A method and system for detecting risks in electrolytic aluminum production

Through a combination of deep learning and the least square vector machine algorithm, a neural network and Hammerstein model are established to detect the risks of electrolytic aluminum production in real time, solving the risk detection problems in electrolytic aluminum production, and improving production safety and prediction accuracy.

CN118747605BActive Publication Date: 2024-12-06SHANDONG XINFA HUAXIN ALUMINUM IND CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411058305.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-02
Publication Date
2024-12-06
Estimated Expiration
2044-08-02

AI Technical Summary

Technical Problem

There are high risks in the production process of electrolytic aluminum, and the existing technology is difficult to efficiently and accurately detect and predict production risks, resulting in accidents and economic losses.

Method used

Deep learning algorithm is used to establish a neural network model and a least square vector machine algorithm to establish a Hammerstein model. By collecting production data in real time, identifying risk types, training the model for prediction, and displaying potential risk factors to the staff.

Benefits of technology

It realizes efficient and accurate detection and prediction of electrolytic aluminum production risks, reduces the probability of accidents, and improves production safety and prediction accuracy of production data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118747605B_ABST
    Figure CN118747605B_ABST
Patent Text Reader

Abstract

The present application relates to a risk detection method and system for electrolytic aluminum production, and to the technical field of risk detection, wherein the detection method includes the steps of first data acquisition, first modeling, second modeling, data processing and data display; the detection system includes a data acquisition module, a model building module, a data processing module, a judgment module and a data display module. The present application obtains the risk type of production data through a neural network model, obtains the historical data of the risk type, trains a Hammerstein model through the historical data, predicts the production data through the trained Hammerstein model, detects and monitors the future trend of the production data in real time, realizes the prediction and prevention of electrolytic aluminum production risks, and improves safety.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of risk detection, and in particular to a risk detection method and detection system for electrolytic aluminum production. Background Art

[0002] In the electrolytic aluminum production process, ensuring production safety is not only related to the life safety and health of employees, but also the cornerstone of maintaining stable operation of enterprises, ensuring product quality and improving market competitiveness. However, the electrolytic aluminum industry faces many challenges due to its unique process characteristics. The electrolytic aluminum process is complex and delicate, and the production environment is extremely harsh, including high-temperature operating areas, high-voltage power systems and highly corrosive media, which greatly increase the uncertainty and risk in the production process.

[0003] The occurrence of various production risks is accompanied by changes in production data, so slight fluctuations in production data may trigger a chain reaction. If they cannot be identified and handled in a timely and effective manner, it will not only lead to production accidents, causing direct economic losses and reputation damage to the company, but also pose a serious threat to the life safety of employees and undermine the company's social image and sustainable development capabilities.

[0004] Therefore, there is an urgent need to develop a method that can efficiently and accurately detect various risks in the electrolytic aluminum production process. Summary of the invention

[0005] In order to efficiently and accurately detect the production risks of electrolytic aluminum, the present application provides an electrolytic aluminum production risk detection method and detection system.

[0006] In the first aspect, the present application provides a method for detecting risks in electrolytic aluminum production, which adopts the following technical solution:

[0007] A method for detecting risks in electrolytic aluminum production comprises the following steps:

[0008] First data collection: real-time collection of production data;

[0009] First modeling: Building a neural network model based on deep learning algorithms;

[0010] Second modeling: Building the Hammerstein model based on the least squares vector machine algorithm;

[0011] Data processing: including classification, second data collection, first training, prediction and first judgment steps;

[0012] Classification: Input the production data into the neural network model to obtain the risk type probability distribution of the production data, and record the risk type corresponding to the maximum probability as the first type;

[0013] Second data collection: collecting the first type of historical data, recorded as the first data;

[0014] First training: using the first data to train the Hammerstein model to obtain a trained Hammerstein model;

[0015] Prediction: Use the production data as the input data of the trained Hammerstein model, and output the predicted data of the production data, which is recorded as the second data;

[0016] First judgment: judging whether the second data meets expectations, if so, executing the first data collection step; if not, executing the data display step;

[0017] Data presentation: present the second data to the staff.

[0018] By adopting the above technical solution, the present application identifies the risk type of production data through a neural network model, provides a direction for subsequent second data collection and model training, uses historical data of specific risk types to train the Hammerstein model, enhances the Hammerstein model's ability to predict such risks, and inputs the production data into the trained Hammerstein model to generate prediction data. The present application obtains future prediction data of the production data by performing a series of processing on the production data, and determines whether the prediction data meets expectations. If not, the second data is displayed to the staff, thereby preventing production risks.

[0019] Optionally, after executing the classification step and before executing the second data collection step, the method further includes:

[0020] First acquisition: obtain n risk types whose risk type probability distribution of production data exceeds a preset probability threshold, and record the nth risk type as the nth risk type;

[0021] Second judgment: judging whether there is a correlation between the first n risk types, if so, executing the first calculation step; if not, executing the second data collection step;

[0022] First calculation: Calculate the risk. The calculation model is as follows:

[0023] ;

[0024] in, is the probability of the i-th risk type; is the weight of the impact of production data on the i-th risk type.

[0025] By adopting the above technical solution, the present application can screen out the most likely risks by obtaining the n risk types whose probability distribution of risk types of production data exceeds the preset probability threshold, which helps to reduce the complexity and calculation amount of subsequent processing, while ensuring that the focus is on those risk types with the greatest potential harm. Next, the correlation between risk types is judged. If there is a correlation between risk types, they will affect each other, resulting in the expansion or complication of production risks. By judging the correlation between risk types, the risks that may occur after the current production risk occurs can be predicted. In the technical solution of the present application, calculating the risk degree is a quantitative process. By comprehensively considering the weight of the impact of the production data on the i-th risk type, a comprehensive risk degree value can be obtained. The risk degree value can more intuitively reflect the severity of the risk and provide a basis for subsequent decision-making. At the same time, the calculation of risk degree also provides the possibility of comparison between different risk types, which helps to determine the risk type to be prioritized (the greater the risk degree, the higher the priority).

