Electrolytic cell data processing method for large model training

Abstract

An electrolytic cell data processing method for large model training relates to the technical field of industrial data processing in the electrolytic aluminum industry, and comprises the following steps in sequence: step S01, collecting static data and dynamic production data of an electrolytic cell; s02, setting a data sampling interval; s03, constructing a data set according to the time dimension; step S04, checking data missing; s05, data missing processing is carried out; step S06, performing continuous data curve analysis on each column of data, and detecting abnormal values existing in the data; and S07, data outlier analysis, data outlier processing, and storage of the processed data as a data set. By collecting, processing and optimizing static and dynamic data in the production process of the electrolytic cell, the quality and reliability of the data are effectively improved, and a high-quality data basis is provided for training and application of a large model.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial data processing in the electrolytic aluminum industry, and in particular to an electrolytic cell data processing method for large model training. Background Art

[0002] Industrial big model is a new technology developed in the context of intelligent manufacturing. It has become an important tool for industrial data analysis, prediction and decision-making. This technology mainly relies on a massive knowledge base and data analysis capabilities to assist production managers in completing complex production management and decision-making tasks.

[0003] As a representative of high-energy-consuming industries, the electrolytic aluminum industry has become particularly critical for refined management of the production process and precise control of the electrolytic cell as the industry develops and upgrades. In recent years, many electrolytic aluminum production companies have established data collection and information management systems, which not only makes it possible to accumulate data on the electrolytic aluminum production process, but also enables a large amount of data on the working status of the electrolytic cell under different working conditions to be saved.

[0004] These data contain a wealth of industry knowledge and information. However, it is difficult to mine and utilize this knowledge and information by relying on traditional data mining methods and small models. Therefore, it has become an important demand of the current electrolytic aluminum industry to introduce large-scale model technology with more powerful performance and stronger associative learning ability to conduct in-depth analysis of data. The generalization and migration capabilities of large models enable them to better realize industrial process analysis, optimization and decision-making.

[0005] Despite this, industrial large models still face many challenges in practical application. Among them, the most prominent problem is the lack of high-quality industrial data as the basis for model training. The insufficient supply of high-quality data will directly affect the training effect and application performance of large models. Summary of the invention

[0006] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides an electrolytic cell data processing method for large model training to improve the training effect and application performance of the large model.

[0007] In order to achieve the above object, the main technical solutions adopted by the present invention include:

[0008] A method for processing electrolyzer data for large model training comprises the following steps in the following order:

[0009] Step S01: collecting static data and dynamic production data of the electrolytic cell and storing them;

[0010] Step S02: setting the data sampling interval;

[0011] Step S03: constructing a data set according to the time dimension;

[0012] Step S04: data missing check;

[0013] Step S05: data missing processing;

[0014] Step S06: Perform continuous data curve analysis on each column of data to detect abnormal values ​​in the data;

[0015] Step S07: analyzing and processing data outliers, and storing the processed data as a data set;

[0016] Step S08: data normalization;

[0017] Step S09: cluster analysis;

[0018] Step S10: data dimension reduction;

[0019] Step S11: Send the data that completes the above steps and meets the large model training requirements into the large model training.

[0020] Furthermore, in step S01, the static data includes the engineering data of the electrolytic cell design phase, the electrolytic cell construction data of the construction phase, and the test data of the electrolytic cell startup phase; the dynamic production data includes the detection and control process data of the electrolytic cell control system, the data obtained by manual measurement, the data obtained by laboratory analysis, and the precise aluminum output, online temperature measurement, production monitoring, and optimization control system data installed on demand by the factory.

[0021] Furthermore, in step S03, a continuous structured data set is constructed with time as an index.

[0022] Furthermore, in step S04, the structured data set is checked for missing data row by row and column by column. When a row of data is detected to be missing, the current missing position is used as the starting point, the data is checked vertically, the missing length of the data is calculated, and the type of missing data is determined according to the following method;

[0023] When the time interval of missing data is τ1≤ΔT loss ≤τ2 is considered instantaneous missing. When the time interval of missing data is within τ2<ΔT loss If the time interval of missing data is within ΔT, then the missing data is within τ3. loss >τ4 means a large number of missing items; where τ1 is the sampling interval selected in step S2, τ2 is the critical value of missing items that does not cause changes in system trends, and τ3 is the critical value of missing items that can predict changes in system trends.

