Industrial data enhancement method based on hierarchical time sequence progressive diffusion network

Abstract

The invention relates to an industrial data enhancement method based on a hierarchical time sequence progressive diffusion network, and relates to the field of data enhancement, and the method comprises the steps: obtaining a time sequence data set in response to an input signal; obtaining time scale state data corresponding to a preset time scale based on the time sequence data set; performing forward diffusion on the time scale state data to generate noise-added time scale state data; obtaining information stage features corresponding to the time scale based on all the noise-added time scale state data; splicing the adjacent information stage features to obtain the adjacent information stage features; generating a fusion label based on the adjacent information stage features; obtaining a de-noised clean virtual sample based on the fusion label; and obtaining reconstructed synthetic data based on the de-noised clean virtual sample and the time sequence data set. The method has the advantages that the generated virtual sample is high in quality and can reflect the overall trend and local features of the industrial time series data at the same time, and the data set is effectively expanded to improve the prediction precision of the soft measurement model.

Description

Technical Field

[0001] This invention relates to the field of data augmentation, and more particularly to an industrial data augmentation method using a hierarchical temporal progressive diffusion network. Background Technology

[0002] In industrial environments, soft measurement techniques have emerged to determine quality-related indicators that are difficult to measure in real time. These techniques indirectly predict quality variables that are difficult to measure directly by modeling readily available process variables. Data-driven soft measurement methods rely on statistical and machine learning techniques, requiring the automatic learning of complex relationships between variables from large amounts of historical data. They are particularly suitable for high-dimensional processes with significant nonlinear characteristics. However, in practice, problems such as missing values, low sampling frequency, or high redundancy often exist, leading to insufficient training data. This "small data" problem affects the model's ability to capture process dynamics, thereby reducing the accuracy of quality variable predictions.

[0003] To address the small sample size problem, virtual sample generation technology effectively augments data by mining its latent distribution and inherent correlations, thereby mitigating model overfitting caused by data scarcity in industrial scenarios. Furthermore, considering the dynamic characteristics of industrial time-series data, researchers have proposed generative models such as Time-series GAN (TimeGAN) and Time-series VAE (TimeVAE), which can preserve the temporal structure of data and synthesize dynamically consistent samples, effectively improving the accuracy and generalization ability of soft measurement models. However, existing methods are mostly based on single-time-scale modeling, making it difficult to adapt to the multi-scale dynamic characteristics of complex industrial systems.

[0004] For the aforementioned technologies, single-scale modeling loses high-frequency details or ignores long-term trends, resulting in an inability to perfectly match changes in industrial processes. Summary of the Invention

[0005] In order to capture data features from multiple time scales and design more robust generative models to improve the practical application effect of the models, this invention provides an industrial data augmentation method based on hierarchical temporal progressive diffusion networks.

[0006] In a first aspect, the present invention provides an industrial data augmentation method using a hierarchical temporal progressive diffusion network, employing the following technical solution: An industrial data augmentation method using hierarchical temporal progressive diffusion networks includes: Step 1: Obtain the time series dataset in response to the input signal; Step 2: Obtain the time scale state data corresponding to the preset time scale based on the time series dataset; Step 3: Perform forward diffusion on the time-scale state data to generate noisy time-scale state data; Step 4: Obtain the information stage features corresponding to the time scale based on all the noisy time scale state data; Step 5: Obtain adjacent information stage features by concatenating adjacent information stage features; Step 6: Generate fused labels based on the adjacent information stage features; Step 7: Obtain denoised and clean virtual samples based on the fused labels; Step 8: Obtain reconstructed synthetic data based on the denoised clean virtual samples and the time series dataset.

[0007] By adopting the above technical solution, and through analyzing the changing characteristics of industrial data at different time scales, data samples are refined from coarse-grained to fine-grained, and global and local information is effectively integrated to synthesize high-quality virtual samples. This solves the problem that traditional methods rely on single-time-scale modeling, making it difficult to adapt to the multi-scale dynamic characteristics of complex industrial systems.