[0026] Optionally, after executing the second determining step and before executing the first calculating step, the method further includes:

[0027] Calculating the correlation degree: including third data collection, second calculation, second acquisition, third judgment, fourth judgment, first expression, second expression and third expression;

[0028] Third data collection: collect historical data of all risk types and record them as third data;

[0029] Second calculation: based on the third data, respectively calculate the correlation coefficients between the current risk type and the remaining n-1 risk types by means of statistical analysis;

[0030] Second acquisition: acquiring the first risk type with which the correlation coefficient is calculated for the current risk type;

[0031] Third judgment: judging whether the absolute value of the correlation coefficient is greater than a preset correlation coefficient threshold, if so, executing the fourth judgment step; if not, executing the third representation step;

[0032] Fourth judgment: judging whether the correlation coefficient is a positive number, if so, executing the first representation step; if not, executing the second representation step;

[0033] The first representation: indicates that the current risk type is positively correlated with the first risk type;

[0034] The second representation: indicates that the current risk type is negatively correlated with the first risk type;

[0035] The third indication: indicates that the current risk type is not related to the first risk type.

[0036] By adopting the above technical solution, calculating the correlation coefficient between the current risk type and the remaining n-1 risk types by means of statistical analysis is a key step in quantifying the correlation between risk types. The correlation coefficient can intuitively reflect the strength of the relationship between the two risk types, providing a scientific basis for subsequent judgments. Judging whether the absolute value of the correlation coefficient is greater than the preset correlation coefficient threshold is an important basis for determining whether there is a significant correlation between the two risk types. If the absolute value of the correlation coefficient is greater than the threshold, it means that there is a strong correlation between the two risk types, and the direction of the correlation needs to be further analyzed; if it is less than the threshold, it means that the correlation is weak and can be regarded as no correlation. After confirming that there is a significant correlation between the two risk types, it is necessary to further determine whether the correlation coefficient is a positive number. This step is to determine the direction of the correlation, that is, positive correlation or negative correlation. Positive correlation means that the change trends of the two risk types are the same, while negative correlation means that the change trends are opposite.

[0037] Optionally, after executing the step of calculating the degree of association and before executing the first calculation step, the method further includes:

[0038] Third acquisition: obtaining p first risk types that are positively correlated with the current risk type;

[0039] Fourth acquisition: acquiring q first risk types that are negatively correlated with the current risk type.

[0040] By adopting the above technical solution, through the third acquisition step, p first risk types that are positively correlated with the current risk type can be clearly identified. The positive correlation means that these first risk types have similar changing trends with the current risk type in terms of probability of occurrence or degree of impact. The fourth acquisition step focuses on q first risk types that are negatively correlated with the current risk type. The negative correlation indicates that these first risk types have opposite changing trends with the current risk type in terms of probability of occurrence or degree of impact.

[0041] Optionally, in the first calculation step, the risk may also adopt the following calculation model:

[0042] ;

[0043] in, is the probability of the current risk type; The weight of the impact of production data on the current risk type; is the probability of the i-th risk type; is the probability of the j-th risk type; is the weight of the impact of production data on the i-th first risk type; is the weight of the impact of production data on the jth first risk type; is the correlation coefficient of the i-th first risk type that is positively correlated with the current risk type; is the correlation coefficient of the jth first risk type that is negatively correlated with the current risk type.

[0044] By adopting the above technical solution, the present application obtains p first risk types that are positively correlated with the current risk type and q first risk types that are negatively correlated with the current risk type, and combines their respective correlation coefficients and the weights of the impact of production data on the risk type to construct a more comprehensive and sophisticated risk calculation model. This calculation model makes the risk assessment process more refined and scientific by introducing parameters such as weights and correlation coefficients. It not only takes into account the probability of occurrence and severity of consequences of a single risk type, but also takes into account the mutual influence between different risk types, so that it can more comprehensively evaluate the overall risk level in the production process.

[0045] Optionally, after executing the third data acquisition step and before executing the second calculation step, the method further includes:

[0046] Data storage: storing the third data in a database;

[0047] First scan: Scan the database, count the number of occurrences of each historical data in each risk type, obtain the historical data whose number of occurrences is greater than the preset threshold, and record it as the first frequent item;

[0048] Establishing an FP tree: constructing an FP tree according to the frequency of occurrence of the first frequent item, wherein a node of the FP tree represents a first frequent item, and a counter of a node in the FP tree represents the number of times the first frequent item appears in the database;

[0049] First construction: Starting from the root node of the FP tree, a conditional pattern base is constructed for each first frequent item;

[0050] Establish a conditional FP tree: construct a conditional FP tree based on the conditional pattern of the first frequent item, the conditional FP tree includes a plurality of child nodes, each child node represents a frequent item associated with the first frequent item, the frequent item associated with the first frequent item is recorded as the second frequent item, and the counter of each child node represents the number of times the second frequent item appears in the database in the current conditional FP tree;

[0051] Fifth judgment: judge whether the conditional FP tree contains only one second frequent item. If so, execute the data merging step; if not, take the conditional FP tree as a new FP tree, take the second frequent items as new first frequent items, and execute the first construction step;

[0052] Data merging: Merge the second frequent items belonging to the same conditional pattern base to obtain the first frequent item set.

[0053] By adopting the above technical solution, the present application scans the database, counts the number of occurrences of each historical data in each risk type, and selects the historical data with a number of occurrences greater than a preset threshold as the first frequent item. This step is intended to identify those historical data that appear frequently in the database. FP tree is constructed according to the frequency of occurrence of the first frequent item. FP tree is an efficient data structure for storing frequent item sets. It can effectively compress data and retain the association information between frequent items. By constructing FP tree, subsequent frequent item set mining can be conveniently performed. Starting from the root node of FP tree, a conditional pattern base is constructed for each first frequent item. The conditional pattern base is the basis for constructing the conditional FP tree, which contains the occurrence of other item sets under the condition of the occurrence of a given item set. This step provides necessary information for the subsequent construction of conditional FP tree. Conditional FP tree is constructed according to the conditional pattern base of the first frequent item. Conditional FP tree is an extension of FP tree under specific conditions, which further mines the association relationship between frequent items under given conditions. By constructing conditional FP tree, detailed patterns in historical data can be gradually deepened. Determine whether the conditional FP tree contains only one second frequent item. If the conditional FP tree contains only one second frequent item, it means that a complete frequent item set has been found; if the conditional FP tree contains multiple second frequent items, the conditional FP tree needs to be used as a new FP tree, and the second frequent items are used as new first frequent items respectively, and the first construction step is continued for iterative processing. This step is the key in the frequent item set mining process, which ensures the integrity and accuracy of the mining process. The second frequent items belonging to the same conditional pattern base are merged to obtain the first frequent item set for processing in subsequent steps.

