An artificial intelligence-based coking coal type automatic identification method

Abstract

The application discloses a kind of based on artificial intelligence's coking coal kind automatic identification method, belong to intelligent coking technical field.The identification method of the present application includes, step one, obtains the parameter information of the coking coal to be identified in multiple dimensions, the parameter information includes ash parameter information, sulfur parameter information, volatile parameter information, reflectivity parameter information and cohesiveness parameter information;Step two, the parameter information in multiple dimensions is input into the identification submodel in each dimension parameter information corresponding identification submodel in pre-trained coking coal kind identification model, and according to confidence is iteratively calculated, until each identification submodel corresponding confidence all meet the confidence threshold requirement, and then output the target kind for the coking coal to be identified.The present application can effectively improve the accuracy and intelligence of coal type determination, reduce the error caused by artificial experience.

Description

Technical Field

[0001] This invention belongs to the field of intelligent coking technology, and more specifically, relates to an automatic identification method for coking coal types based on artificial intelligence. Background Technology

[0002] According to the degree of coalification and caking ability, my country's coal classification standard (GB / T5751-2009) stipulates that long-flame coal, non-caking coal, weakly caking coal, 1 / 2 medium-caking coal, gas coal, gas-rich coal, 1 / 3 coking coal, fat coal, coking coal, lean coal, semi-lean coal, and poor coal belong to the category of bituminous coal. Among them, gas coal, gas-rich coal, 1 / 3 coking coal, fat coal, coking coal, and lean coal are coking coals (Technical Conditions for Coking Coal GB / T397-2009). However, in actual production, all 12 categories of bituminous coal have been used in blending for coking, only the amount used varies. The blending ratio of different types of coking coal determines the properties of the blended coal and the quality of the coke. With the improvement of equipment automation and the widespread application of artificial intelligence, intelligent coking will be increasingly valued. The premise of intelligent coking is intelligent coal blending, and the classification and automatic identification of incoming coal types are the key to realizing intelligent coal blending.

[0003] A search revealed Chinese patent application number 201410195065.6, which discloses a coal quality classification and blending method based on the coking properties of coking coal. The classification method of this application includes: 1) determining the indicators affecting the coking properties of coking coal: the average maximum reflectance of vitrinite, the maximum flowability of Gibbs freewheeling, the solid-soft temperature range, and the optical microstructure of coke are determined as indicators of the coking properties of coking coal; 2) measuring the indicators affecting the coking properties of coking coal; 3) based on the different results of step 2), classifying each type of coal into gas-rich coal, gas coal, fat coal, 1 / 3 coking coal, coking coal, lean coal, and inferior blended coal or coal of special origin.

[0004] For example, application number 201410335178.1 discloses a method for subdividing coking coal based on coking properties and its application in coal blending. The method of this application includes the following steps: 1) determining the indicators affecting the coking properties of coking coal; 2) measuring the average maximum reflectance and coke optical structure of a single type of coking coal, and, depending on the different maximum reflectance, further measuring the solid-soft temperature range, the maximum flowability of the Gibbs free, or not measuring other indicators; 3) based on the measurement results of step 2), subdividing the coking coal into gas coal 1#, gas coal 2#, gas fat coal, 1 / 3 coking coal 1#, 1 / 3 coking coal 2#, 1 / 3 coking coal 3#, fat coal 1#, fat coal 2#, coking coal 1#, coking coal 2#, coking coal 3#, coking coal 4#, lean coal 1#, and lean coal 2#.

[0005] The two applications mentioned above use the average maximum reflectance of vitrinite, the maximum flowability of Kierkegaard, the solid-soft temperature range, and the optical microstructure of coke as indicators of the coking properties of coking coal, and provide the classification range of the indicators, but do not consider the volatile matter content (V) of coal. daf The impact of the indicators. Summary of the Invention

[0006] 1. The problem to be solved

[0007] To address the issues of poor automation and accuracy in existing coking coal type identification technologies, this invention provides an artificial intelligence-based automatic coking coal type identification method. Employing this invention effectively improves the accuracy and intelligence of coal type determination, reducing errors caused by human experience.