[0024] Furthermore, in step S05, when processing continuous time series data, interpolation is used to supplement data for instantaneous missing data and small missing data:

[0025]

[0026] Among them, Vx oss It is T loss The missing value at the moment, V1 is the distance T loss The value of the most recent moment T1, V2 is the distance from T loss The value of the next moment T2;

[0027] For a large amount of missing data, each row of data with missing data is deleted, and the original data set is split into data sets with continuous time series.

[0028] Furthermore, in the step S07, the method for analyzing outliers is that if more than three outliers appear continuously and the data associated with the data fluctuates during the production process of the electrolytic cell, no processing is performed; when the number of outliers that appear is less than or equal to three, and the data associated with the data does not fluctuate during the production process of the electrolytic cell, it is determined that the outlier is dirty data, the dirty data is deleted, and the deleted data is processed as missing values ​​according to steps S04 and S05.

[0029] Furthermore, in step S07, the processed data is stored as a data set. If the data set is continuous time series data, it is divided into several time-continuous data sets and stored as data sets.

[0030] Furthermore, in step S09, process-related indicators in the electrolytic production process are selected for cluster analysis, including cell voltage, series current, cell temperature, molecular ratio, electrolyte level, aluminum liquid level, alumina feed rate, aluminum fluoride feed rate, electrolyte composition, cell age, furnace bottom pressure drop data, and the number of clusters and thresholds are set.

[0031] Furthermore, if the clustering effect of step S09 is a silhouette coefficient <0.5, it is necessary to execute step S10 to reduce the dimension of the data; first use PCA to analyze the data, if the analysis result is valid, that is, the cumulative variance contribution rate of the first 10 principal components exceeds 70%, then reduce the dimension of the selected data according to the PCA result; if the PCA analysis result is invalid, that is, the cumulative variance contribution rate of the first 10 principal components is less than 70%, then manually select process-related indicators for dimensionality reduction; if the clustering effect is a silhouette coefficient ≥0.5, execute step S11.

[0032] The beneficial effects of the present invention are as follows: the present invention effectively improves the quality and reliability of data by collecting, processing and optimizing static and dynamic data in the electrolytic cell production process, provides a high-quality data foundation for the training and application of large models, thereby improving the training effect and application performance of large models, and provides a set of standardized processing procedures for electrolytic cell production data. This technical solution can be used for both discrete structured data and continuous time series data, provides solutions to multiple types of missing data, proposes a method for determining whether outliers are dirty data, and enhances data availability through cluster analysis and two dimensionality reduction methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is the data processing flow of the present invention;

[0034] Figure 2 It is the collection process of electrolytic cell related data of the present invention;

[0035] Figure 3 It is the time series data preprocessing process of the present invention;

[0036] Figure 4 It is the discrete structured data preprocessing process of the present invention;

[0037] Figure 5 This is the process of further processing after data preprocessing of the present invention. DETAILED DESCRIPTION

[0038] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation modes in conjunction with the accompanying drawings.

[0039] like Figure 1 As shown, the present invention provides an electrolytic cell data processing method for large model training, and performs the following steps in sequence to achieve data collection, preprocessing, and further data processing for training the large model, and finally form a data set that can be used for large model training.

[0040] Step S01: Collect and store static data and dynamic production data of the electrolytic cell.

[0041] Specifically, Figure 2As shown, in the step S01, the static data comes from the factory's archival data, including the engineering data of each electrolytic cell design stage, the electrolytic cell construction data of the construction stage, and the test data of the electrolytic cell startup stage; the dynamic production data is collected from the control systems and equipment of the electrolytic cells through the data acquisition gateway and the data acquisition server, including the detection and control process data of each electrolytic cell control system, the data obtained by manual measurement, the data obtained by laboratory testing, the precise aluminum output, online temperature measurement, production monitoring, and optimization control system data installed by the factory on demand. Specifically, it includes the cell voltage, series current, alumina feeding amount, aluminum fluoride feeding amount, needle vibration, swing, feeding times, aluminum output time, under-feed times, over-feed times, molecular ratio, aluminum liquid iron content, aluminum liquid silicon content, cell temperature, electrolyte level, aluminum liquid level, furnace bottom pressure drop, machine change time, busbar lifting time, effect times, etc. of each electrolytic cell. The collected data is structured and stored in a time dimension, and the storage medium and form can be flexibly selected, which will not be repeated here.