[0008] Optionally, a method for obtaining the number of time-scale state data based on the time-series data includes: Step 20: Obtain the corresponding convolutional kernel size and sparsity control factor based on all preset sparse scale numbers; Step 21: Select the currently executing sparse scale number as the current sparse scale number; Step 22: Based on the convolution kernel size, perform a padding operation on the time scale state data corresponding to the previous sparse scale number to obtain padding scale data; Step 23: Perform average pooling on the filled scale data based on the convolution kernel size corresponding to the current sparse scale number to obtain average pooled data; Step 24: Perform sparsification processing on the average pooling data based on the sparse control factor to obtain the time-scale state data; Step 25: Determine the next sparse scale number based on the current sparse scale number, and treat it as the current sparse scale number. Repeat steps 22 to 24 until all sparse scale numbers have been selected.

[0009] By employing the above technical solutions, the dynamic changes of data across multiple time scales are analyzed, solving the problem that single-scale modeling is insufficient to adapt to the multi-scale dynamic characteristics of complex industrial systems. The steps described above obtain state data at different time scales, enabling the capture of data features from multiple time scales. This lays the foundation for subsequent diffusion processing and the fusion of global and local information by the HTRDiff model at different time scales, thereby generating high-quality virtual samples that better reflect the actual changes in industrial processes.

[0010] Optionally, the method for forward-diffusion of the time-scale state data to generate the noisy time-scale state data includes: Step 260: Select a preset initial sparse scale number as the current sparse scale number, and select a preset initial forward diffusion scale number as the current forward diffusion scale number. Step 261: Randomly collect the core noise distribution; Step 262: Calculate the noise scheduling coefficient corresponding to the current forward diffusion scale number; Step 263: Obtain the noisy time-scale state data based on the core noise distribution, the noise scheduling coefficient, and the time-scale state data corresponding to the current sparse scale number; Step 264: Determine the next forward diffusion scale number based on the current forward diffusion scale number, and regard it as the current forward diffusion scale number. Continue to execute steps 262 to 264 until all the forward diffusion scale numbers have been selected. Step 265: Determine the next sparse scale number based on the current sparse scale number, and regard it as the current sparse scale number. Continue to execute steps 261 to 265 until all the sparse scale numbers have been selected.

[0011] By employing the above technical solution, progressive perturbation of multi-scale features is achieved through forward diffusion performed at different time scales. The core of this solution is to provide samples with varying degrees of noise for the subsequent denoising process, thereby enabling the model to learn the ability to recover true data features from noise.

[0012] Optionally, a method for obtaining denoised clean virtual samples based on the fused labels is also included: Step 70: Randomly collect initial noise and treat it as input noise; Step 71: Obtain the preset initial time scale of the highest level and define it as the current time scale; Step 72: Generate intermediate denoising results based on the fusion label corresponding to the current time scale and the input noise; Step 73: Save the intermediate denoising result as the input noise; Step 74: Obtain the next level time scale based on the current time scale, and regard it as the current time scale. Continue to execute steps 72 to 73 until the time scale is the time scale corresponding to the lowest level. Step 75: Obtain the denoised clean virtual sample from the input noise based on the fusion label.

[0013] By employing the above technical solution, the denoising network is controlled by fusing labels (the result of cross-scale information fusion). Starting from the initial noise, denoising is performed layer by layer in order from the highest level to the lowest level, thereby achieving gradual refinement of multi-scale features. Its core function is to transform random noise into virtual samples that closely resemble the distribution of real industrial time-series data, while preserving the multi-scale dynamic characteristics of the data (global trends and local details).

[0014] Optionally, it also includes a method for obtaining the reconstructed synthetic data and quality variables based on the denoised clean virtual samples and the time-series dataset, the method comprising: Step 76: Obtain reconstructed synthetic data based on the preset multi-task neural network and the denoised clean virtual samples; Step 77: Predict quality variables based on the multi-task neural network and the time-series dataset; Step 78: Output the reconstructed synthetic data and the quality variables.

[0015] By employing the above technical solution, a multi-task neural network is used to simultaneously perform a self-supervised reconstruction task on denoised clean virtual samples and a supervised prediction task on quality variables based on the original time-series dataset. This approach not only fully leverages the latent information within the virtual samples through self-supervised learning, enhancing the model's ability to capture the inherent patterns of industrial processes, but also strengthens the correlation learning between process variables and quality variables through supervised tasks, ultimately improving the prediction accuracy and generalization ability of the soft measurement model.

