An industrial data interaction method, an interaction platform and application thereof

Abstract

The application relates to the technical field of industrial internet, solves the technical problems that a model structure in the prior art is simple, long-range and local dependence cannot be considered, a long sequence is roughly processed, and information loss is large, and in particular relates to an industrial data interaction method, an interaction platform and application of the interaction platform, the platform comprises a preliminary processing module, a preprocessing module, a buffer module, an analysis and prediction module, an interaction module and a security module, the preliminary processing module is used for collecting original data of the industrial data interaction platform, the application can dynamically focus on the importance of different time points in a sequence, no matter how long the dependence span is, while ensuring accurate capture of trend and periodic characteristics, thereby greatly enhancing the expression ability of the model to complex and nonlinear time series relationships, and the application can intelligently compress and retain the most critical information according to the characteristics of an input sequence, so that the model has strong robustness for sequences of different lengths.

Description

Technical Field

[0001] This invention relates to the field of industrial internet technology, and in particular to an industrial data interaction method, interaction platform and its application. Background Technology

[0002] The Industrial Internet is a new type of infrastructure, application model, and industrial ecosystem that deeply integrates next-generation information and communication technologies with the industrial economy. By comprehensively connecting people, machines, things, and systems, it builds a brand-new industrial manufacturing system covering all factors, the entire industrial chain, and the entire value chain, providing a way to realize the digital, networked, intelligent, and green development of industry.

[0003] Existing technologies for industrial data processing have several drawbacks. First, existing data processing models have simple structures and are difficult to balance long-range and local dependencies. Second, they are coarse in processing variable-length sequences, resulting in significant information loss. Third, hyperparameter tuning relies on manual experience or inefficient search. Therefore, a new industrial data processing method is needed to solve these problems. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an industrial data interaction method, interaction platform, and its application. It solves the technical problems of existing technologies, such as simple model structure, difficulty in balancing long-range and local dependencies, coarse processing of variable-length sequences, and large information loss. It achieves the goal of comprehensively capturing features of feature data, intelligently processing inputs of different lengths, and efficiently optimizing based on model parameters.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] An industrial data interaction method, comprising the following steps:

[0007] S1. Collect raw data from industrial data interaction, construct a dual-track adaptive processing model based on signal-to-noise ratio perception, and process the raw data through the dual-track adaptive processing model to obtain preliminary processed data.

[0008] S11. Perform anomaly identification on the original data based on dynamic threshold to find the abnormal data, and perform mask generation and moving average interpolation on the abnormal data to obtain masked interpolation data.

[0009] S12. The signal-to-noise ratio of the original data is evaluated to obtain the noise level. The basis function and threshold algorithm of wavelet transform are dynamically selected according to the noise level, and the noise reduction parameters are obtained.

[0010] S13. Based on the mask interpolation processing data and noise reduction parameters, construct a parallel dual-track adaptive processing model for adaptive noise reduction and mask interpolation, and obtain preliminary processed data.

[0011] S2. Preprocess the preliminary data and obtain preprocessed data;

[0012] S3. Based on the preprocessed data, cached data is obtained step by step through mutual information feature selection and dynamic sequence segmentation, and a cache area is constructed based on the cached data;

[0013] S4. Perform sliding window segmentation on the cached data to obtain sequence data. Construct an initial interaction processing model with hyperparameters based on the sequence data. Perform preliminary optimization on the initial interaction processing model to obtain an optimized interaction processing model. Perform secondary optimization on the optimized interaction processing model to obtain the optimal model. Perform predictive analysis on the cached data based on the optimal model to obtain the analysis and prediction results.

[0014] S5. Transmit the analysis and prediction results to the industrial data interaction center to generate interactive decisions.

[0015] Furthermore, in S2, the specific implementation steps are as follows:

[0016] S21. Based on the preliminary processed data, the operating parameters are obtained through the sub-line clustering algorithm. The mean and standard deviation of the preliminary processed data are calculated. The operating parameters are standardized based on the mean and standard deviation to obtain standardized data.

[0017] S22. Construct a trained isolation forest model for predicting standardized data anomaly scores and obtain the parameters of the isolation forest model.

[0018] S23. Define a probability distribution for the parameters of each isolated forest model to provide the search space, pre-set a maximum number of iterations, initialize an empty observation set to store parameter combinations and evaluation results, and fit the observation set with a Gaussian process probability model to obtain the fitted model.

[0019] S24. Define a collection function. By maximizing the collection function, obtain the next candidate parameter combination that is most likely to find the global optimum. Replace the parameters of the original isolated forest model with the candidate parameter combination to obtain the updated forest model.

[0020] S25. Obtain the current iteration number, update the forest model based on the maximum iteration number, and obtain the optimized forest model;

[0021] If the current iteration count is less than the maximum iteration count, return to S23 to continue iterating;

[0022] If the current iteration count equals the maximum iteration count, then the current updated forest model is used as the optimized forest model;

[0023] S26. Input the standardized data into the optimized forest model to obtain outlier data. Smooth the outlier data using the Kalman filter to obtain corrected data. Replace the original outlier data with the corrected data. Set the smoothed standardized data as preprocessed data.

