DTSIGAN data enhancement method for residual life of industrial equipment under unbalanced missing data

Abstract

The invention belongs to the technical field of industrial residual life, discloses a DTSIGAN data enhancement method and system for the residual life of industrial equipment under unbalanced missing data, and solves the problems that a data set is small in scale, data is unbalanced and missing values exist. By providing a solid data basis, the application efficiency of the data-driven artificial intelligence technology in prediction of the residual life of the industrial equipment is improved. According to the method, a time sequence generation adversarial network is optimized by adopting a cyclic dynamics interpolation algorithm, a mask matrix, a time interval matrix and a time decay factor are introduced, missing value features are dynamically captured and analyzed, and a missing value interpolation process is optimized. In order to enhance the stability of the model, the Wasserstein distance and the gradient penalty term are combined, the loss function of the generative network and the discriminant network is optimized, gradient disappearance and explosion are effectively avoided, the stability of the training process is ensured, and the synthesis capability of high-quality time series data is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial remaining life, and in particular relates to a DTSIGAN data enhancement method and system for the remaining life of industrial equipment under unbalanced missing data. Background Art

[0002] In the industrial field, the degradation of equipment performance or the loss of key components often threaten the stability of the entire production process. They not only pose safety risks, but also directly lead to significant economic losses, posing a severe challenge to production efficiency and cost control.

[0003] In recent years, the wave of the Industrial 4.0 revolution has swept in, with its core focusing on the rise of intelligent manufacturing. Its distinctive feature lies in the seamless interconnection of information-physical fusion systems such as artificial intelligence, the Internet of Things, and cloud computing. In this context, the fault prediction and health management system (PHM), as a key technical bridge, is leading the traditional production industry to leap and upgrade to a new intelligent and efficient manufacturing model, becoming a key breakthrough in transformation and upgrading. In the fault prediction and health management system, the remaining service life (RUL) occupies a core and vital position. Through in-depth analysis of the current equipment condition or component status, the future operating status can be accurately predicted, providing solid data support for the formulation of production and operation decisions. Based on this, enterprises can implement targeted preventive maintenance strategies, which not only avoid economic losses caused by sudden equipment failures, but also significantly reduce the risk of safety accidents, thereby ensuring the continuous, stable and efficient operation of the production line. However, in complex actual production environments, data collection faces many challenges. For example, the limitations of the internal space of enclosed equipment make it particularly difficult to deploy a sufficient number of sensors to fully capture data, thereby exacerbating the technical difficulty of data collection. In addition, data collection is expensive, especially in order to obtain the operating data of the complete life cycle of the equipment, which requires a lot of resources and time. The dual pressure of these time or economic costs has led to the scarcity and imbalance of data resources. At the same time, data collection system failures, human operational errors, and uncertainties in the external environment often cause data missing problems. The imbalance and lack of data have become a key factor restricting the in-depth development of artificial intelligence technology based on data-driven modeling in areas such as the remaining life of industrial equipment.

[0004] In the existing technology, researchers have proposed many solutions to the problem of enhancing the remaining life data of industrial equipment:

[0005] (1) Time-frequency transformation (TF). The TF method extracts the periodicity and trend characteristics of the original time series data by mapping the time domain information to the frequency domain. However, when processing high-dimensional time series data, it cannot fully capture the dynamic feature changes of the data, which limits the ability to generate high-quality and effective enhanced data.

[0006] (2) VAE. The VAE method is a deep latent space generative model. The encoder can compress the features of the original time series data into the latent space for learning, and then reconstruct the output similar to the original data through the decoder. However, as an unsupervised learning model, its processing ability is slightly limited when dealing with high-dimensional and complex time series data enhancement tasks, and there is a certain gap in performance compared with supervised learning models.

[0007] (3) Generative Adversarial Network (GAN). GAN is a deep learning model that restores the feature distribution of real data to the greatest extent and enhances data through the dynamic game between the generative network and the discriminative network. The GAN model has shown remarkable results in computer vision. However, unlike image data, in the field of the remaining life of industrial equipment, the data has a more complex time dependency, and GAN has difficulty capturing the correlation between time features and attributes. In addition, the long-term correlation of time series data and its dimensional uncertainty increase the difficulty of data generation. On the other hand, GAN itself has shortcomings such as non-convergence, gradient vanishing, and mode collapse, which makes the model training more difficult.

[0008] Through the above analysis, the problems and defects of the prior art are as follows:

[0009] (1) Current technologies mainly focus on enhancement strategies for imbalanced data, ignoring the common phenomenon of missing data in real application scenarios.

[0010] (2) They are not very effective in processing complex high-dimensional time series data, and there are still deficiencies in the model construction and training process. In particular, in the data augmentation tasks in the field of remaining life of industrial equipment, which are unbalanced and contain missing values, existing methods are difficult to provide effective and reliable solutions.

[0011] (3) Deep learning technologies represented by GAN models have shown excellent performance in dealing with data augmentation problems and have become a current research hotspot. However, the GAN model has different methods for measuring the similarity between real data and generated data, as well as differences in network architecture design or loss function selection, which have led to phenomena such as gradient vanishing and gradient exploding during model training, which has hindered the convergence process of the generated network and further affected the synthesis of effective data. Summary of the invention

[0012] In view of the problems existing in the prior art, the present invention provides a DTSIGAN data enhancement method for the remaining life of industrial equipment under unbalanced missing data.

[0013] The present invention is implemented as follows: a DTSIGAN data enhancement method for the remaining life of industrial equipment under unbalanced missing data includes:

[0014] Step 1: preprocess the unbalanced and missing industrial equipment data collected by the sensor, including data normalization, constructing a mask matrix and a time interval matrix, and splicing them with the original time series to form a data set, which is divided into a training set and a test set in proportion;

[0015] Step 2, by integrating the principle of time dynamics with the structure of the time series generative adversarial network; introducing the cyclic dynamics interpolation algorithm to optimize the embedding network model; the improved strategy maps the real data and the captured missing data features to the latent feature space, and then uses the recovery network to reconstruct the original data based on the features in this space; at the same time, in the loss function design, the original data, the restored data and the missing interpolation data are included in the error calculation;

[0016] Step 3: Random noise is input into the generative network and mapped to the generative feature space; the discriminant network distinguishes true from false in the latent space, and the generative network is iteratively optimized and updated based on the feedback of the discriminant network;

[0017] Step 4: Replace the traditional cross entropy function between the generator and the discriminator with the Wasserstein distance, and introduce a gradient penalty term to impose a Lipschitz constraint while measuring the similarity between the real data and the synthetic data distribution.

