Steel evaluation method, program, and model generation method

Abstract

Provided, inter alia, is a steel evaluation method by which a more-detailed evaluation can be carried out regarding steel satisfying a composition-related standard.　In the steel evaluation method, a steel composition is analyzed, and if composition data obtained from the analysis satisfies a prescribed standard, the composition data is input to a model that has been trained through machine learning so as to output data that is the same as input data. A loss error between output data output from the model and the composition data is then calculated, and it is determined whether the loss error is greater than a prescribed reference value.

Description

Steel evaluation method, program, and model generation method

[0001] The present invention relates to a method, a program, and a method for generating a model of a steel.

[0002] Steel is an alloy of iron with small amounts of additives such as carbon, silicon, manganese, and phosphorus. The properties of steel vary greatly depending on the ratio of each component, i.e., the composition. Therefore, various types of steel, such as carbon steel, tool steel, and spring steel, are manufactured and used. Standards regarding the composition are established for each type of steel (see Patent Document 1).

[0003] Japanese Patent Application Laid-Open No. 2006-152356

[0004] In the above-mentioned standards, the ratio of each component is specified within a range. When manufacturing steel, the content of each component is adjusted so that it falls within the specified range, i.e., so that the component specifications are met. The manufactured steel undergoes various shipping inspections, including compositional analysis, to confirm that the composition specifications are actually met. Steel that passes the shipping inspections is shipped.

[0005] However, it is empirically known that when comparing multiple lots of steel manufactured to meet the same chemical composition specifications, some lots may exhibit superior properties compared to other lots. Detecting the occurrence of such superior steel properties and reflecting this information in the management of the manufacturing process is expected to raise the standard of steel quality.

[0006] In one aspect, an object of the present invention is to provide a steel evaluation method that can evaluate steel that meets standards regarding composition in more detail.

[0007] The steel evaluation method involves analyzing the composition of the steel, and if the composition data obtained by the analysis meets a predetermined standard, inputting the composition data into a machine-learned model that outputs data identical to the input data, calculating the error loss between the output data output from the model and the composition data, and determining whether the error loss exceeds a predetermined reference value.

[0008] In one aspect, it is possible to provide a steel evaluation method that can evaluate steel that meets standards regarding composition in more detail.

[0009] 1 is an explanatory diagram illustrating the process of determining and using a steel evaluation method; FIG. 2 is an explanatory diagram illustrating the utilization stage of a steel evaluation method; FIG. 3 is a table showing the compositions of samples used in the preliminary study stage; FIG. 4 is a graph showing the heat treatment dimensional change rates of steels with different molybdenum contents side by side; FIG. 5 is a graph showing the heat treatment dimensional change rates of steels with different chromium contents side by side; FIG. 6 is a graph showing the heat treatment dimensional change rates of steels with different vanadium contents side by side; FIG. 7 is a graph showing the heat treatment dimensional change rates of steels with different silicon contents side by side; FIG. 8 is a graph showing the heat treatment dimensional change rates of steels with different manganese contents side by side. FIG. 9 is an explanatory diagram illustrating the configuration of an information processing device used to generate a model; FIG. 10 is an explanatory diagram illustrating a model; FIG. 11 is a flowchart illustrating the processing flow of a program for generating a model; FIG. 12 is a graph showing the variation in mean absolute error loss depending on the manufacturing time; FIG. 13 is an explanatory diagram illustrating the configuration of a steel evaluation system;

[0010] [Embodiment 1] Figure 1 is an explanatory diagram illustrating the process of determining and using a steel evaluation method. In this embodiment, an example will be described in which the composition of a steel that is already mass-produced is fine-tuned to stably reproduce its excellent properties, and an evaluation method is determined for stably producing the fine-tuned steel. An overview of the process of determining a steel evaluation method will be described using Figure 1.

[0011] First, a preliminary study is carried out, in which several steel samples with slightly different compositions are prepared. The samples include steels with standard properties among mass-produced steels, as well as steels with intentionally high or low contents of each element.

[0012] Each sample is subjected to a compositional analysis. Furthermore, each sample is subjected to a property evaluation. Based on the results of the compositional analysis and property evaluation, a target composition is established for consistently producing steel with desired properties.