[0024] Furthermore, in S3, the specific implementation steps are as follows:

[0025] S31. Preset a window K with width a, and calculate the window cache data for the preprocessed data using the moving average formula and time decay factor;

[0026] S32. Obtain correlation-optimized data based on window cache data using the mutual information feature selection method;

[0027] S33. Based on the correlation optimization data, obtain cached data through dynamic sequence segmentation, and construct a cache area based on the cached data.

[0028] Furthermore, in S4, the specific implementation steps are as follows:

[0029] S41. Extract cached data from the cache area, and divide the continuous time series in the cached data into fixed-length sequence segments using the sliding window serialization method. Each sequence segment is used as an input sample of the model to obtain sequence data.

[0030] S42. Construct an initial interaction processing model with hyperparameters based on sequence data;

[0031] S43. Convert multiple sets of hyperparameters in the initial interaction processing model into input vectors, use the input vectors as individuals in the population optimization to construct an optimization parameter layer for optimizing the input vectors, and obtain the hyperparameters to be tested. Replace the initial hyperparameters in the initial interaction processing model with the hyperparameters to be tested to generate an optimized interaction processing model.

[0032] S44. The optimal interaction processing model is trained and regularized using cosine annealing and gradient pruning to obtain the optimal model, and a secondary optimization layer is constructed for secondary optimization of the optimal interaction processing model.

[0033] S45. Input the cached data into the optimal model for predictive analysis to obtain the analysis and prediction results.

[0034] Furthermore, in S42, the specific implementation steps are as follows:

[0035] S421. Construct a multi-head attention layer based on the multi-head attention mechanism. Calculate the weight values ​​based on the sequence data using the attention weight formula. Calculate the weighted sequence based on the weight values ​​to highlight key moments and capture long-distance dependencies.

[0036] S422. A bidirectional gated loop layer is constructed using a bidirectional GRU structure. The weighted sequence is processed and temporal features are extracted from both the forward and backward directions using the bidirectional GRU method. The feature data used to enhance feature expression is obtained by splicing the forward and backward temporal features.

[0037] S423. Obtain convolutional data by performing temporal convolution operations on the feature data. Then, fuse local details and global dependencies using the residual calculation formula based on the convolutional data to obtain fused data for representing complex temporal fusion features, and construct a temporal fusion layer.

[0038] S424. Obtain the base window length by dividing the total length of the fused data by the target length to be processed and rounding it up. Based on the base window length, dynamically adjust the last incomplete window and obtain the partitioned data divided by the window. Perform max pooling operation on each partitioned data and obtain compressed fused data of fixed length by concatenating the data. At the same time, construct a dynamic compressed fusion layer.

[0039] S425. Based on the compression fusion data, a fully connected layer is constructed to obtain the analysis results. An initial interactive processing model with hyperparameters is constructed based on a multivariate attention layer, a bidirectional gated loop layer, a temporal fusion layer, a dynamic compression fusion layer, and a fully connected layer.

[0040] Furthermore, in S43, the specific implementation steps are as follows:

[0041] S431. Define the search space of the input vector, divide the cached data into training set and validation set, define the detection threshold, use the cross-entropy loss function as the loss function for optimizing the parameter model, extract an input data in the training set, calculate the loss value relative to the training set, and combine the input vector and the loss value to form the initial data.

[0042] S432. Obtain model parameters by maximizing marginal likelihood. and The surrogate model is obtained by fitting the initial data using a Gaussian process fitting model.

[0043] S433. Obtain the cumulative distribution value and probability density value of the input data in the standard normal distribution. Obtain the predicted mean and standard deviation of the input data according to the surrogate model. Obtain the minimum loss value of the training set based on the loss function. Obtain the test difference by subtracting the minimum loss value from the predicted mean and 0.01. Obtain the collected value by multiplying the test difference by the cumulative distribution value and adding the product of the standard deviation and the probability density value. Optimize the collected value globally in the search space using the L-BFGS algorithm. Take the data point with the largest collected value as the updated data point. The input vector corresponding to the collected value is the vector to be tested.

[0044] S434. Convert the vector to be tested into hyperparameters to be tested, replace the hyperparameters in the initial interactive processing model to obtain the model to be tested, construct the test model based on the mean absolute error formula, and calculate the test value of the model to be tested based on the training set and validation set.

[0045] S435. Test the model to be tested based on the test value and obtain the optimized interactive processing model;

[0046] If the test value is less than the detection threshold, the current model to be tested will be used as the optimized interactive processing model.

[0047] If the test value is greater than or equal to the detection threshold, return to S431 to continue optimizing the model to be tested.

[0048] The technical solution also provides an industrial data interaction platform, which includes: a preliminary processing module, a preprocessing module, a buffer module, an analysis and prediction module, an interaction module, and a security module.

[0049] Furthermore, the preliminary processing module is used to collect raw data from the industrial data interaction platform, construct a dual-track adaptive processing model based on signal-to-noise ratio perception, and process the raw data through the dual-track adaptive processing model to obtain preliminary processed data.

[0050] The preprocessing module is used to preprocess the initially processed data and obtain preprocessed data;

[0051] The buffer module is used to obtain cached data step by step from the preprocessed data through mutual information feature selection and dynamic sequence segmentation, and to build a cache area based on the cached data;

[0052] The analysis and prediction module is used to divide the cached data into sequence data by sliding window segmentation, construct an initial interaction processing model with hyperparameters based on the sequence data, perform preliminary optimization on the initial interaction processing model to obtain an optimized interaction processing model, perform secondary optimization based on the optimized interaction processing model to obtain the optimal model, and perform predictive analysis on the cached data based on the optimal model to obtain the analysis and prediction results.

