Machine learning model

Abstract

To provide a machine learning model which can accurately predict sensibility of purification performance according to a catalyst structure and durability time.SOLUTION: A machine learning model predicts purification performance of an exhaust gas purification catalyst. The machine learning mode is constructed through machine learning which uses information representing a constitution as an explanatory variable and information representing purification performance as an objective variable. The machine learning is given in which learning of a model is performed predetermined times; the model after learning predetermined times is stored as a temporary model; and an update learning is repeatedly executed. The update learning is given in which learning of the temporary mode is performed at once; sensibility of purification performance according to amount of catalytic metal and durability time predicted using the model after learning at once is determined to satisfy a sensibility condition or not; and when the sensibility condition is satisfied, the model after learning at once is overlaid as the temporary model, and when the sensibility condition is not satisfied, the model after learning at once is not overlaid as the temporary model.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The entire contents of Patent Application No. 2023-007974 filed with the Japan Patent Office on January 23, 2023 and the entire contents of Patent Application No. 2023-081090 filed with the Japan Patent Office on May 16, 2023 are incorporated by reference in the disclosure of this application. The present invention relates to a machine learning model for predicting the performance of an exhaust gas purification catalyst. [Background technology]

[0002] Conventionally, various machine learning models for predicting the purification performance of exhaust gas purification catalysts are known. Specifically, for example, Patent Document 1 describes a machine learning model that is trained in advance using the ratio of elements contained in the surfaces of multiple types of catalysts and the specific surface area values ​​of the catalyst surfaces as training data. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2022-185940 A Summary of the Invention [Problem to be solved by the invention]

[0004] Conventionally, it has been difficult to prepare learning data that covers all the factors related to the catalyst configuration and durability history that affect the purification performance, and there is a risk of overlearning due to the large number of explanatory variables that affect the purification performance. Therefore, it has been difficult to accurately predict the sensitivity of the purification performance as expected in conventional knowledge. The present invention has been made in consideration of these points, and its purpose is to provide a machine learning model that can accurately predict the sensitivity of the purification performance according to the catalyst configuration and durability time. [Means for solving the problem]

[0005] In order to solve the above problem, the machine learning model of the present invention is a machine learning model that predicts the purification performance of an exhaust gas purification catalyst, and the machine learning model is constructed by performing machine learning using information having information representing the catalyst configuration as an explanatory variable and information representing the purification performance as a target variable, and the machine learning involves learning a model a predetermined number of times, saving the model after the predetermined number of times of learning as a provisional model, and then repeatedly performing update learning, and in the update learning, learning of the provisional model is performed once, and it is determined whether the sensitivity of the purification performance according to the amount of catalytic metal and durability time predicted using the model after the first learning satisfies a sensitivity condition, and if the sensitivity condition is satisfied, the model after the first learning is overwritten and saved as the provisional model, and if the sensitivity condition is not satisfied, the model after the first learning is not overwritten and saved as the provisional model. Effect of the Invention

[0006] According to the present invention, it is possible to accurately predict the sensitivity of purification performance depending on the catalyst configuration and endurance time. [Brief description of the drawings]

[0007] [Figure 1] 1A is a schematic cross-sectional view showing an enlarged view of the partition walls and catalyst layer of the substrate in each learning sample of the catalyst, and FIG. 1B and FIG. 1C are diagrams explaining a method for acquiring information representing a durability history, a method for acquiring information representing a purification performance evaluation test, and a method for acquiring information representing purification performance, respectively. [Diagram 2] 13A and 13B are process flows for constructing a machine learning model according to an embodiment and a comparative example, respectively. [Diagram 3]The left graph in (a) shows the results of predicting the sensitivity of T50(NOx) according to the endurance time of the endurance pattern A of the medium-performance catalyst by using the machine learning model of the embodiment, together with the results obtained by the actual measurement of the high-performance catalyst and the low-performance catalyst. The right graph in (a) shows the results of predicting the sensitivity of T50(NOx) according to the endurance time of the endurance pattern A of the medium-performance catalyst by using the machine learning model of the comparative example, together with the results obtained by the above actual measurement. The left graph in (b) shows the results of predicting the sensitivity of T50(NOx) according to the endurance time of the endurance pattern B of the low-performance catalyst by using the machine learning model of the embodiment, together with the results obtained by the actual measurement of the high-performance catalyst and the medium-performance catalyst. The right graph in (b) shows the results of predicting the sensitivity of T50(NOx) according to the endurance time of the endurance pattern B of the low-performance catalyst by using the machine learning model of the comparative example, together with the results obtained by the above actual measurement. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0008] The machine learning model according to the embodiment is a prediction model used instead of the prediction model of the prediction unit in the prediction device according to an embodiment of Japanese Patent Application No. 2023-007974. The number of times of learning before the update learning in the machine learning of the machine learning model according to the embodiment is not particularly limited and may be appropriately adjusted so as to improve the prediction accuracy, but if it is too many, overlearning may occur, and if it is too few, the prediction accuracy may decrease. The number of times of repetition of the update learning is not particularly limited, but may be adjusted so as to improve the prediction accuracy, taking into account the number of times of learning before the update learning. According to the machine learning model according to the embodiment, by proceeding with the learning one by one in the update learning, the learning does not progress in a direction that is different from the assumption of conventional knowledge, so that the sensitivity of the purification performance according to the catalyst configuration and the endurance time can be accurately predicted. EXAMPLES