[0016] In summary, this application includes at least one of the following beneficial technical effects: It effectively solves the "small data" problem in industrial scenarios by generating high-quality virtual samples to expand the dataset, alleviate model overfitting, and improve the prediction accuracy of soft measurement models for quality variables; By fully capturing the multi-scale dynamic characteristics of industrial time-series data, and extracting features from coarse to fine granular levels through a hierarchical time-series progressive diffusion network, the virtual samples generated are matched with the multi-level evolution patterns of industrial processes, overcoming the limitations of single-time-scale modeling. Attached Figure Description

[0017] Figure 1 Here is a model structure diagram of HIDiff when S=3 in the embodiments of this application: Figure 2 This is a flowchart of an industrial data augmentation method using a hierarchical temporal progressive diffusion network, as described in an embodiment of this application. Figure 3 This is a conditional network structure diagram in the embodiments of this application; Figure 4 This is a quality prediction framework diagram for multi-task joint learning in the embodiments of this application. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0019] This application discloses an industrial data augmentation method based on a hierarchical temporal progressive diffusion network. (Refer to...) Figure 1 and Figure 2 An industrial data augmentation method based on hierarchical temporal progressive diffusion networks includes: Step 1: Obtain the time series dataset in response to the input signal.

[0020] Time series datasets are sequential data containing various process variables that change over time during industrial production. They are usually associated with quality variables and are used for model training and virtual sample generation.

[0021] Taking data acquisition from a sulfur recovery unit (SRU) in an industrial oil refining process as an example, five relevant process variables were selected as model inputs to predict the SO2 concentration (mass variable) in the tail gas of the SRU: gas flow rate, air flow rate, air flow rate, gas flow rate in the SWS region, and air flow rate in the SWS region. In the experiment, a total of 700 samples were collected, with 500 samples used for training and 200 samples used for testing. The data sets of each variable changing over time in these samples constitute the time-series dataset used in this study.

[0022] Step 2: Obtain the time scale state data corresponding to the preset time scale based on the time series dataset.

[0023] Time scale refers to the scale at which industrial time-series data exhibits different dynamic changes. Different time scales correspond to different feature extraction granularities; by gradually increasing the convolution kernel size, coarser, deeper features can be extracted.

[0024] Time-scale state data, denoted as Trend components at different time scales are extracted sequentially from the original time series data and obtained through average pooling, filling, and sparsification operations to reflect the dynamic characteristics of the data at the corresponding time scale.

[0025] Step 3: Perform forward diffusion on the time-scale state data to generate noisy time-scale state data.

[0026] Forward diffusion refers to the process of gradually adding noise to time-scale state data based on a preset formula at different time scales.

[0027] Noisy timescale state data is data obtained by forward diffusion of timescale state data. In other words, it's data generated by adding noise to the original timescale state data at different timescales according to the forward diffusion formula. express.

[0028] Step 4: Obtain the information stage features corresponding to the time scale based on all noisy time scale state data.

[0029] Information stage characteristics, denoted as It is extracted from noisy time-scale state data, referring to latent features at different time scales, and is a fusion label ( The basis for generation is used to capture the core dynamic characteristics of data at this time scale. yes It is obtained by a 3×1 convolution.

[0030] Step 5: Obtain adjacent information stage features by splicing adjacent information stage features.

[0031] Reference Figure 3 Adjacent information stage features refer to the information stage features corresponding to two adjacent levels on a time scale (i.e., and Adjacent information stage features are obtained through a "concat" operation.

[0032] Step 6: Generate fusion labels based on adjacent information stage features.

[0033] Merge tags, denoted as The method for obtaining the fusion tag is as follows: in This represents the information stage characteristics at the current time scale. For information features at a higher time scale, the two are adjacent in level and are fused into a conditional signal through concatenation and convolution. Through the above process, the fusion label achieves the organic integration of information from different time scales, providing cross-scale guidance for the subsequent denoising process.

[0034] Step 7: Obtain denoised and clean virtual samples based on fused labels.

[0035] A denoised and clean virtual sample, denoted as This refers to a virtual sample that closely resembles the distribution of real industrial time-series data, generated by progressively removing noise from the initial noise through a hierarchical temporal progressive diffusion network (HTRDiff). The process involves: randomly collecting initial noise as input to the denoising process; iteratively denoising layer by layer from the highest to the lowest time scale, with each level guided by fused labels; and finally, after iterative denoising across all levels, the result is the cleaned virtual sample. .

[0036] Step 8: Obtain reconstructed synthetic data based on denoised clean virtual samples and time series datasets.