[0008] 2. Technical Solution

[0009] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0010] The present invention provides an automatic identification method for coking coal types based on artificial intelligence, comprising the following steps:

[0011] Step 1: Obtain parameter information of the coking coal to be identified in multiple dimensions, including ash content, sulfur content, volatile matter, reflectance, and caking properties.

[0012] Step 2: Input the parameter information from multiple dimensions into the identification sub-models corresponding to the parameter information of each dimension in the pre-trained coking coal type identification model, and perform iterative calculations based on the confidence level until the confidence level of each identification sub-model meets the confidence threshold requirement, and then output the target type of coking coal to be identified.

[0013] Furthermore, the adhesive parameters include the adhesion index, adhesive layer thickness, Oya expansion, and Gibbs flowability.

[0014] Furthermore, among the multiple recognition sub-models, the first recognition sub-model corresponding to the ash content parameter information and the second recognition sub-model corresponding to the sulfur content parameter information are trigger models trained according to sub-triggers with different triggering rules.

[0015] Furthermore, the third identification sub-model corresponding to the volatile matter parameter information, the fourth identification sub-model corresponding to the reflectivity information, and the fifth identification sub-model corresponding to the adhesiveness parameter information are obtained through training using an active learning model.

[0016] Furthermore, the training methods for the first and second recognition sub-models are as follows:

[0017] For the aforementioned trigger model, a multi-level sub-trigger is constructed; wherein, the output of each higher-level sub-trigger is connected to the input of three sub-triggers in the next lower level, and the three sub-triggers in the next lower level are configured with corresponding trigger conditions.

[0018] Historical sample data is sequentially input into the top-level sub-trigger, and the sub-triggers are trained layer by layer to obtain the triggering result;

[0019] If a preset number of triggering results are obtained, for any upper-level sub-trigger, if there is a lower-level sub-trigger that does not meet the corresponding preset training rule, the triggering condition of the lower-level sub-trigger is changed, and the historical sample data is re-inputted according to the sub-trigger with the changed triggering condition, until the corresponding preset training rule is met for any sub-trigger.

[0020] Furthermore, the number of sub-flip-flops in each upper level is greater than the number of sub-flip-flops in the corresponding lower level, and the difference between the number of sub-flip-flops in each upper level and the number of sub-flip-flops in the corresponding lower level is the same.

[0021] Furthermore, the preset training rule is that the number of triggering results of the input sub-trigger is not less than the corresponding preset trigger value.

[0022] Furthermore, the triggering condition is that the triggering duration of the previous level exceeds the preset triggering duration threshold, or the preset training rule is not met.

[0023] Furthermore, the training methods for the third, fourth, and fifth recognition sub-models are as follows:

[0024] Historical sample data is input into the reference analysis model to obtain candidate reference results for each historical sample data and the confidence level of each candidate reference result;

[0025] Historical sample data is input into the initial binary classification model to obtain the initial analysis results of each historical sample data;

[0026] The candidate reference results are filtered based on the initial analysis results and the confidence level of the candidate reference results to obtain the target sample data;

[0027] The initial binary classification model is trained based on the target sample data to obtain the corresponding recognition sub-model.

[0028] Furthermore, the acquisition of the target sample data includes the following steps:

[0029] Candidate reference results are divided into multiple sets based on the probability density of the initial analysis results;

[0030] For each set, the set confidence is determined based on the annotation results of historical sample data and the candidate reference results within the set;

[0031] Candidate reference results with built-in confidence levels lower than the set confidence level are removed to obtain the target sample data.

[0032] Furthermore, the training steps of the coking coal type identification model include:

[0033] For any identification sub-model with a confidence level less than the corresponding confidence threshold, the historical sample data corresponding to the identification sub-model is filtered based on the confidence vector propagation method.

