Industrial internet-oriented AIGC generative model training method and data generation method

Abstract

The embodiment of the invention provides an industrial internet-oriented AIGC generation model training method and a data generation method. The method comprises the following steps: determining to-be-processed state data according to an industrial time sequence data sample, a to-be-processed time step and real noise; performing dimension conversion on the to-be-processed state data to obtain to-be-processed sensor data at multiple moments; performing feature extraction through the initial diffusion model to obtain a time sequence dimension feature and a variable cross dimension feature, and outputting first prediction noise; determining predicted industrial time series data according to the first predicted noise and the to-be-processed state data; and based on the first predicted noise, the real noise, the predicted industrial time series data and the industrial time series data sample, adjusting parameters of the initial diffusion model to obtain an industrial Internet-oriented AIGC generation model. The method is used for improving the quality of the AIGC generation model.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence technology and data generation technology, and in particular to a training method and data generation method for the industrial internet. Background Technology

[0002] Training relevant models for industrial equipment requires time-series data of industrial equipment. To train high-quality relevant models, rich and high-quality industrial time-series data samples are required. However, in practical applications, it is difficult to obtain rich and high-quality industrial time-series data samples.

[0003] In related technologies, generative models are used to process a small amount of collected industrial time-series data to generate augmented data to expand the industrial time-series data sample. However, the operation of industrial equipment is complex, and generative models have difficulty learning the complex and ever-changing relationships between industrial time-series data, resulting in poor quality of the generated industrial time-series data. Summary of the Invention

[0004] This application provides an AIGC generative model training method and a data generation method for the Industrial Internet, which aims to improve the quality of the AIGC generative model and the quality of industrial augmented time-series data generated by the AIGC generative model.

[0005] In a first aspect, embodiments of this application provide a method for training an AIGC generative model for the Industrial Internet, including:

[0006] Based on industrial time-series data samples, time steps to be processed, and real noise, stochastic differential equations are used to determine the state data to be processed. The state data to be processed includes time-series data from multiple sensors. Dimensional transformation is performed on the state data to be processed to obtain sensor data at multiple time points. Through an initial diffusion model, features are extracted from the time-series data from multiple sensors, the sensor data at multiple time points, and the time steps to be processed, obtaining time-series dimensional features and variable cross-dimensional features. These features are then processed to obtain the first predictive noise. Based on the first predictive noise and the state data to be processed, the predicted industrial time-series data is determined. Based on the first predictive noise, real noise, predicted industrial time-series data, and industrial time-series data samples, the parameters of the initial diffusion model are adjusted until the initial diffusion model meets the training conditions, resulting in an AIGC generation model for the Industrial Internet.

[0007] Secondly, embodiments of this application provide a data generation method, including:

[0008] Random noise is used as the data to be processed, and the data to be processed is transformed in dimension to obtain the transformed data to be processed.

[0009] The AIGC generation model processes the data to be processed, the data to be transformed, and the time step of the data to be processed to obtain the second prediction noise; the AIGC generation model is the AIGC generation model in the training method of AIGC generation model for industrial Internet.

[0010] The drift term is determined based on the second predicted noise and the diffusion coefficient, and the diffusion term is determined based on the diffusion coefficient.

[0011] Physical sampling is performed based on the data to be processed, the drift term, and the diffusion term to obtain denoised data;

[0012] If the time step is not zero, update the data to be processed with the noise reduction number, return to perform dimensional transformation on the data to be processed, and continue the step of obtaining the transformed data to be processed.

[0013] With a time step of zero, the denoised data is used as industrial augmented time series data.

[0014] Thirdly, embodiments of this application provide an AIGC generative model training device for the Industrial Internet, comprising:

[0015] The processing module is used to determine the state data to be processed based on industrial time-series data samples, time steps to be processed, and real noise using stochastic differential equations; the state data to be processed includes time-series data from multiple sensors.

[0016] The first dimension transformation module is used to transform the dimensions of the state data to be processed, so as to obtain the sensor data to be processed at multiple time points;

[0017] The first prediction module is used to extract features from the time-series data to be processed from multiple sensors, the sensor data to be processed at multiple times, and the time steps to be processed through the initial diffusion model, to obtain time-series dimension features and variable cross-dimensional features, and to process the time-series dimension features and variable cross-dimensional features to obtain the first prediction noise.

[0018] The second prediction module is used to determine the predicted industrial time series data based on the first prediction noise and the state data to be processed.

[0019] The training module is used to adjust the parameters of the initial diffusion model based on the first predicted noise, the real noise, the predicted industrial time series data, and the industrial time series data samples, until the initial diffusion model meets the training conditions and the AIGC generated model is obtained.

[0020] The generation module is used to iteratively denoise random noise using the AIGC generation model to obtain industrial augmented time series data.

[0021] Fourthly, embodiments of this application provide a data generation apparatus, comprising:

[0022] The second dimension transformation module is used to treat random noise as data to be processed and to perform dimension transformation on the data to be processed to obtain transformed data.

[0023] The noise prediction module is used to process the data to be processed, the data to be transformed, and the time step of the data to be processed through the AIGC generation model to obtain the second predicted noise; the AIGC generation model is the AIGC generation model in the AIGC generation model training method for the industrial Internet.

[0024] The physical sampling module is used to determine the drift term based on the second predicted noise and the diffusion coefficient, and to determine the diffusion term based on the diffusion coefficient. It performs physical sampling based on the data to be processed, the drift term, and the diffusion term to obtain denoised data.

[0025] The iteration module is used to update the data to be processed with the denoised data when the time step is not zero, return the data to be processed for dimensional transformation, and continue the process of obtaining the transformed data to be processed; when the time step is zero, the denoised data is used as industrial augmented time series data.

[0026] Fifthly, embodiments of this application provide an electronic device, including: a memory and a processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0027] In a sixth aspect, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0028] In a seventh aspect, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0029] The training method, data generation method, training device, data generation device, electronic device, storage medium, and computer program product for the Industrial Internet provided in this application, based on industrial time-series data samples, time steps to be processed, and real noise, use stochastic differential equations to determine the state data to be processed; by performing dimensional transformation on the state data to be processed, multiple time-step sensor data to be processed are obtained, and then an initial diffusion model is established. Feature extraction is performed on the time-series data to be processed from multiple sensors and the sensor data to be processed from multiple time steps, respectively, to obtain time-series dimension features and variable cross-dimensional features. The time-series dimension features can represent the characteristics of sensor data evolution over time, and the variable cross-dimensional features can be used to represent the spatial distribution and correlation characteristics of multiple sensors at the same time point. Prediction is performed based on the time-series dimension features and variable cross-dimensional features, improving the accuracy of prediction noise; based on the first prediction noise and real noise, the accuracy of prediction noise is improved. By adjusting the parameters of the initial diffusion model using noise, predicted industrial time-series data, and industrial time-series data samples, the model can learn the characteristics of data evolution from the same sensor over time, as well as the spatial distribution and correlation characteristics of multiple sensors at the same time point. This improves the quality of the AIGC-generated model. Furthermore, the AIGC-generated model generates industrial augmented time-series data, imbuing it with the trends and periodic changes found in the original data. This ensures that the generated augmented data better conforms to the physical laws of industrial systems in the time dimension, further improving the quality of the industrial augmented time-series data. In addition, the generated industrial augmented time-series data can be used to train relevant models for industrial equipment, such as fault detection models and life prediction models. This increases the amount of training samples for these models, enhancing their diversity and improving their generalization ability, robustness, and accuracy. Attached Figure Description