[0013] Here, the target composition is determined so that the ratio of a specific component is higher or lower within the range of the content of each component specified in the standard for the mass-produced steel. In other words, the target composition does not deviate from the standard for the mass-produced steel. Details of the preliminary study stage will be described later.

[0014] Next, a data collection stage is carried out. In the data collection stage, multiple lots of steel are produced under manufacturing conditions that are modified based on the target composition established in the preliminary study stage. Note that the composition of the steel actually produced varies between lots.

[0015] A composition analysis is performed on each lot of steel. A training database (DB) 51 is created, in which multiple sets of composition analysis data are recorded. A model 56 is generated using the training DB 51. The model 56 is, for example, an autoencoder, and is trained by machine learning so as to reconstruct and output data identical to the input composition analysis data. Details of the model 56 will be described later.

[0016] Next, a verification stage is carried out. In the verification stage, compositional analysis data measured on multiple lots of steel produced before the preliminary study stage and compositional analysis data measured on multiple lots of steel produced in the data collection stage are used. Each compositional analysis data is input to a model 56, and output data is output from the model 56. An error loss between the compositional analysis data as input data and the output data is calculated.

[0017] In this embodiment, the error loss is defined by the loss mean absolute error (MAE). When the composition of the input data is similar to the training data used to generate the model 56, the error loss is small. Conversely, when the composition of the input data is significantly different from the training data, the error loss is large.

[0018] Therefore, the error loss calculated for steel produced before the data collection stage tends to be larger than the error loss calculated for steel produced during the data collection stage. A reference value for the error loss is established at a level that can distinguish the range of variation in steel produced in a production lot after the data collection stage from the range of variation in steel produced in a production lot before the data collection stage. Details of the verification stage will be described later.

[0019] Finally, an overview of the utilization stage will be explained. As in the verification stage, the error loss is calculated for each steel production lot. The manufactured steel is evaluated based on the established reference values. Details of the utilization stage will be described later.

[0020] This section provides an overview of the significance of establishing a standard value for error loss. Steel with a composition similar to the target composition established in the preliminary study stage is expected to have good properties. However, because the content of individual components changes due to manufacturing variations, it is difficult for manufacturing engineers to determine whether the properties of the measured composition are similar to the target composition when looking at the composition data.

[0021] By using the reference value of the error loss established in the verification stage, it is possible to easily determine whether the steel is similar to the target composition. That is, if the calculated error loss is equal to or less than the reference value, it can be determined that the composition of the produced steel is similar to the steel produced in the data collection stage and has the desired properties. If the calculated error loss exceeds the reference value, it can be determined that the composition of the produced steel, even if it meets the specified specifications, is different from the steel produced in the data collection stage and does not have the particularly desired properties.

[0022] Figure 2 is an explanatory diagram illustrating the stages of use of the steel evaluation method. Newly manufactured steel undergoes various quality inspections, including composition analysis. Based on the results of the quality inspections, a conformity determination is made as to whether the steel meets the established standards. Steel that is determined to be non-conforming is not shipped as it is non-conforming, and is instead used as raw material for manufacturing new steel. Conformity determination for steel quality inspections has been carried out for some time, so a detailed explanation will be omitted.

[0023] For steels that are determined to be suitable (suitable products), the compositional analysis data is input to a model 56. The error loss between the output data of the model 56 and the compositional analysis data is calculated. If the error loss is equal to or less than a reference value, the steel is determined to be a high-quality product with desirable properties. The high-quality product is delivered to, for example, a customer that requires particularly high reliability.

[0024] If the error loss exceeds the standard value, the steel is judged to be a conforming product that meets the specified standards, but not a quality product. In the following explanation, products that conform but are not a quality product will be referred to as a non-quality product. For non-quality products, additional inspections are conducted based on, for example, the intended use at the delivery destination. Based on the results of the additional inspections, the non-quality steel is sorted into standard conforming products and conforming products for specific applications, and shipped to the appropriate customers.