[0009] [Example] An example of a machine learning model according to the embodiment was created.

[0010] 1. Obtaining a data set of explanatory variables and target variables to be used in building a machine learning model (1) Preparation of a study sample for exhaust gas purification catalyst First, 800 study samples of exhaust gas purification catalysts were prepared in the same manner as in Japanese Patent Application No. 2023-007974. As shown in FIG. 1(a), each study sample CS is a straight-flow type catalyst, and includes a honeycomb substrate 10 integrally formed with a frame (not shown) and partition walls 14 that divide the inner space into a honeycomb shape, and a catalyst layer 20 provided on the cell 12 side surface 14c of the partition walls 14 and having at least a first layer and a second layer among the first layer 20a to the fifth layer 20e. Each layer of the catalyst layer 20 contains powders P1 to Px (x is a predetermined integer of 2 or more) and contains a catalyst metal supported on at least one type of powder. The catalyst metal is at least one type of catalyst metals M1 to My (y is a predetermined integer of 2 or more).

[0011] (2) Obtaining explanatory variables and objective variables for each learning sample of the catalyst Next, for each learning sample of the catalyst, a group consisting of information expressing the catalyst configuration, information expressing the durability history, and information expressing the purification performance evaluation test was obtained as explanatory variables, and information expressing the purification performance was obtained as a target variable. From each learning sample of the catalyst, information having information on each layer of the catalyst layer and information on the honeycomb substrate was obtained as information expressing the catalyst configuration, similar to Japanese Patent Application No. 2023-081090. As information on each layer of the catalyst layer, information including the amount [g] of each type of powder P1 to Px and the amount [g] of each type of catalyst metal M1 to My supported on each type of powder P1 to Px, and data on the coating region was obtained. As information on the honeycomb substrate, data on weight, cell shape, and wall thickness (6 dimensions) was obtained. Note that the data on the coating region is data including the coating width and coating position of each layer of the catalyst layer. A durability test was performed on each learning sample of the catalyst in the same manner as in Japanese Patent Application No. 2023-007974. Next, as shown in FIG. 1(b), the catalyst bed temperature (endurance temperature) and the air-fuel ratio (A / F sensor value) with the catalyst were measured throughout the entire duration of each durability test, and were saved as data for each durability test. Next, similar to Patent Application No. 2023-007974, durability history 2D histogram data (image data of 30 pixels x 30 pixels in which pixel values ​​represent the frequency of the combination of air-fuel ratio and catalyst bed temperature) that represents the durability history of each learning sample of the catalyst was created from the data of each durability test. Next, similar to Patent Application No. 2023-081090, the data of the durability history 2D histogram was dimensionally compressed by a convolutional autoencoder to obtain data of latent variables (16 dimensions) that represent the durability history as information that represents the durability history. A purification performance evaluation test was performed on each learning sample of the catalyst deteriorated in each durability test in the same manner as in Patent Application No. 2023-007974, and the test condition data of the purification performance evaluation test was obtained as information expressing the purification performance evaluation test of each learning sample of the catalyst, and the temperature at the catalyst inlet and the NOx concentration at the catalyst outlet in the purification performance evaluation test were obtained per second as information expressing the purification performance. Note that, by using the information expressing the purification performance, T50(NOx) (NOx50% purification temperature) can be calculated as shown in Figure 1(c).

[0012] (3) Obtaining a data set of explanatory variables and target variables For the 800 learning samples of exhaust gas purification catalysts, the explanatory variables and response variables of each learning sample of the catalyst were obtained as described above, thereby obtaining a dataset consisting of 800 sets of data for the explanatory variables and response variables.