[0030] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0031] Figure 1 A flowchart illustrating the training method for the Industrial Internet provided in this application;

[0032] Figure 2 A schematic diagram of the diffusion process provided for this application;

[0033] Figure 3 A schematic diagram of the time-series data to be processed from multiple sensors and the sensor data to be processed at multiple times provided in this application;

[0034] Figure 4 A schematic diagram illustrating the determination and prediction of industrial time-series data provided in this application;

[0035] Figure 5 A schematic diagram of the sampling process provided in this application;

[0036] Figure 6 A schematic diagram of the initial diffusion model provided for this application;

[0037] Figure 7 A schematic diagram of the dual-channel coding unit for temporal evolution and variable crossover provided in this application;

[0038] Figure 8 A flowchart illustrating the data generation method provided in this application;

[0039] Figure 9 A schematic diagram of the noise reduction process for the data to be processed, with a correction procedure, provided for this application;

[0040] Figure 10 A schematic diagram of the industrial time-series data generation device for the diffusion model provided in this application;

[0041] Figure 11 A schematic diagram of the data generation apparatus provided in this application;

[0042] Figure 12 A schematic diagram of the structure of the electronic device provided in this application.

[0043] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0044] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0045] The specific application scenario of this application is to train an initial diffusion model through a forward diffusion process to obtain an AIGC generation model for the Industrial Internet, and then perform a reverse diffusion process based on the trained AIGC generation model to generate industrial augmented time series data. The diffusion process is a process of iteratively adding noise to the original data to obtain Gaussian noise, and the reverse diffusion process is a process of iteratively denoising the Gaussian noise to generate new data.

[0046] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0047] Figure 1 A flowchart illustrating the training method for the Industrial Internet provided in this application. Figure 1 Training methods for the Industrial Internet can be applied to electronic devices, which can be terminals or servers; for example... Figure 1 As shown, training methods for the Industrial Internet include:

[0048] S101. Based on industrial time-series data samples, time steps to be processed, and actual noise, stochastic differential equations are used to determine the state data to be processed; the state data to be processed includes time-series data from multiple sensors.

[0049] The industrial time-series data sample includes time-series data samples from multiple sensors. In practical applications, for each of the multiple sensors, data is collected over a period of time to obtain the time-series data sample for that sensor. The industrial time-series data sample can be time-series data collected through the Industrial Internet.

[0050] In practical applications, multiple sensors can be used, such as multiple sensors for an aircraft engine, including but not limited to: temperature sensors, pressure sensors, speed sensors, flow sensors, etc.

[0051] The time step to be processed can be obtained by random sampling from multiple time steps; the diffusion process can be understood as performing noise addition processing on industrial time series data samples at multiple time steps until pure noise data is obtained, and the time step to be processed is any one of the multiple time steps.

[0052] Stochastic differential equations are used to transform discrete time steps in a diffusion process into continuous time variations, enabling the handling of the interaction between noise and signal at a given time step.

[0053] The pending state data is intermediate data obtained by adding noise to the pending time step of the industrial time series data sample during the diffusion process.

[0054] Specifically, such as Figure 2 As shown, the diffusion process of DDPM (Denoising Diffusion Probabilistic Models) is a fixed forward process whose purpose is to progressively add Gaussian noise to the original data until it becomes completely random noise; this process is driven by a Gaussian noise variance sequence ( ) control, where T is the total number of steps in the diffusion process; for raw data Each step in the diffusion process is related to the previous step. Adding Gaussian noise is represented by formula (1).

[0055] Formula (1): ;

[0056] By continuously adding noise, as long as T is large enough, the final result will be... It will infinitely tend towards a Gaussian random noise. The entire diffusion process is a Markov chain, expressed as Equation (2).

[0057] Formula (2): ;

[0058] By introducing stochastic differential equations (SEDs), the discrete time step t in DDPM is extended to continuous time variables. The continuous process can be described by the SED equation, as shown in formula (3).

[0059] Formula (3): ;

[0060] in, This is the data status data at time t. It is a drift term. It is the diffusion coefficient. Represents Brownian motion (representing random noise), drift term and diffusion coefficient It is the noise added at each time step in DDPM. It's confirmed.

[0061] In practical applications, the diffusion process can be directly based on the original data. For any t steps Sampling was performed to obtain ,make , ; It is the information retention coefficient at time step t. It is the cumulative information retention coefficient at the t-th time step; by using the reparameter technique, the final realization of the diffusion process using stochastic differential equations can be obtained, as shown in formula (4).

[0062] Formula (4): ;in, .

[0063] In this embodiment, It is the pending status data corresponding to the pending time step t. It is an industrial time-series data sample. It is real noise.

[0064] Since the industrial time-series data samples include time-series data samples from multiple sensors, noise is added to the industrial time-series data samples during the diffusion process to obtain the state data to be processed, so that the state data to be processed includes the time-series data to be processed from multiple sensors.

[0065] S102. Perform dimensional transformation on the state data to be processed to obtain sensor data to be processed at multiple time points.

[0066] Specifically, the state data to be processed includes time-series data from multiple sensors. For ease of explanation, let the number of sensors be M, and the number of time points at which the sensor data is collected be F. The state data to be processed can be represented as follows: In other words, the state data to be processed can be understood as a tensor with L rows and F columns; transforming the dimensions of the state data to be processed can be understood as transposing the time dimension and the variable dimension (sensor dimension) to obtain the sensor data to be processed at multiple time points, which can be represented as follows: ,like Figure 3 As shown.

[0067] S103. Using the initial diffusion model, feature extraction is performed on the time-series data to be processed from multiple sensors, the sensor data to be processed at multiple times, and the time steps to be processed to obtain time-series dimension features and variable cross-dimensional features. The time-series dimension features and variable cross-dimensional features are then processed to obtain the first prediction noise.

[0068] Among them, the temporal dimension feature is used to represent the characteristics of sensor data evolution over time; the variable cross dimension feature is used to represent the spatial distribution and correlation characteristics of multiple sensors at the same point in time; and the temporal variable fusion feature integrates the sensor's evolution characteristics over time, spatial distribution, and correlation characteristics.

[0069] Specifically, the time-series data to be processed from multiple sensors Sensor data to be processed at multiple time points The time step t to be processed is input into the initial diffusion model to obtain the predicted noise.

[0070] For example, by using an initial diffusion model, the time-series data to be processed from multiple sensors, the sensor data to be processed at multiple time points, and the time steps to be processed are processed to obtain the first predictive noise, which can be expressed as: ;in, This represents the processing steps for the initial diffusion model. It is time-series data to be processed from multiple sensors. It consists of sensor data to be processed at multiple points in time, where t is the time step to be processed. It is the first prediction noise.