[0025] As explained above, steel that has been determined to be in conformity through conventional quality inspections is sorted into good and second-rate products based on the standard values, and the second-rate products are then sorted according to the customer's intended use through additional inspections, allowing each type of steel to be shipped to the appropriate customer based on its characteristics.

[0026] [Preliminary Study Stage] Figure 3 is a table showing the compositions of the samples used in the preliminary study stage. Eight types of samples were used. Each sample is distinguished by the sample symbol shown on the left side of Figure 3. For each sample, each component is shown in mass percent.

[0027] "Std" indicates steel with standard properties among mass-produced steels. "MoL" indicates steel with a lower molybdenum (Mo) content than "Std". "MoH" indicates steel with a higher molybdenum content than "Std". "CrL" indicates steel with a lower chromium (Cr) content than "Std". "CrH" indicates steel with a higher chromium content than "Std". "VH" indicates steel with a higher vanadium (V) content than "Std". "SiL" indicates steel with a lower silicon (Si) content than "Std". "MnH" indicates steel with a higher manganese (Mn) content than "Std".

[0028] Six rectangular parallelepiped specimens were prepared for each sample shown in Figure 3. Each specimen was cut from a single steel material so that the length direction was the direction in which the steel material was expanded by hot working, the thickness direction was the direction in which the steel material was compressed, and the width direction was the direction perpendicular to both the length direction and the thickness direction. In the graphs used in the following explanation, the length direction is indicated by the symbol L, the width direction by the symbol W, and the thickness direction by the symbol T.

[0029] The dimensional change rate after heat treatment was measured for each sample before and after the heat treatment. The dimensional change rate after heat treatment is the dimensional change rate obtained by dividing the amount of dimensional change due to heat treatment by the dimension before the heat treatment. The measurement of the dimensional change rate after heat treatment is an example of a characteristic evaluation method in the preliminary study stage described using Figure 1.

[0030] The heat treatments were quenching and tempering. The quenching conditions were all the same. Quenching was performed by semi-cooling from 1020°C for 10 minutes. The semi-cooling number indicates the time required to cool to half the temperature of 1020°C (510°C). Tempering was performed under five different conditions: tempering temperatures of 485°C, 500°C, 515°C, 530°C, and 545°C. Tempering was performed twice at each temperature.

[0031] Figure 4 is a graph showing the heat treatment dimensional change rates of steels with different molybdenum contents. In Figure 4, three graphs with a common vertical axis are arranged horizontally. (a) on the left side shows the heat treatment dimensional change rate of "MoL", (b) in the middle shows the heat treatment dimensional change rate of "Std", and (c) on the right side shows the heat treatment dimensional change rate of "MoH".

[0032] The horizontal axis of the graph indicates the tempering temperature, and the vertical axis of the graph indicates the heat treatment dimensional change rate. The heat treatment dimensional change rate on the vertical axis is a value calculated by measuring the sample dimensions before and after heat treatment, and then dividing the dimensional change before and after heat treatment by the dimension before heat treatment. Positive values ​​on the vertical axis indicate expansion due to heat treatment, and negative values ​​on the vertical axis indicate contraction due to heat treatment. "As(Q)" on the horizontal axis indicates samples that were only quenched.

[0033] The black circles in the graph indicate the heat treatment dimensional change rate in the length direction (L). The black diamonds indicate the heat treatment dimensional change rate in the width direction (W). The black triangles indicate the heat treatment dimensional change rate in the thickness direction (T). For each temperature, the average value ave and the sample standard deviation s of the three heat treatment dimensional change rates in the length direction, width direction, and thickness direction were calculated, and the range of ave ± 3s is shown in gray.

[0034] Figure 5 is a graph showing the heat treatment dimensional change rates of steels with different chromium contents. In Figure 5, three graphs with a common vertical axis are arranged horizontally. (a) on the left side shows the heat treatment dimensional change rate of "CrL", (b) in the middle shows the heat treatment dimensional change rate of "Std", and (c) on the right side shows the heat treatment dimensional change rate of "CrH". The vertical axis, horizontal axis, and plot symbols are the same as in Figure 4, so their explanation will be omitted. (b) of Figure 4 and (b) of Figure 5 are the same graph.