[0018] Step 5: Use the trained DTSIGAN to dynamically interpolate and expand the unbalanced missing data set, and then use the prediction model to predict the remaining life of industrial equipment.

[0019] Furthermore, in step 1, the raw data collected by the sensor is mathematically described as a multivariate time series set with a time step of T and a dimension of D, where X = {x1, x2, x3, ..., x T}∈R T×D ; Among them, step t Because there are missing values ​​in the time series, in order to better mark the location of the missing variables in X, the mask matrix M∈R is introduced T×D To express:

[0020]

[0021] in, represents the d-th dimension variable at the t-th step, that is, the d-th eigenvalue at time t;

[0022] For the timestamp vector S = {s1, s2, s3, ..., s T}∈R T , introduce the time interval matrix δ to reflect the time lag distribution between data points; define It represents the time interval between the value observed at time t and the last non-missing observation value, that is:

[0023]

[0024] Furthermore, in step 2, the embedding network E inputs the real data X obtained in step 1, the mask matrix M and the time interval matrix δ, extracts the features of the time series data, and maps them to the potential feature space H; then, the recovery network R restores these features to the original data space and imputed data The reconstruction loss function L is constructed by comparing the differences between the real data, the restored data and the interpolated data. R , and use this to train the embedding network and recovery network; the relevant formula is as follows:

[0025] H=E(X,M,δ)

[0026]

[0027] Among them, the embedding network E is improved and optimized by the cyclic dynamics interpolation algorithm, and the recovery network R is composed of an autoregressive network architecture that obeys the causal order, consisting of multiple layers of GRU units and a fully connected layer.

[0028] Furthermore, the embedding network E will save the hidden state h of the information before time t-1 t-1 Mapped to the predicted value at time t Using the mask matrix obtained in step 1, Replace x t The corresponding missing value in gets the value at time t after filling The relevant calculation formula is as follows:

[0029]

[0030] Among them, W x 、b x is the parameter to be learned by the model, h t-1 is the hidden state at time t-1, the initial hidden state h0 is a vector of all zeros, m t Represents the mask matrix vector corresponding to time t, x t is the real data at time t, ⊙ represents the dot multiplication operation;

[0031] By introducing the time decay factor γ t , this factor shows a decay effect over time to adjust the weight of historical information in the current state; the calculation formula of the time decay function is as follows:

[0032] γ t =exp{-max(0,W γ δ t +b γ )}

[0033] Among them, W γ 、b γ is the corresponding parameter; δ t The larger the value, the larger the data missing time lag distance. t The smaller the t The smaller the data is, the smaller the time lag distance is. t The bigger;

[0034] Based on the attenuation factor γ t With the previous hidden state h t-1 , calculate the hidden state h at time t t ; The relevant calculation formula is as follows:

[0035]

[0036] In the formula, σ is the activation function, Sigmoid is selected as the activation function, W h , U h and b h are the weights and bias parameters that the model needs to learn.

[0037] Further, in step 3, the generating network G maps the random noise Z to the generating feature space The discriminant network D distinguishes true from false in the feature space dimension; both the generative network G and the discriminant network D are composed of multiple layers of GRU units and a fully connected layer.

[0038] In step 4, Wasserstein distance and gradient penalty terms are introduced to ensure that the model will not fail to converge due to gradient explosion or gradient disappearance, thereby improving the stability of the model; the loss function calculation formula between the generating network G and the discriminating network D is as follows:

[0039]

[0040] Where P x is the distribution satisfied by the real data x, P z is the distribution satisfied by random noise z, λ is the penalty weight factor, Indicates that the discriminant network D is in the sample The second normal form of the gradient at Randomly sample from the sample space of synthetic data y and real data x, for The distribution satisfied; the relevant calculation formula is as follows:

[0041]

[0042] Among them, ∈ is a random number in the range [0,1] that follows a uniform distribution;

[0043] Step 5: Use the trained DTSIGAN to dynamically interpolate and expand the unbalanced missing data set, mix the synthetic data with the real data as the data set for the prediction task, and then use the prediction model to predict the remaining life of industrial equipment. The mean absolute error MAE, mean absolute percentage error MAPE and root mean square error RMSE are used as evaluation indicators to evaluate the prediction performance and verify the effect of the DTSIGAN method on the enhancement of unbalanced missing data. The relevant calculation formula is as follows:

[0044]

[0045] In the formula, n represents the sample size, p i represents the true label, Represents the predicted value.

[0046] Another object of the present invention is to provide a DTSIGAN data enhancement system for the remaining life of industrial equipment under unbalanced missing data, comprising:

[0047] The data preprocessing module is used to preprocess the unbalanced and missing industrial equipment data collected by sensors, including data normalization, construction of mask matrix and time interval matrix, and splicing with the original time series to form a data set, which is divided into training set and test set in proportion;

[0048] The adversarial network generation module is used to generate the structure of the adversarial network by integrating the principle of time dynamics with time series; the cyclic dynamics interpolation algorithm is introduced to optimize the embedded network model; the improved strategy maps the real data and the captured missing data features to the latent feature space, and then uses the recovery network to reconstruct the original data based on the features in this space; at the same time, in the loss function design, the original data, the restored data and the missing interpolation data are included in the error calculation;

[0049] The optimization and update module is used to input random noise into the generative network and map it to the generative feature space; the discriminant network distinguishes true from false in the latent space, and based on the feedback of the discriminant network, the generative network can be iteratively optimized and updated;

[0050] The replacement module is used to replace the traditional cross entropy function between the generator and the discriminator with the Wasserstein distance, which measures the similarity between the real data and the synthetic data distribution while introducing a gradient penalty term to impose the Lipschitz constraint;

[0051] The life prediction module is used to dynamically interpolate and expand the unbalanced missing data set using the trained DTSIGAN, and then use the prediction model to predict the remaining life of industrial equipment.