[0071] Specifically, the time-series data to be processed from multiple sensors The data, along with the time step t to be processed, is input into the initial diffusion model. Feature extraction is performed through the initial diffusion model to obtain temporal dimension features. This process combines the sensor data from multiple time points. The time step t to be processed is input into the initial diffusion model. Feature extraction is performed through the initial diffusion model to obtain the cross-dimensional features of variables. The time-series features and cross-dimensional features of variables are fused through the initial diffusion model. Then, the fused features are encoded and decoded to obtain the first prediction noise.

[0072] S104. Based on the first predicted noise and the data of the state to be processed, determine the predicted industrial time series data.

[0073] Among them, the predicted industrial time series data is the denoised data obtained by estimating the first predicted noise based on the initial diffusion model.

[0074] Specifically, based on the first predicted noise and the state data to be processed obtained from the initial diffusion model, the predicted industrial time series data are calculated through stochastic differential equations.

[0075] For example, the final implementation of the diffusion process of SEF is shown in Equation (4). By transforming Equation (4), we can obtain the formula for solving the inverse SED, as shown in Equation (5).

[0076] Formula (5): ;

[0077] in, It is the pending status data corresponding to the pending time step t. It is predictive industrial time-series data. It is the first predicted noise obtained from the initial diffusion model.

[0078] For example, such as Figure 4 As shown, based on industrial time-series data samples, the time step to be processed t, and the actual noise, stochastic differential equations are used to determine the state data to be processed (time-series data from multiple sensors). ; Status data to be processed Perform dimensional transformation to obtain sensor data to be processed at multiple time points. ; Processing time-series data from multiple sensors Sensor data to be processed at multiple time points The time step t to be processed is input into the initial diffusion model to obtain the first predicted noise. Based on the time-series data to be processed from multiple sensors Combined with the first prediction noise, determine the predicted industrial time series data. .

[0079] S105. Based on the first predicted noise, the real noise, the predicted industrial time-series data, and the industrial time-series data samples, adjust the parameters of the initial diffusion model until the initial diffusion model converges to obtain the AIGC generation model for the Industrial Internet.

[0080] Specifically, a first loss value is calculated based on the first predicted noise and the actual noise. A second loss value is calculated based on the predicted industrial time-series data and the industrial time-series data samples. A total loss value is calculated based on the first and second loss values. The parameters of the initial diffusion model are adjusted using the total loss value to determine whether the training conditions are met. If not, the process returns to step S101 for the next training iteration. This iterative training of the initial diffusion model continues until the training conditions are met. The initial diffusion model that meets the training conditions is then used as the AIGC generation model for the Industrial Internet.

[0081] Optionally, the training conditions can be met by either reaching a preset number of iterations or by the initial diffusion model converging.

[0082] Optionally, after obtaining the AIGC generation model, the random noise is iteratively denoised based on the AIGC generation model to obtain industrial augmented time series data.

[0083] like Figure 5 As shown in the embodiment of this application, starting from the Tth time step, the random noise is iteratively denoised for multiple time steps based on the AIGC generation model until the time step equals 0, and industrial augmented time series data is obtained through the sampling process.

[0084] Among them, the AIGC generation model is used to generate time-series data for the Industrial Internet based on pure noise, namely industrial augmented time-series data.

[0085] The total time step of the iterative denoising process is the same as the total time step of the noise addition process during diffusion. Let T be the total time step, and denot the random noise as... Random noise is pure Gaussian noise, which can be randomly drawn from a standard normal distribution. .

[0086] The dimension of the random noise is the same as the dimension of the industrial time-series data samples; for example, the industrial time-series data samples... Let represent data collected from L sensors at F time points, then the dimension of the random noise is... yes .

[0087] It should be noted that random noise Obtained during diffusion All of these are pure noise. To distinguish between noise reduction and noise addition processes, random noise is denoted as... .

[0088] For random noise By performing iterative denoising for T time steps, industrial augmented time series data can be obtained. Specifically, random noise is used as the data to be processed. The data to be processed is dimensionally transformed to obtain the transformed data to be processed. The AIGC generation model processes the data to be processed, the transformed data to be processed, and the time steps of the data to be processed to obtain the second prediction noise. The second prediction noise is removed from the data to be processed to obtain the denoised data. If the time step of the denoised data is not zero, the data to be processed is updated with the denoised data, and the process returns to continue executing the step of dimensionally transforming the data to be processed to obtain the transformed data to be processed, until the time step of the denoised data is zero. The denoised data with the time step of zero is used as the industrial augmented time series data.

[0089] Specifically, the AIGC generation model processes the data to be processed, the data to be transformed, and the time steps of the data to be processed to obtain the second prediction noise. This includes: extracting features from the data to be processed, the data to be transformed, and the time steps of the data to be processed using the AIGC generation model to obtain the first target feature in the time series dimension and the second target feature in the variable cross dimension; and processing the first target feature and the second target feature to obtain the second prediction noise.

[0090] It should be noted that the process of generating the second predictive noise by using the AIGC model to process the data to be processed, the data to be transformed, and the time steps of the data to be processed is similar to the process of generating the first predictive noise by using the initial diffusion model to process the time series data to be processed from multiple sensors, the sensor data to be processed at multiple times, and the time steps to be processed.

[0091] For example, the AIGC generative model is used to address random noise. Perform a denoising step to obtain the denoised data. Then, the AIGC generation model is used to target the denoised data. Perform a denoising step to obtain the denoised data. After performing this denoising process for T time steps, the denoised data is obtained. Denoising data That is, industrial augmentation time series data.

[0092] AIGC generative model for dealing with random noise Perform a denoising step to obtain the denoised data. Let's take an example to illustrate; regarding random noise Perform dimensional transformation to obtain the transformed noise. random noise Noise after conversion The current time step T is input into the AIGC generation model to obtain the second prediction noise. The second prediction noise is then removed from the random noise to obtain the denoised data. For example, as shown in formula (6).

[0093] Formula (6): ;

[0094] in, It is the diffusion coefficient at time step T. It is the second prediction noise output by the AIGC generative model. It is random noise. It's the step length. It is random Gaussian noise.

[0095] According to the treatment of random noise Perform a denoising step to obtain the denoised data. The same process continues for the denoised data. Iterative denoising processing was performed to obtain industrial augmented time series data. .

[0096] It should be noted that the generated industrial augmented time-series data can be used as training samples for relevant models of industrial Internet devices, such as for training fault detection models and life prediction models of industrial equipment. In practical applications, industrial datasets have limited data samples for certain devices or operating conditions, especially industrial data for specific fault types. The embodiments of this application can simulate sensor outputs under different operating conditions and equipment operating states, which can not only improve the generalization ability of existing relevant models, but also increase data diversity. With limited sample size, it can help relevant models learn and predict better in complex industrial scenarios.