[0053] The interaction module is used to transmit the analysis and prediction results to the industrial data interaction center to generate interactive decisions.

[0054] Furthermore, the security module is used to encrypt and sign the data processed by the preliminary processing module, preprocessing module, buffering module, analysis and prediction module, and interaction module, and to terminate the data storage through blockchain, and to perform two-factor authentication by combining device fingerprint and dynamic password.

[0055] The technical solution also provides an application of an industrial data interaction method, which is applied to the semiconductor manufacturing, medical device, smart water management, and metallurgical industries.

[0056] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0057] 1. This invention can integrate the data after the first track of abnormal data processing and the second track of noise reduction processing to obtain preliminary processed data. This integration method takes into account the outliers and noise problems in the data. Through parallel processing of dual tracks and final integration, it can more comprehensively and effectively improve the quality of raw industrial data and provide a reliable foundation for subsequent data analysis and application.

[0058] 2. This invention constructs a precise, stable, and fully automated industrial data anomaly detection pipeline by combining standardized working condition perception, adaptive model optimization, and intelligent post-processing. It avoids the limitations of single threshold processing and overcomes the blindness and suboptimal nature of traditional manual parameter tuning. It significantly improves the sensitivity and accuracy of anomaly identification, outputs smooth and reliable final results, and greatly enhances the robustness and practicality of the entire system in complex industrial environments.

[0059] 3. This invention achieves accurate capture and efficient learning of deep patterns in complex time-series data through multi-level feature extraction, intelligent adaptive processing, efficient hyperparameter optimization, and stable training strategies. This step makes the method perform well in terms of accuracy, stability, and generalization ability, and it is particularly suitable for handling data analysis tasks with complex temporal dependencies in the real world, providing reliable technical support for subsequent data interaction and decision-making.

[0060] 4. This invention can dynamically focus on the importance of different time points in the sequence, regardless of the length of the dependency span, while ensuring accurate capture of trend and periodic features. The innovative design of the spatiotemporal fusion layer integrates local details with global dependencies, and the residual connection ensures that while deepening features, it greatly enhances the model's ability to express complex and nonlinear temporal relationships. It can also intelligently compress and retain the most critical information according to the characteristics of the input sequence, making the model robust to sequences of different lengths. Attached Figure Description

[0061] Figure 1 This is a flowchart of an industrial data interaction method according to the present invention;

[0062] Figure 2 This is a structural block diagram of an industrial data interaction platform according to the present invention.

[0063] In the diagram: 1. Preliminary processing module; 2. Preprocessing module; 3. Buffer module; 4. Analysis and prediction module; 5. Interaction module; 6. Security module. Detailed Implementation

[0064] 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 numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0065] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” used in this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0066] It should be understood that although the terms first, second, third, etc., may be used in this invention to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of this invention, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0067] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0068] Example 1

[0069] Due to the limitations of existing technologies, such as simple model structures, difficulty in balancing long-range and local dependencies, coarse processing of variable-length sequences, and significant information loss, this embodiment proposes an industrial data interaction method. This method can comprehensively capture features from feature data, intelligently process inputs of different lengths, and efficiently optimize based on model parameters. Figure 1 As shown, the method includes the following steps:

[0070] S1. Collect raw data from industrial data exchange, construct a dual-track adaptive processing model based on signal-to-noise ratio (SNR) sensing, and process the raw data through the dual-track adaptive processing model to obtain preliminary processed data. Because industrial data is easily affected by multiple factors during the collection process, such as environmental factors, industrial equipment vibration, and inter-equipment interference, the acquired data contains a large amount of noise, leading to inaccurate data. To solve this problem, this embodiment proposes a more detailed implementation method, as follows:

[0071] S11. Anomaly identification based on dynamic thresholds is performed on the original data to find anomalous data. Masking and moving average interpolation are then applied to the anomalous data to obtain masked and interpolated data. This masked and interpolated data can perform moving average interpolation and mask generation, and can detect and process anomalous data based on dynamic thresholds. This generates the first track of a dual-track adaptive processing model. The dual-track adaptive processing model employs a parallel processing approach, where two processing tracks simultaneously process the original industrial data in different ways, ultimately integrating them to obtain preliminary processed data. This parallel architecture fully leverages the advantages of different processing methods, improving the comprehensiveness and accuracy of data processing. It comprehensively considers outliers and noise issues in the data. Through dual-track parallel processing and final integration, the quality of the original industrial data can be improved more comprehensively and effectively, providing a reliable foundation for subsequent data analysis and applications.

[0072] S12. The signal-to-noise ratio (SNR) of the original data is evaluated to determine the noise level. Based on the noise level, the basis functions and thresholding algorithm of the wavelet transform are dynamically selected, and denoising parameters are obtained. The original data is then subjected to wavelet transform denoising processing using the selected basis functions and thresholding algorithm. Wavelet transform can decompose the signal into different frequency sub-bands. Thresholding removes noise components from the noise sub-bands, and wavelet reconstruction yields the denoised data. This processing step primarily addresses the noise problem in the data, improving the SNR and making the data clearer and more accurate. It obtains denoising parameters based on basis functions and thresholding algorithms, generating the second track of the dual-track adaptive processing model.

