Cleaning performance prediction device

The device enhances the prediction of exhaust gas purification catalyst performance by using an autoencoder to compress and expand durability history data, improving the accuracy of predicting sensitivity to temperature and time variations.

JP2024165171AActive Publication Date: 2024-11-28TOYOTA JIDOSHA KK

Patent Information

Application Number
JP2023081090
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-05-16
Publication Date
2024-11-28
Estimated Expiration
2043-05-16

AI Technical Summary

Technical Problem

Conventional purification performance prediction devices for exhaust gas purification catalysts lack accuracy in predicting the sensitivity of purification performance to durability temperature due to the use of durability history data collected under predetermined conditions, limiting the ability to accurately model the catalyst's performance under varying conditions.

Method used

A purification performance prediction device that utilizes an autoencoder to dimensionally compress durability history data, expanding it in terms of durability temperature and time, and employs a machine learning model to predict purification performance based on latent variable data, incorporating catalyst configuration and evaluation test information as explanatory variables.

Benefits of technology

The device achieves high-accuracy prediction of the sensitivity of purification performance to durability temperature, enabling precise estimation of catalyst performance post-durability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024165171000001_ABST
    Figure 2024165171000001_ABST
Patent Text Reader

Abstract

To provide a cleaning performance prediction device predicting highly accurately sensitivity of cleaning performance with respect to an endurance temperature of catalyst.SOLUTION: A cleaning performance prediction device predicts cleaning performance of a catalyst after endurance based on an endurance history, and includes: a dimensional compression part acquiring data of potential variables by dimensionally compressing endurance history data by an auto encoder; and a prediction part acquiring information presenting the cleaning performance from an information group including data of potential variables by using a prediction model. The auto encoder is an auto encoder having subjected to machine learning using a data set including original data acquired from condition data of an endurance test as the endurance history data and expansion data acquired by expanding data in dimensions of endurance temperature and endurance time. The prediction model is a learning model having subjected to machine learning with information group including data of potential variables used as explanatory variables and setting information presenting the cleaning performance used as target variables.SELECTED DRAWING: Figure 3
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to a purification performance prediction device for predicting the purification performance of an exhaust gas purification catalyst. [Background technology]

[0002] Conventionally, a device that predicts the purification performance of an exhaust gas purification catalyst using a prediction model based on machine learning has been used as a purification performance prediction device. Purification performance is determined by catalyst configuration, durability history, purification performance evaluation test conditions, etc., so the prediction model predicts purification performance by acquiring information (objective variable) that expresses purification performance from information (explanatory variable) that expresses catalyst configuration, durability history, purification performance evaluation test conditions, etc. As such a prediction device, for example, a device that estimates purification performance by applying information about a catalyst that purifies harmful substances in an exhaust flow path to a catalyst state estimation model constructed by machine learning is known (Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2021-76112 Summary of the Invention [Problem to be solved by the invention]

[0004] In the conventional purification performance prediction device, as information expressing the durability history to be used as an explanatory variable when the prediction model is machine-learned, for example, durability history data such as a durability history 2D histogram expressing the frequency of a combination of an air-fuel ratio and an endurance temperature (catalyst bed temperature) obtained from the condition data of the durability test may be used as a latent variable data obtained by dimensional compression using an autoencoder. In this case, it is possible to predict purification performance under various conditions such as the endurance time and endurance temperature of the durability history using the prediction model. However, conventionally, the set of durability history data used in machine learning is a set of data obtained in a durability test performed under predetermined conditions assuming a predetermined deterioration of a vehicle, and therefore the endurance temperature is a set of data in which the endurance temperature is discrete. For this reason, when predicting the purification performance of an exhaust gas purification catalyst using a prediction model machine-learned using a set of latent variable data obtained by dimensionally compressing such a set of durability history data using an autoencoder as a data set of explanatory variables, the sensitivity of the purification performance to the endurance temperature could not be predicted with high accuracy.

[0005] The present invention has been made in consideration of the above-mentioned points, and an object of the present invention is to provide a purification performance prediction device that can predict with high accuracy the sensitivity of the purification performance of an exhaust gas purification catalyst to its endurance temperature. [Means for solving the problem]

[0006] In order to solve the above problem, the purification performance prediction device of the present invention is a purification performance prediction device that predicts the purification performance of an exhaust gas purification catalyst after durability testing from a durability history, and includes: a dimensional compression unit that acquires data of a latent variable that represents the durability history by dimensionally compressing the durability history data using an autoencoder that has been machine-learned to reproduce the input data using durability history data that represents the durability history; and a prediction unit that acquires information that represents the purification performance from an information group including the data of the latent variables by using a prediction model, wherein the autoencoder is an autoencoder that has been machine-learned using a data set after extension of the durability history data, the data set including, as the durability history data, original data acquired from condition data of a durability test and extended data acquired by data extension of the original data in the dimensions of durability temperature and durability time, and the prediction model is a learning model that has been machine-learned using an information group including the data of the latent variables as explanatory variables and information that represents the purification performance as a target variable. Effect of the Invention

[0007] According to the purification performance prediction device of the present invention, it is possible to predict with high accuracy the sensitivity of the purification performance of an exhaust gas purification catalyst to the endurance temperature. [Brief description of the drawings]