[0037] Reconstructing synthetic data, denoted as In a multi-task learning framework, this refers to samples reconstructed from denoised and clean virtual samples through self-supervised learning. The aim is to enable the model to learn the intrinsic structure and potential patterns among process variables. The specific methods will be described in detail in step 80 below.

[0038] Methods for obtaining time-scale state data based on time-series data include: Step 20: Obtain the corresponding convolution kernel size and sparsity control factor based on all preset sparse scale numbers.

[0039] The sparse scale number, denoted as s, is a number used to distinguish different processing stages at different time scales during the process of processing industrial time-series data and constructing state data at different time scales. Its value range is s=1, 2, 3..., S-1, gradually increasing from 1 to a maximum of S-1, where S represents the total number of time scales.

[0040] The kernel size, denoted as τ s , is the convolution kernel parameter used when extracting state data at different time scales. By gradually increasing its size, coarser, deeper features are extracted, corresponding to the feature extraction granularity of data at different time scales.

[0041] The sparsity control factor, denoted as α, is a parameter used to sparsify the data after average pooling. It controls the sparsity of the time-scale state data through sparsification operations to highlight key features and suppress redundant information.

[0042] Step 21: Select the currently executing sparse scale number and define it as the current sparse scale number.

[0043] The current sparse scale number refers to the specific scale number being processed during the iterative process of extracting multi-timescale state data. It is used to identify the current timescale processing stage, and its value range is the sparse scale number range.

[0044] The current sparse scale number will continuously increase as the program runs, and its increasing order is consistent with the requirement of extracting data features "from fine-grained to coarse-grained". The sparse scale with a smaller number corresponds to a smaller convolution kernel size, which is used to extract high-frequency details of the data; the sparse scale with a larger number corresponds to a larger convolution kernel size, which is used to extract long-term trends of the data.

[0045] Step 22: Based on the kernel size, perform a padding operation on the time scale state data corresponding to the previous sparse scale number to obtain the padding scale data.

[0046] Time-scale state data, denoted as The time-scale state data corresponding to the previous sparse scale number is denoted as... .

[0047] Padding is a preprocessing operation performed on data from a previous time scale during the extraction of time-scale state data to maintain data consistency. Its parameter is the kernel size corresponding to the current scale. Specifically... The processing formula is: .

[0048] Filled scale data is the time-scale state data corresponding to the previous sparse scale number. The intermediate data obtained after performing the fill operation is the input for the average pooling operation.

[0049] Step 23: Perform average pooling on the filled scale data based on the kernel size corresponding to the current sparse scale number to obtain average pooled data.

[0050] Average pooling is a feature extraction operation performed on filled-scale data during the extraction of time-scale state data. Its purpose is to extract the overall features of a specified region by calculating the average value of the data within that region, thereby reducing data dimensionality while preserving key trend information. The formula for performing average pooling on filled-scale data is: .

[0051] Average pooling data is intermediate data obtained after performing average pooling on the filled-scale data. It is the result of averaging the regions corresponding to the kernel size of the filled-scale data, and provides input for subsequent sparsification processing.

[0052] Step 24: Perform sparsification processing on the average pooling data based on the sparsity control factor to obtain time-scale state data.

[0053] The sparse control factor, denoted as The sparsity control factor is a parameter used to regulate the degree of sparsity processing. By controlling the sparsity operation on the average pooling data, it filters and retains key features and suppresses redundant information, making it an important control variable for generating time-scale state data. The sparsity control factor is preset, and the sparsity intensity is manually adjusted.

[0054] Sparsification, denoted as It is a feature filtering operation performed on average pooled data. It is based on the sparsity control factor. By removing or weakening secondary information, it retains key features that are meaningful to the current time scale, making the data more sparsity while maintaining the core trend. It is the final step in generating time-scale state data.

[0055] The complete formula for obtaining time-scale state data is: , where s=1,2...S-1.

[0056] Step 25: Determine the next sparse scale number based on the current sparse scale number, and treat it as the current sparse scale number. Repeat steps 22 to 24 until all sparse scale numbers have been selected.

[0057] Determining the next number based on the current number (incrementing sequentially) is to extract multi-scale features of the data step by step from fine-grained to coarse-grained, ensuring that state data at each time scale can be generated sequentially.

[0058] Steps 22 to 24 are executed in a loop, that is, each sparse scale number is processed sequentially. The core significance is to generate a complete set of multi-time scale state data through layer-by-layer iteration.