[0034] The filtered historical sample data is input into the corresponding recognition sub-model to obtain the new confidence level. If the new confidence level is less than the corresponding confidence level threshold, the confidence level of the recognition sub-model is determined to meet the confidence level threshold. The coking coal type recognition model is trained until the confidence levels of all recognition sub-models meet the confidence level threshold.

[0035] 3. Beneficial effects

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] This invention provides an automatic identification method for coking coal types based on artificial intelligence. By configuring identification sub-models corresponding to parameter information in various dimensions within a pre-trained coking coal type identification model, corresponding identification results can be obtained for parameter information in multiple dimensions. Compared with traditional identification methods, this invention uses two sets of indicators—volatile matter and reflectivity—to define the degree of coal metamorphism, which is more accurate than a single indicator. Simultaneously, it categorizes indicators reflecting coal caking properties, such as the caking index, plastic layer thickness, Oya expansion, and Gibbs freeness, into caking parameter information, which has a more significant distinguishing effect than a single indicator. For adjacent types of coking coal with similar metamorphism and caking parameters, machine learning and iterative model training improve the accuracy of coal type classification. Furthermore, the use of multiple models for artificial intelligence identification of coking coal types improves the accuracy and intelligence of coal type determination, reducing errors caused by human experience. Attached Figure Description

[0038] Figure 1 This is a flowchart of an automatic coking coal type identification method based on artificial intelligence according to the present invention;

[0039] Figure 2 This is a flowchart illustrating the training process of the identification sub-model corresponding to ash content parameter information and sulfur content parameter information in this invention.

[0040] Figure 3This is a flowchart illustrating the training process of the identification sub-model corresponding to the volatile matter parameter information, reflectivity parameter information, and adhesiveness parameter information in this invention. Detailed Implementation

[0041] The present invention will be further described below with reference to specific embodiments.

[0042] This embodiment presents an automatic identification method for coking coal types based on artificial intelligence.

[0043] First, obtain parameter information of the coking coal to be identified in multiple dimensions, including ash content, sulfur content, volatile matter, reflectance, and caking properties. Among them, the caking properties include caking index, plastic layer thickness, Oya expansion degree, and Gibbs flowability.

[0044] Then, the parameter information from multiple dimensions is input into the pre-trained coking coal type identification model to obtain the target type of coking coal to be identified from the output of the coking coal type identification model.

[0045] Specifically, the pre-trained coking coal type identification model includes identification sub-models corresponding to parameter information in each dimension. The first identification sub-model corresponding to the ash content parameter information and the second identification sub-model corresponding to the sulfur content parameter information are trigger models trained according to sub-triggers with different triggering rules.

[0046] The third identification sub-model corresponding to the volatile matter parameter information, the fourth identification sub-model corresponding to the reflectivity information, and the fifth identification sub-model corresponding to the adhesiveness parameter information are obtained by actively learning the model to filter historical sample data to obtain target sample data, and then training the binary classification model based on the target sample data.

[0047] The pre-trained coking coal type identification model is obtained by iteratively calculating the confidence level of the identification results output by each identification sub-model until the confidence level of each identification sub-model meets the confidence level threshold.

[0048] The first and second recognition sub-models were trained using the following method:

[0049] For the aforementioned trigger model, a multi-level sub-trigger is constructed. The output of each higher-level sub-trigger is connected to the inputs of three sub-triggers in the next lower level, and these three sub-triggers are configured with corresponding trigger conditions.

[0050] Historical sample data is sequentially input into the top-level sub-trigger, and the sub-triggers are trained layer by layer to obtain the triggering results. Specifically, based on whether the triggering result of each sub-trigger satisfies the triggering condition of the corresponding sub-trigger in the next level, the target sub-trigger is determined from the three sub-triggers in the next level.