[0097] The training method for the Industrial Internet provided in this application uses stochastic differential equations to determine the state data to be processed based on industrial time-series data samples, time steps to be processed, and real noise. By performing dimensional transformation on the state data to be processed, multiple time-step sensor data to be processed are obtained. An initial diffusion model is then established, and features are extracted from the time-series data of multiple sensors and the sensor data at multiple time steps to obtain time-series dimensional features and variable cross-dimensional features. The time-series dimensional features represent the evolution of sensor data over time, while the variable cross-dimensional features represent the spatial distribution and correlation characteristics of multiple sensors at the same time point. Prediction is performed based on the time-series dimensional features and variable cross-dimensional features, improving the accuracy of the predicted noise. Based on the first predicted noise, real noise, predicted industrial time-series data, and industrial time-series data samples, adjustments are made... The parameters of the initial diffusion model enable it to learn the characteristics of data evolution from the same sensor over time, as well as the spatial distribution and correlation characteristics of multiple sensors at the same time point. This improves the quality of the AIGC-generated model. Furthermore, the AIGC-generated model generates industrial augmented time-series data that retains the trends and periodic changes found in the original data. This ensures that the generated augmented data better conforms to the physical laws of industrial systems in the time dimension, further improving the quality of the industrial augmented time-series data. In addition, the generated industrial augmented time-series data can be used to train relevant models for industrial equipment, such as fault detection models and life prediction models. This increases the amount of training samples for these models, enhances their diversity, and improves their generalization ability, robustness, and accuracy.

[0098] In some embodiments, the initial diffusion model includes a dual-channel encoding unit for temporal evolution and variable cross-cutting, other encoding units, and a decoder. Using the initial diffusion model, features are extracted from the unprocessed temporal data of multiple sensors, the unprocessed sensor data at multiple time points, and the unprocessed time steps to obtain temporal dimension features and variable cross-cutting dimension features. These features are then processed to obtain first prediction noise, including: inputting the unprocessed time steps and the unprocessed temporal data of multiple sensors into the first channel of the dual-channel encoding unit to obtain temporal dimension features; inputting the unprocessed time steps and the unprocessed sensor data at multiple time points into the second channel of the dual-channel encoding unit to obtain variable cross-cutting dimension features; fusing the temporal dimension features and variable cross-cutting dimension features through the fusion layer of the dual-channel encoding unit to obtain temporal variable fusion features; and encoding and decoding the temporal variable fusion features through other encoding units and a decoder to obtain the first prediction noise.

[0099] The dual-channel coding unit for temporal evolution and variable crossover includes a first channel, a second channel, and a fusion layer. Both the first and second channels are used for coding.

[0100] Specifically, the time step to be processed and the time series data from multiple sensors to be processed are input into the first channel of the dual-channel encoding unit to obtain the time series dimension features; the time step to be processed and the sensor data to be processed at multiple moments are input into the second channel of the dual-channel encoding unit to obtain the variable cross dimension features.

[0101] In practical applications, the initial diffusion model can be implemented based on UNet1D. UNet1D includes an encoder and a decoder. The encoder includes multiple coding units (downsampling units), and the decoder includes multiple decoding units (upsampling units). There are skip connections between the encoder and decoder. The encoder also includes an input layer, and the decoder also includes an output layer. Figure 6 As shown, in this embodiment of the application, the first coding unit of UNet1D is changed to a dual-channel coding unit with temporal evolution and variable crossover to obtain the initial diffusion model.

[0102] The dual-channel coding unit for temporal evolution and variable crossover includes a first channel, a second channel, and a fusion layer. For example... Figure 7 As shown, the first channel of the dual-channel coding unit for temporal evolution and variable crossover includes: a first fully connected layer, a first coding module, a first diffusion conversion module, and a second fully connected layer; the second channel includes a third fully connected layer, a second coding module, a second diffusion conversion module, and a fourth fully connected layer; the dual-channel coding unit also includes a time-step coding layer.

[0103] Specifically, the time-series data to be processed from the time step to be processed and from multiple sensors are input into the first channel of the dual-channel encoding unit for time-series evolution and variable cross-processing to obtain time-series dimensional features, including:

[0104] The time step to be processed is encoded by the time step encoding layer to obtain the time step feature; the time series data to be processed from multiple sensors is processed by the first fully connected layer to obtain the first feature to be processed; the first feature to be processed is superimposed with the position encoding to obtain the second feature to be processed; the second feature to be processed is processed by the first encoding module to obtain the third feature to be processed; the time step feature and the third feature to be processed are input into the first diffusion conversion module to obtain the fourth feature to be processed; the fourth feature to be processed is input into the second fully connected layer to obtain the time series dimension feature.

[0105] The sensor data to be processed at the desired time step and at multiple time points are input into the second channel of the dual-channel encoding unit for time-series evolution and variable cross-interaction to obtain variable cross-dimensional features, including:

[0106] The time step encoding layer encodes the time step to be processed to obtain the time step feature. The third fully connected layer processes the sensor data to be processed at multiple time points to obtain the fifth feature to be processed. The fifth feature to be processed is processed by the second encoding module to obtain the sixth feature to be processed. The time step feature and the sixth feature to be processed are input into the second diffusion transformation module to obtain the seventh feature to be processed. The seventh feature to be processed is input into the fourth fully connected layer to obtain the variable cross-dimensional feature.

[0107] To enhance the temporal sequence information, position encoding is added to the input of the first channel. The time-series data from multiple sensors are processed through a first fully connected layer to obtain a first feature to be processed. This first feature is then superimposed with the position encoding to obtain a second feature to be processed. Finally, the second feature is input to the first encoding module to obtain a third feature to be processed. This can be represented as: ;in, It is the third feature to be processed. This is pending status data. It is a positional encoding. It is used to implement linear mapping, Map from M×F dimensions to M×H dimensions.

[0108] The time step encoding layer encodes the time step t to be processed, obtaining the time step features; the third feature to be processed... The time-step features are input into the first diffusion transformation module, and then the output of the first diffusion transformation module is processed by the second fully connected layer to obtain the time-series dimension features.

[0109] The first diffusion transformation module includes a multi-layer Transformer Encoder. Each Transformer Encoder uses a multi-head attention mechanism and a feedforward network, and ensures smooth information transmission through residual connections, which can be expressed as formula (7).

[0110] Formula (7): ;

[0111] in, It is the first in a multi-layer Transformer Encoder The hidden state of the layer output, It is a multi-layer Transformer Encoder. The hidden state of the layer output, It is the total number of layers in the multi-layer Transformer Encoder. This indicates a multi-head attention mechanism. Represents a feedforward network, the first The data output by the layer is the time-series dimension feature; the time-series dimension feature is denoted as... ,in, To hide the dimensions, this representation contains both local temporal patterns and captures long-term temporal dependencies.

[0112] It should be noted that the multi-layer Transformer Encoder uses positional encoding to enhance temporal information, enabling the initial diffusion model to recognize temporal order. At the same time, the self-attention mechanism captures long-term temporal evolution features, and the feedforward network further enhances the data representation capability, thereby improving the quality of temporal dimension features.

[0113] Using the same method as extracting temporal dimension features through the first channel, the sensor data to be processed at multiple time points is also processed through the second channel. The variables are processed at the time step t to obtain the cross-dimensional features; the cross-dimensional features are denoted as... .

[0114] Let the temporal dimension feature of the first channel output be denoted as... Cross-dimensional features of variables in the second channel output Feature fusion is performed to integrate time and variable dependencies to obtain cross-dimensional features of variables, as shown in formula (8).

