JP7635739B2 - Quality prediction model generating device, quality prediction device, quality prediction model generating method, and quality prediction method - Google Patents

Description

The present invention relates to a quality prediction model generation device, a quality prediction device, a quality prediction model generation method, and a quality prediction method.

Patent Document 1 discloses a method for predicting the quality of steel products, which uses a neural network model trained from the past manufacturing performance data and quality performance data of the slab to predict the presence or absence of defects in the slab.

JP 2020-157333 A

However, in the method of Patent Document 1, the feature discrimination layer is formed without relying on the input structure of the neural network. Therefore, it is not possible to add input structural improvements that would emphasize characteristic relationships that may be latent in multiple manufacturing performance data. In addition, in the method of Patent Document 1, improving the accuracy of the model relies solely on adjusting the hyperparameters of the neural network, so there is a possibility that sufficient prediction accuracy will not be achieved.

The present invention has been made in consideration of the above, and aims to provide a quality prediction model generation device, a quality prediction device, a quality prediction model generation method, and a quality prediction method that are capable of predicting product quality with high accuracy.

In order to solve the above-mentioned problems and achieve the objective, the quality prediction model generation device of the present invention includes: performance data storage means for storing performance data linking the operating conditions for manufacturing a product with the quality of the manufactured product in a database; quality prediction model construction means for constructing a quality prediction model using the performance data to predict the quality using the operating conditions as input data, the quality prediction model learning the structure of the input data as a feature pattern; prediction accuracy evaluation means for evaluating the prediction accuracy of the quality prediction model using evaluation data; and prediction accuracy optimization means for optimizing the prediction accuracy of the quality prediction model by changing the structure of the input data based on the evaluation result of the prediction accuracy obtained by the prediction accuracy evaluation means.

In addition, in the quality prediction model generating device according to the present invention, in the above invention, the performance data storage means creates performance data linking the operating conditions and the quality of the product for each specific position of the product when the quality of the product is evaluated for each specific position of the product, and creates performance data linking the operating conditions for each manufacturing process and the quality of the product for each specific position of the product when the quality of the product is evaluated for each specific position of the product and the product is manufactured through multiple manufacturing processes.

In addition, in the quality prediction model generating device according to the present invention, in the above invention, the structure of the input data that is changed by the prediction accuracy optimization means is the order of the input data that corresponds to the order of the operating conditions.

The quality prediction model generating device according to the present invention is the above invention, in which the quality prediction model is a neural network having a function of extracting features of the input data.

To solve the above problems and achieve the above objectives, the quality prediction device according to the present invention includes a quality prediction means for predicting the quality of a manufactured product based on a quality prediction model generated by the above quality prediction model generation device.

In order to solve the above-mentioned problems and achieve the object, the quality prediction model generating method of the present invention is a quality prediction model generating method executed by an apparatus constructed by a computer, and includes: a performance data storage step in which a performance data storage means provided in the computer stores performance data linking the operating conditions when manufacturing a product with the quality of the manufactured product in a database; a quality prediction model construction step in which a quality prediction model construction means provided in the computer uses the performance data to construct a quality prediction model that predicts the quality using the operating conditions as input data and learns the structure of the input data as a feature pattern; a prediction accuracy evaluation step in which a prediction accuracy evaluation means provided in the computer uses evaluation data to evaluate the prediction accuracy of the quality prediction model; and a prediction accuracy optimization step in which a prediction accuracy optimization means provided in the computer optimizes the prediction accuracy of the quality prediction model by changing the structure of the input data based on the evaluation result of the prediction accuracy obtained in the prediction accuracy evaluation step.

In order to solve the above-mentioned problems and achieve the above-mentioned objectives, the quality prediction method according to the present invention is a quality prediction method executed by a device constructed by a computer, and includes a quality prediction step of predicting the quality of a manufactured product based on a quality prediction model generated by the above-mentioned quality prediction model generation method.

In the quality prediction model generation device, quality prediction device, quality prediction model generation method, and quality prediction method according to the present invention, when constructing a quality prediction model that learns the structure of input data as a feature pattern, the structure of the input data is changed to optimize prediction accuracy. This makes it possible to generate a quality prediction model that can predict product quality with high accuracy, and to predict product quality with high accuracy using the quality prediction model.