[0073] S13. Based on the mask interpolation data and noise reduction parameters, a parallel adaptive dual-track adaptive processing model for noise reduction and mask interpolation is constructed to obtain preliminary processed data. After outlier processing (the first track) and noise reduction processing (the second track), the data is integrated to obtain preliminary processed data. This integration method comprehensively considers outliers and noise issues in the data. Through parallel dual-track processing and final integration, the quality of the original industrial data can be improved more comprehensively and effectively, providing a reliable foundation for subsequent data analysis and applications.

[0074] S2. Preprocess the initial data to obtain preprocessed data; the denoised data still contains a large amount of outlier data, requiring further anomaly identification and preprocessing for more accurate state analysis. To solve this problem, the detailed implementation steps are as follows:

[0075] S21. Based on the preliminary processed data, the operating parameters are obtained through the sub-line clustering algorithm. The mean and standard deviation of the preliminary processed data are calculated. The operating parameters are standardized based on the mean and standard deviation to obtain standardized data.

[0076] S22. Construct a trained isolation forest model for predicting standardized data anomaly scores and obtain the parameters of the isolation forest model.

[0077] S23. Define a probability distribution for the parameters of each isolated forest model to provide the search space, pre-set a maximum number of iterations, initialize an empty observation set to store parameter combinations and evaluation results, and fit the observation set with a Gaussian process probability model to obtain the fitted model. The Gaussian process probability model is a mathematical model generated based on the Gaussian process probability formula. It is a type of stochastic process in probability theory and mathematical statistics, which is a combination of a series of normally distributed random variables in an exponential set. The fitted model is the fitted model obtained by fitting standardized data based on the Gaussian process probability model.

[0078] S24. Define a collection function. By maximizing the collection function, obtain the next candidate parameter combination most likely to find the global optimum. Replace the parameters of the original isolated forest model with the candidate parameter combination to obtain the updated forest model. The advantage of fitting the observation set with a Gaussian process probability model is that it can provide a probabilistic model of the evaluation result of the objective function, i.e., the parameter combination. It not only predicts the performance value but also quantifies the uncertainty, thus achieving efficient exploration-utilization trade-offs in the Bayesian optimization process. This method is data-efficient, highly adaptable, does not require assuming the form of the objective function, and can quickly converge to the global optimum parameter combination with a small number of iterations. Combining the advantages of the optimized forest model, i.e., the isolated forest model after Bayesian optimization and parameter tuning, includes the advantages of automated parameter optimization improving the accuracy and robustness of anomaly detection, while retaining the efficiency and unsupervised learning advantages of the isolated forest algorithm itself.

[0079] S25. Obtain the current iteration number, update the forest model based on the maximum iteration number, and obtain the optimized forest model;

[0080] If the current iteration count is less than the maximum iteration count, return to S23 to continue iterating;

[0081] If the current iteration count equals the maximum iteration count, then the current updated forest model is used as the optimized forest model;

[0082] S26. Standardized data is input into the optimized forest model to obtain abnormal data. The abnormal data is then smoothed and corrected using Kalman filtering to obtain corrected data. The original abnormal data is replaced with the corrected data, and the smoothed and corrected standardized data is collected as preprocessed data. Through the synergy of condition-aware standardization, adaptive model optimization, and intelligent post-processing, a precise, stable, and fully automated industrial data anomaly detection pipeline is constructed. This method first enables the model to adapt to multiple operating conditions through sub-line clustering and standardization, avoiding the limitations of single-threshold processing. Then, a Gaussian process based on Bayesian optimization is used to automatically search for the optimal parameter combination for the isolated forest, overcoming the blindness and suboptimal nature of traditional manual parameter tuning, and significantly improving the sensitivity and accuracy of anomaly identification. Finally, Kalman filtering is introduced to smooth the initial anomaly scores, effectively suppressing prediction fluctuations caused by the randomness of the isolated forest, outputting smooth and reliable final results, and greatly enhancing the robustness and practicality of the entire system in complex industrial environments.

[0083] S3. Based on the preprocessed data, cached data is obtained step by step through mutual information feature selection and dynamic sequence segmentation, and a cache area is constructed based on the cached data;

[0084] S31. Preset a window K with width a. Calculate the window cache data for the preprocessed data using the moving average formula and the time decay factor. The time decay factor can be obtained through logistic regression, which is a common method for obtaining decay factors based on the logistic regression model. The moving average formula is a commonly used formula for calculating the moving average.

[0085] S32. Based on the window cache data, obtain the correlation-optimized data through the mutual information feature selection method. The mutual information feature selection method is a feature selection method based on the joint probability calculation formula and conditional entropy. It can capture nonlinear relationships, is not sensitive to data types, is applicable to discrete and continuous variables, and does not require assumptions about data distribution.

[0086] S33. Based on the correlation optimization data, cached data is obtained through dynamic sequence segmentation, and a cache area is constructed based on the cached data. Dynamic sequence segmentation is a time series segmentation technique based on dynamic programming, used to divide time series data into continuous segments with similar statistical characteristics. The cached data is obtained step by step through the segment difference measurement formula, dynamic programming recursive formula, and optimal segmentation point backtracking method.