[0042] Step S02: Setting the data sampling interval.

[0043] According to the function and modeling purpose of the large model, combined with the situation of data collection, select a suitable sampling interval. According to the slow change of electrolytic cell production, the sampling interval can usually be set to 1 day. There are many data sources involved in the electrolytic cell production process (as described in step S01). Different data sources have different data sampling intervals. Small sampling interval data should be aligned with large sampling interval data.

[0044] Specifically, the selection of sampling intervals should fully consider the total amount of data and the frequency of data collection of each indicator, and select the sampling interval of the sampling frequency data. Optionally, for data with a low collection frequency (sampling interval greater than 1 day) (such as molecular ratios tested every 2 to 3 days, furnace bottom pressure drop measured once a week or so), interpolation processing is performed to increase the amount of available data. Data interpolation can use linear interpolation or soft measurement calculation interpolation.

[0045] More specifically, the collected data has the characteristics of multiple time scales, so it is necessary to establish a suitable data sampling interval in step S02. Electrolytic production is a slow process. In order to facilitate stable analysis, the data with a collection interval of less than 1 day, such as cell voltage, series current, etc., need to calculate the average value. For data with a collection interval of more than 1 day, such as electrolyte composition test data, furnace bottom pressure drop, etc., interpolation and soft measurement are used to complete the data.

[0046] Step S03: construct a data set according to the time dimension.

[0047] Specifically, in step S03, a continuous structured data set is constructed with time as an index.

[0048] Step S04: Data missing check.

[0049] like Figure 3 , 4 As shown, the multi-source data of electrolytic cell production collected may cause missing data values ​​due to communication interruption, system abnormality, human error, etc. It is necessary to locate and judge the missing values ​​of the data. The structured data set is checked for missing values ​​row by row and column by column. Data missing can be divided into instantaneous missing, small missing, and large missing. When it is detected that a row of data (at a certain time) is missing, the current missing position is used as the starting point, the data is checked vertically, the length of the missing data is calculated, and the type of data missing is judged according to the following method;

[0050] When the time interval of missing data is τ1≤ΔT loss ≤τ2 is considered instantaneous missing. When the time interval of missing data is within τ2<ΔT loss If the time interval of missing data is within ΔT, then the missing data is within τ3. loss >τ4 means a large number of missing data; where τ1 is the sampling interval selected in step S2, τ2 is the critical value of missing data that does not cause system trend changes, and τ3 is the critical value of missing data that can predict system trend changes. τ2 and τ3 are determined by electrolysis process experts according to different data types.

[0051] Step S05: Data missing processing.

[0052] Specifically, according to the missing value and classification results of the data check in step S04, an appropriate method for handling missing values ​​is selected.

[0053] When processing continuous time series data:

[0054] For instantaneous missing and small amount of missing data, interpolation is used to supplement the data:

[0055]

[0056] Among them, V loss It is T loss The missing value at the moment, V1 is the distance T loss The value of the most recent moment T1, V2 is the distance from T loss The value of the next moment T2;

[0057] For a large amount of missing data, each row of data with missing data is deleted, and the original data set is split into data sets with continuous time series.

[0058] For a small amount of missing data, regression and soft measurement methods are selected according to the electrolytic aluminum production process conditions. If the missing values ​​can be predicted, data filling is used; if the missing values ​​cannot be predicted, the processing method is the same as that for a large amount of missing data.

[0059] When dealing with non-time series data:

[0060] For data sets with sufficient data (deleting all missing data can still make the large model fully converge), delete all data in the time range of missing data.

[0061] For data sets with insufficient data (deleting all missing data cannot make the large model fully converge), instantaneous missing data and a small amount of missing data can refer to the data filling method for the above-mentioned time series data.