[0035] Figure 6 is a graph showing the heat treatment dimensional change rates of steels with different vanadium contents. In Figure 6, two graphs with a common vertical axis are arranged horizontally. (a) on the left side shows the heat treatment dimensional change rate of "Std", and (b) on the right side shows the heat treatment dimensional change rate of "VH". The vertical axis, horizontal axis, and plot symbols are the same as those in Figure 4, so their explanation will be omitted. (b) of Figure 4 and (a) of Figure 6 are the same graph.

[0036] Figure 7 is a graph showing the heat treatment dimensional change rates of steels with different silicon contents. In Figure 7, two graphs with a common vertical axis are arranged horizontally. (a) on the left side shows the heat treatment dimensional change rate of "SiL", and (b) on the right side shows the heat treatment dimensional change rate of "Std". The vertical axis, horizontal axis, and plot symbols are the same as in Figure 4, so their explanation will be omitted. (b) of Figure 4 and (b) of Figure 7 are the same graph.

[0037] Figure 8 is a graph showing the heat treatment dimensional change rates of steels with different manganese contents. In Figure 8, two graphs with a common vertical axis are arranged horizontally. (a) on the left side shows the heat treatment dimensional change rate of "Std", and (b) on the right side shows the heat treatment dimensional change rate of "MnH". The vertical axis, horizontal axis, and plot symbols are the same as those in Figure 4, so their explanation will be omitted. (b) of Figure 4 and (a) of Figure 8 are the same graph.

[0038] When processing steel with large anisotropy in the heat treatment dimensional change rate, differences in shape occur due to dimensional change caused by heat treatment after machining, which increases the machining cost and processing man-hours required for machining to correct the shape. Therefore, it is desirable that the anisotropy in the heat treatment dimensional change rate for each steel material is small, that is, the range of ave ±3s shown in gray in the graphs of Figures 4 to 8 is narrow.

[0039] Furthermore, it is difficult to consider dimensional changes due to heat treatment because the temperature for heat treatment, such as tempering, is appropriately selected depending on the equipment used and the mechanical properties of the machine parts to be manufactured. Therefore, it is desirable that the difference in dimensional changes due to heat treatment depending on the tempering temperature is small, that is, that there is little change in the vertical axis direction in the graphs of Figures 4 to 8.

[0040] Considering the above criteria, it can be seen that the steel with a relatively high chromium content shown in Fig. 5(c) and the steel with a relatively high vanadium content shown in Fig. 6(b) show less variation in heat treatment dimensional change depending on the direction and temperature and exhibit good properties. Therefore, in the preliminary study stage, a target composition was established that had relatively high contents of chromium and vanadium.

[0041] [Data Collection Stage] Based on the target composition, 18 lots of steel with higher chromium and vanadium contents within the specifications were manufactured (after the target composition was established). Composition analysis was performed on each lot of steel, and a training DB 51 recording the composition analysis data was created (see FIG. 9, described below). The record layout of the training DB 51 is similar to the table showing the composition of the samples described using FIG. 3, and the content of each component is recorded for each lot. The record layout of the training DB 51 is not shown.

[0042] 9 is an explanatory diagram illustrating the configuration of the information processing device 10 used to generate the model 56. The information processing device 10 includes a control unit 11, a main storage device 12, an auxiliary storage device 13, a communication unit 14, an output unit 15, an input unit 16, a reading unit 19, and a bus.

[0043] The control unit 11 is an arithmetic and control device that executes the program of this embodiment. The control unit 11 uses one or more central processing units (CPUs), graphics processing units (GPUs), tensor processing units (TPUs), multi-core CPUs, etc. The control unit 11 is connected to each hardware unit that constitutes the information processing device 10 via a bus.

[0044] The main memory device 12 is a storage device such as an SRAM (Static Random Access Memory), a DRAM (Dynamic Random Access Memory), a flash memory, etc. The main memory device 12 temporarily stores information required during processing performed by the control unit 11 and programs currently being executed by the control unit 11.