[0087] S4. The cached data is segmented using a sliding window to obtain sequence data. An initial interactive processing model with hyperparameters is constructed based on the sequence data. This initial interactive processing model is then preliminarily optimized to obtain an optimized interactive processing model. A second optimization is performed based on the optimized interactive processing model to obtain the optimal model. The optimal model is then used to perform predictive analysis on the cached data to obtain the prediction results. The specific method used to solve this problem is as follows:

[0088] S41. Extract cached data from the cache area. Based on the sliding window serialization method, the continuous time series in the cached data is divided into fixed-length sequence segments by a sliding window. Each sequence segment is used as an input sample of the model to obtain the sequence data. The sliding window serialization method is a method for processing cached data based on the sequence segmentation formula. This method ensures the temporal continuity of the sequence by defining the window size and step size, such as the window size being R and the sliding step size being a, and provides a standardized input format for subsequent model processing.

[0089] S42. Construct an initial interaction processing model with hyperparameters based on sequence data; the initial interaction processing model needs further optimization. To address the above issues, the specific implementation steps are as follows:

[0090] S421. Construct a multi-head attention layer based on the multi-head attention mechanism. Calculate the weight values ​​based on the sequence data using the attention weight formula. Calculate the weighted sequence based on the weight values ​​to highlight key moments and capture long-distance dependencies. The weighted sequence is a sequence that has been attention-weighted, preserving the global dependencies of the input sequence. The attention weight formula is a commonly used formula for calculating weights. The weighted sequence is obtained by multiplying the weight values ​​by the sequence data.

[0091] S422. A bidirectional gated recurrent layer is constructed using a bidirectional GRU structure. The weighted sequence is processed and temporal features are extracted from both the forward and backward directions using the bidirectional GRU method. The feature data used to enhance feature representation is obtained by concatenating the forward and backward temporal features. The bidirectional GRU method is a very powerful and commonly used model in natural language processing, speech recognition, time series analysis and other fields. It is used for bidirectional gated recurrent processing and is a common method for obtaining temporal features. The temporal features can be periodic or trend features. The concatenation of forward and backward temporal features can be obtained by direct concatenation. The feature data is a sequence containing forward and backward information, and each time step corresponds to a feature data with a fused hidden state.

[0092] S423. Based on the feature data, convolutional data is obtained through temporal convolution operation. Based on the convolutional data, local details and global dependencies are fused using the residual calculation formula to obtain fused data for representing complex temporal fusion features, and a temporal fusion layer is constructed. Temporal convolution operation is a commonly used convolution calculation method, calculated based on the convolution calculation formula. The residual calculation formula is a commonly used residual calculation method. Combined with convolution operation, local details and global dependencies of temporal data can be fused. The fused data is a sequence after the fusion of local and global features, which enhances the integration of spatiotemporal information.

[0093] S424. Obtain the base window length by dividing the total length of the fused data by the target length to be processed and rounding up. Based on the base window length, dynamically adjust the last incomplete window to obtain the partitioned data of the window. Perform max pooling on each partitioned data and concatenate the data to obtain compressed fused data of a fixed length. At the same time, construct a dynamic compression fusion layer. Adjust the pooling strategy dynamically according to the length of the fused data. You can choose adaptive average pooling or max pooling. In this step, you choose max pooling. Through dynamic max pooling, you can retain key time series information and reduce information loss. Compress the fused data into vector data of a fixed length, i.e., compressed fused data.

[0094] S425. Based on the compression-fusion data, a fully connected layer is constructed to obtain the analysis results. An initial interactive processing model with hyperparameters is then built based on a multi-source attention layer, a bidirectional gated recurrent layer, a temporal fusion layer, a dynamic compression fusion layer, and a fully connected layer. This initial interactive processing model is the initial data processing model, completing data processing before the interactive method. By organically integrating attention, bidirectional recurrent, spatiotemporal convolution, and adaptive aggregation, a powerful feature learning engine is formed. This engine can more accurately and stably model the complex inherent laws of real-world data in tasks such as time series prediction and classification, thus exhibiting excellent comprehensive performance and providing more accurate and stable analytical data for data interaction. Specifically, the multi-source attention layer and bidirectional gated recurrent layer... The collaborative work of the layers can dynamically focus on the importance of different time points in the sequence, regardless of the length of the dependency span; while the bidirectional GRU can deeply understand the dynamic patterns of the time series from both past and future directions, ensuring accurate capture of trend and periodic features. The innovative design of the spatiotemporal fusion layer integrates local details with global dependencies, and the residual connection ensures that while deepening features, the important global information learned by previous layers is not lost, thus greatly enhancing the model's ability to express complex and nonlinear temporal relationships. The adaptive aggregation layer abandons the information loss that may be caused by the traditional fixed pooling method, and can intelligently compress and retain the most critical information according to the characteristics of the input sequence, making the model robust to sequences of different lengths.

[0095] S43. Convert the multiple sets of hyperparameters in the initial interaction processing model into input vectors. Use the input vectors as individuals in the population optimization to construct an optimization parameter layer for optimizing the input vectors, and obtain the hyperparameters to be tested. Replace the initial hyperparameters in the initial interaction processing model with the hyperparameters to be tested to generate an optimized interaction processing model. In order to enable secondary optimization of the model, the specific implementation steps are as follows:

[0096] S431. Define the search space of the input vector, divide the cached data into training and validation sets, define the detection threshold, use the cross-entropy loss function as the loss function for optimizing the parameter model, extract an input data from the training set, and calculate the loss value relative to the training set. Combine the input vector and the loss value to form the initial data. The loss value relative to the training set can be calculated using the cross-entropy loss function, which is a common method for calculating cross-entropy.