[0115] Formula (8): ;

[0116] in, It is a feature of the cross-dimensional dimension of variables. The linear fusion weights for the time channel and the variable channel are used to optimize the contribution ratio of different feature information in the generation process. For learnable bias, Used to increase nonlinear transformation capability It is an activation function used to control the gating mechanism of channel fusion. Additional weights for gating fusion, It is an element-wise multiplication, achieving a weighted fusion equivalent to the attention mechanism.

[0117] It should be noted that in this embodiment, the first coding unit of UNet1D is changed to a dual-channel coding unit of temporal evolution and variable crossover to obtain the initial diffusion model. Therefore, the initial diffusion model includes the dual-channel coding unit of temporal evolution and variable crossover, as well as other coding units and decoders in UNet1D. That is to say, in this embodiment, the initial diffusion module includes an encoder and a decoder, and the encoder includes the dual-channel coding unit of temporal evolution and variable crossover and other coding units.

[0118] The cross-dimensional features of the variables are input into other subsequent coding units. After being processed by other coding units, they are input into the decoder, and the first prediction noise is obtained by the decoder.

[0119] The encoder of the initial diffusion model extracts higher-level features through progressive downsampling, while the decoder recovers the original resolution through progressive upsampling. This structure can learn both local temporal patterns and global temporal dependencies simultaneously.

[0120] The skip connections in the initial diffusion model directly pass early high-resolution features from the encoder to the corresponding layers in the decoder. This mechanism ensures that local details in the time series can be recovered during decoding, rather than just inferred from high-level abstract features.

[0121] The initial diffusion model can learn multi-scale temporal features through its progressive downsampling and upsampling process, including: fine-grained features, that is, in the shallow convolution of the encoder, the initial diffusion model can capture the short-term dependencies of time series data, such as instantaneous changes or trends in a short period of time; and long-term features, that is, through deep convolution and downsampling processes, the model can capture the long-term dependencies of time series, which is very important for state prediction and fault detection in industrial scenarios.

[0122] Furthermore, downsampling reduces the resolution on the time axis, extracting higher-dimensional abstract features. Upsampling restores the original temporal resolution, ensuring that the generated data is aligned with the original data in the time dimension in the final generation or prediction task.

[0123] In the above embodiments, in order to adapt to the characteristics of industrial time series data such as multivariates, strong dependencies, and complex structures, a dual-channel feature extraction method is adopted. The first channel captures the evolution pattern of the sequence in the time dimension, including short-term fluctuations and long-term trends, while the second channel models the correlation between multiple variables. This is particularly suitable for the collaborative relationship between multiple sensors and control parameters in industrial scenarios, so that the initial diffusion model can fully capture the temporal evolution of time series data and the statistical relationship between variables during the training process. Subsequently, the AIGC generation model obtained from the training is used to generate data, which improves the authenticity and statistical consistency of the data.

[0124] In some embodiments, the training method for the Industrial Internet further includes: acquiring a first generation condition corresponding to an industrial time-series data sample; randomly discarding the first generation condition to obtain a target condition; and extracting features from the unprocessed time-series data of multiple sensors, the unprocessed sensor data at multiple times, and the unprocessed time steps through an initial diffusion model to obtain time-series dimension features and variable cross-dimensional features, including: extracting features from the unprocessed time-series data of multiple sensors, the unprocessed sensor data at multiple times, the unprocessed time steps, and the target condition through an initial diffusion model to obtain time-series dimension features and variable cross-dimensional features.

[0125] This embodiment introduces a Classifier-Free Guidance (CFG) diffusion process; in order to introduce CFG during the generation process, a conditional diffusion model needs to be trained during the training phase.

[0126] The first generation condition can be a text description or a category label, used to guide the development in the direction indicated by the first generation condition, so as to obtain results that are more in line with the intent; the first generation condition can be preset.

[0127] Specifically, during training, the first generation condition corresponding to the industrial time series data sample is determined, and the first generation condition is randomly discarded with a certain probability. If the first generation condition is discarded, the given first generation condition is replaced with a null value. In this way, the initial diffusion model is conditionally generated in some training rounds and unconditionally generated in other training rounds. Finally, an AIGC generation model that can achieve both conditional and unconditional prediction can be obtained.

[0128] For example, industrial time-series data samples are obtained through sampling. The corresponding first generation condition y is discarded with a fixed probability to obtain the target condition. The target condition may be y or a null value. The fixed probability is predetermined, and this application does not limit it in the embodiments.

[0129] When the target condition is y, the initial diffusion model processes the time-series data from multiple sensors, the sensor data at multiple time points, the time steps to be processed, and the target condition to obtain the first prediction noise, which can be expressed as: .

[0130] When the target condition is null, the initial diffusion model processes the time-series data from multiple sensors, the sensor data at multiple time points, the time steps to be processed, and the target condition to obtain the first prediction noise, which can be expressed as: .

[0131] In the above embodiments, the generation quality and control capabilities are improved by introducing classifier-free guidance.

[0132] In some embodiments, adjusting the parameters of an initial diffusion model based on a first predicted noise, actual noise, predicted industrial time-series data, and industrial time-series data samples includes: determining a noise loss value based on the first predicted noise and actual noise; determining a population distribution preservation loss value based on the predicted industrial time-series data and industrial time-series data samples; determining a total loss value based on the noise loss value and the population distribution preservation loss value; and adjusting the parameters of the initial diffusion model using the total loss value.

[0133] Specifically, the noise loss value is determined based on the first predicted noise and the actual noise, and can be determined by formula (9).

[0134] Formula (9): ;

[0135] in, This is the noise loss value. It is the noise standard deviation corresponding to the time step t to be processed. It's real noise. It is the first prediction noise. It is the expectation operator.

[0136] Optionally, the population distribution preservation loss value is determined based on the predicted industrial time-series data and the industrial time-series data samples. This can be achieved by determining the similarity between the predicted industrial time-series data and the industrial time-series data samples, calculating the difference between 1 and the similarity, obtaining the difference degree, and using the difference degree as the population distribution preservation loss value.

[0137] Optionally, the predicted industrial time-series data includes predicted time-series data from multiple sensors; the industrial time-series data sample includes time-series data samples from multiple sensors; determining the population distribution preservation loss value based on the predicted industrial time-series data and the industrial time-series data samples includes: obtaining predicted time-series data from two sensors from the predicted time-series data of multiple sensors according to a preset sequence group, and determining a first correlation between the predicted time-series data of the two sensors; obtaining time-series data samples from the time-series data samples of multiple sensors according to a preset sequence group, and determining a second correlation between the time-series data samples of the two sensors; determining the mean difference corresponding to the preset sequence group based on the first correlation and the second correlation; and determining the population distribution preservation loss value based on the mean differences corresponding to the multiple preset sequence groups respectively.

[0138] There are multiple preset number groups, each of which includes two preset numbers. The two preset numbers included in each of the multiple preset number groups correspond to multiple sensors. In other words, the numbers of multiple sensors are predetermined, and a preset number group is constructed based on the numbers of any two sensors among the multiple sensors, resulting in multiple preset number groups. For example, the preset number group is (1,2), which corresponds to the sensors with numbers 1 and 2.