[0045] The auxiliary storage device 13 is a storage device such as an SRAM, a flash memory, a hard disk, or a magnetic tape. The auxiliary storage device 13 stores a training DB 51, a model 56 being generated, a program to be executed by the control unit 11, and various data required for executing the program. The training DB 51 may be stored in an external large-capacity storage device connected to the information processing device 10. The communication unit 14 is an interface for communication between the information processing device 10 and a network.

[0046] The output unit 15 is, for example, a liquid crystal display device or an organic EL (Electro Luminescence) display device. The input unit 16 is, for example, an input device such as a keyboard, a mouse, a trackball, or a microphone. The output unit 15 and the input unit 16 may be stacked together to form a touch panel.

[0047] The portable recording medium 96 is, for example, a USB (Universal Serial Bus) memory, a CD-ROM (Compact Disc Read Only Memory), a magneto-optical disk medium, other optical disk medium, an SD memory card, etc. The portable recording medium 96 stores a program 97, which will be described later.

[0048] The reading unit 19 is an interface capable of reading a portable recording medium 96, such as a USB connector, a CD-ROM drive, or an SD memory reader. The semiconductor memory 98 stores a program 97 and is a memory that can be attached inside the information processing device 10.

[0049] The information processing device 10 is a general-purpose personal computer, a tablet, a mainframe computer, a virtual machine running on a mainframe computer, or a quantum computer. The information processing device 10 may be configured with hardware such as multiple personal computers or mainframe computers that perform distributed processing. The information processing device 10 may be configured with a cloud computing system. The information processing device 10 may be configured with hardware such as multiple personal computers or mainframe computers that operate in cooperation with each other.

[0050] The program 97 is recorded on a portable recording medium 96. The control unit 11 reads the program 97 via the reading unit 19 and stores it in the auxiliary storage device 13. The control unit 11 may also read out the program 97 stored in a semiconductor memory 98. Furthermore, the control unit 11 may download the program 97 from another server computer (not shown) connected via the communication unit 14 and a network (not shown) and store it in the auxiliary storage device 13.

[0051] The program 97 is installed as a control program for the information processing device 10, and is loaded into and executed by the main storage device 12. The program 97 of this embodiment is an example of a program product.

[0052] The program 97 may be constructed and used on an open software platform, which allows the program 97 to be easily ported to various hardware.

[0053] FIG. 10 is an explanatory diagram illustrating the model 56. As described above, the model 56 is a machine learning model trained to output data identical to the input composition analysis data. Note that the composition analysis data is a one-dimensional matrix in which numerical values ​​indicating the content rate of each component are arranged, as described using FIG. 3. In other words, the model 56 accepts an input of a one-dimensional matrix and outputs a one-dimensional matrix of the same size.

[0054] The model 56 has the structure of an autoencoder, which is a combination of an encoder and a decoder, for example, and may be generated using an algorithm such as Transformers or LSTM (Long Short-term Memory).

[0055] Fig. 11 is a flowchart illustrating the processing flow of a program for generating the model 56. Prior to execution of the program of Fig. 11, for example, an untrained autoencoder model is prepared. The parameters of the untrained model are adjusted by the program of Fig. 11, and the model 56 described using Fig. 10 is generated.

[0056] The control unit 11 acquires a training record from the training DB 51 (step S601). The control unit 11 inputs the composition analysis data included in the acquired training record into the model under training 56 and acquires an output one-dimensional matrix (step S602). In the following description, the matrix output from the model under training 56 is referred to as "output data under training."

[0057] The control unit 11 adjusts the parameters of the model 56 using a technique such as backpropagation so that the difference between the composition analysis data input to the model 56 and the training output data output from the model 56 is reduced (step S603).

[0058] The control unit 11 determines whether to end the parameter adjustment (step S604). For example, the control unit 11 determines to end the parameter adjustment when learning has been completed using all of the training data recorded in the training DB 51. If the number of training data recorded in the training DB 51 is sufficiently large, the control unit 11 may determine to end the process when learning has been repeated a predetermined number of times defined by the hyperparameters.

[0059] If it is determined not to end the process (NO in step S604), the control unit 11 returns to step S601. If it is determined to end the process (YES in step S604), the control unit 11 records the model 56 including the adjusted parameters in the auxiliary storage device 13 (step S605). Thereafter, the control unit 11 ends the process. This completes the generation of the model 56.