[0097] S432. Obtain model parameters by maximizing marginal likelihood. and A surrogate model is obtained by fitting the initial data using a Gaussian process fitting model. The expression of the surrogate model is as follows:

[0098]

[0099] in, Dl Indicates weight, and They represent the first i and the i -1 input data; Maximizing marginal likelihood is a commonly used method to obtain model parameters.

[0100] S433. Obtain the cumulative distribution value and probability density value of the input data in the standard normal distribution. Obtain the predicted mean and standard deviation of the input data according to the surrogate model. Obtain the minimum loss value of the training set based on the loss function. Obtain the test difference by subtracting the minimum loss value from the predicted mean and 0.01. Obtain the collected value by multiplying the test difference by the cumulative distribution value and adding the product of the standard deviation and the probability density value. Optimize the collected value globally in the search space using the L-BFGS algorithm. Use the data point with the largest collected value as the updated data point. The input vector corresponding to the collected value is the vector to be tested. The collected value is a chain structure with multiple data points, each containing the hyperparameter to be tested. The cumulative distribution value and probability density value are obtained by substituting the input data into the cumulative distribution function formula and the probability density function formula, respectively.

[0101] S434. Convert the vector to be tested into hyperparameters to be tested, replace the hyperparameters in the initial interactive processing model to obtain the model to be tested, construct the testing model based on the mean absolute error formula, and calculate the test value of the model to be tested based on the training set and the validation set. Converting the vector to be tested into hyperparameters to be tested can be achieved by decomposing the elements in the vector to obtain the hyperparameters to be tested. This method is a common method for vector decomposition.

[0102] S435. Test the model to be tested based on the test value and obtain the optimized interactive processing model;

[0103] If the test value is less than the detection threshold, the current model to be tested will be used as the optimized interactive processing model.

[0104] If the test value is greater than or equal to the detection threshold, return to S431 to continue optimizing the model to be tested.

[0105] S44. The optimal interactive processing model is trained and regularized using cosine annealing and gradient clipping to obtain the optimal model. A secondary optimization layer is then constructed to further optimize the interactive processing model. Cosine annealing is a dynamic learning rate adjustment method that periodically decays the learning rate using a cosine function to improve model convergence performance. Gradient clipping is a method used in deep learning to prevent gradient explosion. By optimizing the model through the parameter layer and the secondary optimization layer, the optimization performance of the model can be ensured, and the accuracy of the model in data processing can be significantly improved.

[0106] S45. Input the cached data into the optimal model for predictive analysis to obtain the analysis and prediction results. This step constructs a powerful and flexible time series analysis system. Through multi-level feature extraction, intelligent adaptive processing, efficient hyperparameter optimization, and stable training strategies, it achieves accurate capture and efficient learning of deep patterns in complex time series data. This step makes the method perform well in terms of accuracy, stability, and generalization ability, and is particularly suitable for handling data analysis tasks with complex time series dependencies. It provides reliable technical support for subsequent data interaction and decision-making. The analysis and prediction results contain data with complex time series dependencies. For example, in wafer yield prediction analysis in semiconductor manufacturing, a wafer needs to go through hundreds of processes, such as photolithography, etching, thin film deposition, and ion implantation, to be completed. The final wafer yield is the core lifeline indicator. The analysis and prediction results contain data with time series dependencies of multiple processes, which can be used to analyze and predict the yield.

[0107] S5. The analysis and prediction results are transmitted to the industrial data interaction center to generate interactive decisions. The industrial data interaction center is the core hub of modern intelligent manufacturing and the industrial internet system. It is not a simple database, but an integrated data operation and value transformation platform. For example, in semiconductor manufacturing, it is a processing device that receives data from multiple stages of semiconductor manufacturing inspection and quality control, processes data, and outputs strategies. It can be manually operated or a digitally integrated system. The interactive decision is an instruction given based on the analysis and prediction results. For example, if the yield rate of wafers is unqualified, then an instruction to deem the wafers unqualified and recall them is required.

[0108] Example 2

[0109] Due to the limitations of existing technology models, such as their simple structure, difficulty in balancing long-range and local dependencies, coarse processing of variable-length sequences, and significant information loss, this embodiment proposes an industrial data interaction platform based on Embodiment 1. This platform can comprehensively capture features from feature data, intelligently process inputs of different lengths, and efficiently optimize based on model parameters. Figure 2 As shown, the platform includes: a preliminary processing module 1, a preprocessing module 2, a buffer module 3, an analysis and prediction module 4, an interaction module 5, and a security module 6;

[0110] Among them, the preliminary processing module 1 is used to collect the raw data from the industrial data interaction platform, construct a dual-track adaptive processing model based on signal-to-noise ratio perception, and process the raw data through the dual-track adaptive processing model to obtain the preliminary processed data.