FIG. 1 is a diagram showing a schematic configuration of an information processing device which realizes a quality prediction model generating device and a quality prediction device according to first and second embodiments of the present invention. FIG. 2 is a diagram showing performance data created by linking operation condition data and quality data in the quality prediction model generating device and the quality prediction device according to the first embodiment of the present invention. FIG. 3 is a diagram for explaining a procedure for searching for the order of operation condition sequences in input data in the quality prediction model generating device and the quality prediction device according to the first embodiment of the present invention. FIG. 4 is a flowchart showing the steps of a quality prediction model generating method executed by the quality prediction model generating device according to the first embodiment of the present invention. FIG. 5 is a diagram showing performance data created by linking operation condition data and quality data in the quality prediction model generating device and the quality prediction device according to the second embodiment of the present invention. FIG. 6 is a diagram for explaining a procedure for searching for the order of operation condition sequences in input data in the quality prediction model generating device and the quality prediction device according to the second embodiment of the present invention. FIG. 7 is a flowchart showing the steps of a quality prediction model generating method executed by the quality prediction model generating device according to the second embodiment of the present invention.

A quality prediction model generating device (hereinafter referred to as the "model generating device"), a quality prediction device, a quality prediction model generating method (hereinafter referred to as the "model generating method"), and a quality prediction method according to embodiments of the present invention will be described with reference to the drawings. Note that the components in the following embodiments include those that are replaceable and easy for a person skilled in the art, or those that are substantially the same.

(Embodiment 1)
A model generating device, a quality predicting device, a model generating method, and a quality predicting method according to a first embodiment of the present invention will be described with reference to FIGS. 1 to 4. FIG.

[Device configuration]
1 shows an example of the configuration of an information processing device 1 for realizing the quality prediction device according to this embodiment. The information processing device 1 includes an input unit 10, a storage unit 20, a calculation unit 30, and a display unit 40. The model generation device according to this embodiment is realized by the components of the information processing device 1 excluding the quality prediction unit 33 of the calculation unit 30.

The input unit 10 is an input means for the calculation unit 30, and is realized by an input device such as a keyboard, a mouse pointer, or a numeric keypad.

The storage unit 20 is composed of recording media such as an EPROM (Erasable Programmable ROM), a hard disk drive (HDD), and removable media. Examples of removable media include a universal serial bus (USB) memory, a compact disc (CD), a digital versatile disc (DVD), and a Blu-ray (registered trademark) disc (BD). The storage unit 20 can store an operating system (OS), various programs, various tables, various databases, and the like. The storage unit 20 also stores an operation database (DB) 21.

In the operation DB21, performance data of products manufactured in the past is stored in a searchable manner. Here, the "product" in this embodiment includes steel products manufactured through multiple processes. For example, in the case of steel products, the "multiple processes" mentioned above include one or more of the following processes: iron-making process, steel-making process, hot rolling process, cold rolling process, annealing process, and surface treatment process.

Furthermore, the "performance data" includes, for example, the operating conditions in multiple processes when a product was manufactured in the past (hereinafter referred to as "operating condition data") and the quality of the product inspected after the final process (hereinafter referred to as "quality data"). Furthermore, the operating condition data includes, for example, manufacturing condition data and sensor data in each process.

For example, in the case of steel products, operational condition data includes data on product thickness, product width, product length, steel components, casting speed, etc. in the steelmaking process, data on product thickness, product width, product length, product temperature, rolling speed, etc. in the hot rolling process, data on product thickness, product width, product length, rolling speed, etc. in the cold rolling process, data on product thickness, product width, product length, temperature in the surface treatment furnace, surface treatment speed, etc. in the surface treatment process. Quality data includes data on the results of surface defect inspection at the end of the surface treatment process.

The calculation unit 30 is realized by a processor such as a CPU (Central Processing Unit) and a memory (main storage unit) such as a RAM (Random Access Memory) and a ROM (Read Only Memory).

The calculation unit 30 loads a program into the working area of the main memory and executes it, and controls each component part through the execution of the program to realize functions that meet a predetermined purpose. Through the execution of the program, the calculation unit 30 functions as a data linking unit 31, a model construction unit 32, and a quality prediction unit 33. Note that while FIG. 1 shows an example in which the functions of each part are realized by a single computer, the specific method of realizing the functions of each part is not particularly limited, and for example, the functions of each part may be realized by multiple computers.