[0060] [Verification Stage] Figure 12 is a graph showing the variation of mean absolute error loss depending on the production time. The horizontal axis indicates the production time of the steel. The vertical axis indicates the mean absolute error loss calculated from the composition analysis data of the steel and the output data output from the model 56 after the composition analysis data was input into the model 56. The mean absolute error loss is defined by equation (1). Based on the target composition described in the [Data Collection Stage], 18 lots of steel with higher chromium and vanadium contents within the specification range were produced (18 black circles, after the target composition was established).

[0061] N is the number of components included in the composition analysis data, Xi is the content rate of the i-th component in the composition analysis data, and Yi is the i-th data in the matrix output from the model 56.

[0062] FIG. 12 is a graph plotting the calculated mean absolute error loss when the composition analysis data for each lot of steel is input into the completed model 56. The horizontal axis indicates the production date of the steel. The scale on the horizontal axis indicates the number of years since the start of production of the steel according to this embodiment. The scale lines indicate January 1 of each year. The vertical axis indicates the mean absolute error loss. Each black circle indicates data for one lot of steel.

[0063] As shown in the graph of Figure 12, the steel used in this embodiment has been manufactured since the end of the first year, and production frequency increased from the first half of the sixth year. From the end of the first year to the middle of the eighth year, production was carried out based on the same specifications. A preliminary study stage was carried out in the second half of the eighth year, and a target composition was determined. Samples in the preliminary study stage are not plotted in the graph of Figure 12.

[0064] From the beginning of the ninth year, a data collection phase was started, and production was carried out under production conditions modified based on the target composition. Samples from the data collection phase are plotted in the graph of Figure 12. Composition analysis data of steels produced according to the target specifications from the beginning of the ninth year to around the autumn of the tenth year was recorded in the training DB 51 and used to generate the model 56.

[0065] When comparing steel lots produced before the target composition was established with steel lots produced after the target composition was established, it can be seen that the mean absolute error loss is clearly lower for the steel lots produced after the target composition was established.

[0066] The dashed line indicates the reference value of the mean absolute error loss determined by experts such as manufacturing engineers based on the data plotted in Figure 12. In Figure 12, the reference value is set to a value slightly larger than the maximum value of the mean absolute error loss after the target composition is determined. In other words, the reference value is set to a value smaller than the maximum value of the mean absolute error loss before the target loss is determined.

[0067] 13 is an explanatory diagram illustrating the configuration of a steel evaluation system 40. The steel evaluation system 40 includes an information processing device 20 and an analyzing device 30. The analyzing device 30 is, for example, an ICP-OES (Inductively Coupled Plasma Optical Emission Spectrometer), and is used to analyze the composition of steel.

[0068] The information processing device 20 includes a control unit 21, a main memory device 22, an auxiliary memory device 23, a communication unit 24, an output unit 25, an input unit 26, a reading unit 29, and a bus.

[0069] The control unit 21 is an arithmetic and control device that executes the program of this embodiment. The control unit 21 uses one or more CPUs, GPUs, multi-core CPUs, etc. The control unit 21 is connected to each hardware unit that constitutes the information processing device 20 via a bus.

[0070] The main memory device 22 is a storage device such as an SRAM, a DRAM, a flash memory, etc. The main memory device 22 temporarily stores information required during processing performed by the control unit 21 and programs being executed by the control unit 21.

[0071] The auxiliary storage device 23 is a storage device such as an SRAM, a flash memory, a hard disk, or a magnetic tape. The auxiliary storage device 23 stores the model 56, the programs to be executed by the control unit 21, and various data required for executing the programs. The communication unit 24 is an interface for communication between the information processing device 20 and a network.

[0072] The output unit 25 is, for example, a liquid crystal display device or an organic EL display device. The input unit 26 is, for example, an input device such as a keyboard, a mouse, a trackball, or a microphone. The output unit 25 and the input unit 26 may be stacked together to form a touch panel.