[0111] Preprocessing module 2 is used to preprocess the preliminary processed data and obtain preprocessed data;

[0112] Buffer module 3 is used to obtain cached data step by step from the preprocessed data through mutual information feature selection method and dynamic sequence segmentation method, and to build a cache area based on the cached data;

[0113] The analysis and prediction module 4 is used to perform sliding window segmentation on the cached data to obtain sequence data, construct an initial interaction processing model with hyperparameters based on the sequence data, perform preliminary optimization on the initial interaction processing model to obtain an optimized interaction processing model, perform secondary optimization based on the optimized interaction processing model to obtain the optimal model, and perform predictive analysis on the cached data based on the optimal model to obtain the analysis and prediction results.

[0114] Interaction module 5 is used to transmit the analysis and prediction results to the industrial data interaction center to generate interactive decisions;

[0115] Security module 6 is used to encrypt and sign the data processed by preliminary processing module 1, preprocessing module 2, buffering module 3, analysis and prediction module 4 and interaction module 5, and terminate the data storage through blockchain, and perform two-factor authentication by combining device fingerprint and dynamic password.

[0116] Example 3

[0117] This embodiment also proposes an application of an industrial data interaction method based on Embodiments 1 and 2. The industrial data interaction method of the industrial data interaction platform is applied to the semiconductor manufacturing, medical device, smart water and metallurgical industries.

[0118] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0119] The various embodiments in this invention are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The device embodiments described above are merely exemplary. The modules described as separate components may or may not be physically separate. When implementing the present invention, the functions of each module can be implemented in one or more software and / or hardware. Alternatively, some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0120] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. An industrial data interaction method, characterized in that, The method includes the following steps: S1. Collect raw data from industrial data interaction, construct a dual-track adaptive processing model based on signal-to-noise ratio perception, and process the raw data through the dual-track adaptive processing model to obtain preliminary processed data. S11. Perform anomaly identification on the original data based on dynamic threshold to find the abnormal data, and perform mask generation and moving average interpolation on the abnormal data to obtain masked interpolation data. S12. The signal-to-noise ratio of the original data is evaluated to obtain the noise level. The basis function and threshold algorithm of wavelet transform are dynamically selected according to the noise level, and the noise reduction parameters are obtained. S13. Based on the mask interpolation processing data and noise reduction parameters, construct a parallel dual-track adaptive processing model for adaptive noise reduction and mask interpolation, and obtain preliminary processed data. S2. Preprocess the preliminary data and obtain preprocessed data; S3. Based on the preprocessed data, cached data is obtained step by step through mutual information feature selection and dynamic sequence segmentation, and a cache area is constructed based on the cached data; S4. Perform sliding window segmentation on the cached data to obtain sequence data. Construct an initial interaction processing model with hyperparameters based on the sequence data. Perform preliminary optimization on the initial interaction processing model to obtain an optimized interaction processing model. Perform secondary optimization on the optimized interaction processing model to obtain the optimal model. Perform predictive analysis on the cached data based on the optimal model to obtain the analysis and prediction results. The specific implementation steps of S4 are as follows: S41. Extract cached data from the cache area, and divide the continuous time series in the cached data into fixed-length sequence segments using the sliding window serialization method. Each sequence segment is used as an input sample of the model to obtain sequence data. S42. Construct an initial interaction processing model with hyperparameters based on sequence data; S43. Convert multiple sets of hyperparameters in the initial interaction processing model into input vectors, use the input vectors as individuals in the population optimization to construct an optimization parameter layer for optimizing the input vectors, and obtain the hyperparameters to be tested. Replace the initial hyperparameters in the initial interaction processing model with the hyperparameters to be tested to generate an optimized interaction processing model. S44. The optimal interaction processing model is trained and regularized using cosine annealing and gradient pruning to obtain the optimal model, and a secondary optimization layer is constructed for secondary optimization of the optimal interaction processing model. S45. Input the cached data into the optimal model for predictive analysis to obtain the analysis and prediction results; S5. Transmit the analysis and prediction results to the industrial data interaction center to generate interactive decisions.

2. The industrial data interaction method according to claim 1, characterized in that, In S2, the specific implementation steps are as follows: S21. Based on the preliminary processed data, the operating parameters are obtained through the sub-line clustering algorithm. The mean and standard deviation of the preliminary processed data are calculated. The operating parameters are standardized based on the mean and standard deviation to obtain standardized data. S22. Construct a trained isolation forest model for predicting standardized data anomaly scores and obtain the parameters of the isolation forest model. S23. Define a probability distribution for the parameters of each isolated forest model to provide the search space, pre-set a maximum number of iterations, initialize an empty observation set to store parameter combinations and evaluation results, and fit the observation set with a Gaussian process probability model to obtain the fitted model. S24. Define a collection function. By maximizing the collection function, obtain the next candidate parameter combination that is most likely to find the global optimum. Replace the parameters of the original isolated forest model with the candidate parameter combination to obtain the updated forest model. S25. Obtain the current iteration number, update the forest model based on the maximum iteration number, and obtain the optimized forest model; If the current iteration count is less than the maximum iteration count, return to S23 to continue iterating; If the current iteration count equals the maximum iteration count, then the current updated forest model is used as the optimized forest model; S26. Input the standardized data into the optimized forest model to obtain outlier data. Smooth the outlier data using the Kalman filter to obtain corrected data. Replace the original outlier data with the corrected data. Set the smoothed standardized data as preprocessed data.