[0139] For each preset sequence number group, from the time-series data samples of multiple sensors, obtain the time-series data samples of the two sensors corresponding to the preset sequence number group; for example, obtain the time-series data samples corresponding to the preset sequence number group (i,j). and From the prediction time-series data of multiple sensors, obtain the prediction time-series data of two sensors corresponding to a preset sequence number group. For example, obtain the prediction time-series data corresponding to the preset sequence number group (i,j). and ; Calculate the first correlation between the time series data samples of the two sensors, and calculate the second correlation between the predicted time series data of the two sensors; determine the mean difference corresponding to the preset number group based on the first correlation and the second correlation; determine the population distribution preservation loss value based on the mean difference corresponding to the multiple preset number groups respectively; as shown in formulas (10)-(12).

[0140] Formula (10): ;

[0141] Formula (11): ;

[0142] Formula (12): ;

[0143] in, It is a population distribution preservation loss; preset sequence group (i,j), It is the first of multiple predicted time series data sets A predictive time series data set, It is the first of multiple predicted time series data sets One predicted time series data set; yes and The first degree of correlation between them; It is the first of multiple time-series data samples from various sensors. Time-series data samples from each sensor, It is the first Time-series data samples from each sensor; yes and The second correlation between them; It is the difference in mean value corresponding to the preset sequence number group (i,j).

[0144] After calculating the noise loss value and the population distribution maintenance loss value, the noise loss value and the population distribution maintenance loss value are weighted and summed to obtain the total loss value. The weights of the noise loss value and the population distribution maintenance loss value can be set according to actual needs. In this embodiment, the noise loss value and the population distribution maintenance loss value are not limited; as shown in formula (13).

[0145] Formula (13): ;

[0146] in, This is the total loss value. This is the noise loss value. It is the population distribution preservation loss value. It is the weight that preserves the loss value of the group distribution.

[0147] In related technologies, only the authenticity of individual data is usually considered, while the distribution of the group is ignored. In this application, the accuracy of a single sample in the time dimension is ensured by noise loss, and the group distribution preservation loss is introduced to ensure that the statistical attributes of the generated data will not be shifted. The maximum mean difference is used to measure the statistical shift between the generated data and the real data in terms of variable distribution and quantitative correlation, which effectively suppresses the distribution drift problem generated by the model.

[0148] The training method for the Industrial Internet provided in this application uses stochastic differential equations to determine the state data to be processed based on industrial time-series data samples, time steps to be processed, and real noise. By performing dimensional transformation on the state data to be processed, multiple time-step sensor data to be processed are obtained. An initial diffusion model is then established, and features are extracted from the time-series data and the sensor data at multiple time steps from multiple sensors to obtain time-series features and variable cross-dimensional features. The time-series features represent the characteristics of sensor data evolution over time, while the variable cross-dimensional features represent the spatial distribution and correlation characteristics of multiple sensors at the same time point. Prediction is performed based on the time-series features and variable cross-dimensional features, improving the accuracy of the predicted noise. Based on the first predicted noise, real noise, predicted industrial time-series data, and industrial time-series data samples, the parameters of the initial diffusion model are adjusted so that the initial diffusion model can learn the characteristics of data evolution of the same sensor over time, as well as the spatial distribution and correlation characteristics of multiple sensors at the same time point, improving the quality of the AIGC-generated model.

[0149] Figure 8 This is a flowchart illustrating the data generation method provided in this application. The data generation method can be applied to an electronic device, which can be a terminal or a server; for example... Figure 8 As shown, the data generation methods include:

[0150] S801. Take random noise as the data to be processed, and perform dimensional transformation on the data to be processed to obtain the transformed data to be processed.

[0151] S802. The AIGC generation model processes the data to be processed, the data to be transformed, and the time step of the data to be processed to obtain the second prediction noise. The AIGC generation model is the AIGC generation model in the AIGC generation model training method for the industrial Internet.

[0152] S803. Determine the drift term based on the second predicted noise and the diffusion coefficient, and determine the diffusion term based on the diffusion coefficient;

[0153] S804. Perform physical sampling based on the data to be processed, the drift term, and the diffusion term to obtain denoised data;

[0154] S805. If the time step is not zero, update the data to be processed with the noise reduction number, return to perform dimensional transformation on the data to be processed, and continue to execute the step of obtaining the transformed data to be processed.

[0155] S806. When the time step is zero, the denoised data is used as industrial augmentation time series data.

[0156] Specifically, such as Figure 9 As shown, random noise is used as the data to be processed, and the data to be processed is dimensionally transformed to obtain the transformed data to be processed; the AIGC generation model is used to process the data to be processed, the transformed data to be processed, and the time step of the data to be processed to obtain the second predicted noise. The drift term is determined according to the second predicted noise and the diffusion coefficient, and the diffusion term is determined according to the diffusion coefficient; physical sampling is performed according to the data to be processed, the drift term, and the diffusion term to obtain the initial denoised data, and the Langevin dynamics with physical sensed log gradient is used to correct the initial denoised data to obtain the denoised data; as shown in formulas (14) and (15).

[0157] Formula (14): ;

[0158] in, This is the initial denoised data. This is data to be processed. This is data to be processed and transformed. It is a time step. It is a time step diffusion coefficient, It's the step length. It is random Gaussian noise; It is a drift term. It is a diffusion term.

[0159] Formula (15): ;

[0160] in, It is denoised data. This is the initial denoised data. It's the step length. It is a time step diffusion coefficient, It is the logarithmic gradient of the data distribution.

[0161] Calibration steps Starting with the initial denoised data, multiple fine-tuning gradient ascents are performed to fine-tune the initial denoised data, resulting in denoised data that better fits the data distribution and satisfies physical constraints, thereby enhancing the physical consistency of the samples and the usability of the data; the denoised data... As data to be processed and based on Continue iterative denoising processing until industrial augmented time series data is obtained.

[0162] Optionally, if CFG is introduced during the training of the initial diffusion model, the random noise is iteratively denoised using an AIGC-generated model to obtain industrial augmented time-series data, including:

[0163] Random noise is used as the data to be processed, and the data to be processed is dimensionally transformed to obtain the transformed data to be processed; a second generation condition is obtained; wherein, the second generation condition can be randomly sampled from multiple preset conditions; the data to be processed, the transformed data to be processed, and the time step of the data to be processed are processed by the AIGC generation model to obtain the second prediction noise; the data to be processed, the transformed data to be processed, the time step of the data to be processed, and the second generation condition are processed by the AIGC generation model to obtain the third prediction noise; the target prediction noise is calculated based on the second prediction noise and the third prediction noise; the stock intention is determined based on the target prediction noise and the diffusion coefficient; the diffusion term is determined based on the diffusion coefficient; and physical sampling is performed based on the data to be processed, the drift term, and the diffusion term to obtain the initial denoised data.

[0164] The initial denoised data is dimensionally transformed to obtain initial denoised transformed data. The initial denoised data, initial denoised transformed data, and time step are processed by the AIGC generation model to obtain the first corrected noise. The initial denoised data, initial denoised transformed data, time step, and second generation conditions are processed by the AIGC generation model to obtain the second corrected noise. The target corrected noise is determined based on the first and second corrected noise. The initial denoised data is corrected using the target corrected noise to obtain denoised data.