[0051] If a preset number of triggering results are obtained, for any upper-level sub-trigger, if a lower-level sub-trigger does not meet the corresponding preset training rule, the triggering condition of the lower-level sub-trigger is modified. Historical sample data is then input again based on the sub-triggers with modified triggering conditions until the corresponding preset training rule is met for all sub-triggers.

[0052] The preset training rule is that the number of triggering results of the input sub-triggers is not less than the corresponding preset triggering value. The triggering condition is that the triggering duration of the upper level exceeds the preset triggering duration threshold, or the preset training rule is not met. In addition, the number of sub-triggers in each upper level is greater than the number of sub-triggers in the corresponding lower level, and the difference between the number of sub-triggers in each upper level and the number of sub-triggers in the corresponding lower level is the same.

[0053] The third, fourth, and fifth recognition sub-models were trained using the following method:

[0054] Historical sample data is input into the reference analysis model to obtain the candidate reference results for each historical sample data and the confidence level of each candidate reference result. The historical sample data is selected from the historical sample dataset used to train the reference analysis model, which is built based on an active learning model.

[0055] Input historical sample data into the initial binary classification model to obtain the initial analysis results of each historical sample data output by the initial binary classification model;

[0056] The candidate reference results are filtered based on the initial analysis results and the confidence level of the candidate reference results to obtain the target sample data;

[0057] The initial binary classification model is trained based on the target sample data to obtain the corresponding recognition sub-model.

[0058] The process of filtering candidate reference results based on the initial analysis results and the confidence level of the candidate reference results to obtain target sample data includes the following specific steps:

[0059] Candidate reference results are divided into multiple sets based on the probability density of the initial analysis results;

[0060] For each set, the set confidence is determined based on the annotation results of historical sample data and the candidate reference results within the set;

[0061] Candidate reference results with built-in confidence levels lower than the set confidence level are removed to obtain the target sample data.

[0062] The pre-training of the coking coal type identification model specifically includes the following steps:

[0063] For any identification sub-model with a confidence level less than the corresponding confidence threshold, the historical sample data corresponding to the identification sub-model is filtered based on the confidence vector propagation method.

[0064] The filtered historical sample data is input into the corresponding recognition sub-model to obtain the new confidence level of the recognition sub-model for the filtered historical sample data; if the new confidence level is less than the corresponding confidence level threshold, it is determined that the confidence level of the recognition sub-model meets the confidence level threshold, and the coking coal type recognition model is trained until the confidence level of all recognition sub-models meets the confidence level threshold.

[0065] This embodiment presents an automatic identification method for coking coal types based on artificial intelligence. By configuring identification sub-models corresponding to parameter information of each dimension in a pre-trained coking coal type identification model, the method can obtain corresponding identification results for parameter information in multiple dimensions. Furthermore, the first identification sub-model and the second identification sub-model are trigger models trained according to sub-triggers with different triggering rules. With the help of trigger models with different triggering rules, ash content and sulfur content can be accurately determined.

[0066] The third, fourth, and fifth identification sub-models respectively obtain target sample data by filtering historical sample data through an active learning model. The binary classification model is then trained based on this target sample data. The pre-trained coking coal type identification model is obtained by iteratively calculating the confidence scores of the identification results output by each sub-model until the confidence scores of each sub-model all meet the confidence threshold. Combining multiple models for coking coal type identification not only improves the accuracy of coal type determination and reduces errors caused by relying on human experience, but also enhances the intelligence of coking coal type determination.

[0067] The present invention and its embodiments have been described above illustratively. This description is not restrictive, and the figures shown are only one embodiment of the present invention; the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. An automatic identification method for coking coal types based on artificial intelligence, characterized in that, Includes the following steps: Step 1: Obtain parameter information of the coking coal to be identified in multiple dimensions, including ash content, sulfur content, volatile matter, reflectance, and caking properties. Step 2: Input the parameter information from multiple dimensions into the corresponding identification sub-models of the pre-trained coking coal type identification model, and iteratively calculate based on the confidence level until the confidence level of each identification sub-model meets the confidence threshold requirement, and then output the target type of coking coal to be identified; where, Among the multiple recognition sub-models, the first recognition sub-model corresponding to the ash content parameter information and the second recognition sub-model corresponding to the sulfur content parameter information are trigger models trained according to sub-triggers with different triggering rules. The third identification sub-model corresponding to the volatile matter parameter information, the fourth identification sub-model corresponding to the reflectivity information, and the fifth identification sub-model corresponding to the adhesiveness parameter information are obtained through active learning model training.