3. The industrial data interaction method according to claim 1, characterized in that, In S3, the specific implementation steps are as follows: S31. Preset a window K with width a, and calculate the window cache data for the preprocessed data using the moving average formula and time decay factor; S32. Obtain correlation-optimized data based on window cache data using the mutual information feature selection method; S33. Based on the correlation optimization data, obtain cached data through dynamic sequence segmentation, and construct a cache area based on the cached data.

4. The industrial data interaction method according to claim 1, characterized in that, In S42, the specific implementation steps are as follows: S421. Construct a multi-head attention layer based on the multi-head attention mechanism. Calculate the weight values ​​based on the sequence data using the attention weight formula. Calculate the weighted sequence based on the weight values ​​to highlight key moments and capture long-distance dependencies. S422. A bidirectional gated loop layer is constructed using a bidirectional GRU structure. The weighted sequence is processed and temporal features are extracted from both the forward and backward directions using the bidirectional GRU method. The feature data used to enhance feature expression is obtained by splicing the forward and backward temporal features. S423. Obtain convolutional data by performing temporal convolution operations on the feature data. Then, fuse local details and global dependencies using the residual calculation formula based on the convolutional data to obtain fused data for representing complex temporal fusion features, and construct a temporal fusion layer. S424. Obtain the base window length by dividing the total length of the fused data by the target length to be processed and rounding it up. Based on the base window length, dynamically adjust the last incomplete window and obtain the partitioned data divided by the window. Perform max pooling operation on each partitioned data and obtain compressed fused data of fixed length by concatenating the data. At the same time, construct a dynamic compressed fusion layer. S425. Based on the compression fusion data, a fully connected layer is constructed to obtain the analysis results. An initial interactive processing model with hyperparameters is constructed based on a multivariate attention layer, a bidirectional gated loop layer, a temporal fusion layer, a dynamic compression fusion layer, and a fully connected layer.

5. The industrial data interaction method according to claim 1, characterized in that, In S43, the specific implementation steps are as follows: S431. Define the search space of the input vector, divide the cached data into training set and validation set, define the detection threshold, use the cross-entropy loss function as the loss function for optimizing the parameter model, extract an input data in the training set, calculate the loss value relative to the training set, and combine the input vector and the loss value to form the initial data. S432. Obtain model parameters by maximizing marginal likelihood. and The surrogate model is obtained by fitting the initial data using a Gaussian process fitting model. S433. Obtain the cumulative distribution value and probability density value of the input data in the standard normal distribution. Obtain the predicted mean and standard deviation of the input data according to the surrogate model. Obtain the minimum loss value of the training set based on the loss function. Obtain the test difference by subtracting the minimum loss value from the predicted mean and 0.

01. Obtain the collected value by multiplying the test difference by the cumulative distribution value and adding the product of the standard deviation and the probability density value. Optimize the collected value globally in the search space using the L-BFGS algorithm. Take the data point with the largest collected value as the updated data point. The input vector corresponding to the collected value is the vector to be tested. S434. Convert the vector to be tested into hyperparameters to be tested, replace the hyperparameters in the initial interactive processing model to obtain the model to be tested, construct the test model based on the mean absolute error formula, and calculate the test value of the model to be tested based on the training set and validation set. S435. Test the model to be tested based on the test value and obtain the optimized interactive processing model; If the test value is less than the detection threshold, the current model to be tested will be used as the optimized interactive processing model. If the test value is greater than or equal to the detection threshold, return to S431 to continue optimizing the model to be tested.

6. The industrial data interaction method according to claim 1, characterized in that, The industrial data interaction method described above is applied in the semiconductor manufacturing, medical device, smart water management, and metallurgical industries.

7. An industrial data interaction platform based on the industrial data interaction method according to any one of claims 1 to 6, characterized in that, The platform includes: a preliminary processing module (1), a preprocessing module (2), a buffer module (3), an analysis and prediction module (4), an interaction module (5), and a security module (6). Among them, the preliminary processing module (1) is used to collect the raw data of the industrial data interaction platform, construct a dual-track adaptive processing model based on signal-to-noise ratio perception, and process the raw data through the dual-track adaptive processing model to obtain the preliminary processed data. The preprocessing module (2) is used to preprocess the preliminary processed data and obtain preprocessed data; The buffer module (3) is used to obtain cached data step by step from the preprocessed data through mutual information feature selection method and dynamic sequence segmentation method, and to build a cache area based on the cached data; The analysis and prediction module (4) is used to perform sliding window segmentation on the cached data to obtain sequence data, construct an initial interaction processing model with hyperparameters based on the sequence data, perform preliminary optimization on the initial interaction processing model to obtain an optimized interaction processing model, perform secondary optimization based on the optimized interaction processing model to obtain the optimal model, and perform prediction analysis on the cached data based on the optimal model to obtain the analysis and prediction results. The interaction module (5) is used to transmit the analysis and prediction results to the industrial data interaction center to generate interactive decisions.

8. An industrial data interaction platform according to claim 7, characterized in that, The security module (6) is used to encrypt and sign the data processed by the preliminary processing module (1), preprocessing module (2), buffer module (3), analysis and prediction module (4) and interaction module (5), and end the data storage through blockchain, and perform two-factor authentication by combining device fingerprint and dynamic password.

Images