[0059] Optionally, the method for forward-diffusion of time-scale state data to generate noisy time-scale state data includes: Step 260: Select a preset initial sparse scale number as the current sparse scale number, and select a preset initial forward diffusion scale number as the current forward diffusion scale number.

[0060] The initial sparse scale number is a pre-set starting sparse scale number during forward diffusion of the time-scale state data. It serves as the first sparse scale number processed in the entire forward diffusion process. During forward diffusion, the range of values ​​for the sparse scale number differs from the range used when obtaining the time-scale state data: s = 0, 1, 2...S-1 (where S is the total number of time scales). The initial sparse scale number is typically the starting value within this range (i.e., S = 0), and it is the starting point for iteratively processing all sparse scale numbers.

[0061] The initial forward diffusion scale number is a pre-set starting forward diffusion scale number when performing forward diffusion at each sparse scale, serving as the starting point for the forward diffusion process at the current sparse scale. The initial forward diffusion scale number is usually the starting value of the forward diffusion time step (i.e., t=1), and it is the starting point for iteratively processing all forward diffusion scale numbers, used to control the gradual and progressive process of noise addition at the same sparse scale.

[0062] Step 261: Randomly collect the core noise distribution.

[0063] Core noise distribution , is the distribution state of noise data randomly sampled from the N(0,I) Gaussian distribution when forward diffusion is performed on the time-scale state data.

[0064] Step 262: Calculate the noise scheduling coefficient corresponding to the current forward diffusion scale number.

[0065] Noise dispatch coefficient, , which is the noise scheduling coefficient corresponding to the current forward diffusion scale number t, used to control the degree of noise addition.

[0066] Step 263: Obtain the noisy timescale state data based on the core noise distribution, noise scheduling coefficient, and the timescale state data corresponding to the current sparse scale number.

[0067] Noisy timescale state data, denoted as Noise is generated by adding noise to the time-scale state data corresponding to the current sparse scale number during the forward diffusion process of time-scale state data. It is used to simulate the gradual change of data from the original state to the noisy state. It is an intermediate product of the forward diffusion process and reflects the noisy characteristics under a specific forward diffusion scale.

[0068] The specific formula is as follows: ; This is noisy timescale state data; The noise scheduling coefficient corresponding to the current forward diffusion scale number t; This refers to the time-scale state data corresponding to the current sparse scale number s; This represents the core noise distribution.

[0069] The physical meaning of this formula is: by assigning weights to the original time-scale state data and the core noise distribution respectively. and The data are summed to generate noisy time-scale state data corresponding to the forward diffusion scale t and the sparse scale s, thus achieving the gradual addition of noise.

[0070] Step 264: Determine the next forward diffusion scale number based on the current forward diffusion scale number, and treat it as the current forward diffusion scale number. Continue to execute steps 262 to 264 until all forward diffusion scale numbers have been selected.

[0071] The next forward diffusion scale number is the next forward diffusion scale number to be processed after the current forward diffusion scale number has been processed. It is used to continue the forward diffusion process until all forward diffusion scale numbers are covered.

[0072] The next forward diffusion scale is numbered in ascending order, from t=1 to t=T, where T is the total number of forward diffusion scales.

[0073] Step 265: Determine the next sparse scale number based on the current sparse scale number, and treat it as the current sparse scale number. Continue to execute steps 261 to 265 until all sparse scale numbers have been selected.

[0074] The next sparse scale number is the next sparse scale number to be processed determined based on the current number after the forward diffusion corresponding to the current sparse scale number is completed. It is used to perform forward diffusion on the state data of the next time scale until all sparse scale numbers are covered.

[0075] The next sparse scale number is determined in ascending order, from s=1 to s=S-1, consistent with the generation order of the time-scale state data.

[0076] It also includes a method for obtaining denoised clean virtual samples based on the fused labels: Step 70: Randomly collect initial noise and treat it as input noise.

[0077] Input noise is the noise data that serves as the network input during the denoising process. Its initial value is randomly collected initial noise, and it will be updated to the intermediate denoising results of the previous level during the subsequent iterative denoising process. It is used to gradually generate denoised virtual samples and is the core input variable of the hierarchical temporal progressive diffusion network (HTRDiff) denoising process.

[0078] The process of randomly acquiring initial noise refers to randomly extracting noise data from Gaussian noise.

[0079] Step 71: Obtain the preset highest-level initial time scale and define it as the current time scale.