[0165] If the time step is not zero, update the data to be processed with the denoised data, return to perform dimensional transformation on the data to be processed, and continue the process to obtain the transformed data to be processed; if the time step is zero, use the denoised data as industrial augmented time series data.

[0166] In this embodiment, industrial augmented time-series data is generated through an AIGC generation model. This allows the industrial augmented time-series data to retain the trend and periodic changes characteristics of the original data, ensuring that the generated augmented data is more consistent with the physical laws of the industrial system in the time dimension, thus improving the quality of the industrial augmented time-series data. In addition, the generated industrial augmented time-series data can be used to train relevant models of industrial equipment, such as fault detection models and life prediction models. This increases the amount of training samples for the relevant models, enhances the diversity of training samples, and thus improves the generalization ability, robustness, and accuracy of the relevant models.

[0167] The embodiments of this application have the following technical effects:

[0168] 1. Data generation is achieved; the embodiments of this application are based on time-series data (industrial time-series data samples) collected from multiple sensors to train the initial diffusion model, thereby obtaining an AIGC generation model for industrial networks, and then generating industrial augmented time-series data through the AIGC generation model; the industrial augmented time-series data is used to expand the training samples of related models.

[0169] It should be noted that publicly available industrial datasets have limited data samples for certain equipment or operating conditions, especially industrial data for specific fault types. The embodiments of this application can simulate sensor outputs under different operating conditions and equipment operating states, which can not only improve the generalization ability of related models, but also increase data diversity. With limited sample size, this helps related models to learn and predict better in complex industrial scenarios.

[0170] 2. Improved data diversity: Through the embodiments of this application, industrial augmented time-series data that is similar to the distribution of real data but has greater diversity can be generated. The generated augmented data is used to supplement the training set, especially when some sensor failures or system failures result in less data. Increasing the diversity of data helps to improve the robustness and accuracy of the model under different operating conditions.

[0171] 3. It can simulate new working conditions; by adjusting the generation conditions, such as adjusting the fault occurrence event or equipment operating conditions, corresponding industrial augmented time series data can be generated, thereby providing more diverse training samples for equipment fault diagnosis and predictive maintenance. The simulated data can also include data on extreme cases or special fault modes, helping relevant models learn a wider range of operating modalities and improve their adaptability to scenarios.

[0172] 4. Data augmentation and improvement of few-shot learning performance: Since the sample data of some devices or working conditions is limited, the embodiments of this application can generate rich and large amounts of augmented data, effectively perform data augmentation, so that the model can still maintain good performance when facing few samples and reduce the risk of overfitting.

[0173] 5. Improve data quality; This application embodiment captures the long-term and short-term dependencies in the data by denoising and gradually generating time-series data. The generated augmented data has the characteristics of trends and periodic changes in the original data, ensuring that the generated augmented data is more in line with the physical laws of the industrial system in the time dimension.

[0174] 6. Improve model quality: By introducing the generated industrial augmented time-series data, the relevant models can learn more diverse features during training. This diversity not only helps improve the model's performance under known operating conditions, but also helps the relevant models to be more robust when facing unseen failure modes or operating conditions. In particular, the generated data provides additional support for fault diagnosis and remaining service life prediction, significantly improving the predictive performance of the relevant models.

[0175] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0176] Figure 10 A schematic diagram of the AIGC generative model training device for the Industrial Internet provided in this application is shown below. Figure 10 As shown, the AIGC generative model training device 100 for the Industrial Internet provided in this embodiment includes:

[0177] The processing module 1001 is used to determine the state data to be processed based on industrial time-series data samples, time steps to be processed, and real noise using stochastic differential equations; the state data to be processed includes time-series data to be processed from multiple sensors.

[0178] The first dimension transformation module 1002 is used to perform dimension transformation on the state data to be processed to obtain sensor data to be processed at multiple time points;

[0179] The first prediction module 1003 is used to extract features from the time-series data to be processed from multiple sensors, the sensor data to be processed at multiple times, and the time steps to be processed through the initial diffusion model, to obtain time-series dimension features and variable cross-dimensional features, and to process the time-series dimension features and variable cross-dimensional features to obtain the first prediction noise.

[0180] The second prediction module 1004 is used to determine the predicted industrial time series data based on the first prediction noise and the state data to be processed.

[0181] The training module 1005 is used to adjust the parameters of the initial diffusion model based on the first predicted noise, the real noise, the predicted industrial time series data, and the industrial time series data samples, until the initial diffusion model meets the training conditions, thereby obtaining the AIGC generation model for the Industrial Internet.

[0182] In one possible implementation, the initial diffusion model includes a dual-channel encoding unit for temporal evolution and variable cross-cutting, other encoding units, and a decoder; the first prediction module is further configured to input the time-series data to be processed from the time step to be processed and multiple sensors into the first channel of the dual-channel encoding unit to obtain temporal dimension features; input the sensor data to be processed from the time step to be processed and multiple time points into the second channel of the dual-channel encoding unit to obtain variable cross-cutting dimension features; perform feature fusion on the temporal dimension features and variable cross-cutting dimension features through the fusion layer of the dual-channel encoding unit to obtain temporal variable fusion features; and perform encoding and decoding processing on the temporal variable fusion features through other encoding units and a decoder to obtain the first prediction noise.

[0183] In one possible implementation, the first prediction module is further used to obtain the first generation condition corresponding to the industrial time series data sample; randomly discard the first generation condition to obtain the target condition; and extract features from the time series data to be processed from multiple sensors, the sensor data to be processed at multiple times, the time steps to be processed, and the target condition through an initial diffusion model to obtain time series dimension features and variable cross dimension features.

[0184] In one possible implementation, a training module is used to determine a noise loss value based on a first predicted noise and the actual noise; determine a population distribution preservation loss value based on predicted industrial time-series data and industrial time-series data samples; determine a total loss value based on the noise loss value and the population distribution preservation loss value; and adjust the parameters of the initial diffusion model using the total loss value.

[0185] In one possible implementation, the predicted industrial time-series data includes predicted time-series data from multiple sensors; the industrial time-series data samples include time-series data samples from multiple sensors; a training module is used to acquire predicted time-series data from two sensors based on a preset sequence number group, and determine a first correlation between the predicted time-series data from the two sensors; acquire time-series data samples from the time-series data samples from multiple sensors based on the preset sequence number group, and determine a second correlation between the time-series data samples from the two sensors; determine the mean difference corresponding to the preset sequence number group based on the first and second correlations; and determine the population distribution preservation loss value based on the mean differences corresponding to the multiple preset sequence number groups respectively.

[0186] Figure 11 A schematic diagram of the data generation apparatus provided in this application is shown below. Figure 11 As shown, the data generation device 110 provided in this embodiment includes:

[0187] The second dimension transformation module 1101 is used to treat random noise as data to be processed and to perform dimension transformation on the data to be processed to obtain transformed data to be processed.

[0188] The noise prediction module 1102 is used to process the data to be processed, the data to be transformed, and the time step of the data to be processed through the AIGC generation model to obtain the second predicted noise; the AIGC generation model is the AIGC generation model in the AIGC generation model training method for the industrial Internet.

[0189] The physical sampling module 1103 is used to determine the drift term based on the second predicted noise and the diffusion coefficient, and to determine the diffusion term based on the diffusion coefficient. It performs physical sampling based on the data to be processed, the drift term, and the diffusion term to obtain denoised data.