[0062] Step S06: Perform continuous data curve analysis on each column of data to detect abnormal values ​​in the data. The detection method may adopt statistical analysis, time series mutation detection, model or machine learning detection.

[0063] Optional, noise detection method 1: statistical method, through statistical analysis of the data distribution, to detect whether there are outliers or data points that do not conform to the normal distribution.

[0064] Optional, noise detection method 2: time series method, analyzes the time series characteristics of the data and detects abnormal fluctuations or mutations in the data.

[0065] Optional,Noise Detection Method 3: Model Comparison Method,compares the data with a known model to detect whether there is data that does not match the model,which may be caused by noise.

[0066] Optional, Noise Detection Method 4: Machine Learning Method, using machine learning algorithms to train models and identify abnormal patterns in data.

[0067] Step S07: Data outlier analysis, data outlier processing, and storing the processed data as a data set.

[0068] Specifically, the detected data curve noise is called an outlier. In industrial production, not all outliers are dirty data. The present invention uses the following method to determine outliers:

[0069] (1) When there are multiple consecutive (usually more than 3) outliers, and during the production process of the electrolytic cell, the data associated with the data also fluctuates, then the outlier value is data with characteristic significance, not dirty data, and is not processed.

[0070] (2) When the number of outliers is small (usually less than or equal to 3), and the data associated with the data does not fluctuate during the electrolytic cell production process, the outlier is determined to be dirty data.

[0071] According to the outlier analysis results, the data is processed. If the outlier is dirty data, it should be deleted, and the deleted data is processed as missing values ​​according to steps S04 and S05. In addition, the noise value is usually short, and only instantaneous missing and small missing data are processed. If the outlier is not dirty data, it is retained.

[0072] Further, the processed data is stored as a data set. Optionally, if the data set is continuous time series data, it is divided into multiple time-continuous data sets and stored as data sets.

[0073] Step S08: Data normalization.

[0074] Normalize the data of the dataset according to each indicator dimension (by column):

[0075]

[0076] Among them, n is the sum of the dimensions of the indicators in the data set, For x i，j The normalized value, and is the variable x i The maximum and minimum values ​​in the dimension.

[0077] Step S09: cluster analysis.

[0078] Specifically, in step S09, process-related indicators in the electrolytic production process are selected for cluster analysis, including cell voltage, series current, cell temperature, molecular ratio, electrolyte level, aluminum liquid level, alumina feed rate, aluminum fluoride feed rate, electrolyte composition, cell age, and furnace bottom pressure drop data. Cluster analysis is performed on these data, and the number of clusters and thresholds are set according to the production status of the electrolytic cell and the evaluation method.

[0079] Specifically, if the clustering effect of step S09 is a silhouette coefficient <0.5, step S10 needs to be executed to reduce the dimension of the data; if the clustering effect is a silhouette coefficient ≥0.5, it is determined that the current data quality is high, and step S11 is executed to use the data for large model training.

[0080] Step S10: Data dimensionality reduction. Figure 5As shown, if the data dimension is too high, resulting in poor clustering effect in step S09, it proves that there are many data with low correlation with electrolytic production in the selected data, which not only affects the clustering effect, but also may affect the model training, and the data is reduced in dimension. First, PCA is used to analyze the data. If the analysis result is valid (the cumulative variance contribution rate of the first 10 principal components exceeds 70%), the selected data is reduced in dimension according to the PCA result; if the PCA analysis result is invalid (the cumulative variance contribution rate of the first 10 principal components is less than 70%), the dimension is reduced by manually selecting process-related indicators.

[0081] Step S11: The data that completes the above steps and meets the requirements of large model training is sent to large model training. The final processed data is stored and backed up to facilitate effect reproduction and model tuning.

[0082] The above steps are used to solve the problem that the data quality is not high when training a large model, which affects the training effect and performance. In particular, it is used to process the data related to the electrolytic cell production for the large model of electrolytic aluminum production.

[0083] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. Alterations, modifications, substitutions and variations of the above embodiments by a person skilled in the art are all within the scope of the present invention.