[0054] Optionally, after executing the data merging step and before executing the second calculating step, the method further includes:

[0055] The third calculation: calculate the probability of the current first frequent item set and the remaining first frequent item sets occurring at the same time, recorded as the first probability;

[0056] First deletion: deleting the first probability lower than the preset first probability threshold and the first remaining frequent item set corresponding to the first probability, to obtain the current first frequent item set and the second frequent item set;

[0057] Fourth calculation: Calculate the probability that the second frequent item set will also occur when the first frequent item set occurs, recorded as the second probability;

[0058] Second deletion: deleting the second probability lower than the preset second probability threshold and the second frequent item set corresponding to the second probability, to obtain the current first frequent item set and the third frequent item set;

[0059] Data update: Combine the first frequent item set and the third frequent item set as the new third data.

[0060] By adopting the above technical solution, by calculating the probability (first probability) of the current first frequent item set and the remaining first frequent item set occurring at the same time, the co-occurrence between different frequent item sets can be evaluated, which is helpful to identify those item sets that often appear together, representing that there is a certain correlation between the two. Deleting the first probability below the preset first probability threshold and the corresponding first remaining frequent item set can delete those item sets with low co-occurrence probability and unimportant, so as to reduce the complexity and amount of calculation of subsequent analysis. After determining the current first frequent item set and the second frequent item set, the probability (second probability) of the second frequent item set also occurring when the current first frequent item set occurs is calculated, which further refines the conditional probability relationship between frequent item sets and helps to more accurately evaluate the correlation between frequent item sets. Deleting the second probability below the preset second probability threshold and the corresponding second frequent item set is similar to the first deletion step. This step is also to remove those item sets that are not strongly correlated and unimportant, so as to further improve the purity of the data and the accuracy of the analysis results. The first frequent item set and the third frequent item set (i.e., the item set remaining after two deletions) that have been screened and calculated are combined as new third data, which reduces the amount of calculation in the second calculation step.

[0061] Optionally, the conditional pattern base includes:

[0062] A prefix path, wherein the prefix path includes a set of all items on a node path from a first frequent item to a certain first frequent item, and the set of all frequent items does not include the certain first frequent item;

[0063] A suffix set, wherein the suffix set includes a first frequent item after a first frequent item in the FP tree.

[0064] By adopting the above technical solution, the prefix path is used to construct a conditional FP tree for a specific frequent item (a first frequent item). When constructing the conditional FP tree, it is necessary to consider the frequent items that appear before all nodes containing a first frequent item in the original FP tree, and these frequent items constitute the prefix part of the conditional FP tree. The suffix set is used to construct the part after the node where the specific frequent item (a first frequent item) is located in the conditional FP tree. When constructing the conditional FP tree, it is necessary to reorganize the frequent items (i.e., the suffix set) that appear after all the nodes of the specific frequent item (a first frequent item) in the original FP tree in a certain way (such as counting sorting) and connect them to the corresponding child nodes of the node of the specific frequent item (a first frequent item) in the conditional FP tree. The conditional pattern base is actually a "slice" of the specific frequent item in the original FP tree, which contains all the context information (i.e., the prefix path) and subsequent information (i.e., the suffix set) related to the specific frequent item in the original FP tree. By constructing the conditional pattern base, a more in-depth analysis can be performed on the specific frequent item.

[0065] Optionally, after executing the second data collection step and before executing the first training step, the method further includes:

[0066] Sixth judgment: judging whether the amount of historical data exceeds the preset historical data amount threshold, if yes, executing the first training step; if no, executing the data supplementation step;

[0067] Supplementary data: including data generation, first assessment and seventh judgment steps;

[0068] Generate data: Based on historical data, use the Borderline SMOTE algorithm to generate new historical data;

[0069] First evaluation: Use the clean factor to evaluate the number of historical data around each historical data, recorded as the fourth data;

[0070] Seventh judgment: judging whether the fourth data exceeds a preset quantity threshold, if so, deleting the current historical data; if not, retaining the current historical data.

[0071] By adopting the above technical solution, the present application ensures that there is enough historical data to train the Hammerstein model by judging whether the amount of historical data collected is sufficient (i.e., whether it exceeds the preset historical data amount threshold), thereby improving the generalization ability and accuracy of the Hammerstein model. If the historical data is lower than the preset historical data amount threshold, new historical data is generated to supplement the historical data to increase the diversity and quantity of historical data, thereby improving the effect of model training.

[0072] In the second aspect, the present application provides an electrolytic aluminum production risk detection system, which adopts the following technical solution:

[0073] An electrolytic aluminum production risk detection system, comprising:

[0074] The data collection module is used to collect production data in real time; it is also used to collect historical data based on the probability distribution of risk types of production data;

[0075] Establish model module, used to establish neural network model and Hammerstein model;

[0076] The data processing module includes a first data processing unit, a second data processing unit and a third data processing unit, wherein the first data processing unit is communicatively connected to the data acquisition module, the second data processing unit is communicatively connected to the data acquisition module, and the third data processing unit is communicatively connected to the second data processing unit and the judgment module;

[0077] A first data processing unit is used to input the production data into the neural network model to obtain a risk type probability distribution of the production data;

[0078] A second data processing unit, used for training the Hammerstein model using historical data;

[0079] A third data processing unit is used to predict the production data using the trained Hammerstein model to obtain second data;

[0080] A judging module, used to judge whether the second data meets expectations;

[0081] The data display module is connected to the judgment module for displaying the second data to the staff.

[0082] By adopting the above technical solution, the data acquisition module is responsible for collecting current production data from the production environment in real time and collecting the first type of historical data, which is the basis for subsequent analysis and prediction. This system combines the neural network model and the Hammerstein model, and uses historical data to train the Hammerstein model to improve the accuracy of prediction of future production data. This multi-model fusion method can more comprehensively capture the complex characteristics of the production process. When the prediction result does not meet expectations, the data display module will intuitively display the second data to the staff, which helps the staff to quickly understand the risk situation and take timely countermeasures, thereby reducing the probability of accidents.