[0080] The highest-level initial time scale is the preset starting time scale during the denoising process. It is the highest-level scale among all time scales and corresponds to the coarsest-grained time features.

[0081] The current timescale is the timescale in which the denoising process is being processed during the denoising iteration, and it is used to identify the current denoising level.

[0082] Step 72: Generate intermediate denoising results based on the fusion label and input noise corresponding to the current time scale.

[0083] The intermediate denoising result is the intermediate data obtained by the denoising network based on the fused label and input noise corresponding to the current time scale during the hierarchical denoising process. It is an intermediate state from the input noise to the final denoised clean virtual sample. As the time scale gradually decreases from the highest level to the lowest level, the intermediate denoising result will be continuously refined and gradually approach the real data distribution.

[0084] The method for generating intermediate denoising results is to use a denoising network designed by the staff. The denoising network is a denoising network based on Pyramid Multi-scale CNN Transformer (PMC-former), which combines the local feature extraction capability of CNN with the global dependency modeling capability of Transformer, and realizes the capture of local detailed features and global dynamic features.

[0085] The generation of intermediate denoising results depends on the guidance of the fusion label to the denoising network. The fusion label serves as conditional information, enabling the denoising network to remove some noise from the input noise based on the feature associations of the current time scale, while retaining features related to the real data.

[0086] Step 73: Save the intermediate denoising result as input noise.

[0087] Saving the intermediate denoising results generated at the current level as the input noise for the next level allows the next level's denoising network to further refine features based on the existing denoising, gradually remove noise and enhance details, and finally generate high-quality denoised clean virtual samples from the initial noise.

[0088] Step 74: Obtain the next level time scale based on the current time scale and treat it as the current time scale. Continue to execute steps 72 to 73 until the time scale is the time scale corresponding to the lowest level scale.

[0089] The next level of time scale is a lower-level time scale determined based on the current time scale during the denoising iteration process. It belongs to the same time scale system as the current time scale and is arranged in descending order of level. It is used to gradually introduce more refined local features during the denoising process, realizing progressive data generation from coarse-grained to fine-grained.

[0090] The lowest time scale corresponds to the finest-grained features and is the carrier of high-frequency dynamic information in industrial time series data. Completing denoising at this level means that the virtual sample retains both the global trend and local details, which conforms to the characteristic of real industrial data that "global dynamics and local features coexist."

[0091] Step 75: Obtain denoised clean virtual samples from input noise based on fused labels.

[0092] A denoised and clean virtual sample, denoted as Unlike intermediate denoising results, which are transitional data during the iteration process and only complete denoising at some levels, leaving some residual noise and incomplete feature refinement, making them unusable as valid samples, intermediate denoising results, when the time scale corresponds to the lowest level, result in a fully denoised and feature-complete virtual sample after completing denoising iterations at all levels. This marks the end of the entire denoising process and signifies the completion of virtual sample generation.

[0093] Reference Figure 4 It also includes a method for obtaining reconstructed synthetic data and quality variables based on denoised clean virtual samples and time-series datasets, the method comprising: Step 76: Obtain reconstructed synthetic data based on a preset multi-task neural network and denoised clean virtual samples.

[0094] Multi-task neural networks are improved neural networks based on LSTM architecture, and they differ significantly from ordinary LSTMs. Their core lies in simultaneously executing two tasks through a joint training mechanism: a self-supervised task and a supervised task. The self-supervised task involves reconstructing denoised clean virtual samples to learn the intrinsic structure and potential patterns between process variables to generate reconstructed synthetic data.

[0095] Step 77: Predict quality variables based on multi-task neural networks and time-series datasets.

[0096] Mass variables are critical indicators in industrial production processes that are difficult to measure in real time but are essential for production safety and efficiency. For example, to measure the sulfur dioxide concentration in the exhaust gas of a sulfur recovery unit, the sulfur dioxide concentration can be indirectly predicted by measuring parameters such as gas flow rate and air flow rate. The supervised task, on the other hand, predicts mass variables based on time-series datasets and denoised clean virtual samples to establish a mapping relationship between process variables and mass variables. By inputting denoised clean virtual samples, time-series datasets, and mixed data into the neural network, this processing method can fully extract the potential information in the virtual samples and combine it with the features of the original data, enhancing the model's ability to capture the complex dynamic characteristics of industrial processes, thereby improving the prediction accuracy and generalization ability of the soft measurement model for mass variables.

