UTS-GMN-GAN model proposed in the present invention and other methods, where the data in the first row represents the evaluation results based on real data and can be used as a reference benchmark for authenticity. From the experimental results, it can be seen that UTS-GMN-GAN achieved the best performance in indicators such as FID, discriminant score (Discriminative Score) and predictive score (Predictive Score), and is closest to the real data. Table 6 further shows the computing resource comparison of different generative models on the LC-UTS dataset, including key indicators such as training time, inference time, model parameter amount and computational amount, in order to comprehensively evaluate the practical application value of each method. According to the experimental results, the UTS-GMN-GAN model proposed in the present invention has the least computational amount, suboptimal training and inference time, and relatively few parameters, and has certain advantages among the compared models.

[0121] Table 5 Performance comparison of different time series generation models (UTS)

[0122]

[0123]

[0124] Table 6 Comparison of computing resources of different time series generation models (UTS)

[0125]

[0126] When evaluating the quality of generated data, existing research generally uses a quantity of generated data equivalent to the training data. However, in real applications, the amount of generated data is often adjusted based on actual needs. Therefore, this paper focuses on analyzing the impact of different data augmentation factors (i.e., the ratio of the amount of generated data to the amount of original data) on the quality of generated data. The experimental results are shown in Table 7.

[0127] Table 7 Effect of different data augmentation multiples on the generation quality of UTS-GMN-GAN

[0128]

[0129] As shown in Table 7, at lower augmentation factors (e.g., 1x and 2x), the metrics of the generated data are relatively close to those of the original data. Although the FID increases slightly (indicating a slight increase in the distribution gap between the generated samples and the real samples), the overall change is not significant. The discriminant score and prediction score show some volatility, which may reflect their high sensitivity to subtle changes in the data distribution.

[0130] As the augmentation factor increases (4x and above), the discriminability score decreases significantly, reaching its lowest value of 0.007 at 8x augmentation. This indicates that the difference between generated and real samples in the discriminant space is minimal, and the generation quality is relatively high. The prediction score also achieves optimal results at 8x augmentation, demonstrating that the generated data strongly supports downstream prediction tasks. However, it is important to note that this trend may also be affected by the imbalance in the ratio of real to generated data, which to some extent affects the accuracy and generalization of the evaluation results. The FID achieves optimal results at both 1x and 10x augmentation, with overall stable variation. This indicates that this metric is robust to changes in the amount of generated data and is not easily affected by changes in the augmentation factor. Overall, the UTS-GMN-GAN demonstrates relatively stable generation quality at different augmentation factors, and the above metrics accurately reflect the overall performance of the model at 1x augmentation.

[0131] In summary, when evaluating quality using generated data of comparable size to the training data (i.e., 1x augmentation), various metrics can effectively reflect the model's performance. Therefore, using 1x augmentation for quality evaluation of generated data is reasonable, ensuring both objectivity and fairness. It also aligns with mainstream practices in existing literature and demonstrates high rationality and practical value.

[0132] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions of the present invention, and all such changes should be included in the scope of the claims of the present invention.

Claims

1. A generative adversarial network method for electronic nose univariate time series data enhancement, characterized by: The following steps are involved: I) Fit a Gaussian mixture model for each category of data in the electronic nose dataset. Each Gaussian mixture model is expressed as: Among them, V i is the Gaussian mixture model corresponding to the i-th category, i∈{1,2,…,C} represents the category number, k i Indicates V i kth i Gaussian components, K i The Gaussian mixture model representing category i has a total of K i Gaussian components, is the kth Gaussian mixture model of category i i The weights of the Gaussian components, The representative mean is The covariance is Gaussian distribution; Each sample in the electronic nose dataset is assigned an index pair (i, k i ), used for noise generation of different patterns within a class; II) performing adversarial training on a generative adversarial network, which includes a generator and a discriminator. The adversarial training process is divided into two stages, and each stage of adversarial training includes: 1) Gaussian mixture model is based on the label and index (i, k) of the sample i ) Find the corresponding Gaussian component k i , for Gaussian component k i Sampling Gaussian mixture noise: in, is Gaussian mixture noise; 2) Using Gaussian mixture noise as the input noise of the generator, the generator generates fake samples; 3) The fake samples generated by the generator and the real samples from the real data are input into the discriminator, which judges the fake samples and the real samples to obtain the discrimination results; In the first stage of the adversarial training process, the total loss L is introduced latent Acting on the generator and introducing the total loss L R3 Acts on both the generator and the discriminator; in the second stage of the anti-training training process, only the total loss L is introduced R3 Acts on both the generator and the discriminator; Total loss L latent The expression is as follows: THE latent =L cst +λL spec (3) Among them, L cst is the consistency loss, L spec is spectrum loss; is a weight hyperparameter used to balance the contribution of consistency and spectral loss; Consistency loss L cst Used to constrain the distribution between the data generated by the generator and the real data, L cst The expression is as follows: in, For the generator Fake samples generated for input, is the category i in the real data and belongs to the kth i True samples of clusters; Spectrum loss L spec It is used to constrain the spectrum similarity between the data generated by the generator and the real data, L spec The expression is as follows: Where STFT(·) represents the short-time Fourier transform operation; During the training process, the total loss L cst Backward pass to the generator to constrain the distribution consistency between the latent space and the generated sample space; Total loss L R3 The expression is as follows: L RpGAN The loss is described as follows: Among them, p z (z) is the distribution of noise z, p data (x) is the distribution of the real sample x, D(G(z)) is the discriminant output of the discriminator for the generated sample G(z), D(x) is the discriminant output of the discriminator for the real sample x, and function f is the Softplus function; Gradient penalty term L R1 The description is as follows: Where L is the length of the input time series, d is the number of sensors, is the partial derivative of the discriminator with respect to the value of the j-th variable at the l-th time step, that is, the gradient of the data point; Gradient penalty term L R2 The description is as follows: in, Represents the partial derivative of the discriminator with respect to the value of the j-th variable at the l-th time step in the generated data, that is, the gradient of the data point; γ is a hyperparameter used to control the gradient penalty term L R1 and L R2 Strength; During the training process, the total loss L R3 Backward pass to the generator and discriminator; III) The Gaussian mixture model finds the corresponding Gaussian components according to the labels and indices of the samples in the electronic nose dataset, samples the Gaussian components to obtain Gaussian mixture noise, and inputs the Gaussian mixture noise into the generative adversarial network trained in step II), and the generator generates enhanced data.

2. The generative adversarial network method for enhancing univariate time series data of an electronic nose according to claim 1, characterized in that: The generator in step II consists of an encoder and a decoder.

Images