[0008] [Figure 1] FIG. 1A is a diagram for explaining a prediction device according to an embodiment, and FIG. 1B is a diagram for showing a configuration of a computer that realizes the prediction device according to an embodiment. [Diagram 2] (a) 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. (b) is an explanatory diagram of a method for acquiring durability history data that is dimensionally compressed into data of latent variables expressing durability history. (c) is an explanatory diagram of a method for acquiring information expressing a purification performance evaluation test and information expressing purification performance. [Diagram 3](a) is a schematic diagram for explaining data expansion of original data of a durability history 2D histogram according to an embodiment. (b1) and (b2) are graphs plotting the number of durability history data before and after data expansion using a pseudo durability temperature. (c1) and (c2) are graphs plotting the number of durability history data before and after data expansion using a pseudo durability time. [Figure 4] 1A is an explanatory diagram of dimensionality reduction by a convolutional autoencoder according to Example 1 and Comparative Example 1. FIG. 1B is an explanatory diagram of dimensionality reduction by a flattened autoencoder according to Example 2 and Comparative Example 2. [Diagram 5] 13 is a graph in which N pieces of data of latent variables obtained by dimensionally compressing the N pieces of histogram data obtained in Example 1 using the adopted models of the autoencoders according to Examples 1 and 2 and Comparative Examples 1 and 2 are plotted on XY coordinates. [Figure 6] 1 is a graph showing the results of prediction of changes in NOx-T50 with respect to the average endurance temperature in durability histories A and B by the prediction devices of Examples 1 and 2 and Comparative Examples 1 and 2. [Figure 7] 1A is a two-dimensional graph showing the results of predictions of the change in NOx-T50 versus the endurance time and average endurance temperature of durability history A using the prediction devices of Examples 1 and 2 and Comparative Examples 1 and 2, and FIG. 1B is a three-dimensional graph showing the same results. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] Hereinafter, an embodiment of the purification performance prediction device of the present invention will be described. First, the outline of the purification performance prediction device (hereinafter, sometimes abbreviated as "prediction device") according to the embodiment will be described by taking the prediction device according to one embodiment as an example. Fig. 1(a) is a diagram for explaining the prediction device according to one embodiment, and Fig. 1(b) is a diagram for showing the configuration of a computer that realizes the prediction device according to one embodiment.

[0010] The prediction device 1 according to an embodiment shown in FIG. 1(a) is a prediction device that predicts the purification performance of an exhaust gas purification catalyst after durability testing from its durability history. The exhaust gas purification catalyst (hereinafter, sometimes abbreviated as "catalyst") is, for example, a catalyst shown in FIG. 2(a) described later, and includes a catalyst layer having a first layer (upper layer) and a second layer (lower layer) including a plurality of types of powder (material). The prediction device 1 includes an input unit 2, a storage unit 4, a processing unit 6 including a dimensional compression unit 6a, a prediction unit 6b, and a calculation unit 6c, and an output unit 8. The prediction device 1 is realized, for example, by a computer 100 shown in FIG. 1(b). The computer 100 includes a processing device (calculator) 110, a storage device 120, an input device 130, an output device 140, an input / output interface (I / F) 150, and the like.

[0011] The input unit 2 of the prediction device 1 inputs information expressing the catalyst configuration, durability history data expressing the durability history, and information expressing the purification performance evaluation test. The storage unit 4 stores an autoencoder 4a, a prediction model 4b, and other data 4c. The other data 4c is various data used in the prediction device 1, and includes, for example, information expressing the catalyst configuration, durability history data expressing the durability history, and information expressing the purification performance evaluation test.

[0012] The information expressing the catalyst configuration includes information on each layer of the catalyst layer and information on the honeycomb substrate. The information on each layer of the catalyst layer includes the amount of each type of multiple types of powder contained in each layer of the catalyst layer, the amount of each type of multiple types of catalytic metal supported on each type of multiple types of powder, and data on the coating region. The information on the honeycomb substrate includes data on the weight, cell shape, and wall thickness. The durability history data expressing the durability history is not particularly limited as long as it is data expressing the durability history that the catalyst has undergone, and for example, a durability history 2D histogram expressing the frequency (durability history) of the combination of air-fuel ratio and durability temperature (catalyst bed temperature) obtained from the condition data of the durability test performed on the catalyst can be mentioned. The information expressing the purification performance evaluation test is the condition data of the purification performance evaluation test performed on the catalyst.

[0013] The dimensional compression unit 6a of the processing unit 6 uses the durability history data as input data, and the autoencoder 4a, which has been machine-learned to reproduce the input data, compresses the durability history data to obtain data of latent variables representing the durability history. The autoencoder 4a is an autoencoder that has been machine-learned using a data set after extension of the durability history data, including original data obtained from the condition data of the durability test and extended data obtained by data extension of the original data in the dimensions of endurance temperature and endurance time, as the durability history data. The extended data obtained by data extension of the original data in the dimensions of endurance temperature and endurance time is not particularly limited, but is preferably, for example, extended data obtained by data extension of the original data of the durability history 2D histogram (image data of A pixels x B pixels (A and B are predetermined integers of 2 or more)) by the pseudo endurance temperature and pseudo endurance time. Examples of such extended data include extended data of one or more levels of pseudo-endurance temperature, each of which is obtained by one or more processes in which a predetermined percentage (e.g., 1 / 10) of the pixel value of each pixel of the original data of the durability history 2D histogram is moved to an adjacent pixel on the lower or higher temperature side, and extended data of two or more levels of pseudo-endurance time, each of which is obtained by one or more processes in which the pixel value of each pixel of the original data and the extended data of the pseudo-endurance temperature is lowered or increased by a predetermined percentage (e.g., 1 / 50).

[0014] The prediction unit 6b of the processing unit 6 uses the prediction model 4b to obtain information expressing the purification performance (objective variable) from a group of information (explanatory variables) including information expressing the catalyst configuration, data of latent variables expressing the durability history, and information expressing the purification performance evaluation test. The information expressing the purification performance is data per second of the temperature at the inlet of the catalyst, the NOx concentration at the inlet of the catalyst, and the NOx concentration at the outlet of the catalyst in the purification performance evaluation test.

[0015] The prediction model 4b is a learning model that has been machine-learned using a data set consisting of multiple sets of explanatory variables and multiple sets of data on explanatory variables and multiple target variables obtained for multiple learning samples of the catalyst.

[0016] The calculation unit 6c of the processing unit 6 calculates the NOx 50% purification temperature (NOx-T50) from the information representing the purification performance acquired by the prediction unit 6b. The output unit 8 outputs the information representing the purification performance and the calculation result (NOx-T50). This allows the purification performance of the catalyst after durability testing to be predicted.