[0073] The portable recording medium 91 is, for example, a USB memory, a CD-ROM, a magneto-optical disk medium, other optical disk medium, an SD memory card, etc. The portable recording medium 91 stores a program 92, which will be described later.

[0074] The reading unit 29 is an interface capable of reading a portable recording medium 91, such as a USB connector, a CD-ROM drive, or an SD memory reader. The semiconductor memory 93 stores a program 92 and is a memory that can be attached inside the information processing device 20.

[0075] The information processing device 20 is a general-purpose personal computer, a tablet, a mainframe computer, a virtual machine running on a mainframe computer, or a quantum computer. The information processing device 20 may be configured with hardware such as multiple personal computers or mainframe computers that perform distributed processing. The information processing device 20 may be configured with a cloud computing system. The information processing device 20 may be configured with hardware such as multiple personal computers or mainframe computers that operate in cooperation with each other.

[0076] The information processing device 20 may be configured integrally with the analysis device 30 and may also serve as a control device for the analysis device 30. The analysis device 30 may be configured to automatically transmit composition analysis data to the information processing device 20. A user may load composition analysis data recorded in the analysis device 30 into the information processing device 20 via, for example, a network drive or any recording medium.

[0077] The program 92 is recorded on a portable recording medium 91. The control unit 21 reads the program 92 via the reading unit 29 and stores it in the auxiliary storage device 23. The control unit 21 may also read out the program 92 stored in the semiconductor memory 93. Furthermore, the control unit 21 may download the program 92 from another server computer (not shown) connected via the communication unit 24 and a network (not shown) and store it in the auxiliary storage device 23.

[0078] The program 92 is installed as a control program for the information processing device 20, and is loaded into and executed by the main storage device 22. The program 92 of this embodiment is an example of a program product.

[0079] 14 is a flowchart illustrating the flow of program processing in the utilization stage. The control unit 21 acquires composition analysis data from the analyzer 30, a network drive, or the like (step S501). The control unit 21 determines whether the content of each component is within the range specified by the standard (step S502).

[0080] If it is determined that there is a component that is not within the range specified by the standard (NO in step S502), the control unit 21 notifies the user that the steel related to the composition analysis data acquired in step S501 is a non-conforming product (step S511). The notification may be sent to, for example, an automated machine, which then marks the steel as unshippable. The notification may also be sent to a manufacturing engineer.

[0081] If it is determined that all components are within the ranges defined by the standards (YES in step S502), the control unit 21 inputs the composition analysis data acquired in step S501 into the model 56 and acquires output data output from the model 56 (step S503). The control unit 21 calculates the error loss between the composition analysis data and the output data based on, for example, equation (1) (step S504).

[0082] The control unit 21 determines whether the calculated error loss is equal to or less than a reference value (step S505). If it is determined that the error loss is equal to or less than the reference value (YES in step S505), the control unit 21 notifies the user that the steel related to the composition analysis data acquired in step S501 is a high-quality product (step S512). The notification is sent to, for example, an automated machine, which then marks the steel as a high-quality product. The notification may also be sent to a manufacturing engineer.

[0083] If it is determined that the value is greater than the reference value (NO in step S505), the control unit 21 notifies the user that the steel related to the composition analysis data acquired in step S501 is neither a good product nor a non-conforming product, but a near-good product (step S513). This notification is sent to, for example, a manufacturing engineer. The manufacturing engineer then performs various additional inspections as described with reference to FIG. 2.

[0084] After step S511, step S512, or step S513, the control unit 21 ends the process.

[0085] According to this embodiment, a steel evaluation method can be provided that can evaluate in more detail steel that is judged to be conforming by existing inspection methods. Since it is possible to easily distinguish between high-quality steel that meets standards and has particularly good properties and other second-quality steel, it is possible to prioritize the delivery of high-quality steel to customers who require particularly high reliability, for example.

[0086] According to this embodiment, by setting a target composition, it is possible to increase the frequency of producing high-quality products, thereby raising the standard of steel quality.

[0087] According to this embodiment, by utilizing composition analysis data of steel produced before the target composition is determined in the determination of the reference value, it is not necessary to produce and test samples for comparison after the data collection stage is completed, and the reference value can be determined quickly.