[0013] 2. Building a Machine Learning Model Next, a machine learning model was constructed by performing machine learning using the data set of explanatory variables and objective variables in the process flow shown in FIG. 2(a). First, the data set was randomly divided into three sets of a training data set, a test data set, and a validation data set, and a multilayer perceptron model with two hidden layers was constructed ("pre-learning process"). At this time, a model was constructed in which random values ​​were given to the hyperparameters (the number of dimensions of the latent variables in the first hidden layer, the number of dimensions of the latent variables in the second hidden layer, the learning rate, dropout, and epsilon). Next, the multilayer perceptron model was trained for 500 iterations ("training x 500 times"). At this time, in each training, the model was trained on the training data set while verifying the accuracy of the model on the test data set. Then, the model was trained on the test data set so that the RMSE (Root Mean Squared Error) was small. Next, the model after training was saved as a provisional model ("save provisional model"). Next, the provisional model was loaded ("Provisional model load"), and the provisional model was trained in one iteration ("Training x 1"), after which a sensitivity test was performed to determine whether the sensitivity condition was satisfied ("Sensitivity test condition satisfied?"). In one training, the model was trained on the training data set while verifying the accuracy of the model on the test data set. Then, the model was trained so that the RMSE for the test data set would be small. In the sensitivity test, 1) first, data on one type of exhaust gas purification catalyst was selected from the data set. 2) Next, from the explanatory variables of the data on one type of exhaust gas purification catalyst, explanatory variables whose durability history was durability pattern A or durability pattern B were selected as base explanatory variables. 3) Next, explanatory variables in which the amount of a specific type of catalytic metal supported on a specific type of powder of the catalyst configuration was changed in the base explanatory variables, and explanatory variables in which the durability time of the durability history was changed in the base explanatory variables were created. 4) Next, using a provisional model, the sensitivity of the purification performance to the amount of a specific type of catalytic metal supported on a specific type of powder and the sensitivity of the purification performance to the durability time of the durability history were predicted from these explanatory variables.In the judgment of the sensitivity condition, if the sensitivity of the purification performance according to the amount of catalytic metal is in the direction of improvement where the purification performance improves as the amount of catalytic metal increases, and if the sensitivity of the purification performance according to the durability time of the durability history is in the direction of deterioration where the purification performance decreases as the durability time increases, the sensitivity condition is judged to be satisfied, and if not, it is judged to be not satisfied. If the sensitivity condition is satisfied, the model after the one learning is overwritten and saved as a provisional model ("temporary model overwritten and saved") to update the provisional model, and if the sensitivity condition is not satisfied, the model after the one learning is not overwritten and saved as a provisional model, and the original provisional model is maintained as it is. A loop process was performed to repeatedly execute the above process (update learning) 200 times. Next, after the loop process, the provisional model was saved as the actual model ("actual model saved"). Next, the actual model was used to make a prediction on the validation data set, and the RMSE was calculated ("error calculation"). The above series of operations from "pre-learning process" to "error calculation" were repeated 100 times by changing the values ​​given to the hyperparameters. Then, for each of the 100 operations, the number of times the sensitivity conditions were satisfied in the judgment of "Is the sensitivity test condition satisfied?" in the loop processing (the number of YES results) and the RMSE calculated in "error calculation" were used to calculate the number of YES results / RMSE.The model obtained for the operation with the largest number of YES results / RMSE was then adopted as the machine learning model.

[0014] [Comparative Example] First, a data set similar to the data set described in item 1 of the above [Example] was obtained from the learning sample of the exhaust gas purification catalyst as a data set of explanatory variables and objective variables used to construct a machine learning model. Next, a machine learning model was constructed by performing machine learning using the data set of explanatory variables and objective variables in the processing flow shown in FIG. 2(b). First, similar to the "pre-learning processing" described in item 2 of the above [Example], the data set was randomly divided into three, a training data set, a test data set, and a verification data set, and a multilayer perceptron model was constructed ("pre-learning processing"). Next, the multilayer perceptron model was trained for 10,000 iterations ("learning x 10,000 times"). At this time, each learning was performed in the same manner as the "learning x 500 times" described in item 2 of the above [Example]. Then, EarlyStopping with patience set to 2000 times was applied. Next, the model after learning was saved ("model saving"). Next, the saved model was used to make predictions on the validation data set and the RMSE was calculated ("error calculation"). The above series of operations from "pre-learning processing" to "error calculation" were repeated 100 times while changing the values ​​given to the hyperparameters. Then, the model obtained by the operation that gave the smallest RMSE calculated by "error calculation" out of the 100 operations was adopted as the machine learning model.