[0017] In the prediction device according to one embodiment, the autoencoder 4a is an autoencoder that has been machine-trained using a data set of the extended durability history data, including original data acquired from the condition data of the durability test and extended data acquired by data extension of the original data in the dimensions of the endurance temperature and the endurance time, as the durability history data. Therefore, in the dimension compression unit 6a of the processing unit 6, the data of the latent variables expressing the durability history acquired by dimensionally compressing the durability history data by such an autoencoder 4a can accurately express the endurance temperature of the durability history. As a result, in the dimension compression unit 6a of the processing unit 6, by using the prediction model 4b, when the purification performance of the catalyst after durability is predicted by acquiring information (objective variable) expressing the purification performance from an information group (explanatory variable) including data of the latent variable expressing the durability history, the sensitivity of the purification performance to the endurance temperature of the catalyst can be predicted with high accuracy.

[0018] Next, the prediction device according to the embodiment will be described in further detail. The autoencoder that compresses the dimension of the durability history data in the dimension compression unit of the prediction device is not particularly limited, but for example, a convolutional autoencoder or the like is preferable, because it can predict the sensitivity of the purification performance to the durability temperature of the catalyst with particularly high accuracy.

[0019] The prediction model used in the prediction unit of the prediction device is not particularly limited as long as it is a learning model in which a group of information including data of latent variables representing durability history is used as an explanatory variable and information representing purification performance is used as an objective variable, and machine learning is performed using the model. For example, as in one embodiment, a group of information including information representing the catalyst configuration, data of latent variables representing durability history, and information representing purification performance evaluation test is used as an explanatory variable. The prediction unit of the prediction device is not particularly limited as long as it uses a prediction model to obtain information representing purification performance (objective variable) from a group of information including data of latent variables representing durability history. For example, as in one embodiment, a prediction model is used to obtain information representing purification performance from a group of information including information representing the catalyst configuration, data of latent variables representing durability history, and information representing purification performance evaluation test. EXAMPLES

[0020] Hereinafter, the prediction device according to the embodiment will be described more specifically with reference to examples and comparative examples.

[0021] [Example 1] An example of a prediction device according to an embodiment was manufactured as the prediction device of Example 1. The manufacturing procedure of the prediction device of Example 1 will be described below.

[0022] 1. Obtaining a data set of explanatory variables and target variables First, a data set of explanatory variables and objective variables used to construct a prediction model was obtained from a learning sample of a catalyst. The method of obtaining the data is described below.

[0023] (1) Preparation of catalyst study samples First, S (S: 800) study samples of the catalyst were prepared. FIG. 2(a) is a schematic cross-sectional view showing an enlarged partition wall and catalyst layer of the substrate in each study sample of the catalyst. As shown in FIG. 2(a), each study sample CS is a straight-flow type catalyst, and includes a honeycomb substrate 10 and a catalyst layer 20. The honeycomb substrate 10 is a substrate in which a cylindrical frame (not shown) and partition walls 14 that divide the space inside the frame into a honeycomb shape are integrally formed. The partition walls 14 are porous bodies that define a plurality of cells 12 extending from the inlet end face to the outlet end face of the honeycomb substrate 10. The catalyst layer 20 is provided on the cell 12 side surface 14c of the partition walls 14 (substrate 10), and includes at least the first layer 20a and the second layer 20b of the first layer 20a to the fifth layer 20e. The n-th layer to the 1st layer (n is an integer of 2 to 5) are laminated in this order on the surface 14c of the partition wall 14 on the cell 12 side. Each of the n-th layer to the 1st layer contains powders P1 to Px (x is a predetermined integer of 2 or more) and contains a catalytic metal supported by at least one of the contained powders. The powders P1 to Px are types of powder of a support or a promoter (for example, an OSC material) and include AZ (alumina-zirconia binary composite oxide) species and alumina species. The AZ species and the alumina species are types belonging to AZ and types belonging to alumina, respectively. The catalytic metal is at least one of catalytic metals M1 to My (y is a predetermined integer of 2 or more). The catalytic metals M1 to My are types of catalytic metals and include Pd species and Rh species. The Pd species and the Rh species are types belonging to Pd and Rh, respectively.

[0024] (2) Obtaining explanatory variables and objective variables for each learning sample of the catalyst Next, for each learning sample of the catalyst, a group of information including information expressing the catalyst configuration, data of latent variables expressing the durability history, and information expressing the purification performance evaluation test was acquired as explanatory variables, and information expressing the purification performance was acquired as a target variable. FIG. 2(b) is an explanatory diagram of a method of acquiring durability history data that is dimensionally compressed into a latent variable expressing the durability history. FIG. 2(c) is an explanatory diagram of a method of acquiring information expressing the purification performance evaluation test and information expressing the purification performance. The following description will be given with reference to these figures.

[0025] (2-1) Acquisition of information describing catalyst structure From each learning sample of the catalyst, information was obtained that contains information about each layer of the catalyst layer and information about the honeycomb substrate as information expressing the catalyst configuration, as shown in Table 1 below. The information about each layer of the catalyst layer includes the amount [g] of each type of powder P1-Px and the amount [g] of each type of catalytic metal M1-My supported on each type of powder P1-Px, as well as data about the coating region. Data (six dimensions) about the weight, cell shape, and wall thickness was obtained as information about the honeycomb substrate. The data about the coating region includes the coating width and coating position of each layer of the catalyst layer.

[0026] [Table 1]

[0027] (2-2) Acquisition of data on latent variables representing durability history a. Durability test Durability tests were conducted on each learning sample of the catalyst. In each durability test, the catalyst was installed in the left and right banks of the exhaust system of a 4300cc V8 gasoline engine, and the endurance temperature (catalyst bed temperature), endurance time, and operating pattern were set under predetermined conditions to deteriorate the catalyst. Then, as shown in Figure 2(b), the endurance temperature throughout the entire test and the air-fuel ratio (A / F sensor value) with the catalyst were measured during each durability test, and these were obtained as condition data for each durability test.