[0083] In summary, the present application includes at least one of the following beneficial technical effects:

[0084] 1. This application identifies the risk type of production data through a neural network model, provides direction for subsequent second data collection and model training, uses historical data of specific risk types to train the Hammerstein model, enhances the model's ability to predict such risks, inputs production data into the trained Hammerstein model, generates prediction data, and determines whether further data analysis is needed based on whether the prediction data meets expectations, which helps to promptly discover abnormal situations in the production process. When the prediction data does not meet expectations, the prediction data is displayed to the staff. This application obtains future prediction data for production data by performing a series of processing on production data, and determines whether the prediction data meets expectations. If it does not meet expectations, the dangerous factors are displayed to the staff, thereby preventing production risks.

[0085] 2. The p first risk types that are positively correlated with the current risk type and the q first risk types that are negatively correlated with the current risk type obtained in this application are combined with their respective correlation coefficients and the weights of the impact of production data on the current risk type to construct a more comprehensive and sophisticated risk calculation model. This calculation model makes the risk assessment process more refined and scientific by introducing parameters such as weights and correlation coefficients. It not only takes into account the probability of occurrence and severity of consequences of a single risk type, but also takes into account the mutual influence between different risk types, so that the overall risk level in the production process can be more comprehensively evaluated.

[0086] 3. The data collection module is responsible for collecting current production data from the production environment in real time and collecting the first type of historical data, which is the basis for subsequent analysis and prediction. This system combines the neural network model and the Hammerstein model, and uses historical data to train the Hammerstein model to improve the accuracy of predictions for future production data. This multi-model fusion method can more comprehensively capture the complex characteristics of the production process. When the prediction results do not meet expectations, the data display module displays the second data to the staff, which helps the staff quickly understand the risk situation and take timely countermeasures, thereby reducing the probability of accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Figure 1 is a flow chart of Example 1 of the present application;

[0088] Figure 2 This is a flowchart of step S32 of embodiment 1 of the present application: first data processing;

[0089] Figure 3 This is a flowchart of step S322 of embodiment 1 of the present application: calculating the degree of association;

[0090] Figure 4 It is a structural diagram of Example 2 of the present application. DETAILED DESCRIPTION

[0091] The following combination Figures 1 to 4 This application is described in further detail.

[0092] Example 1: This example discloses a method for detecting risks in electrolytic aluminum production. Figure 1 The detection method includes: step S1: first data collection, step S2: modeling, step S3: data processing, step S4: data display, firstly, real-time production data is collected, then a neural network model and a Hammerstein model are established, then the neural network model is used to classify the production data, and the risk type corresponding to the production data is obtained, then the historical data of the risk type is extracted, and then the Hammerstein model is trained with the historical data to improve the prediction ability of the Hammerstein model, and then the trained Hammerstein model is used to predict the production data, and then it is determined whether the predicted data meets expectations. If it does not meet expectations (subsequent production data may cause accidents), the second data is displayed to the staff to prompt the production risks that will occur in the future. The specific steps of this embodiment are as follows:

[0093] Step S1: First data collection, real-time collection of production data during the electrolytic aluminum production process, the production data includes:

[0094] Current and voltage data: The current and voltage data are core indicators for evaluating the operating status and efficiency of the electrolytic cell. When the current and voltage data are abnormal, such as excessive voltage or unstable current, it will cause damage to the electrolytic cell and related electrical equipment, and then cause electric shock accidents, especially at the connection point between the high-voltage power supply and the electrolysis workshop network, the risk of electric shock is particularly serious.

[0095] Temperature data: For high-temperature equipment such as electrolytic cells and medium-frequency induction furnaces in anode assembly workshops, if operators fail to take proper protection or operate improperly during work, they may come into contact with high-temperature components or materials, resulting in scalding accidents. In areas such as electrolysis workshops and casting workshops, the temperature of molten aluminum alloy solutions and residues is extremely high and easy to splash. If protection is inadequate or operation is wrong, it may cause scalding. For example: Install video surveillance and temperature alarm systems in electrolysis workshops, casting workshops and other areas. When abnormally high temperatures or splashing of molten materials are detected, the alarm is immediately triggered and the operator is notified.

[0096] Electrolyte composition and liquid level: Use spectral analysis, liquid level sensors and other technical means to monitor the key chemical composition content and liquid level changes in the electrolyte in real time, providing a basis for timely adjustment of the electrolyte formula and maintaining the appropriate liquid level. If the electrolyte composition and liquid level data are abnormal, it indicates the risk of leakage in the electrolytic cell. Once the leaked high-temperature aluminum liquid contacts air or water, it may cause a fire or explosion.

[0097] Gas emission data: A large amount of waste gas, such as carbon dioxide and fluoride, is generated during the production of electrolytic aluminum. The emission, composition and concentration of these waste gases are collected and analyzed in real time through gas detection devices. If harmful gases leak, they can cause poisoning or suffocation to workers.

[0098] Equipment status data: monitor the operating status, fault warning and maintenance requirements of key equipment such as electrolytic cells, agitators, cooling systems, etc. There are a large number of lifting machinery in areas such as electrolysis workshops and foundry workshops. If the equipment fails, the binding and lifting are not done properly, illegal operations or operating errors occur, mechanical injuries or lifting injuries will occur.

[0099] The step S2: modeling, includes step S21: first modeling and step S22: second modeling.

[0100] Step S21: First modeling, establishing a neural network model based on a deep learning algorithm.

[0101] Step S22: Second modeling, establishing a Hammerstein model based on the least squares vector machine algorithm.

[0102] The Hammerstein model first processes the input data through the nonlinear part to generate an internal state representation. The calculation model is as follows:

[0103] ;

[0104] ;

[0105] in, Represents the input data (i.e. production data) of the Hammerstein model; is a nonlinear function, indicating and The mapping relationship; For forecast data; is the model parameter.

[0106] The step S3: data processing includes step S31: classification and step S32: first data processing.

[0107] Step S31: Classification, input the production data into the neural network model, obtain the risk type probability distribution of the production data, and record the risk type corresponding to the maximum probability as the first type.

[0108] like Figure 2 As shown, the step S32: first data processing includes step S320: first acquisition, step S321: second judgment, step S322: calculation of association, step S323: acquisition of data, step S324: first calculation, step S325: second data collection, step S326: addition of data, step S327: first training, step S328: prediction and step S329: first judgment.

[0109] Step S320: First, obtain n risk types whose risk type probability distribution of production data exceeds a preset probability threshold. These risk types are considered to be the objects that need to be focused on in the current production environment. To facilitate subsequent processing, these risk types are sorted from high to low according to their probability, and the nth risk type is recorded as the nth risk type.