[0190] The iteration module 1104 is used to update the data to be processed with the denoised data when the time step is not zero, return the data to be processed for dimensional transformation, and continue to execute the step of obtaining the transformed data to be processed; when the time step is zero, the denoised data is used as industrial augmented time series data.

[0191] The AIGC generative model training device for the Industrial Internet provided in this embodiment can execute the AIGC generative model training method for the Industrial Internet provided in the above method embodiment, and its implementation principle and technical effect are similar; the data generation device provided in this embodiment can execute the data generation method provided in the above method embodiment, and its implementation principle and technical effect are similar; this embodiment will not be described in detail here.

[0192] Figure 12 A schematic diagram of the structure of the electronic device provided in this application. Figure 12As shown, the electronic device 120 provided in this embodiment includes at least one processor 1201 and a memory 1202. Optionally, the device 120 further includes a communication component 1203. The processor 1201, memory 1202, and communication component 1203 are connected via a bus.

[0193] In a specific implementation, at least one processor 1201 executes computer execution instructions stored in memory 1202, causing at least one processor 1201 to perform the above-described method.

[0194] The specific implementation process of processor 1201 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0195] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0196] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0197] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0198] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0199] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0200] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0201] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0202] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0203] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0204] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0205] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0206] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0207] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A training method for AIGC generative models for the Industrial Internet, characterized in that, include: Based on industrial time-series data samples, time steps to be processed, and real noise, stochastic differential equations are used to determine the state data to be processed. The status data to be processed includes time-series data from multiple sensors. The state data to be processed is transformed in dimension to obtain sensor data to be processed at multiple time points; Using an initial diffusion model, features are extracted from the time-series data to be processed from the multiple sensors, the sensor data to be processed at the multiple time points, and the time steps to be processed, to obtain time-series dimension features and variable cross-dimensional features. The time-series dimension features and the variable cross-dimensional features are then processed to obtain the first prediction noise. Based on the first predicted noise and the state data to be processed, predict the industrial time series data; Based on the first predicted noise, the real noise, the predicted industrial time-series data, and the industrial time-series data samples, the parameters of the initial diffusion model are adjusted until the initial diffusion model meets the training conditions, thereby obtaining an AIGC generation model for the Industrial Internet.

2. The method according to claim 1, characterized in that, The initial diffusion model includes a dual-channel coding unit with temporal evolution and variable crossover, other coding units, and a decoder; The process involves using an initial diffusion model to extract features from the time-series data to be processed from the multiple sensors, the sensor data to be processed at the multiple time points, and the time steps to be processed, obtaining time-series dimension features and variable cross-dimensional features. These time-series dimension features and variable cross-dimensional features are then processed to obtain the first prediction noise, including: The time step to be processed and the time series data to be processed from the multiple sensors are input into the first channel of the dual-channel encoding unit to obtain the time series dimension features; The data from the time step to be processed and the sensor data to be processed at the multiple times are input into the second channel of the dual-channel encoding unit to obtain the variable cross-dimensional features; The temporal dimension features and the variable cross-dimensional features are fused through the fusion layer of the dual-channel coding unit to obtain temporal variable fusion features; The temporal variable fusion features are encoded and decoded by the other encoding units and the decoder to obtain the first prediction noise.

3. The method according to claim 1, characterized in that, The step of adjusting the parameters of the initial diffusion model based on the first predicted noise, the actual noise, the predicted industrial time-series data, and the industrial time-series data samples includes: The noise loss value is determined based on the first predicted noise and the actual noise; Based on the predicted industrial time-series data and the industrial time-series data sample, determine the population distribution preservation loss value; The total loss value is determined based on the noise loss value and the population distribution preservation loss value; The parameters of the initial diffusion model are adjusted using the total loss value.

4. The method according to claim 3, characterized in that, The predicted industrial time-series data includes predicted time-series data from multiple sensors; the industrial time-series data sample includes time-series data samples from multiple sensors. The step of determining the population distribution preservation loss value based on the predicted industrial time-series data and the industrial time-series data sample includes: Based on a preset sequence number group, the predicted time series data of two sensors are obtained from the predicted time series data of multiple sensors, and the first correlation between the predicted time series data of the two sensors is determined. Based on a preset sequence number group, time-series data samples of two sensors are obtained from time-series data samples of multiple sensors, and the second correlation between the time-series data samples of the two sensors is determined. The mean difference corresponding to the preset sequence group is determined based on the first correlation and the second correlation. The population distribution preservation loss value is determined based on the mean differences corresponding to the multiple preset number groups.

5. The method according to claim 1, characterized in that, The method further includes: Obtain the first generation condition corresponding to the industrial time-series data sample; The first generated condition is randomly discarded to obtain the target condition; The process involves using an initial diffusion model to extract features from the time-series data to be processed from the multiple sensors, the sensor data to be processed at the multiple time points, and the time steps to be processed, resulting in time-series dimension features and variable cross-dimensional features, including: Using an initial diffusion model, features are extracted from the time-series data to be processed from the multiple sensors, the sensor data to be processed at the multiple time points, the time steps to be processed, and the target conditions to obtain time-series dimension features and variable cross-dimensional features.

6. A data generation method, characterized in that, include: Random noise is used as the data to be processed, and the data to be processed is transformed in dimension to obtain the transformed data to be processed. The AIGC generation model is used to process the data to be processed, the transformed data to be processed, and the time step of the data to be processed to obtain the second prediction noise; the AIGC generation model is the AIGC generation model described in any one of claims 1 to 5. The drift term is determined based on the second predicted noise and the diffusion coefficient, and the diffusion term is determined based on the diffusion coefficient; Physical sampling is performed based on the data to be processed, the drift term, and the diffusion term to obtain denoised data; If the time step is not zero, update the data to be processed with the denoised number, and return to the step of performing dimensional transformation on the data to be processed to obtain the transformed data to be processed and continue execution; When the time step is zero, the denoised data is used as industrial augmented time series data.

7. A training device for an AIGC generative model for the Industrial Internet, characterized in that, The device includes: The processing module is used to determine the state data to be processed based on industrial time-series data samples, time steps to be processed, and real noise using stochastic differential equations; the state data to be processed includes time-series data to be processed from multiple sensors. The first dimension conversion module is used to perform dimension conversion on the state data to be processed to obtain sensor data to be processed at multiple time points. The first prediction module is used to extract features from the time-series data to be processed from the multiple sensors, the sensor data to be processed at the multiple times, and the time steps to be processed through an initial diffusion model, to obtain time-series dimension features and variable cross-dimensional features, and to process the time-series dimension features and the variable cross-dimensional features to obtain the first prediction noise. The second prediction module is used to determine the predicted industrial time series data based on the first prediction noise and the state data to be processed. The training module is used to adjust the parameters of the initial diffusion model based on the first predicted noise, the real noise, the predicted industrial time-series data, and the industrial time-series data samples, until the initial diffusion model meets the training conditions, thereby obtaining an AIGC generation model for the Industrial Internet.

8. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, Includes computer execution instructions, which, when executed by a processor, implement the method as described in any one of claims 1 to 6.

Images