[0052] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the DTSIGAN data enhancement method for the remaining life of industrial equipment under the unbalanced missing data.

[0053] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the DTSIGAN data enhancement method for the remaining life of industrial equipment under the unbalanced missing data.

[0054] Another object of the present invention is to provide an information data processing terminal, which is used to implement the DTSIGAN data enhancement system for the remaining life of industrial equipment under the unbalanced missing data.

[0055] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0056] First, the present invention can solve the problems of small data set size, unbalanced data distribution and missing values ​​in the task of predicting the remaining life of industrial equipment. By providing a solid data foundation, it promotes the application efficiency of data-driven artificial intelligence technology in this field, effectively exerts the representation performance and self-learning ability of the model, and thus can complete the task of predicting the remaining life of industrial equipment more accurately and efficiently.

[0057] This paper adopts the cyclic dynamics interpolation algorithm to improve the model structure of the time series generative adversarial network. By introducing multiple mechanisms such as mask matrix, time interval matrix and time decay factor, the missing value related features are captured and analyzed, and the missing value interpolation process is integrated into the optimization process of the time series data generation model to achieve dynamic interpolation of missing values, thereby improving the model's ability to generate unbalanced missing time series data.

[0058] The present invention combines Wasserstein distance and gradient penalty term, and optimizes and replaces the traditional loss functions of the generative network and the discriminative network. This improved strategy not only improves the difference measurement between the distribution of synthetic data and real data, but also prevents the phenomenon of gradient vanishing and gradient exploding during training, thereby enhancing the stability and reliability of the model.

[0059] Second, the expected benefits and commercial value after the transformation of the technical solution of the present invention are as follows: The present invention provides a DTSIGAN data enhancement method for the remaining life of industrial equipment under unbalanced missing data, which can solve the problems of small data set size, unbalanced data distribution and missing values ​​caused by factors such as time or economic cost in the task of predicting the remaining life of industrial equipment. The present invention provides a solid data foundation, promotes the application efficiency of data-driven artificial intelligence technology in this field, effectively exerts the representation performance and self-learning ability of artificial intelligence models, enables artificial intelligence models to perform the task of predicting the remaining life of industrial equipment more accurately and efficiently, and provides strong support for industrial maintenance and decision-making.

[0060] The technical solution of the present invention solves the technical problem that people have been eager to solve but have never succeeded in solving: the present invention provides a DTSIGAN data enhancement method for the remaining life of industrial equipment under unbalanced missing data, and innovatively optimizes the model structure of the time series generative adversarial network by adopting a cyclic dynamics interpolation algorithm. The introduction of multiple mechanisms such as mask matrix, time interval matrix and time decay factor aims to capture and analyze the characteristics related to missing values, integrate the missing value interpolation process into the optimization process of the time series data generation model, and realize the dynamic interpolation of missing values; at the same time, in order to further enhance the stability of the model and the effectiveness of data synthesis, the present invention combines the Wasserstein distance with the gradient penalty term, optimizes the loss function of the generation network and the discriminant network, effectively avoids the problem of gradient vanishing and gradient explosion, and ensures the stability of the model training process, thereby greatly improving the ability to synthesize high-quality time series data.

[0061] Third, an embodiment of the present invention provides a method for predicting the remaining life of industrial equipment under unbalanced missing data based on virtual reality. First, in step S101, the sensor data is collected and mathematically transformed into a multivariate time series set X with a time step of T and a dimension of D. Since there are missing values ​​in the time series, in order to better mark the location of the missing data, a mask matrix M is constructed, and a time interval matrix δ is introduced to reflect the time delay distribution between data points. The mask matrix uses 1 and 0 to respectively identify whether the data is missing or not, and the time interval matrix is ​​used to calculate the time interval between adjacent non-missing values. This step lays the data foundation for subsequent processing.

[0062] In step S102, the preprocessed data (including real data X, mask matrix M and time interval matrix δ) is input into the embedding network E, and the time series features are extracted and mapped to the latent feature space H. The restoration network R uses these latent features to reconstruct the original data X~ and the interpolated data Y~, and constructs the reconstruction loss function L_R by comparing the differences between the real data and the restored data and the interpolated data, so as to train the embedding network and the restoration network. The network structure combines the cyclic dynamics interpolation algorithm, uses multi-layer GRU units and fully connected layers, and can efficiently handle the missing problem of sequence data.

[0063] In step S103, the generator network G generates synthetic data similar to real data by inputting random noise Z and mapping it to the generated feature space H~. The discriminant network D discriminates the authenticity of the data in the latent feature space and distinguishes synthetic data from real data. Both the generator network and the discriminant network are composed of multi-layer GRU units and fully connected layers, and the model is gradually optimized through adversarial training.

[0064] In step S104, Wasserstein distance and gradient penalty terms are introduced to ensure that the model does not experience gradient explosion or gradient disappearance during training, thereby improving the stability of the model. The loss function between the generator network and the discriminator network is calculated using the second normal form of the gradient to effectively control the data generation and discrimination process, ensuring that the generated data has a high similarity to the real data distribution. Finally, through step S105, the trained DTSIGAN model is used to dynamically interpolate the unbalanced missing data, and combined with the remaining life prediction model of industrial equipment, the accuracy of the prediction is further improved to ensure the enhancement effect in the case of unbalanced data. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a flow chart of the DTSIGAN data enhancement method for the remaining life of industrial equipment under unbalanced missing data provided by an embodiment of the present invention.

[0066] Figure 2 It is a system structure block diagram of the DTSIGAN data enhancement method for the remaining life of industrial equipment under unbalanced missing data provided by an embodiment of the present invention.

[0067] Figure 3 Detailed flow chart of the method for data enhancement of unbalanced missing data provided by an embodiment of the present invention.

[0068] Figure 4 An embedded network structure diagram provided by an embodiment of the present invention.

[0069] Figure 5 Structural diagram of the data enhancement model DTSIGAN provided in an embodiment of the present invention.

[0070] Figure 6 An embodiment of the present invention provides a visualization comparison diagram of data synthesized by different models in an embodiment.