[0028] b. 2D histogramming As shown in Figure 2(b), original data (durability history data) for a durability history 2D histogram expressing the durability history (frequency of combinations of air-fuel ratio and durability temperature (catalyst bed temperature)) of each learning sample of the catalyst was obtained from the condition data of each durability test. The air-fuel ratio was in the range of 12 to 18 with a bin of 0.2. The durability temperature was in the range of 700°C to 1090°C with a bin of 13°C. In this way, 30 pixel x 30 pixel image data was obtained as the original data for the durability history 2D histogram, with the pixel value of each pixel expressing the frequency of combinations of air-fuel ratio and durability temperature.

[0029] c. Data augmentation of the original data of the durability history 2D histogram As shown in Fig. 2(b), the original data of the durability history 2D histogram of each learning sample of the catalyst was expanded in the dimensions of durability temperature and durability time. Fig. 3(a) is a schematic diagram illustrating the data expansion of the original data of the durability history 2D histogram according to the embodiment.

[0030] In data expansion in the dimensions of endurance temperature and endurance time, data expansion was performed using pseudo endurance temperature and pseudo endurance time, as shown in Figure 3(a). In data expansion using pseudo endurance temperature, low-temperature processing was performed 49 times, in which 1 / 10 of the pixel value of each pixel of the original data was moved to an adjacent pixel on the low-temperature side, and the data of the endurance history 2D histogram obtained each time was treated as one level of expanded data, generating expanded data of 49 levels of pseudo endurance temperature. Figures 3(b1) and (b2) are graphs plotting the number of durability history data with the same average endurance temperature before and after data expansion using pseudo endurance temperature, and the combinations of those average endurance temperatures. In the data expansion by pseudo endurance time, the original data of the durability history 2D histogram of one level and the expanded data of the pseudo endurance temperature of 49 levels were subjected to a time shortening process of reducing the pixel value of each pixel by 1 / 50 for each data of 50 levels in total, and the data of the durability history 2D histogram obtained each time was treated as the expanded data of one level, thereby generating the expanded data of 49 levels of pseudo endurance time. Note that Fig. 3 (c1) and (c2) are graphs plotting the number of durability history data with the same durability time before and after the data expansion by pseudo endurance time and the combination of the durability times. As a result, the data set after the expansion of the durability history 2D histogram of each learning sample of the catalyst was obtained as 2500 levels (50 levels x 50 levels) of durability history 2D histogram data including the original data of one level and the expanded data (durability history data) of 2499 levels. This resulted in obtaining N pieces of durability history 2D histogram data (S pieces x 2500 levels) as an expanded data set of S durability history 2D histograms, each of which represents the durability history of S learning samples of the catalyst.

[0031] d. Dimensionality reduction using autoencoder The data of N durability history 2D histograms was subjected to dimensional compression by a convolutional autoencoder. Fig. 4(a) is an explanatory diagram of dimensional compression by a convolutional autoencoder according to Example 1 and Comparative Example 1 described later.

[0032] In this case, we first constructed a convolutional autoencoder. As shown in Figure 4(a), the convolutional autoencoder takes each data (30 pixel x 30 pixel image data) of N durability history 2D histogram data as input data, and processes the input data in the first convolutional layer, the first pooling layer, the second convolutional layer, and the second pooling layer in order. Specifically, the first convolutional layer obtains (30 pixel x 30 pixel image data) x 4 images from the input data by filtering (feature extraction). The first pooling layer obtains (8 pixel x 8 pixel image data) x 4 images from the data obtained in the first convolutional layer by pooling (image compression). The second convolutional layer obtains (8 pixel x 8 pixel image data) x 4 images from the data obtained in the first pooling layer by filtering. The second pooling layer obtains (2 pixel x 2 pixel image data) x 4 images from the data obtained in the second convolutional layer by pooling. Therefore, in a convolutional autoencoder, the features of the input data are compressed in dimension, and ultimately extracted as latent variable (16-dimensional) data ((2 pixels x 2 pixels) x 4 images).

[0033] Next, the data of N durability history 2D histograms was divided into a teacher data set and a test data set. Then, for the convolutional autoencoder, while verifying the accuracy of the autoencoder with the test data set, the teacher data set was used to machine-train the autoencoder to reproduce the input data using the durability history 2D histogram data (durability history data). At this time, the autoencoder was tuned so that the mean squared logarithmic error was minimized for the test data set and the teacher data set. In addition, PReLU was used as the activation function, rmsprop was used as the optimization function, and the iteration was set to 10,000 times. Then, the convolutional autoencoder model at the iteration when the mean squared logarithmic error of the test data was minimized and the correlation coefficient was closest to 1 was adopted.

[0034] Next, as shown in Figure 4(a), the model adopted for the convolutional autoencoder was used to compress the dimensions of each data item in the N durability history 2D histogram data items, thereby obtaining N pieces of data ((2 pixels x 2 pixels) x 4 sheets x N items) of latent variable (16 dimensions) representing the durability history. Next, from among the N pieces of latent variable data items, S pieces of latent variable data items in which the original data items in the S durability history 2D histogram data items were compressed in dimensions were selected as S pieces of data items of latent variable representing the durability history of the S learning samples of the catalyst. In this way, latent variable (16 dimensions) data representing the durability history included in the explanatory variables was obtained for each learning sample of the catalyst.

[0035] (2-3) Acquisition of information expressing purification performance evaluation tests and information expressing purification performance Purification performance evaluation tests were conducted on each learning sample of the catalyst deteriorated in each durability test, and information expressing the purification performance evaluation test and information expressing the purification performance of each learning sample of the catalyst were obtained. Specifically, a heat exchanger and a deteriorated catalyst were installed in the exhaust system of a gasoline engine with a displacement of 2000cc. Then, the engine output was kept constant, and the temperature at the catalyst inlet was raised from 200°C at 10°C / min to 600°C by controlling the heat exchanger. The condition data of this purification performance evaluation test, specifically, the data per second of the exhaust gas flow rate, the temperature at the catalyst inlet, the CO concentration at the catalyst inlet, the NOx concentration at the catalyst inlet, the HC concentration at the catalyst inlet, and the air-fuel ratio (A / F sensor value) at the catalyst inlet were obtained as information expressing the purification performance evaluation test. In addition, the data per second of the temperature at the catalyst inlet, the NOx concentration at the catalyst inlet, and the NOx concentration at the catalyst outlet in this purification performance evaluation test were obtained as information expressing the purification performance. Furthermore, by using information expressing the purification performance, as shown in Figure 2(c), the NOx purification amount can be calculated by subtracting the NOx concentration out of the catalyst from the NOx concentration in the catalyst, and the temperature in the catalyst when the NOx purification amount reaches half the NOx concentration in the catalyst can be calculated as the NOx 50% purification temperature (NOx-T50).