[0088] According to this embodiment, by using the reference value, deviations in composition during the manufacturing process can be quickly detected, and steel of stable quality can be manufactured.

[0089] The loss mean squared error (LOS MSE) may be used as the error loss. The loss mean squared error loss is defined by equation (2).

[0090] N is the number of components included in the composition analysis data, Xi is the content rate of the i-th component in the composition analysis data, and Yi is the i-th data in the matrix output from the model 56.

[0091] Even when the error loss is defined as the mean squared error loss, the error loss will be small if the composition of the input data is similar to the training data used to generate the model 56. Conversely, the error loss will be large if the composition of the input data is significantly different from the training data.

[0092] Other loss functions that may be used for generating and evaluating machine learning models include the square root of the mean squared error loss, the loss mean squared logarithmic error (LOSS MSLE), the square root of the mean squared logarithmic error loss, the Huber loss, the Poisson loss, the hinge loss, or the Kullback-Leibler divergence (KLD).

[0093] The loss function used when adjusting the parameters of the model 56 through machine learning using the training DB 51 (step S603 in FIG. 11 ) and the error loss used when formulating the reference value in the verification stage may be the same function or different functions. The same function is used for the error loss used when formulating the reference value and the error loss calculated in the utilization stage (step S504 in FIG. 14 ).

[0094] A computer program can be deployed to be executed on a single computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected by a communication network.

[0095] The technical features (constituent elements) described in each embodiment can be combined with each other, and by combining them, new technical features can be formed. The embodiments disclosed herein are illustrative in all respects and should be considered not to be limiting. The scope of the present invention is defined by the claims, not by the meaning described above, and is intended to include all modifications within the meaning and scope of the claims.

[0096] Independent and dependent claims may be combined with each other in any combination, regardless of the reference format. Furthermore, while the claims use a format in which a claim references two or more other claims (multiple claim format), this is not a limitation. Multiple claims that reference at least one other claim (multi-multi claim format) may also be used.

[0097] REFERENCE SIGNS LIST 10 Information processing device 11 Control unit 12 Main memory device 13 Auxiliary memory device 14 Communication unit 15 Output unit 16 Input unit 19 Reading unit 20 Information processing device 21 Control unit 22 Main memory device 23 Auxiliary memory device 24 Communication unit 25 Output unit 26 Input unit 29 Reading unit 30 Analysis device 40 Steel evaluation system 51 Training DB 56 Model 91 Portable recording medium 92 Program 93 Semiconductor memory 96 Portable recording medium 97 Program 98 Semiconductor memory

Claims

1. A method for evaluating steel, comprising: analyzing the composition of steel; and, if composition data obtained by the analysis meets a predetermined standard, inputting the composition data into a machine-learned model that outputs data identical to the input data; calculating an error loss between output data output from the model and the composition data; and determining whether or not the error loss exceeds a predetermined reference value.

2. The method for evaluating steel according to claim 1, further comprising the step of: performing an additional inspection on the steel if it is determined that the error loss exceeds a predetermined reference value.

3. The method for evaluating steel according to claim 1, wherein the model is an autoencoder model including an encoder and a decoder.

4. A method for evaluating steel as described in any one of claims 1 to 3, wherein the model is machine-trained using composition data of steel produced after a change in manufacturing process as training data, and the reference value is set to a value that is greater than the maximum error loss between the composition data included in the training data and the output data output from the model when the composition data included in the training data is input into the model after machine learning is completed, and is smaller than the maximum error loss between the output data and the composition data output from the model when the composition data of steel produced before a change in manufacturing process is input into the model after machine learning is completed.

5. A program that causes a computer to execute the following process: acquiring composition data of steel; if the composition data satisfies a predetermined standard, inputting the composition data into a machine-learned model that outputs data identical to the input data; calculating the error loss between the output data output from the model and the composition data; and determining whether the error loss exceeds a predetermined reference value.

6. A method for generating a model, comprising: acquiring training data that records a large number of sets of composition data showing the content rate of each component of steel; and, based on the acquired training data, generating a model through machine learning that outputs data identical to the input composition data when composition data is input.

Images