Claims

1. A method for processing electrolytic cell data for large model training, characterized in that: The steps are as follows: Step S01: collecting static data and dynamic production data of the electrolytic cell and storing them; Step S02: setting the data sampling interval; Step S03: constructing a data set according to the time dimension; Step S04: data missing check; Step S05: data missing processing; Step S06: Perform continuous data curve analysis on each column of data to detect abnormal values ​​in the data; Step S07: analyzing and processing data outliers, and storing the processed data as a data set; Step S08: data normalization; Step S09: cluster analysis; Step S10: data dimension reduction; Step S11: Send the data that completes the above steps and meets the large model training requirements into the large model training.

2. The electrolytic cell data processing method for large model training according to claim 1, characterized in that: In step S01, the static data includes the engineering data of the electrolytic cell design phase, the electrolytic cell construction data of the construction phase, and the test data of the electrolytic cell startup phase; the dynamic production data includes the detection and control process data of the electrolytic cell control system, the data obtained by manual measurement, the data obtained by laboratory testing, and the precise aluminum output, online temperature measurement, production monitoring, and optimization control system data installed on demand by the factory.

3. The electrolytic cell data processing method for large model training according to claim 1, characterized in that: In the step S03, a continuous structured data set is constructed with time as an index.

4. The electrolytic cell data processing method for large model training according to claim 1, characterized in that: In the step S04, the structured data set is checked for missing data row by row and column by column. When a row of data is detected to be missing, the current missing position is used as the starting point, the data is checked vertically, the missing length of the data is calculated, and the type of missing data is determined according to the following method; When the time interval of missing data is τ1≤ΔT loss ≤τ2 is considered instantaneous missing. When the time interval of missing data is within τ2<ΔT loss If the time interval of missing data is within ΔT, then the missing data is within τ3. loss >τ4 means a large number of missing items; where τ1 is the sampling interval selected in step S2, τ2 is the critical value of missing items that does not cause changes in system trends, and τ3 is the critical value of missing items that can predict changes in system trends.

5. The electrolytic cell data processing method for large model training according to claim 1, characterized in that: In step S05, when processing continuous time series data, interpolation is used to supplement data for instantaneous missing data and small missing data: Among them, V loss It is T loss The missing value at the moment, V1 is the distance T loss The value of the most recent moment T1, V2 is the distance from T loss The value of the next moment T2; For a large amount of missing data, each row of data with missing data is deleted, and the original data set is split into data sets with continuous time series.

6. The electrolytic cell data processing method for large model training according to claim 1, characterized in that: In the step S07, the method for analyzing outliers is that if more than three outliers appear continuously and the data associated with the data fluctuate during the production process of the electrolytic cell, no processing is performed; when the number of outliers that appear is less than or equal to three and the data associated with the data does not fluctuate during the production process of the electrolytic cell, it is determined that the outlier is dirty data, the dirty data is deleted, and the deleted data is processed as missing values ​​according to steps S04 and S05.

7. The electrolytic cell data processing method for large model training according to claim 1 or 6, characterized in that: In step S07, the processed data is stored as a data set. If the data set is continuous time series data, it is divided into several time-continuous data sets and stored as data sets.

8. The electrolytic cell data processing method for large model training according to claim 1, characterized in that: In step S09, process-related indicators in the electrolytic production process are selected for cluster analysis, including cell voltage, series current, cell temperature, molecular ratio, electrolyte level, aluminum liquid level, alumina feed rate, aluminum fluoride feed rate, electrolyte composition, cell age, furnace bottom pressure drop data, and the number of clusters and threshold values ​​are set.

9. The electrolytic cell data processing method for large model training according to claim 1, characterized in that: If the clustering effect of step S09 is that the silhouette coefficient is <0.5, it is necessary to execute step S10 to reduce the dimension of the data; first use PCA to analyze the data. If the analysis result is valid, that is, the cumulative variance contribution rate of the first 10 principal components exceeds 70%, then reduce the dimension of the selected data according to the PCA result; if the PCA analysis result is invalid, that is, the cumulative variance contribution rate of the first 10 principal components is less than 70%, then manually select process-related indicators to reduce the dimension; if the clustering effect is that the silhouette coefficient is ≥0.5, execute step S11.

Images