[0036] (3) Obtaining a data set of explanatory variables and target variables For S training samples of catalysts, the explanatory variables and response variables of each training sample of the catalyst were obtained as described above, thereby obtaining a dataset consisting of S sets of data for the explanatory variables and response variables.

[0037] 2. Building a predictive model Next, a predictive model was constructed by machine learning using a data set of explanatory variables and objective variables.

[0038] First, the data set was randomly divided into three: a training data set, a test data set, and a validation data set. Next, a multilayer perceptron model with two hidden layers was constructed. At this time, a model was constructed in which values ​​were randomly 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, for the multilayer perceptron model, machine learning was performed on the training data set while verifying the accuracy of the model with the test data set. At this time, the model was machine-learned so that the RMSE (Root Mean Squared Error) was small for the training data set and the test data set. In addition, the iteration was set to 1000 times. Next, predictions were made on the validation data set using the learned model, and the RMSE with the actual measured value was calculated. The series of operations of constructing the model, learning, and calculating the RMSE for the validation data set using the learned model were repeated 100 times by changing the values ​​given to the hyperparameters. As a result, a data set consisting of 100 sets of data for the hyperparameters and RMSE was obtained.

[0039] Next, the values ​​of the hyperparameters that minimized the RMSE were selected from the data set consisting of 100 sets of data for the hyperparameters and RMSE obtained as described above. Next, a multilayer perceptron model was constructed in which the selected values ​​were given to the hyperparameters, and this was used as the prediction model.

[0040] [Example 2] The procedure for producing the prediction device of the second embodiment will be described below.

[0041] 1. Obtaining a data set of explanatory variables and target variables First, a data set of explanatory variables and objective variables used to construct a prediction model was obtained from a learning sample of a catalyst. In this case, a data set consisting of S sets of data of explanatory variables and objective variables was obtained by the same method as that described in "1." of [Example 1], except for the method of obtaining data of latent variables expressing the durability history of explanatory variables.

[0042] In the method of acquiring data of latent variables expressing durability history, data of latent variables (16 dimensions) expressing durability history for each learning sample of catalyst was acquired by the same method as that described in the section "1.(2)(2-2)" of [Example 1], except that data of N durability history 2D histograms was dimensionally compressed by a Flatten autoencoder instead of a convolutional autoencoder. Fig. 4(b) is an explanatory diagram of dimensionality compression by a Flatten autoencoder according to Example 2 and Comparative Example 2 described later.

[0043] When the data of N durability history 2D histograms was dimensionally compressed by the flattened autoencoder, a flattened autoencoder was first constructed. As shown in Figure 4(b), the flattened autoencoder takes each data (30 pixel x 30 pixel image data) of N durability history 2D histograms as input data, and performs flattening and dimensionality compression on the input data in order. Specifically, in the flattening process, each data is converted into 900 columns x 1 row data, and N pieces of data are converted into 900 columns x N rows of data. In the dimensionality compression process, 16 columns x 1 row data is obtained from the 900 columns x 1 row data, and 16 columns x N rows of data is obtained from the 900 columns x N row data. Therefore, in the flattened autoencoder, the features of the input data are dimensionally compressed, and finally extracted into latent variable (16-dimensional) data (16 columns).

[0044] Next, the data of the N durability history 2D histograms was divided into a teacher data set and a test data set, and the accuracy of the flattened autoencoder was verified using the test data set while the data of the durability history 2D histograms (durability history data) was used as input data to machine-learn the autoencoder so as to reproduce the input data using the teacher data set. In this case, the autoencoder was machine-learned using a method similar to that described in item "d." of "1.(2)(2-2)" in [Example 1], and a model of the flattened autoencoder in the iteration when the mean squared logarithmic error of the test data was minimized and the correlation coefficient was closest to 1 was adopted.

[0045] Next, as shown in Figure 4(b), the adopted model of the Flatten autoencoder was used to compress the dimensions of each data of the N durability history 2D histogram data, thereby obtaining N data (16 columns x N rows) of latent variables (16 dimensions) representing the durability history. Next, from among the N data of latent variables, S data of latent variables obtained by compressing the dimensions of the original data of the S durability history 2D histograms were selected as S data of latent variables representing the durability history of the S learning samples of the catalyst. In this way, data of latent variables (16 dimensions) representing the durability history included in the explanatory variables was obtained for each learning sample of the catalyst.

[0046] 2. Building a predictive model Next, a predictive model was constructed in a manner similar to that described in “2.” of [Example 1], except that the data sets of explanatory variables and dependent variables obtained in “1.” of [Example 2] were used as the data sets of explanatory variables and dependent variables.

[0047] [Comparative Example 1] The procedure for producing the prediction device of Comparative Example 1 will be described below.

[0048] 1. Obtaining a data set of explanatory variables and target variables First, a data set of explanatory variables and objective variables used to construct a prediction model was obtained from a learning sample of a catalyst. In this case, a data set consisting of S sets of data of explanatory variables and objective variables was obtained by the same method as that described in "1." of [Example 1], except for the method of obtaining data of latent variables expressing the durability history of explanatory variables.