[0071] Figure 7 A comparison chart of prediction results of synthetic data using different models in an embodiment provided by the present invention. DETAILED DESCRIPTION

[0072] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0073] like Figure 1 As shown, a DTSIGAN data enhancement method for the remaining life of industrial equipment under unbalanced missing data provided by an embodiment of the present invention includes the following steps:

[0074] S101, preprocessing the unbalanced and missing industrial equipment data collected by sensors, including data normalization, constructing a mask matrix and a time interval matrix, and splicing them with the original time series to form a data set, which is divided into a training set and a test set in proportion;

[0075] S102, by integrating the principle of time dynamics with the structure of time series generative adversarial network; introducing the cyclic dynamics interpolation algorithm to optimize the embedding network model; improving the strategy to map the real data and the captured missing data features into the latent feature space, and then using the recovery network to reconstruct the original data based on the features in this space; at the same time, in the loss function design, the original data, the restored data and the missing interpolation data are all included in the error calculation;

[0076] S103, by inputting random noise into the generative network, mapping it to the generative feature space; the discriminative network distinguishes true from false in the latent space, and based on the feedback of the discriminative network, the generative network is iteratively optimized and updated;

[0077] S104, the traditional cross entropy function between the generative network and the discriminative network is replaced by the Wasserstein distance, and a gradient penalty term is introduced to impose a Lipschitz constraint while measuring the similarity between the real data and the synthetic data distribution;

[0078] S105, using the trained DTSIGAN to dynamically interpolate and expand the unbalanced missing data set, and then using the prediction model to predict the remaining life of industrial equipment.

[0079] The method first performs data preprocessing on the unbalanced missing industrial equipment data collected by sensors. The data preprocessing step includes data normalization so that data with different features can be compared on the same scale. Then, a mask matrix and a time interval matrix are constructed to identify the missing parts in the data and represent the time intervals between various time points. Finally, these matrices are concatenated with the original time series data into a complete dataset and divided into training and test sets in proportion for subsequent model training and validation.

[0080] Then, the method introduces the principle of time dynamics and the structure of time series generative adversarial network (DTSIGAN), and optimizes the embedded network model through the cyclic dynamics interpolation algorithm. This improved strategy maps the real data and missing data features captured by the sensor to the latent feature space, and the recovery network reconstructs the original data based on the features in this space. In this process, the loss function design is very important, especially the errors of the original data, the restored data and the missing interpolation data are included in the calculation together to improve the accuracy of data recovery.

[0081] The generative network inputs random noise, maps it into the generated feature space, and optimizes it based on the feedback of the discriminative network. The discriminative network distinguishes the generated fake data from the real data in the latent space, and promotes the continuous iteration and optimization of the generative network through this feedback mechanism. The generative network gradually learns more realistic data distribution characteristics, thereby generating more realistic missing data and filling the missing parts of the data set.

[0082] In order to improve the quality of data generation, this method replaces the traditional cross entropy loss function between the generative network and the discriminative network with the Wasserstein distance, which can more accurately measure the similarity between the real data and the generated data distribution. At the same time, the method also introduces a gradient penalty term and imposes a Lipschitz constraint to ensure the smoothness and stability of the generative network. Finally, the trained DTSIGAN is used to dynamically interpolate and expand the unbalanced missing data set, which improves the integrity and balance of the data. Finally, the remaining life of industrial equipment is accurately predicted by combining the prediction model.

[0083] S101 provided in the embodiment of the present invention mathematically describes the raw data collected by the sensor as a multivariate time series set with a time step of T and a dimension of D, where X = {x1, x2, x3, ..., x T}∈R T×D ; Among them, step t Because there are missing values ​​in the time series, in order to better mark the location of the missing variables in X, the mask matrix M∈R is introduced T×D To express:

[0084]

[0085] in, represents the d-th dimension variable at the t-th step, that is, the d-th eigenvalue at time t;

[0086] For the timestamp vector S = {s1, s2, s3, ..., s T}∈R T , introduce the time interval matrix δ to reflect the time lag distribution between data points; define It represents the time interval between the value observed at time t and the last non-missing observation value, that is:

[0087]

[0088] In S102 provided in the embodiment of the present invention, the embedding network E inputs the real data X obtained in step 1, the mask matrix M and the time interval matrix δ, extracts the features of the time series data, and maps them to the potential feature space H; then, the recovery network R restores these features to the original data space and imputed data The reconstruction loss function L is constructed by comparing the differences between the real data, the restored data and the interpolated data. R , and use this to train the embedding network and recovery network; the relevant formula is as follows:

[0089] H=E(X,M,δ)

[0090]

[0091] Among them, the embedding network E is improved and optimized by the cyclic dynamics interpolation algorithm, and the recovery network R is composed of an autoregressive network architecture that obeys the causal order, consisting of multiple layers of GRU units and a fully connected layer.

[0092] The embedding network E provided in the embodiment of the present invention saves the hidden state h of the information before time t-1 t-1 Mapped to the predicted value at time t Using the mask matrix obtained in S101, Replace x t The corresponding missing value in gets the value at time t after filling The relevant calculation formula is as follows:

[0093]

[0094]

[0095] Among them, W x 、b x is the parameter to be learned by the model, h t-1is the hidden state at time t-1, the initial hidden state h0 is a vector of all zeros, m t Represents the mask matrix vector corresponding to time t, x t is the real data at time t, ⊙ represents the dot multiplication operation;

[0096] By introducing the time decay factor γ t , this factor shows a decay effect over time to adjust the weight of historical information in the current state; the calculation formula of the time decay function is as follows:

[0097] γ t =exp{-max(0,W γ δ t +b γ )}

[0098] Among them, W γ 、b γ is the corresponding parameter; δ t The larger the value, the larger the data missing time lag distance. t The smaller the t The smaller the data is, the smaller the time lag distance is. t The bigger;

[0099] Based on the attenuation factor γ t With the previous hidden state h t-1 , calculate the hidden state h at time t t ; The relevant calculation formula is as follows:

[0100]

[0101] Wherein, σ is the activation function. In this embodiment, Sigmoid is selected as the activation function. h , U h and b h are the weights and bias parameters that the model needs to learn.