The data linking unit 31 links the operation condition data with the quality data. First, the data linking unit 31 collects, from the operation DB 21, operation condition data measured by multiple measuring devices during a multi-step manufacturing process, for example, and quality data of the product actually inspected after the final step.

Here, the operating condition data X for the production of a certain product has, for example, the following elements:

X=[x ₁ , x ₂ ,..., x _m ,..., x _M ] (1≦m≦M)

However, the above "x _m " indicates one actual value of one manufacturing facility or manufacturing device in process k (1≦k≦K) of K processes when manufacturing a product through K processes from the first process to the Kth process. In other words, "x _m " indicates the mth (1≦m≦M) actual value out of a total of M types of actual values in all processes.

Furthermore, quality data Y for a certain product may have the following elements:

Y = [y]

Here, "y" above indicates the quality inspection result of the product after process K, i.e., the final process, when manufacturing a product through K processes from the first process to the Kth process. The quality of the product may be a plurality of different quality evaluations. In other words, a plurality of outputs may be considered as _yi for each quality category.

The data linking unit 31 treats the product manufactured in the final process (final product) as one unit, links the operating condition data in each process with the quality data of the product inspected after the final process, and creates performance data in a tabular format, for example, as shown in Figure 2.

In the example of performance data shown in Fig. 2, data for N products is arranged in the row direction, and quality data Y and operation condition data _Xm (1 < m < M) are arranged in the column direction. The performance data shown in the figure is input data when constructing a quality prediction model described below. The data linking unit 31 stores the performance data that has been linked in this way, for example, in the operation DB 21. In this way, the data linking unit 31 functions as performance data storage means that stores in the operation DB 21 performance data that links operation condition data when manufacturing a product with quality data of the manufactured product.

The model construction unit 32 uses the performance data created by the data linking unit 31 to construct a quality prediction model (mathematical model) that predicts product quality for any operating conditions. This quality prediction model is a model that predicts product quality using operating condition data as input data, and is a model that learns the structure of the input data as a feature pattern. Specifically, the quality prediction model is composed of an input part, a feature discrimination part, a prediction judgment part, and an output part.

In the input section, as shown in Figure 2 above, input data having a vector structure with multiple operating conditions (operating condition strings) as elements is entered.

In the feature discrimination section, the model is trained to capture patterns of features in the structure of the input data and discriminate features. This forms a feature discrimination layer that reacts to specific feature patterns and transmits information. This feature discrimination layer is formed in multiple layers, and transmits information from one layer to the next in sequence. Also, information is transmitted from the layer connected to the input part that passes on information to the layer connected to the prediction and judgment part that receives information.

The prediction and judgment part outputs quality judgment prediction results for the input data based on the complex patterns of multiple operating conditions that affect quality.

Specifically, in this embodiment, a quality prediction model is constructed using a Convolutional Neural Network (hereinafter referred to as "CNN"), which is a type of deep learning algorithm that has the function of extracting features of input data. This CNN is a machine learning algorithm that mimics the network structure of nerve cells (neurons) called the neural network of the human brain and its learning mechanism. In addition, the CNN is composed of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer.

Input data is input to the input layer. In the convolutional layer, a filter called a kernel is applied to each region of the input data to convolve the data and extract features to construct a feature map (also called an "element"). There may be multiple filters, in which case multiple feature maps will be constructed. Furthermore, when CNN trains, the filter parameters are learned automatically.

In the pooling layer, the size of each feature map obtained by convolution is reduced. The convolutional layers and pooling layers may be configured with the same number of layers, alternating. In the fully connected layer, classification is performed to classify the combination of feature maps into specific prediction results. For example, in the case of steel products, the aforementioned "specific prediction results" include the presence or absence of quality defects such as surface defects.