[0049] In the method of acquiring data of latent variables expressing durability history, original data of durability history 2D histogram expressing durability history of each learning sample of the catalyst was acquired by the method described in the items "a." and "b." of "1.(2)(2-2)" of [Example 1]. Then, unlike the method described in the item "c." of "1.(2)(2-2)" of [Example 1], data expansion was performed only by pseudo durability time for the original data of durability history 2D histogram of each learning sample of the catalyst. In the data expansion only by pseudo durability time, a time shortening process was performed 49 times for the original data of one level of durability history 2D histogram, which reduces the pixel value of each pixel by 1 / 50, and the data of the durability history 2D histogram acquired each time was used as one level of expanded data, thereby generating expanded data of 49 levels of pseudo durability time. As a result of the above, 50 levels of durability history 2D histogram data, including 1 level of original data and 49 levels of expanded data, were obtained as a data set after expanding the durability history 2D histograms of each learning sample of the catalyst. In this way, M pieces of (S pieces × 50 levels) of durability history 2D histogram data were obtained as a data set after expanding S durability history 2D histograms each representing the durability history of the S learning samples of the catalyst.

[0050] Then, the data of the M durability history 2D histograms was dimensionally compressed by a convolutional autoencoder. In this case, first, the convolutional autoencoder described in the item "d." of "1.(2)(2-2)" of [Example 1] was constructed. Next, the data of the M durability history 2D histograms was divided into a teacher data set and a test data set, and the convolutional autoencoder was machine-learned by the method described in the item "d." of "1.(2)(2-2)" of [Example 1]. Then, the model of the convolutional autoencoder in the iteration when the mean square logarithmic error of the test data was minimized and the correlation coefficient was closest to 1 was adopted. Next, as shown in FIG. 4(a), the adopted model of the convolutional autoencoder was used to dimensionally compress each data of the M durability history 2D histograms, thereby obtaining M data ((2 pixels x 2 pixels) x 4 sheets x M pieces) of latent variables (16 dimensions) expressing the durability history. Next, from among the M pieces of latent variable data, S pieces of latent variable data obtained by dimensionally compressing the original data of the S durability history 2D histograms were selected as S pieces of latent variable data expressing the durability history of the S learning samples of the catalyst. In this way, for each learning sample of the catalyst, data of the latent variable (16 dimensions) expressing the durability history included in the explanatory variables was obtained.

[0051] 2. Building a predictive model Next, a predictive model was constructed in the same manner as described in "2." of [Example 1], except that the data sets of explanatory variables and dependent variables obtained in "1." of [Comparative Example 1] were used as the data sets of explanatory variables and dependent variables.

[0052] [Comparative Example 2] The procedure for producing the prediction device of Comparative Example 2 will be described below.

[0053] 1. Obtaining a data set of explanatory variables and target variables First, a data set of explanatory variables and objective variables used to construct a prediction model was obtained from a learning sample of a catalyst. In this case, a data set consisting of S sets of data of explanatory variables and objective variables was obtained by the same method as that described in "1." of [Comparative Example 1], except for the method of obtaining data of latent variables expressing the durability history of explanatory variables.

[0054] The method for obtaining data on latent variables representing durability history was similar to that described in section "1." of [Comparative Example 1], except that the data on M durability history 2D histograms was dimensionally compressed using a Flattened autoencoder instead of a convolutional autoencoder, to obtain data on latent variables (16 dimensions) representing durability history for each learning sample of the catalyst.

[0055] When the data of M durability history 2D histograms was dimensionally compressed by the Flatten autoencoder, first, the Flatten autoencoder described in the section "1." of [Example 2] was constructed. Next, the data of M durability history 2D histograms was divided into a teacher data set and a test data set, and the Flatten autoencoder was machine-trained by the method described in the section "1." of [Example 2]. Then, the model of the Flatten autoencoder in the iteration when the mean square logarithmic error of the test data was minimized and the correlation coefficient was closest to 1 was adopted. Next, as shown in FIG. 4(b), the adopted model of the Flatten autoencoder was used to dimensionally compress each data of the data of M durability history 2D histograms, thereby obtaining M data (16 columns × M rows) of latent variables (16 dimensions) expressing durability history. Next, from among the M data of latent variables, S data of latent variables in which the original data of S durability history 2D histograms was dimensionally compressed was selected as S data of latent variables expressing the durability history of S learning samples of catalysts. This allowed us to obtain data on latent variables (16 dimensions) expressing the durability history contained in the explanatory variables for each learning sample of the catalyst.

[0056] 2. Building a predictive model Next, a predictive model was constructed in the same manner as described in “2.” of [Example 1], except that the data sets of explanatory variables and dependent variables obtained in “1.” of [Comparative Example 2] were used as the data sets of explanatory variables and dependent variables.

[0057] [evaluation] 1. Evaluation of the autoencoder model used to reduce the dimension of the durability history 2D histogram In the model using the autoencoder employed for dimensional compression of the durability history 2D histogram in Examples 1 and 2 and Comparative Examples 1 and 2, an evaluation was performed as to whether or not the latent variables can accurately express the durability time and durability temperature of the durability history. Specifically, the evaluation was performed by investigating how the data of the latent variables obtained by dimensionally compressing the data of the N durability history 2D histograms acquired in Example 1 changes in response to changes in the durability time and durability temperature of the durability history using the model using the autoencoder in each example.

[0058] In FIG. 5, the upper left graph is a graph in which N data of latent variables obtained in Example 1 by the adoption model of the convolutional autoencoder according to Example 1 are plotted on XY coordinates, with the data of N histograms being dimensionally compressed. The upper right graph is a graph in which N data of latent variables obtained in Example 1 by the adoption model of the Flatten autoencoder according to Example 2 are plotted on XY coordinates, with the data of the same N histograms being dimensionally compressed. The lower left graph is a graph in which N data of latent variables obtained in Example 1 by the adoption model of the convolutional autoencoder according to Comparative Example 1 are plotted on XY coordinates, with the data of the same N histograms being dimensionally compressed. The lower right graph is a graph in which N data of latent variables obtained in Example 1 by the adoption model of the Flatten autoencoder according to Comparative Example 2 are plotted on XY coordinates, with the data of the same N histograms being dimensionally compressed.