[0102] S103 provided in the embodiment of the present invention, the generating network G maps the random noise Z to the generating feature space The discriminant network D distinguishes true from false in the feature space dimension; both the generative network G and the discriminant network D are composed of multiple layers of GRU units and a fully connected layer.

[0103] S104 provided in the embodiment of the present invention introduces Wasserstein distance and gradient penalty term to ensure that the model will not fail to converge the generating network due to gradient explosion or gradient disappearance, thereby improving the stability of the model; the loss function calculation formula between the generating network G and the discriminating network D is as follows:

[0104]

[0105] Where P x is the distribution satisfied by the real data x, P z is the distribution satisfied by random noise z, λ is the penalty weight factor, Indicates that the discriminant network D is in the sample The second normal form of the gradient at Randomly sample from the sample space of synthetic data y and real data x, for The distribution satisfied; the relevant calculation formula is as follows:

[0106]

[0107] Among them, ∈ is a random number in the range [0,1] that follows a uniform distribution;

[0108] S105, use the trained DTSIGAN to dynamically interpolate and expand the unbalanced missing data set, mix the synthetic data with the real data as the data set for the prediction task, and then use the prediction model to predict the remaining life of industrial equipment. The mean absolute error MAE, mean absolute percentage error MAPE and root mean square error RMSE are used as evaluation indicators to evaluate the prediction performance and verify the effect of the DTSIGAN method on the enhancement of unbalanced missing data. The relevant calculation formula is as follows:

[0109]

[0110] In the formula, n represents the sample size, p i represents the true label, Represents the predicted value.

[0111] like Figure 2 As shown, a DTSIGAN data enhancement system for the remaining life of industrial equipment under unbalanced missing data provided by an embodiment of the present invention includes:

[0112] The data preprocessing module is used to preprocess the unbalanced and missing industrial equipment data collected by sensors, including data normalization, construction of mask matrix and time interval matrix, and splicing with the original time series to form a data set, which is divided into training set and test set in proportion;

[0113] The adversarial network generation module is used to generate the structure of the adversarial network by integrating the principle of time dynamics with time series; the cyclic dynamics interpolation algorithm is introduced to optimize the embedded network model; the improved strategy maps the real data and the captured missing data features to the latent feature space, and then uses the recovery network to reconstruct the original data based on the features in this space; at the same time, in the loss function design, the original data, the restored data and the missing interpolation data are included in the error calculation;

[0114] The optimization and update module is used to input random noise into the generative network and map it to the generative feature space; the discriminant network distinguishes true from false in the latent space, and based on the feedback of the discriminant network, the generative network can be iteratively optimized and updated;

[0115] The replacement module is used to replace the traditional cross entropy function between the generator and the discriminator with the Wasserstein distance, which measures the similarity between the real data and the synthetic data distribution while introducing a gradient penalty term to impose the Lipschitz constraint;

[0116] The life prediction module is used to dynamically interpolate and expand the unbalanced missing data set using the trained DTSIGAN, and then use the prediction model to predict the remaining life of industrial equipment.

[0117] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the DTSIGAN data enhancement method for the remaining life of industrial equipment under the unbalanced missing data.

[0118] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the DTSIGAN data enhancement method for the remaining life of industrial equipment under the unbalanced missing data.

[0119] Another object of the present invention is to provide an information data processing terminal, which is used to implement the DTSIGAN data enhancement system for the remaining life of industrial equipment under the unbalanced missing data.

[0120] The present invention is specifically implemented:

[0121] like Figure 3 As shown, the DTSIGAN data enhancement method for the remaining life of industrial equipment under unbalanced missing data provided by an embodiment of the present invention includes the following steps:

[0122] Step 1: Perform data normalization to eliminate dimensional differences between data; then, construct a mask matrix to clearly identify the missing parts in the data, and construct a time interval matrix to reflect the impact of these missing parts on the integrity of the time series; next, effectively splice these matrices with the original time series data to form a comprehensive data set that contains both complete information and reflects the missing characteristics; finally, according to the scientific principle of proportional allocation, divide the data set into training set and test set for model training and performance evaluation.

[0123] Step 2: By integrating the principle of time dynamics with the structure of the time series generative adversarial network, the present invention introduces a cyclic dynamics interpolation algorithm to optimize the embedded network model. This improved strategy maps the real data and the captured missing data features to the latent feature space, and then uses the recovery network to reconstruct the original data based on the features in this space. At the same time, in the loss function design, the original data, the restored data, and the missing interpolation data are included in the error calculation, which not only realizes the dynamic filling of the missing parts in the original data, but also promotes the construction of the latent feature space, so that it can show a richer and deeper representation ability for the original data.

[0124] Step 3: Input random noise into the generative network and map it to the generative feature space. Then, the discriminant network distinguishes true from false in the latent space. Based on the feedback from the discriminant network, the generative network can be iteratively optimized and updated to continuously improve the quality and authenticity of its synthetic data.

[0125] Step 4: Replace the traditional cross entropy function between the generator and the discriminator with the Wasserstein distance, and introduce a gradient penalty term to impose the Lipschitz constraint while measuring the similarity between the real data and the synthetic data distribution.

[0126] Step 5: By using the trained DTSIGAN model, dynamic interpolation and data expansion are performed on the unbalanced data set with missing values. Then, the prediction model is used to complete the task of predicting the remaining life of industrial equipment based on the expanded data set.

[0127] Furthermore, in step 1, the raw data collected by the sensor is mathematically described as a multivariate time series set with a time step of T and a dimension of D, denoted by X = {x1, x2, x3, ..., x T}∈R T×D Among them, the tth step Because there are missing values ​​in the time series, in order to better mark the location of the missing variables in X, the present invention introduces the mask matrix M∈R T×D To express:

[0128]

[0129] in, Represents the d-th dimensional variable at the t-th step, that is, the d-th eigenvalue at time t.