By applying CNN, it is possible to extract characteristic patterns from input data and obtain classifications that are identified with high accuracy based on those patterns as output data. In this embodiment, complex patterns of multiple operating conditions that affect quality are identified by a CNN (one-dimensional CNN) that outputs the classification probability of the occurrence or absence of quality defects for each one-dimensional input of each operating condition. In this case, a threshold for the classification probability is set so that if the classification probability of the occurrence of quality defects is less than an arbitrary value p (0≦p≦1), it is classified as "no quality defect," and if the classification probability is equal to or greater than an arbitrary value p, it is classified as "quality defect."

The model construction unit 32 uses the operating conditions in multiple processes when manufacturing a product as input data and the defect occurrence probability score for each quality category of the manufactured product as output data to train the CNN. Note that the "defect occurrence probability score for each quality category" refers to the classification probability regarding the presence or absence of multiple types of defects, such as "the probability of occurrence of defect A is 70%" and "the probability of occurrence of defect B is 30%."

The model construction unit 32 may perform CNN training in each process using the operating conditions of parts corresponding to multiple parts in the product manufactured in the final process as input data and the defect occurrence probability scores of multiple parts in the product manufactured in the final process as output data. This makes it possible to trace the operating conditions corresponding to a specific part of the final product back through each process and build a quality prediction model.

The data created by the data linking unit 31, i.e., the vector-structured input data that is input when constructing a quality prediction model, can be considered to contain, as latent information, information regarding the relationship between operating conditions that are in close proximity in the data structure of the input data.

Now consider a case where multiple operating condition strings, i.e., the order of each element of the input data (input vector), have a relationship to the quality of the product. In this case, depending on the order of the operating condition strings, a situation may arise where the feature discrimination layer, which reacts to specific patterns and transmits information, can more easily discern information patterns of relationships between operating condition data.

The model construction unit 32 then uses the evaluation data to evaluate the prediction accuracy of the quality prediction model. Based on the evaluation results of the prediction accuracy obtained at that time, the structure of the input data is changed to optimize the prediction accuracy of the quality prediction model. Here, the above-mentioned "input data structure" specifically refers to the order of the input data that corresponds to the order of the operating conditions.

The model construction unit 32 rearranges the sequence of operation conditions in the input data, for example, as shown in FIG. 3, and searches for the sequence of operation conditions that will maximize the prediction accuracy. For example, in the first search, the model construction unit 32 uses input data arranged in the order of "operation condition data X ₁ ...X ₂ ...X _m ...X _M " to train the CNN and evaluate the prediction accuracy (see (a) in the figure). In the second search, the model construction unit 32 uses input data arranged in the order of "operation condition data X ₂ ...X ₁ ...X _m ...X _M " to train the CNN and evaluate the prediction accuracy (see (b) in the figure). In the i-th search, the model construction unit 32 uses input data arranged in the order of "operation condition data X _M ...X ₁ ...X _m ...X _M-1 " to train the CNN and evaluate the prediction accuracy (see (c) in the figure).

In this way, the model construction unit 32 evaluates the prediction accuracy for each order while rearranging the order of the operation condition columns in the input data. When rearranging the order of the operation condition columns, the model construction unit 32 also rearranges the operation condition data in the column direction of the performance data shown in FIG. 2, that is, in column units (process units). Note that, for ease of explanation, in FIG. 3, some of the symbols indicating the operation condition data or quality data are shown with different symbols than those in FIG. 2.

In this embodiment, we deal with a model that outputs the classification of the input as a prediction result. Therefore, specifically, "Accuracy" is used as an index for evaluating the prediction accuracy in the model construction unit 32. In other words, the model construction unit 32 searches for the order of the operating condition string in the input data that maximizes the value of the evaluation formula shown in the following formula (1).

However, in the above formula (1), TP (True Positive) is the total number of prediction results where the quality prediction model predicts "Positive" and the correct answer is "Positive". Also, TN (True Negative) is the total number of prediction results where the quality prediction model predicts "Negative" and the correct answer is "Negative". Also, FP (False Positive) is the total number of prediction results where the quality prediction model predicts "Positive" but the correct answer is "Negative". Also, FN (False Negative) is the total number of prediction results where the quality prediction model predicts "Negative" but the correct answer is "Positive".

In addition, the above indicators can be evaluated each time the operating condition string in the input data is rearranged using a method called hold-out validation or a method called cross-validation.