[0059] In each graph, a combination of two predetermined dimensions (hereinafter, sometimes referred to as "latent variable 1" and "latent variable 2") in a latent variable (16 dimensions) is plotted on the XY coordinate. In each graph, among all the plots of latent variables (plots of combinations of latent variables 1 and 2), the plots of latent variables in which the original data of a predetermined one durability history 2D histogram and each data of its extended data are dimension-compressed are classified into the "A" group. Among the plots of "A", the plots of latent variables in which the original data and each data of its extended data of pseudo durability time (average durability temperature: 925°C) are dimension-compressed are classified into the "Aa" group. Among the plots of "A", the plots of latent variables in which the extended data of pseudo durability temperature (extended data in which the pixel value of each pixel is translated by 2 pixels to the low temperature side) obtained by performing low-temperature processing of the original data 20 times and each data of its extended data of pseudo durability time (average durability temperature: 899°C) are dimension-compressed are classified into the "Ab" group. The extended data for pseudo endurance temperature obtained by performing low-temperature processing on the original data 30 times (extended data in which the pixel value of each pixel is shifted 3 pixels toward the low temperature side) and each data of the extended data for pseudo endurance time (average endurance temperature: 886°C) were dimension-compressed and plotted as a latent variable in the "Ac" group. The extended data for pseudo endurance temperature obtained by performing low-temperature processing on the original data 40 times (extended data in which the pixel value of each pixel is shifted 4 pixels toward the low temperature side) and each data of the extended data for pseudo endurance time (average endurance temperature: 873°C) were dimension-compressed and plotted as a latent variable in the "Ad" group. Plots of all latent variables that are not classified as group "A" are indicated by "P".

[0060] In the graph of Example 1 in FIG. 5, when comparing the plots of the latent variables in each group of "Aa" to "Ad", the position of the plots is continuously displaced from a position far from the origin of the coordinate system toward the origin. Then, when referring to the data of the durability history 2D histogram in which the feature amount is extracted for the latent variable, it was confirmed that the latent variable whose plot is closer to the origin has a shorter durability time represented in the data. When comparing the "Aa" to "Ad" groups, the "Aa" to "Ad" groups are arranged in order in the positive direction of the X-axis. Then, the group located closer to the positive direction of the X-axis has a lower durability temperature represented in the data in which the feature amount is extracted for the latent variable of the plot. From these facts, it is considered that the latent variables in the adopted model of Example 1 can accurately represent the continuous changes in the durability time and durability temperature of the durability history.

[0061] In the graph of Example 2 in FIG. 5, when comparing the plots of 50 latent variables in each group of "Aa" to "Ad", the positions of the plots are continuously displaced from a position far from the origin of the coordinates to a position approaching the origin. It was confirmed that the endurance time represented in the data is shorter for a latent variable whose plot is closer to the origin. On the other hand, when comparing the "Aa" to "Ad" groups, the "Aa" to "Ad" groups are not arranged in order in one direction, but are discretely arranged so as to surround the origin of the coordinates. From these facts, it is considered that in the adopted model of Example 2, the latent variables can accurately express the continuous change in the endurance time of the endurance history, but cannot sufficiently accurately express the continuous change in the endurance temperature of the endurance history. In the graphs of Comparative Examples 1 and 2 in FIG. 5, the "Aa" to "Ac" groups are arranged overlapping each other. It is considered that the adopted models of Comparative Examples 1 and 2 cannot accurately express the continuous change in the endurance temperature of the endurance history.

[0062] 2.Evaluation of prediction accuracy Using the prediction devices of Examples 1 and 2 and Comparative Examples 1 and 2, the prediction accuracy was evaluated when predicting the purification performance of a catalyst when the endurance time and endurance temperature of the endurance history of the catalyst were changed. A specific description will be given below.

[0063] (1) Preparing explanatory variable data for prediction First, as data on explanatory variables to be used when predicting purification performance when the durability time and durability temperature of the catalyst's durability history change, data on explanatory variables for durability history A (durability time: 100 hours), which is a representative durability history, and data on explanatory variables for durability history B (durability time: 50 hours), which is a representative durability history, were prepared.

[0064] When preparing the explanatory variable data of the durability history A, the input unit of the prediction device inputted information expressing a predetermined catalyst configuration, information expressing a predetermined purification performance evaluation test, and 2500 levels of data (durability history data) obtained by performing data expansion of the original data of the durability history 2D histogram of the durability history A by the pseudo durability temperature and pseudo durability time. At this time, the extension of the durability history 2D histogram was performed by the method described in the item "c." of "1. (2) (2-2)" of [Example 1]. Next, in the dimensional compression unit of the processing unit of the prediction device, the dimensional compression of each data of the 2500 levels of data (durability history data) of the durability history 2D histogram was performed by the adoption model of the autoencoder, thereby obtaining 2500 pieces of data of latent variables (16 dimensions) expressing the durability history. As a result, predetermined common information was used as the explanatory variable data of the durability history A as the information expressing the catalyst configuration and the information expressing the purification performance evaluation test, and 2500 pieces of data of explanatory variables were prepared in which 2500 pieces of data of latent variables were used as the data of the latent variable expressing the durability history A. Furthermore, when preparing the data on the explanatory variables of durability history B, a similar method was used in which predetermined common information was used as the information representing the catalyst configuration and the information representing the purification performance evaluation test as the data on the explanatory variables of durability history B, and 2500 pieces of data on explanatory variables were prepared, each of which used 2500 pieces of data on latent variables as data on latent variables representing durability history B.