[0130] In actual working conditions, we often encounter the situation where the features of continuous timestamps are missing. This phenomenon is often accompanied by the trend that the mutual influence between the features gradually weakens as the time interval lengthens, that is, the time lag correlation phenomenon. Therefore, for the timestamp vector S = {s1, s2, s3, ..., s T}∈R T , the present invention introduces the time interval matrix δ to reflect the time lag distribution between data points. Definition It represents the time interval between the value observed at time t and the last non-missing observation value, that is:

[0131]

[0132] Furthermore, in step 2, the embedding network E inputs the real data X obtained in step 1, the mask matrix M and the time interval matrix δ, extracts the features of the time series data, and maps them to the latent feature space H. Subsequently, the recovery network R restores these features to the original data space and imputed data The reconstruction loss function L is constructed by comparing the differences between the real data, the restored data and the interpolated data. R , and use this to train the embedding network and recovery network. The relevant formula is as follows:

[0133] H=E(X,M,δ)

[0134]

[0135] The embedding network E is improved and optimized by the cyclic dynamics interpolation algorithm, and the recovery network R is composed of a network architecture that is autoregressive and obeys the causal order. In the embodiment, it is composed of multiple layers of GRU units and one fully connected layer.

[0136] Furthermore, the embedding network E will save the hidden state h of the information before time t-1 t-1 Mapped to the predicted value at time t Using the mask matrix obtained in step 1, Replace x t The corresponding missing value in gets the value at time t after filling The relevant calculation formula is as follows:

[0137]

[0138] Among them, W x 、b x is the parameter to be learned by the model, ht-1 is the hidden state at time t-1, the initial hidden state h0 is a vector of all zeros, m t Represents the mask matrix vector corresponding to time t, x t is the real data at time t, and ⊙ represents the dot product operation.

[0139] At the same time, considering the randomness of missing information in the original time series, that is, irregular sampling occurs. As time goes by, the impact of historical information on the current state usually weakens. By introducing the time decay factor γ t , this factor shows a decay effect over time to adjust the weight of historical information in the current state. The calculation formula related to the time decay function is as follows:

[0140] γ t =exp{-max(0,W γ δ t +b γ )}

[0141] Among them, W γ 、b γ is the corresponding parameter. t The larger the value, the larger the data missing time lag distance. t The smaller the t The smaller the data is, the smaller the time lag distance is. t That is, the farther away the historical information is from the current time point, the lower its weight is through the attenuation effect of γ, so that the historical data can be used more reasonably for inference.

[0142] Based on the attenuation factor γ t With the previous hidden state h t-1 , we can calculate the hidden state h at time t t The relevant calculation formula is as follows:

[0143]

[0144] Wherein, σ is the activation function. In this embodiment, Sigmoid is selected as the activation function. h , U h and b h are the weights and bias parameters that the model needs to learn.

[0145] Furthermore, in step 3, the generator network G maps the random noise Z to the generated feature space The discriminant network D distinguishes true from false in the feature space dimension. In the embodiment, both the generator network G and the discriminant network D are composed of multiple layers of GRU units and one fully connected layer.

[0146] Furthermore, in step 4, in the traditional GAN ​​model, JS divergence is used to measure the similarity distance between the distribution of real data and synthetic data. However, the characteristic of JS divergence is that when the difference between the two distributions is large, its value tends to be constant, which will cause the gradient to disappear and make the generated network unable to learn. In addition, the generated network will also fail to converge due to gradient explosion caused by improper initialization weight setting, model structure and other reasons.

[0147] The present invention introduces Wasserstein distance and gradient penalty term to ensure that the model will not fail to converge due to gradient explosion or gradient disappearance, thereby improving the stability of the model. The loss function calculation formula between the generating network G and the discriminating network D is as follows:

[0148]

[0149] Where P x is the distribution satisfied by the real data x, P z is the distribution satisfied by random noise z, λ is the penalty weight factor, Indicates that the discriminant network D is in the sample The second normal form of the gradient at Randomly sample from the sample space of synthetic data y and real data x, for The distribution satisfied. The relevant calculation formula is as follows:

[0150]

[0151] Here, ∈ is a random number in the range [0,1] that follows a uniform distribution.

[0152] Furthermore, in step 5, the trained DTSIGAN is used to dynamically interpolate and expand the unbalanced missing data set, and the synthetic data is mixed with the real data as the data set for the prediction task. The prediction model is then used to predict the remaining life of industrial equipment, and the mean absolute error MAE, mean absolute percentage error MAPE and root mean square error RMSE are used as evaluation indicators to evaluate the prediction performance, thereby verifying the effect of the DTSIGAN method on the enhancement of unbalanced missing data. The relevant calculation formula is as follows:

[0153]

[0154] In the formula, n represents the sample size, p i represents the true label, Represents the predicted value.

[0155] Figure 4 An embedded network structure diagram provided by an embodiment of the present invention.

[0156] Figure 5 Structural diagram of the data enhancement model DTSIGAN provided in an embodiment of the present invention.

[0157] Application examples: In order to prove the creativity and technical value of the technical solution of the present invention, this section provides application examples of the claimed technical solution on specific products or related technologies.

[0158] To verify the effectiveness and practicality of the DTSIGAN method proposed in the present invention in the problem of unbalanced missing data expansion. This embodiment conducts experiments on public data sets. At the same time, three popular improved algorithms are selected for experimental results comparison. By comparing the performance of each method, the reliability and stability of the method proposed in the present invention are verified. The comparison methods include: 1) Raw-Data (original data, only used for prediction experiments), 2) C-RNN-GAN, 3) TimeGAN, 4) RCGAN.

[0159] The evaluation indicators for the experimental results include:

[0160] Visualization: Principal component analysis (PCA) and t-distribution-stochastic neighbor embedding (t-SNE) are used to perform dimensionality reduction and visualization analysis on the data, so as to more intuitively observe the differences between real data and synthetic data;

[0161] Prediction: The data synthesized by the above methods are used as the prediction task dataset, and then prediction experiments are carried out on the synthetic datasets respectively. The effectiveness of the synthetic data is verified by comparing the prediction performance.

[0162] The hardware environment configuration of all experiments involved in this invention is as follows: the CPU uses 11th Gen Intel Core i7-11800H@2.30GHZ, the running memory is 16GB, and the graphics card is NVIDIA GeForce RTX 3060Laptop. The software environment is programmed in Python language, the compilation environment is Python3.8, and the tools relied on include PyTorch, Numpy, Pandas, and Scipy.