In hold-out validation, the performance data used for learning is divided into training data for training a quality prediction model and evaluation data for evaluating the quality prediction model. An index is then calculated based on the results for the evaluation data. In cross-validation, the performance data used for learning is divided into K groups, and in the kth round (1≦k≦K), the kth group is used as the evaluation data, and all the other groups are used as the learning data to repeatedly perform learning and evaluation. The indexes calculated each time are then averaged to calculate the final index.

The number of possible input data structures, i.e., possible orderings of the operating condition strings in the input data, is equal to the factorial of the total number of items in the operating condition data, so it is not realistic to calculate the value of the evaluation function for all input data structures due to the combinatorial explosion problem. Methods for solving these optimization problems include, for example, metaheuristics, genetic algorithms, and mathematical programming, and any of these may be used.

The order of the sequence of operating conditions in the input data obtained here does not need to be the optimal solution; conditions may be set for the calculation time, and the best solution obtained within that time may be used. Alternatively, all solutions obtained within a preset calculation time may be saved and then all output at the end.

In this way, the model construction unit 32 functions as a means for constructing a quality prediction model, and also functions as a prediction accuracy evaluation means for evaluating the prediction accuracy of the quality prediction model, and a prediction accuracy optimization means for optimizing the prediction accuracy of the quality prediction model based on the evaluation of the prediction accuracy.

The quality prediction unit 33 outputs quality prediction results by inputting vector-structured input data consisting of multiple operating condition sequences for the product whose quality is to be predicted into the quality prediction model constructed by the model construction unit 32. "Output of quality prediction results" specifically refers to calculating a defect occurrence probability score for the target product.

Then, the quality prediction unit 33 displays the defect occurrence probability score, which is the quality prediction result, on the display unit 40. In this case, the quality prediction unit 33 may display a circle mark in the case of a prediction result of no quality defect, and a cross mark in the case of a prediction result of a quality defect, on the display unit 40 where information on the target product can be viewed, or may display a statement such as "The target product is predicted to have a quality defect" only in the case of a prediction result of a quality defect.

The display unit 40 is realized by a display device such as an LCD display or a CRT display. Based on the display signal input from the calculation unit 30, the display unit 40 displays, for example, the quality prediction results (defect occurrence probability score) by the quality prediction unit 33 in the form of characters, figures, etc.

[Model generation method]
A model generation method by the model generation device according to the first embodiment will be described with reference to Fig. 4. The quality prediction method includes a data collection step (step S1), a data linking step (step S2), and a model construction step (step S3).

In the data collection step, the data linking unit 31 collects operation condition data and quality data for multiple processes from the operation DB 21 (step S1). Next, in the data linking step, the data linking unit 31 creates performance data by linking the operation condition data with the quality data (step S2). Next, in the model construction step, the model construction unit 32 constructs a quality prediction model in which the operation conditions for multiple processes when manufacturing a product are used as input data and the defect occurrence probability score for each quality category of the manufactured product is used as output data (step S3).

In the model construction step, when constructing a quality prediction model, a prediction accuracy evaluation step and a prediction accuracy optimization step are carried out. In the prediction accuracy evaluation step, the model construction unit 32 evaluates the prediction accuracy of the quality prediction model using evaluation data. In the prediction accuracy optimization step, the model construction unit 32 optimizes the prediction accuracy of the quality prediction model by changing the structure of the input data (specifically, the order of the input data corresponding to the order of the operating conditions) based on the evaluation results of the prediction accuracy obtained in the prediction accuracy evaluation step.

[Quality prediction method]
In the quality prediction method by the quality prediction device according to the first embodiment, the quality prediction unit 33 performs a quality prediction step of predicting the quality of a manufactured product based on the quality prediction model generated by the above-mentioned model generation method. After this quality prediction step, the quality prediction unit 33 may perform a prediction result display step of displaying the quality prediction result on the display unit 40.

(Embodiment 2)
A model generating device, a quality predicting device, a model generating method, and a quality predicting method according to a second embodiment of the present invention will be described with reference to FIG. 1 and FIG. 5 to FIG.

[Device configuration]
The configurations of the model generating device and the quality predicting device according to this embodiment are the same as those of the first embodiment (see FIG. 1 ). Therefore, the following will describe the specific processing contents of the data linking unit 31, the model constructing unit 32, and the quality predicting unit 33, and will omit descriptions of the other configurations.