[0110] Step S321: Second judgment, judging whether the n risk types mentioned in step S320: First acquisition are related, that is, judging whether they belong to different links of the same risk chain, or are aggravated at the same time due to some common factors. If the judgment result shows that there is a correlation between the first n risk types, then execute step S322: calculate the degree of correlation. On the contrary, if the judgment result shows that there is no correlation between the risk types, the system will execute step S325: Second data collection, collecting the historical data of the risk type corresponding to the maximum probability in step S320: First acquisition, for training the Hammerstein model in step S327: First training.

[0111] like Figure 3 As shown, the step S322: calculating the association degree includes step S3220: third data collection, step S3221: mining frequent items, step S3222: updating the third data, step S3223: second calculation, step S3224: second acquisition, step S3225: third judgment, step S3226: fourth judgment, step S3227: first representation, step S3228: second representation and step S3229: third representation.

[0112] Step S3220: third data collection, collecting historical data of all risk types, recorded as third data.

[0113] The step S3221: mining frequent items includes step S32211: data storage, step S32212: first scanning, step S32213: establishing FP tree, step S32214: first construction, step S32215: establishing conditional FP tree, step S32216: fifth judgment and step S32217: data merging.

[0114] Step S32211: Data storage, storing the third data in the database.

[0115] Step S32212: First scan, scan the database, count the number of occurrences of each historical data in each risk type, obtain the historical data whose number of occurrences is greater than a preset threshold, and record it as the first frequent item.

[0116] Step S32213: Establish an FP tree. According to the frequency of occurrence of the first frequent item, construct an FP tree (Frequent Pattern Tree). A node of the FP tree represents a first frequent item. The connection between the nodes reflects the co-occurrence relationship of these first frequent items in the database. The counter of the node in the FP tree represents the number of times the first frequent item appears in the database.

[0117] Step S32214: First construction, starting from the root node of the FP tree, constructing a conditional pattern base for each first frequent item, the conditional pattern base includes:

[0118] A prefix path, wherein the prefix path includes a set of all items on a node path from a first frequent item to a certain first frequent item, wherein the set of all items does not include the first frequent item.

[0119] A suffix set, wherein the suffix set includes a first frequent item after a node where a first frequent item is located in the FP tree.

[0120] The prefix path is truncated during the construction process, and only the frequent items on the path from the root node to a specified first frequent item are retained, which helps to convert the original data set into a series of more compact and easier to handle forms, preparing for the subsequent conditional FP tree construction.

[0121] Step S32215: Establish a conditional FP tree. The conditional FP tree is constructed according to the conditional pattern base of the first frequent item. The conditional FP tree includes multiple child nodes, each child node represents a frequent item associated with the first frequent item, and the frequent item associated with the first frequent item is recorded as the second frequent item. The counter of each child node represents the number of times the second frequent item appears in the database in the current conditional pattern base.

[0122] Step S32216: The fifth judgment is to judge whether the conditional FP tree contains only one second frequent item. If so, it means that a simple and direct association rule has been found, and then step S32217: data merging is executed; if not, the conditional FP tree is used as a new FP tree, and the second frequent items are used as new first frequent items, and step S32214: first construction and step S32215: establishing the conditional FP tree are repeated. This process will continue until there is only one second frequent item left in the conditional FP tree.

[0123] Step S32217: Data merging: merging the second frequent items belonging to the same conditional pattern base to obtain a first frequent item set.

[0124] The step S3222: updating the third data includes step S32221: third calculation, step S32222: first deletion, step S32223: fourth calculation, step S32224: second deletion and step S32225: data update.

[0125] Step S32221: the third calculation, this step is to evaluate the association strength between the current first frequent itemset and each frequent itemset in the remaining first frequent itemsets in the database.

[0126] Calculate the probability of each pair of combinations, that is, the current first frequent itemset and a frequent itemset in the remaining first frequent itemsets, occurring at the same time. This probability value is marked as the first probability. If the first frequent itemset is frequent itemset A, and the remaining first frequent itemsets include frequent itemsets B, C, D, etc., then the first probability of frequent itemsets A and B, the first probability of frequent itemsets A and C, and the first probability of frequent itemsets A and D will be calculated. Then, take frequent itemset B as the first frequent itemset, and the remaining first frequent itemsets include frequent itemsets A, C, and D. Then calculate the first probability of frequent itemsets B and A, the first probability of frequent itemsets B and C, the first probability of frequent itemsets B and D, and so on.

[0127] This process is based on statistical principles and quantifies the correlation between frequent itemsets by comparing the product of their co-occurrence times in the database and their independent occurrence times. Taking frequent itemsets A and B as examples, the product of their co-occurrence times in the database and their independent occurrence times is compared.

[0128] ;

[0129] in, represents the number of times frequent item set A and frequent item set B appear simultaneously in the third data, represents the number of times the frequent item set A appears alone in the third data, represents the number of times the frequent item set B appears alone in the third data, and N represents the total number of third data in the database.

[0130] If Association Strength>1, it indicates that the correlation between frequent item sets A and B is stronger than when they appear independently, that is, they tend to appear together. If Association Strength=1, it indicates that the correlation between frequent item sets A and B is the same as expected when they appear independently, that is, no special correlation is shown. If Association Strength<1, it indicates that the correlation between frequent item sets A and B is weaker than when they appear independently, that is, they tend not to appear together.

[0131] Step S32222: first deletion, deleting the first probability lower than the preset first probability threshold and the first remaining frequent itemsets corresponding to the first probability, to obtain the current first frequent itemsets and the second frequent itemsets.

[0132] In this embodiment, the first probability threshold is set to 2. When the Association Strength result calculated between the first frequent itemset and one of the remaining frequent itemsets is less than 2, the remaining frequent itemset will be deleted. For example, if the first frequent itemset is frequent itemset A, and the Association Strength result calculated between frequent itemset A and frequent itemset B in the remaining frequent itemsets is less than 2, frequent itemset B will be deleted from the remaining frequent itemsets to obtain the first frequent itemset (frequent itemset A) and the second frequent itemset (frequent itemset C, frequent itemset D).

[0133] Step S32223: The fourth calculation is to calculate the probability that the second frequent itemset also occurs when the first frequent itemset occurs, which is recorded as the second probability. Taking the first frequent itemset being frequent itemset A and the second frequent itemset being frequent itemset C as an example, the calculation model of the second probability is as follows:

[0134] ;

[0135] in, Indicates the number of third data that contains both frequent item sets A and frequent item sets C. Represents the number of times the frequent item set A appears alone in the third data.