[0163] The PCA and t-SNE visualization analysis of the experimental results are shown in Figure 6 As shown in the figure, the blue color is the distribution of synthetic data features, and the red color is the distribution of real data features. It can be seen from the figure that the data synthesized by the method proposed in the present invention is closer to the real distribution. The experimental results are shown in Table 1 Synthetic data prediction experimental results. The experimental results show that compared with other methods, the data generated by the DTSIGAN method proposed in the present invention is more accurate in the prediction task, and the effectiveness of the synthetic data is better.

[0164] Table 1. Experimental results of synthetic data prediction

[0165]

[0166] The DTSIGAN data enhancement method for the remaining life of industrial equipment under unbalanced missing data proposed in the present invention can be used for data enhancement in the industrial field. For example, an industrial equipment remaining life prediction system. In an industrial production environment, the efficient and stable operation of equipment is crucial to ensuring production continuity and improving production efficiency. However, as the equipment is used for a longer time, its performance will gradually decline until it eventually fails. Therefore, accurately predicting the remaining life of industrial equipment so that maintenance or replacement can be arranged in a timely manner is of great significance for reducing downtime and reducing maintenance costs. The method of the present invention is applied to the remaining life prediction system of industrial equipment, which can solve the problem of decreased prediction accuracy caused by incomplete or missing historical data of the equipment.

[0167] The system first collects the operating data of the equipment, including but not limited to sensor data such as temperature, vibration, current, voltage, as well as maintenance records, fault reports, etc. Subsequently, using the method of the present invention, the system can automatically analyze the missing patterns and distribution characteristics in the data, and synthesize high-quality synthetic data that conforms to the real distribution. Based on the enhanced data set, the remaining life prediction system for industrial equipment can build a more accurate and reliable prediction model. By applying the remaining life prediction system for industrial equipment, enterprises can achieve accurate management of equipment status, thereby optimizing maintenance plans, reducing unplanned downtime, reducing maintenance costs, and improving overall production efficiency. In addition, the system also helps enterprises better plan equipment update and upgrade strategies to ensure the continuous and stable operation of production lines.

[0168] The DTSIGAN data enhancement method for the remaining life of industrial equipment under unbalanced missing data proposed in the present invention adopts a cyclic dynamics interpolation algorithm to innovatively optimize the time series generative adversarial network model, integrates the mask matrix, time interval matrix and attenuation factor, dynamically captures the missing value characteristics, and realizes the synchronization of interpolation and model optimization. The Wasserstein distance and gradient penalty are combined to optimize the loss function, avoid gradient problems, enhance model stability and data synthesis quality, and significantly improve the ability to generate high-quality time series data. It can be seen from the PCA and t-SNE visualization analysis results in Application Example 1 that the data synthesized by the method proposed in the present invention is closer to the real distribution, effectively enriching the sample size of unbalanced missing data; further, the prediction experiment results using the synthetic data in Application Example 1 show that the method proposed in the present invention can synthesize a large amount of high-quality time series data, which are used to train more efficient prediction models and achieve more accurate predictions, providing strong support for intelligent analysis and decision-making in the industrial field.

[0169] Example 1: Remaining Life Prediction in Industrial Equipment Maintenance

[0170] In industrial fields such as oil and gas, the remaining life prediction of equipment is crucial to production efficiency and safety. The operating data of industrial equipment collected by sensors is often incomplete and there are data missing, which leads to a decrease in the accuracy of the prediction model. This invention uses the DTSIGAN method to dynamically interpolate and expand the missing data to improve the integrity of the data. By preprocessing the multivariate time series data collected by the sensor and applying a generative adversarial network to fill in the missing parts, the remaining life prediction model is further trained in combination with the generated complete data, thereby improving the prediction accuracy. This method can significantly reduce downtime during equipment maintenance, optimize the maintenance cycle of the equipment, and extend the service life of the equipment.

[0171] Example 2: Equipment health management in smart grid

[0172] In the equipment management of smart grids, the health status prediction of key equipment such as transformers is crucial to the stable operation of the power grid. However, due to the limitations of data collection, sensors cannot accurately obtain the operating status of equipment at all times, resulting in a large amount of data missing. The DTSIGAN method of this invention dynamically interpolates these unbalanced and missing data through a generative adversarial network to generate a high-quality equipment operation data set. By analyzing the complete data after generation, the smart grid management system can more accurately predict the health status of the equipment and prevent power outages caused by equipment failures. This application can significantly improve the reliability and safety of power grid operation while reducing maintenance costs.

[0173] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0174] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A DTSIGAN data enhancement method for the remaining life of industrial equipment under unbalanced missing data, characterized in that: The following steps are involved: Step 1, preprocessing the unbalanced and missing industrial equipment data collected by the sensor; This includes data normalization, constructing mask matrices and time interval matrices, and concatenating them with the original time series to form a data set, which is then divided into training and test sets in proportion; Step 2, generate the structure of adversarial network by integrating the principle of time dynamics and time series; A cyclic dynamics interpolation algorithm is introduced to optimize the embedding network model. The improved strategy maps the real data and the captured missing data features into the latent feature space, and then uses the recovery network to reconstruct the original data based on the features in this space. At the same time, in the loss function design, the original data, the restored data and the missing interpolation data are all included in the error calculation. Step 3, input random noise into the generative network and map it to the generative feature space; The discriminant network distinguishes true from false in the latent space. Based on the feedback from the discriminant network, the generative network can be iteratively optimized and updated. Step 4: Replace the traditional cross entropy function between the generator and the discriminator with the Wasserstein distance, and introduce a gradient penalty term to impose a Lipschitz constraint while measuring the similarity between the real data and the synthetic data distribution. Step 5: Use the trained DTSIGAN to dynamically interpolate and expand the unbalanced missing data set, and then use the prediction model to predict the remaining life of industrial equipment.