[0097] Step 78: Output the reconstructed synthetic data and quality variables.

[0098] Reconstructed synthetic data can be used to expand the training dataset, providing richer learning samples for soft sensor models. Quality variables, as key indicators in industrial production, can reflect the real-time status of the production process, providing operators with accurate decision-making support, facilitating timely adjustments to process parameters, and ensuring production safety and efficiency.

[0099] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An industrial data augmentation method based on hierarchical temporal progressive diffusion networks, characterized in that, include: Step 1: Obtain the time series dataset in response to the input signal; Step 2: Obtain the time scale state data corresponding to the preset time scale based on the time series dataset; Step 3: Perform forward diffusion on the time-scale state data to generate noisy time-scale state data; Step 4: Obtain the information stage features corresponding to the time scale based on all the noisy time scale state data; Step 5: Obtain adjacent information stage features by concatenating adjacent information stage features; Step 6: Generate fused labels based on the adjacent information stage features; Step 7: Obtain denoised and clean virtual samples based on the fused labels; Step 8: Obtain reconstructed synthetic data based on the denoised clean virtual samples and the time series dataset.

2. The industrial data augmentation method based on a hierarchical temporal progressive diffusion network according to claim 1, characterized in that, A method for obtaining the time-scale state data of the specified number of scales based on the time-series data includes: Step 20: Obtain the corresponding convolutional kernel size and sparsity control factor based on all preset sparse scale numbers; Step 21: Select the currently executing sparse scale number as the current sparse scale number; Step 22: Based on the convolution kernel size, perform a padding operation on the time scale state data corresponding to the previous sparse scale number to obtain padding scale data; Step 23: Perform average pooling on the filled scale data based on the convolution kernel size corresponding to the current sparse scale number to obtain average pooled data; Step 24: Perform sparsification processing on the average pooling data based on the sparse control factor to obtain the time-scale state data; Step 25: Determine the next sparse scale number based on the current sparse scale number, and treat it as the current sparse scale number. Repeat steps 22 to 24 until all sparse scale numbers have been selected.

3. The industrial data augmentation method based on a hierarchical temporal progressive diffusion network according to claim 2, characterized in that, The method for generating the noisy time-scale state data by forward diffusion of the aforementioned time-scale state data includes: Step 260: Select a preset initial sparse scale number as the current sparse scale number, and select a preset initial forward diffusion scale number as the current forward diffusion scale number. Step 261: Randomly collect the core noise distribution; Step 262: Calculate the noise scheduling coefficient corresponding to the current forward diffusion scale number; Step 263: Obtain the noisy time-scale state data based on the core noise distribution, the noise scheduling coefficient, and the time-scale state data corresponding to the current sparse scale number; Step 264: Determine the next forward diffusion scale number based on the current forward diffusion scale number, and regard it as the current forward diffusion scale number. Continue to execute steps 262 to 264 until all the forward diffusion scale numbers have been selected. Step 265: Determine the next sparse scale number based on the current sparse scale number, and regard it as the current sparse scale number. Continue to execute steps 261 to 265 until all the sparse scale numbers have been selected.

4. The industrial data augmentation method based on a hierarchical temporal progressive diffusion network according to claim 1, characterized in that, It also includes a method for obtaining denoised clean virtual samples based on the fused labels: Step 70: Randomly collect initial noise and treat it as input noise; Step 71: Obtain the preset initial time scale of the highest level and define it as the current time scale; Step 72: Generate intermediate denoising results based on the fusion label corresponding to the current time scale and the input noise; Step 73: Save the intermediate denoising result as the input noise; Step 74: Obtain the next level time scale based on the current time scale, and regard it as the current time scale. Continue to execute steps 72 to 73 until the time scale is the time scale corresponding to the lowest level. Step 75: Obtain the denoised clean virtual sample from the input noise based on the fusion label.

5. The industrial data augmentation method based on a hierarchical temporal progressive diffusion network according to claim 4, characterized in that, It also includes a method for obtaining the reconstructed synthetic data and quality variables based on the denoised clean virtual samples and the time-series dataset, the method comprising: Step 76: Obtain reconstructed synthetic data based on the preset multi-task neural network and the denoised clean virtual samples; Step 77: Predict quality variables based on the multi-task neural network and the time-series dataset; Step 78: Output the reconstructed synthetic data and the quality variables.

Images