[0136] From the above calculation model, it can be seen that the value range of the second probability is [0, 1]. When the value is 0, it means that there is no association between frequent item sets A and frequent item sets C; when the value is 1, it means that frequent item sets A and frequent item sets C are completely related; when the value is between 0-1, it means that frequent item sets A and frequent item sets C have a certain association.

[0137] Step S32224: second deletion, deleting the second probability lower than the preset second probability threshold and the second frequent itemsets corresponding to the second probability, and obtaining the current first frequent itemsets and the third frequent itemsets.

[0138] Similar to the first deletion step, by comparing the second probability with the preset second probability threshold, those second frequent item sets with weaker correlation under the current conditions are eliminated. Any second probability below this threshold and its corresponding second frequent item set will be excluded. In this embodiment, the second probability threshold is 0.6. If it is lower than 0.6, the frequent itemset C is deleted, and the first frequent itemset (frequent itemset A) and the third frequent itemset (frequent itemset D) remain.

[0139] Step S32225: data update, combining the first frequent itemset and the third frequent itemset as new third data, that is, merging the historical data in the first frequent itemset and the historical data in the third frequent itemset as new third data, and executing step S3223: second calculation.

[0140] Step S3223: second calculation, based on step S32225: new third data in data updating, by means of statistical analysis, respectively calculate the correlation coefficients between the current risk type and the remaining n-1 risk types.

[0141] Step S3224: Second acquisition, acquiring the first risk type whose correlation coefficient is calculated with the current risk type.

[0142] Step S3225: The third judgment is to judge whether the absolute value of the correlation coefficient is greater than the preset correlation coefficient threshold. This step can preliminarily judge whether the correlation strength between the two risk types has reached a level that requires special attention by comparing the correlation coefficient with the preset correlation coefficient threshold. If so, it indicates that there is a significant mutual influence between the two, and execute step S3226: the fourth judgment; if not, execute step S3229: the third expression.

[0143] Step S3226: The fourth judgment is to determine whether the correlation coefficient is a positive number. If so, it indicates that the two risk types are positively correlated, and one risk type will trigger another risk type. Step S3227: the first expression needs to be executed. If not, it indicates that the two risk types are unrelated, and step S3228: the second expression is executed.

[0144] Step S3227: The first representation indicates that the current risk type is positively correlated with the first risk type, which means that their changing trends are the same. When one risk occurs, the other risk also tends to occur.

[0145] Step S3228: The second representation indicates that the current risk type is negatively correlated with the first risk type, which means that their changing trends are opposite. When one risk occurs, the other risk tends not to occur.

[0146] Step S3229: The third representation indicates that the current risk type is not related to the first risk type, and whether one risk occurs or not will not affect the occurrence of another risk.

[0147] The step S323: acquiring data includes step S3231: third acquisition and step S3232: fourth acquisition.

[0148] Step S3231: The third acquisition is to obtain p first risk types that are positively correlated with the current risk type by means of statistical analysis.

[0149] Step S3232: fourth acquisition, using statistical analysis to acquire q first risk types that are negatively correlated with the current risk type.

[0150] Step S324: First calculation, calculate the risk, the calculation model is as follows:

[0151] ;

[0152] in, is the probability of the i-th risk type; is the weight of the impact of production data on the i-th risk type.

[0153] In other embodiments, the risk is also calculated using the following calculation model:

[0154] ;

[0155] in, is the probability of the current risk type, which can be obtained through the probability distribution output by the neural network model; The weight of the impact of production data on the current risk type; is the probability of the i-th risk type; is the probability of the j-th risk type; is the weight of the impact of production data on the i-th first risk type; is the weight of the impact of production data on the jth first risk type; is the correlation coefficient of the i-th first risk type that is positively correlated with the current risk type; is the correlation coefficient of the jth first risk type that is negatively correlated with the current risk type. and The weight of the impact of the production data on the risk type is obtained through step S322: calculating the correlation degree.

[0156] Table 1 Weights of the impact of production data on risk types

[0157]

[0158] The risk types and weights mentioned above can be adjusted according to the specific circumstances of the electrolytic aluminum production enterprises (such as equipment conditions, management level, personnel training, etc.).

[0159] Step S325: second data collection, collecting historical data of the first type, recorded as first data.

[0160] The step S326: adding data includes step S3261: the sixth judgment and step S3262: supplementing data.

[0161] Step S3261: Sixth judgment, count the number of historical data stored in the current database, and compare it with the preset historical data quantity threshold. If the number of historical data exceeds the preset historical data quantity threshold, it means that the amount of historical data is rich enough to support the training of the Hammerstein model, and step S37 will be executed: first training, that is, using these data for preliminary training and optimization of the Hammerstein model. On the contrary, if the number of historical data does not reach the preset historical data quantity threshold, it means that the amount of historical data is insufficient, which will affect the training effect and generalization ability of the Hammerstein model, and the supplementary data step S3262 will be executed instead: supplementary data to increase the amount of historical data and ensure the smooth progress of Hammerstein model training.

[0162] This step is used to evaluate whether the amount of historical data currently available is sufficient to support the subsequent model training process.

[0163] Step S3262: supplementing data, including step S32621: generating data, step S32622: first evaluation and step S32623: seventh judgment.

[0164] Step S32621: Generate data using the Borderline SMOTE (Synthetic Minority Over-sampling Technique for imbalanced datasets) algorithm. This algorithm is specifically designed to process imbalanced datasets. It generates new, representative samples by analyzing the characteristics of minority class samples (i.e., risk classes with fewer historical data) and their neighboring samples in historical data. These newly generated samples can enrich historical data, especially increase the number of minority class historical data, thereby improving the balance of the dataset and the training effect of the Hammerstein model.

[0165] Step S32622: First evaluation: after generating new historical data, in order to ensure the quality and validity of the historical data, it is necessary to evaluate whether to retain the new historical data. This evaluation process uses the clean factor as an indicator to evaluate the data density (the number of historical data) around each new historical data. The clean factor can reflect the "cleanliness" or "purity" around the new historical data, that is, the density of similar historical data around it, and the evaluation result is recorded as the fourth data.