[0065] (2) Prediction of the sensitivity of purification performance to the average endurance temperature of the endurance history Using the prediction devices of each of the examples of Examples 1 and 2 and Comparative Examples 1 and 2, the sensitivity of the purification performance of the catalyst to the average endurance temperature of the durability history A was predicted from the explanatory variable data of the durability history A. At this time, five pieces of explanatory variable data were selected from the explanatory variable data of the durability history A, in which five pieces of data with an endurance time of 50 hours and an average endurance temperature of 867°C to 919°C were used as latent variable data expressing the durability history. Then, in the prediction section of the processing section, by using a prediction model, information expressing the purification performance (objective variable) was obtained from each data of the selected explanatory variables. Furthermore, the calculation section of the processing section calculated NOx-T50 from the information expressing the purification performance. As a result, the change in NOx-T50 with respect to the average endurance temperature of the durability history A was predicted. Furthermore, using the prediction device of each example, the sensitivity of the purification performance with respect to the average endurance temperature of the durability history B was predicted from the explanatory variable data of the durability history B. At this time, from the data of explanatory variables of the durability history B, five data of explanatory variables were selected, in which five data of the durability time of 25 hours and the average durability temperature of 867°C to 919°C were used as data of latent variables expressing the durability history. Then, the prediction unit and the calculation unit of the processing unit obtained information expressing the purification performance (objective variable) from each data of the selected explanatory variables, and calculated NOx-T50. As a result, the change in NOx-T50 with respect to the average durability temperature of the durability history B was predicted. FIG. 6 is a graph showing the results of prediction of the change in NOx-T50 with respect to the average durability temperature of the durability history A and B by the prediction devices of Examples 1 and 2 and Comparative Examples 1 and 2.

[0066] As shown in FIG. 6, in the prediction results of both the durability history A and the durability history B by the prediction device of Example 1, the NOx-T50 generally monotonically decreases with the decrease in the average endurance temperature. These results show the inherent tendency that the purification performance improves with the decrease in the average endurance temperature. On the other hand, in the prediction results of the durability history A by the prediction device of Example 2, the NOx-T50 generally monotonically decreases with the decrease in the average endurance temperature, but in the prediction results of the durability history B by the prediction device of Example 2, there are some parts where the NOx-T50 does not decrease with the decrease in the average endurance temperature. In the prediction results of both the durability history A and the durability history B by the prediction device of Comparative Example 1, the NOx-T50 does not monotonically decrease with the decrease in the average endurance temperature. The prediction results by the prediction device of Comparative Example 2 are also similar to the prediction results by the prediction device of Comparative Example 1.

[0067] (3) Prediction of sensitivity of purification performance to endurance time and average endurance temperature of endurance history Using the prediction devices of each of the examples of Examples 1 and 2 and Comparative Examples 1 and 2, the sensitivity of the purification performance of the catalyst to the endurance time and average endurance temperature of the durability history A was predicted from the explanatory variable data of the durability history A in the range of 2 hours to 100 hours and 861 ° C to 925 ° C. In this case, the prediction unit of the processing unit used a prediction model to obtain information (objective variable) expressing the purification performance from each data of the explanatory variables. Furthermore, the calculation unit of the processing unit calculated NOx-T50 from the information expressing the purification performance. As a result, the change in NOx-T50 with respect to the endurance time and average endurance temperature of the durability history A was predicted. FIG. 7(a) is a two-dimensional graph showing the results of the prediction of the change in NOx-T50 with respect to the endurance time and average endurance temperature of the durability history A by the prediction devices of Examples 1 and 2 and Comparative Examples 1 and 2, and FIG. 7(b) is a three-dimensional graph showing the same results.

[0068] As shown in Figs. 7(a) and (b), in the prediction results of Example 1, NOx-T50 monotonically decreases with shortening of the endurance time, and generally monotonically decreases with decreasing average endurance temperature. This result shows the original tendency that the purification performance improves with shortening of the endurance time, and that the purification performance improves with decreasing average endurance temperature. On the other hand, in the prediction results of Example 2, although NOx-T50 monotonically decreases with shortening of the endurance time, there are some parts where NOx-T50 does not decrease with decreasing average endurance temperature. In the prediction results of Comparative Example 1, although NOx-T50 monotonically decreases with shortening of the endurance time, NOx-T50 does not monotonically decrease with decreasing average endurance temperature, and a valley is formed in the three-dimensional graph where NOx-T50 greatly decreases and then greatly increases with decreasing average endurance temperature. The prediction results of Comparative Example 2 are also similar to those of Comparative Example 1.

[0069] Although the embodiment of the prediction device of the present invention has been described in detail above, the present invention is not limited to the above embodiment, and various design modifications can be made without departing from the spirit of the present invention described in the claims. [Explanation of symbols]

[0070] 1: Purification performance prediction device, 4: Memory unit, 4a: Autoencoder, 4b: Prediction model, 4c Other data, 6: Processing unit, 6a: Dimension compression unit, 6b: Prediction unit

Claims

1. A purification performance prediction device that predicts the purification performance of an exhaust gas purification catalyst after durability testing from a durability history, a dimensionality compression unit that acquires data of latent variables representing the durability history by compressing the dimensionality of the durability history data using an autoencoder that has been machine-learned to use durability history data representing the durability history as input data and reproduce the input data; and A prediction unit that acquires information expressing the purification performance from an information group including data of the latent variables by using a prediction model, The autoencoder is an autoencoder that has been machine-learned using a data set of the extended durability history data, the data set including, as the durability history data, original data obtained from condition data of a durability test and extended data obtained by data extension of the original data in the dimensions of an endurance temperature and an endurance time, The purification performance prediction device, characterized in that the prediction model is a learning model that has been machine-learned using a group of information including data on the latent variables as explanatory variables and information expressing the purification performance as a target variable.

2. 2. The purification performance prediction device according to claim 1, wherein the autoencoder is a convolutional autoencoder.

Citation Information

Patent Citations

  • Total evaluation device for exhaust emission control device

    JP2005048642A

  • Calculating automobile exhaust gas emissions

    JP2020537087A

  • Program, method, and device for estimating state of crop

    JP2021157767A

  • Learning model generation device, state quantity estimation device, learning model generation method, and computer program

    JP2022144626A

  • Performance prediction device

    JP2024103993A

Cited By

  • Copper-ceramic bonded body and insulated circuit board

    CN121693482A

  • COMPOUND COPPER / CERAMIC BODY AND INSULATED CONDUCTOR PLATE

    DE112024004177T5