2. The method for automatic identification of coking coal types based on artificial intelligence according to claim 1, characterized in that: The adhesive parameters include the adhesion index, adhesive layer thickness, Oya expansion, and Gibbs flowability.

3. The method for automatic identification of coking coal types based on artificial intelligence according to claim 2, characterized in that, The training methods for the first and second recognition sub-models are as follows: For the aforementioned trigger model, a multi-level sub-trigger is constructed; wherein, the output of each higher-level sub-trigger is connected to the input of three sub-triggers in the next lower level, and the three sub-triggers in the next lower level are configured with corresponding trigger conditions. Historical sample data is sequentially input into the top-level sub-trigger, and the sub-triggers are trained layer by layer to obtain the triggering result; If a preset number of triggering results are obtained, for any upper-level sub-trigger, if there is a lower-level sub-trigger that does not meet the corresponding preset training rule, the triggering condition of the lower-level sub-trigger is changed, and the historical sample data is re-inputted according to the sub-trigger with the changed triggering condition, until the corresponding preset training rule is met for any sub-trigger.

4. The method for automatic identification of coking coal types based on artificial intelligence according to claim 3, characterized in that: The number of sub-flip-flops in each upper level is greater than the number of sub-flip-flops in the corresponding lower level, and the difference between the number of sub-flip-flops in each upper level and the number of sub-flip-flops in the corresponding lower level is the same.

5. The method for automatic identification of coking coal types based on artificial intelligence according to claim 4, characterized in that: The preset training rule is that the number of trigger results of the input sub-trigger is not less than the corresponding preset trigger value.

6. The method for automatic identification of coking coal types based on artificial intelligence according to claim 5, characterized in that: The triggering condition is that the triggering duration of the previous level exceeds the preset triggering duration threshold, or the preset training rule is not met.

7. The method for automatic identification of coking coal types based on artificial intelligence according to claims 1-6, characterized in that, The training methods for the third, fourth, and fifth recognition sub-models are as follows: Historical sample data is input into the reference analysis model to obtain candidate reference results for each historical sample data and the confidence level of each candidate reference result; Historical sample data is input into the initial binary classification model to obtain the initial analysis results of each historical sample data; The candidate reference results are filtered based on the initial analysis results and the confidence level of the candidate reference results to obtain the target sample data; The initial binary classification model is trained based on the target sample data to obtain the corresponding recognition sub-model.

8. The method for automatic identification of coking coal types based on artificial intelligence according to claim 7, characterized in that, The acquisition of the target sample data includes the following steps: Candidate reference results are divided into multiple sets based on the probability density of the initial analysis results; For each set, the set confidence is determined based on the annotation results of historical sample data and the candidate reference results within the set; Candidate reference results with built-in confidence levels lower than the set confidence level are removed to obtain the target sample data.

9. The method for automatic identification of coking coal types based on artificial intelligence according to claim 8, characterized in that: The training steps of the coking coal type identification model include: For any identification sub-model with a confidence level less than the corresponding confidence threshold, the historical sample data corresponding to the identification sub-model is filtered based on the confidence vector propagation method. The filtered historical sample data is input into the corresponding recognition sub-model to obtain the new confidence level. If the new confidence level is less than the corresponding confidence level threshold, the confidence level of the recognition sub-model is determined to meet the confidence level threshold. The coking coal type recognition model is trained until the confidence levels of all recognition sub-models meet the confidence level threshold.

Images