[0166] Step S32623: The seventh judgment, based on the result of the first evaluation in step S32622, i.e., the fourth data, this step will further determine whether to keep or remove each new historical data. If the number of similar historical data around a new historical data (i.e., the fourth data) exceeds the preset number threshold, it means that the historical data is in a relatively dense area and may be redundant or over-concentrated. At this time, the current historical data will be deleted to avoid overfitting of the data or affecting the generalization ability of the model. On the contrary, if the fourth data does not exceed the preset threshold, it means that the data distribution around the data point is relatively reasonable and should be retained.

[0167] After executing step S32623: the seventh judgment step, the first data collected in step S325: the second data collection and the new historical data retained in step S32623: the seventh judgment are merged as new first data, and step S327: the first training is executed.

[0168] Step S327: First training, using the first data to train the Hammerstein model to obtain a trained Hammerstein model. In this step, the parameters of the nonlinear part are adjusted through an optimization algorithm to minimize the error between the predicted output and the actual output.

[0169] Step S328: prediction, using the production data as input data of the trained Hammerstein model, and outputting predicted data of the production data, recorded as the second data.

[0170] Step S329: first judgment, judging whether the second data meets expectations, if so, executing step S1: first data collection; if not, performing data analysis on the second data, and performing step S4: data display.

[0171] The data analysis of the second data is specifically to analyze the hazardous factors in the production process based on the production data using the HAZOP analysis method. During the analysis, guide words (such as "none", "more", "less", "reverse", "partial", "accompanying", etc.) are combined with process parameters (such as temperature, pressure, flow, liquid level, etc.) or operating conditions to systematically guide the team to think about various possible deviation scenarios. Through the combination of guide words and parameters, the HAZOP analysis can identify various unexpected operating conditions or failure modes that may occur in the production process, and then analyze the possible consequences and causes of these deviations, as well as whether the existing safety measures are sufficient and effective. For example, in the analysis of the electrolytic aluminum process, the deviation of "too high temperature" will be considered, and then explore how this deviation affects the reaction rate, product quality, and even the risk of explosion, and whether the cooling system is sufficient to prevent such situations from occurring.

[0172] Step S4: Data display, displaying the obtained risk factors and the risk level calculated in step S324 to the staff.

[0173] This embodiment discloses a method for detecting risks in electrolytic aluminum production. The method first collects production data in real time, and constructs a neural network model for preliminary risk classification to determine the main risk type. Subsequently, historical data is collected for the risk type, and a Hammerstein model is established using the least squares vector machine algorithm for accurate prediction. The Hammerstein model is trained and applied to real-time production data to generate prediction data. Whether the production process is stable is determined based on whether the predicted data meets expectations. If an abnormality occurs, the HAZOP analysis method is further used to deeply analyze the potential risk factors, and the analysis results are intuitively displayed to the staff so that timely measures can be taken to ensure production safety.

[0174] Example 2: This example discloses a risk detection system for electrolytic aluminum production, referring to Figure 4 , the detection system comprises:

[0175] Data acquisition module, this module is equipped with high-precision sensors and real-time data processing technology, which can continuously capture various key parameters of the production site, such as temperature, pressure, flow, equipment status and other production data. In addition, this module can also store historical data of various risk types.

[0176] Establish a model module. This module will establish a neural network model. The model automatically learns complex nonlinear relationships from production data through a deep learning algorithm, and then accurately identifies the risk type of the production data, providing a basis for subsequent early warning and intervention. At the same time, in order to further improve the prediction accuracy, this module will also build a Hammerstein model, which is a composite model that combines a linear dynamic system and a static nonlinear link. It is suitable for processing nonlinear dynamic systems commonly used in industrial production. In this embodiment, the Hammerstein model is used to predict production data to obtain predicted data.

[0177] The data processing module includes a first data processing unit, a second data processing unit and a third data processing unit. The first data processing unit is communicatively connected to the data acquisition module, the second data processing unit is communicatively connected to the data acquisition module, and the third data processing unit is communicatively connected to the second data processing unit and the judgment module.

[0178] The first data processing unit directly receives the real-time production data from the data acquisition module, preprocesses the real-time production data, uses the preprocessed production data as new real-time production data, and sends it as input to the neural network model. After rapid calculation, the model outputs the risk type information of the production data.

[0179] The second data processing unit focuses on the training and optimization of the Hammerstein model. It uses historical data of various risk types to solve and optimize model parameters until the model can accurately capture the dynamic changes in the production process.

[0180] The third data processing unit uses the trained Hammerstein model and the output of the data acquisition module to predict the real-time production data and generate the second data, namely the predicted production status or performance index.

[0181] A judgment module evaluates the second data output by the third data processing unit to determine whether it meets the expected production standard or safety threshold.

[0182] The data display module displays the second data to the staff in various forms including charts, reports, alarm information, etc., to help the staff quickly understand the risk status of the production site, take timely response measures, and ensure the safety and stability of the production process.

[0183] This embodiment provides a detection system for electrolytic aluminum production risks. The data acquisition module of this system collects production data and historical data in real time, and conducts in-depth analysis and prediction of production data by establishing a neural network model and a Hammerstein model in the model module. The three units in the data processing module work together. First, the production data is input into the neural network model to determine the risk type, and then the Hammerstein model is trained with historical data to predict the production status. The judgment module evaluates the prediction results. If it does not meet expectations, the data display module presents the second data to the staff in a clear manner so that timely countermeasures can be taken to ensure production safety.