2. The DTSIGAN data enhancement method for the remaining life of industrial equipment under unbalanced missing data according to claim 1 is characterized in that: In step 1, the raw data collected by the sensor is mathematically described as a multivariate time series set with a time step length of T and a dimension of D, where X = {x1, x2, x3, ..., x T }∈R T×D ; Among them, step t Because there are missing values ​​in the time series, in order to better mark the location of the missing variables in X, the mask matrix M∈R is introduced T×D To express: in, represents the d-th dimension variable at the t-th step, that is, the d-th eigenvalue at time t; For the timestamp vector S = {s1, s2, s3, ..., s T }∈R T , introduce the time interval matrix δ to reflect the time lag distribution between data points; define It represents the time interval between the value observed at time t and the last non-missing observation value, that is:

3. The DTSIGAN data enhancement method for the remaining life of industrial equipment under unbalanced missing data according to claim 1, characterized in that: In step 2, the embedding network E inputs the real data X obtained in step 1, the mask matrix M and the time interval matrix δ, extracts the features of the time series data, and maps them to the potential feature space H; Subsequently, the recovery network R restores these features to the original data space and imputed data The reconstruction loss function L is constructed by comparing the differences between the real data, the restored data and the interpolated data. R , and use this to train the embedding network and recovery network; the relevant formula is as follows: H=E(X,M,δ) Among them, the embedding network E is improved and optimized by the cyclic dynamics interpolation algorithm, and the recovery network R is composed of an autoregressive network architecture that obeys the causal order, consisting of multiple layers of GRU units and a fully connected layer.

4. The DTSIGAN data enhancement method for the remaining life of industrial equipment under unbalanced missing data as claimed in claim 3, characterized in that: The embedding network E will save the hidden state h of the information before time t-1 t-1 Mapped to the predicted value at time t Using the mask matrix obtained in step 1, Replace x t The corresponding missing value in gets the value at time t after filling The relevant calculation formula is as follows: Among them, W x , b x is the parameter to be learned by the model, h t-1 is the hidden state at time t-1, the initial hidden state h0 is a vector of all zeros, m t Represents the mask matrix vector corresponding to time t, x t is the real data at time t, ⊙ represents the dot multiplication operation; By introducing the time decay factor γ t , this factor shows a decay effect over time to adjust the weight of historical information in the current state; the calculation formula of the time decay function is as follows: c t =exp{-max(0,W γ d t +b γ )} Among them, W γ , b γ is the corresponding parameter; δ t The larger the value, the larger the data missing time lag distance. t The smaller the t The smaller the data is, the smaller the time lag distance is. t The bigger; Based on the attenuation factor γ t With the previous hidden state h t-1 , calculate the hidden state h at time t t ; The relevant calculation formula is as follows: Wherein, σ is the activation function. In this embodiment, Sigmoid is selected as the activation function. h , U h and b h are the weights and bias parameters that the model needs to learn.

5. The DTSIGAN data enhancement method for the remaining life of industrial equipment under unbalanced missing data according to claim 1, characterized in that: In step 3, the generating network G maps the random noise Z to the generating feature space The discriminant network D distinguishes true from false in the feature space dimension; both the generative network G and the discriminant network D are composed of multiple layers of GRU units and a fully connected layer.

6. The DTSIGAN data enhancement method for the remaining life of industrial equipment under unbalanced missing data according to claim 1, characterized in that: In step 4, Wasserstein distance and gradient penalty terms are introduced to ensure that the model will not fail to converge due to gradient explosion or gradient disappearance, thereby improving the stability of the model; the loss function calculation formula between the generating network G and the discriminating network D is as follows: Where P x is the distribution satisfied by the real data x, P z is the distribution satisfied by random noise z, λ is the penalty weight factor, Indicates that the discriminant network D is in the sample The second normal form of the gradient at Randomly sample from the sample space of synthetic data y and real data x, for The distribution satisfied; the relevant calculation formula is as follows: Among them, ∈ is a random number in the range [0,1] that follows a uniform distribution; Step 5: Use the trained DTSIGAN to dynamically interpolate and expand the unbalanced missing data set, mix the synthetic data with the real data as the data set for the prediction task, and then use the prediction model to predict the remaining life of industrial equipment. The mean absolute error MAE, mean absolute percentage error MAPE and root mean square error RMSE are used as evaluation indicators to evaluate the prediction performance and verify the effect of the DTSIGAN method on the enhancement of unbalanced missing data. The relevant calculation formula is as follows: In the formula, n represents the sample size, p i represents the true label, Represents the predicted value.

7. A DTSIGAN data enhancement system for the remaining life of industrial equipment under unbalanced missing data that implements the DTSIGAN data enhancement method for the remaining life of industrial equipment under unbalanced missing data as described in any one of claims 1 to 6, characterized in that: The DTSIGAN data enhancement system for the remaining life of industrial equipment under unbalanced missing data includes: The data preprocessing module is used to preprocess the unbalanced and missing industrial equipment data collected by sensors, including data normalization, construction of mask matrix and time interval matrix, and splicing with the original time series to form a data set, which is divided into training set and test set in proportion; The adversarial network generation module is used to generate the structure of the adversarial network by integrating the principle of time dynamics with time series; the cyclic dynamics interpolation algorithm is introduced to optimize the embedded network model; the improved strategy maps the real data and the captured missing data features to the latent feature space, and then uses the recovery network to reconstruct the original data based on the features in this space; at the same time, in the loss function design, the original data, the restored data and the missing interpolation data are included in the error calculation; The optimization and update module is used to input random noise into the generative network and map it to the generative feature space; the discriminant network distinguishes true from false in the latent space, and based on the feedback of the discriminant network, the generative network can be iteratively optimized and updated; The replacement module is used to replace the traditional cross entropy function between the generator and the discriminator with the Wasserstein distance, which measures the similarity between the real data and the synthetic data distribution while introducing a gradient penalty term to impose the Lipschitz constraint; The life prediction module is used to dynamically interpolate and expand the unbalanced missing data set using the trained DTSIGAN, and then use the prediction model to predict the remaining life of industrial equipment.

8. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the DTSIGAN data enhancement method for the remaining life of industrial equipment under unbalanced missing data as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the DTSIGAN data enhancement method for the remaining life of industrial equipment under unbalanced missing data as described in any one of claims 1 to 6.

10. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the DTSIGAN data enhancement system for the remaining life of industrial equipment under unbalanced missing data as described in claim 7.

Images