[0015] [Prediction accuracy evaluation 1] In particular, the prediction accuracy of the sensitivity of the purification performance according to the durability time of the durability history of the durability pattern A was evaluated. In this case, as the exhaust gas purification catalyst to be measured, a high-performance catalyst (honeycomb substrate volume: 1.1 L, catalyst metal loading density: high) and a low-performance catalyst (honeycomb substrate volume: 0.8 L, catalyst metal loading density: low) that were used in learning and were assumed to have relatively high and low purification performance, respectively, were prepared. In contrast, as the exhaust gas purification catalyst to be predicted, a medium-performance catalyst (honeycomb substrate volume: 0.9 L, catalyst metal loading density: high) that was not used in learning and was assumed to have relatively medium purification performance was prepared. Then, by actual measurement, the purification performance sensitivity according to the durability time of the durability history of the durability pattern A was obtained for the high-performance catalyst and the low-performance catalyst. Specifically, for the high-performance catalyst and the low-performance catalyst, a purification performance evaluation test under specified test conditions was performed after each durability time in the durability pattern A, and the T50 (NOx) sensitivity according to the durability time of the durability pattern A was obtained. In response to this, the machine learning models of the examples and comparative examples were used to predict the sensitivity of the purification performance according to the durability time of the durability history of the durability pattern A for the medium performance catalyst. Specifically, a group consisting of information expressing the catalyst configuration of the medium performance catalyst, information expressing the durability history at each durability time of the durability pattern A, and information expressing the purification performance evaluation test under the above-mentioned predetermined test conditions was obtained as explanatory variables at each durability time of the durability pattern A. Then, using the machine learning model, information expressing the purification performance at each durability time of the durability pattern A (objective variable) was obtained from the explanatory variables at each durability time of the durability pattern A, and the T50 (NOx) at each durability time of the durability pattern A was calculated to predict the sensitivity of T50 (NOx) according to the durability time of the durability pattern A. As shown in FIG. 3(a), the line showing the predicted result of the sensitivity of the medium performance catalyst by the machine learning model of the example passes within the expected range between the line showing the actual measurement result of the sensitivity of the high performance catalyst and the line showing the actual measurement result of the sensitivity of the low performance catalyst, regardless of the durability time.

[0016] [Prediction accuracy evaluation 2] In particular, the prediction accuracy of the sensitivity of the purification performance according to the durability time of the durability history of the durability pattern B was evaluated. In this case, as the exhaust gas purification catalyst to be measured, a high-performance catalyst (honeycomb substrate volume: 0.9 L, catalyst metal loading density: high) and a medium-performance catalyst (honeycomb substrate volume: 0.9 L, catalyst metal loading density: medium) that were used in learning and were assumed to have relatively high and medium purification performance, respectively, were prepared. In contrast, as the exhaust gas purification catalyst to be predicted, a low-performance catalyst (honeycomb substrate volume: 0.8 L, catalyst metal loading density: low) that was not used in learning and was assumed to have relatively low purification performance was prepared. Then, by actual measurement, the purification performance sensitivity according to the durability time of the durability history of the durability pattern B was obtained for the high-performance catalyst and the medium-performance catalyst. Specifically, for the high-performance catalyst and the medium-performance catalyst, a purification performance evaluation test under specified test conditions was performed after each durability time in the durability pattern B, and the T50 (NOx) sensitivity according to the durability time of the durability pattern B was obtained. In response to this, the machine learning models of the examples and comparative examples were used to predict the sensitivity of the purification performance according to the durability time of the durability history of the durability pattern B for the low-performance catalyst. Specifically, a group consisting of information expressing the catalyst configuration of the low-performance catalyst, information expressing the durability history at each durability time of the durability pattern B, and information expressing the purification performance evaluation test under the above-mentioned predetermined test conditions was obtained as explanatory variables at each durability time of the durability pattern B. Then, using the machine learning model, information expressing the purification performance at each durability time of the durability pattern B (objective variable) was obtained from the explanatory variables at each durability time of the durability pattern B, and the T50(NOx) at each durability time of the durability pattern B was calculated to predict the sensitivity of T50(NOx) according to the durability time of the durability pattern B. As shown in FIG. 3(b), in the prediction result of the sensitivity of the low-performance catalyst by the machine learning model of the example, as expected from conventional knowledge, the performance deteriorates with the passage of the durability time at a slope even greater than the actual measurement result of the sensitivity of the medium-performance catalyst.

Claims

[Claim 1] A machine learning model for predicting the purification performance of an exhaust gas purification catalyst, The machine learning model is constructed by performing machine learning using information having information expressing a catalyst configuration as an explanatory variable and information expressing a purification performance as a target variable, In the machine learning, a model is learned a predetermined number of times, and the model after the predetermined number of learnings is stored as a provisional model, and then update learning is repeatedly performed; This machine learning model is characterized in that, in the update learning, the provisional model is learned once, and it is determined whether the sensitivity of the purification performance according to the amount of catalytic metal and durability time predicted using the model after the first learning satisfies a sensitivity condition.If the sensitivity condition is satisfied, the model after the first learning is overwritten and saved as the provisional model, and if the sensitivity condition is not satisfied, the model after the first learning is not overwritten and saved as the provisional model.

Images