[0184] The above are all preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. Therefore, any equivalent changes made according to the structure, shape, and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for detecting risks in electrolytic aluminum production, characterized in that: include: First data collection: real-time collection of production data, including temperature data, electrolyte composition and liquid level, gas emission data, equipment status data, and current and voltage data during electrolytic cell operation, including electrolytic cell status data, agitator status data, and cooling system status data; First modeling: Building a neural network model based on deep learning algorithms; Second modeling: Building the Hammerstein model based on the least squares vector machine algorithm; Data processing: including classification, second data collection, first training, prediction and first judgment steps; Classification: Input the production data into the neural network model to obtain the probability distribution of risk types of the production data. The risk types include: electrical risk, fire and explosion risk, scalding risk, mechanical injury risk, chemical risk and environmental risk. The risk type corresponding to the maximum probability is recorded as the first type. Second data collection: collecting the first type of historical data, recorded as the first data; First training: using the first data to train the Hammerstein model to obtain a trained Hammerstein model; Prediction: Use the production data as the input data of the trained Hammerstein model, and output the predicted data of the production data, which is recorded as the second data; First judgment: judging whether the second data meets expectations, if so, executing the first data collection step; if not, performing data analysis on the second data and executing the data display step; The data analysis of the second data includes: based on the production data, using the HAZOP analysis method to analyze the risk factors existing in the production process; Data display: Display the risk factors and risk levels to staff; After executing the classification step and before executing the second data collection step, the method further includes: First acquisition: obtain n risk types whose risk type probability distribution of production data exceeds a preset probability threshold, and record the nth risk type as the nth risk type; Second judgment: judging whether there is a correlation between the first n risk types, if so, executing the first calculation step; if not, executing the second data collection step; First calculation: Calculate the risk. The calculation model is as follows: ; in, is the probability of the i-th risk type; is the weight of the impact of production data on the i-th risk type; After executing the second determination step and before executing the first calculation step, the method further includes: Calculating the correlation degree: including third data collection, second calculation, second acquisition, third judgment, fourth judgment, first expression, second expression and third expression; Third data collection: collect historical data of all risk types and record them as third data; Second calculation: based on the third data, respectively calculate the correlation coefficients between the current risk type and the remaining n-1 risk types by means of statistical analysis; Second acquisition: acquiring the first risk type with which the correlation coefficient is calculated for the current risk type; Third judgment: judging whether the absolute value of the correlation coefficient is greater than a preset correlation coefficient threshold, if so, executing the fourth judgment step; if not, executing the third representation step; Fourth judgment: judging whether the correlation coefficient is a positive number, if so, executing the first representation step; if not, executing the second representation step; The first representation: indicates that the current risk type is positively correlated with the first risk type; The second representation: indicates that the current risk type is negatively correlated with the first risk type; The third representation: indicates that the current risk type is not related to the first risk type; After executing the step of calculating the degree of association and before executing the first calculation step, the method further includes: Third acquisition: obtaining p first risk types that are positively correlated with the current risk type; Fourth acquisition: acquiring q first risk types that are negatively correlated with the current risk type; In the first calculation step, the risk may also be calculated using the following calculation model: ; in, is the probability of the current risk type; The weight of the impact of production data on the current risk type; is the probability of the i-th risk type; is the probability of the j-th risk type; is the weight of the impact of production data on the i-th first risk type; is the weight of the impact of production data on the jth first risk type; is the correlation coefficient of the i-th first risk type that is positively correlated with the current risk type; is the correlation coefficient of the jth first risk type that is negatively correlated with the current risk type; After executing the third data acquisition step and before executing the second calculation step, the method further includes: Data storage: storing the third data in a database; First scan: Scan the database, count the number of occurrences of each historical data in each risk type, obtain the historical data whose number of occurrences is greater than the preset threshold, and record it as the first frequent item; Establishing an FP tree: constructing an FP tree according to the frequency of occurrence of the first frequent item, wherein a node of the FP tree represents a first frequent item, and a counter of a node in the FP tree represents the number of times the first frequent item appears in the database; First construction: Starting from the root node of the FP tree, a conditional pattern base is constructed for each first frequent item; Establish a conditional FP tree: construct a conditional FP tree based on the conditional pattern of the first frequent item, the conditional FP tree includes a plurality of child nodes, each child node represents a frequent item associated with the first frequent item, the frequent item associated with the first frequent item is recorded as the second frequent item, and the counter of each child node represents the number of times the second frequent item appears in the database in the current conditional FP tree; Fifth judgment: judge whether the conditional FP tree contains only one second frequent item. If so, execute the data merging step; if not, take the conditional FP tree as a new FP tree, take the second frequent items as new first frequent items, and execute the first construction step; Data merging: Merge the second frequent items belonging to the same conditional pattern base to obtain the first frequent item set.

2. The electrolytic aluminum production risk detection method according to claim 1, characterized in that: After executing the data merging step and before executing the second calculation step, the method further includes: The third calculation: calculate the probability of the current first frequent item set and the remaining first frequent item sets occurring at the same time, recorded as the first probability; First deletion: deleting the first probability lower than the preset first probability threshold and the first remaining frequent item set corresponding to the first probability, to obtain the current first frequent item set and the second frequent item set; Fourth calculation: Calculate the probability that the second frequent item set also occurs when the first frequent item set occurs, recorded as the second probability; Second deletion: deleting the second probability lower than the preset second probability threshold and the second frequent item set corresponding to the second probability, to obtain the current first frequent item set and the third frequent item set; Data update: Combine the first frequent item set and the third frequent item set as the new third data.

3. The electrolytic aluminum production risk detection method according to claim 1, characterized in that: The conditional pattern base includes: A prefix path, wherein the prefix path includes a set of all frequent items on a node path from a first frequent item to a certain first frequent item; A suffix set, wherein the suffix set includes a first frequent item after a first frequent item in the FP tree.

4. The electrolytic aluminum production risk detection method according to claim 1, characterized in that: After executing the second data collection step and before executing the first training step, the method further includes: Sixth judgment: judging whether the amount of historical data exceeds the preset historical data amount threshold, if yes, executing the first training step; if no, executing the data supplementation step; Supplementary data: including data generation, first assessment and seventh judgment steps; Generate data: Based on historical data, use the Borderline SMOTE algorithm to generate new historical data; First evaluation: Use the clean factor to evaluate the number of historical data around each historical data, recorded as the fourth data; Seventh judgment: judging whether the fourth data exceeds a preset quantity threshold, if so, deleting the current historical data; if not, retaining the current historical data.

5. An electrolytic aluminum production risk detection system, the detection system is used to execute the detection method according to any one of claims 1 to 4, characterized in that: include: The data collection module is used to collect production data in real time; it is also used to collect historical data based on the probability distribution of risk types of production data; Establish model module, used to establish neural network model and Hammerstein model; The data processing module includes a first data processing unit, a second data processing unit and a third data processing unit, wherein the first data processing unit is communicatively connected to the data acquisition module, the second data processing unit is communicatively connected to the data acquisition module, and the third data processing unit is communicatively connected to the second data processing unit and the judgment module; A first data processing unit is used to input the production data into the neural network model to obtain a risk type probability distribution of the production data; A second data processing unit, used for training the Hammerstein model using historical data; A third data processing unit is used to predict the production data using the trained Hammerstein model to obtain second data; A judging module, used to judge whether the second data meets expectations; The data display module is connected to the judgment module for displaying the second data to the staff.