The data linking unit 31 links the operation condition data with the quality data. First, the data linking unit 31 collects, from the operation DB 21, operation condition data measured by multiple measuring devices during a multi-step manufacturing process, for example, and quality data of the product actually inspected after the final step.

Here, the operating condition data X for the production of a certain product has elements such as those shown in the following formula (2).

However, "x _mn " in the above formula (2) indicates the mth (1≦m≦M) value of a set of N types of set values or output values of a large number of manufacturing facilities or manufacturing devices that exist throughout a plurality of processes for manufacturing a product, arranged from the first to the Nth, as manufacturing condition information for a specific position (area) in a specific direction of the product (e.g., the length direction of the product). Also, "x _mn " indicates the nth (1≦n≦N) value of a set of N types of set values or output values of a large number of manufacturing facilities or manufacturing devices that exist throughout a plurality of processes for manufacturing a product, arranged from the first to the Nth, as manufacturing condition information for a specific position (area) in a specific direction of the product (e.g., the length direction of the product). Alternatively, "x _mn " indicates the nth (1≦n≦N) value of a set of N types of measurement values, etc. obtained by measuring the operating state, arranged from the first to the Nth.

Furthermore, quality data Y for a certain product may have the following elements:

Y = [y]

However, the above "y" indicates the quality inspection judgment result of the product after process K of K processes, i.e., the final process, when manufacturing a product through K processes from the first process to the Kth process. The quality of the product may be a plurality of different quality evaluations. In other words, a plurality of outputs can be considered as y _i for each quality category. In addition, the quality data of the final product is given as a judgment result based on the tallying results, in which the quality inspected in any length unit in a specific direction of the product is tallied for each predetermined length (e.g., 100 m). For example, if there is one or more parts acquired as quality defects within the predetermined length, the judgment result within the predetermined length is "quality defect present", and conversely, if there is no part acquired as quality defects within the predetermined length, the judgment result within the predetermined length is "quality defect absent".

The data linking unit 31 creates input data to be input when creating a quality prediction model, which will be described later. This input data is performance data (two-dimensional data) in which one or more pieces of operating condition data for each process are linked to the quality data of the final product inspected after the final process for each specific position in a specific direction of the final product after the final process, and is composed of multiple two-dimensional matrices, for example, as shown in FIG. 5.

The two-dimensional data illustrated in FIG. 5 is composed of data D1, D2, and D3, which are two-dimensional matrices. The columns of the data D1, D2, and D3 are arranged with quality data Y and operation condition data for each process. The rows of the data D1, D2, and D3 are arranged with data relating to a specific position (specific area) in a specific direction of the product.

The aforementioned "data relating to a specific position in a specific direction of the product" can be, for example, a number or coordinates indicating a position when the final product is divided in a specific direction (e.g., the length direction). The two-dimensional data as shown in FIG. 5 serves as input data when constructing a quality prediction model, which will be described later. The data linking unit 31 also stores the thus created performance data (two-dimensional data) in, for example, the operation DB 21. In this way, the data linking unit 31 functions as performance data storage means that stores in the operation DB 21 performance data that links the operation condition data when manufacturing the product with the quality data of the manufactured product.

The data D1, D2, and D3 shown in FIG. 5 may be data relating to the same product (e.g., a coil) or may be data relating to different products. When the data D1, D2, and D3 are data relating to the same product, the data D1 is, for example, data in which the operation condition data for each process up to the product positions "1-100" is arranged in the row direction. The data D2 is data in which the operation condition data for each process up to the product positions "101-200" is arranged in the row direction. The data D3 is data in which the operation condition data for each process up to the product positions "201-300" is arranged in the row direction.

The operating condition data in the length direction of the final product after the final process that the data linking unit 31 uses when creating the two-dimensional data is collected for each predetermined unit length, and M in the above formula (2) is set to a predetermined length. For example, if the "unit length" is set to 1 m and the "predetermined length" is set to 100 m, then "M = 100" and two-dimensional data is created in a matrix format for every 100 m in the length direction of the final product, in other words, for every 100 data rows. In this case, for example, the length corresponding to A in Figure 5 is 1 m, and the length corresponding to B is 100 m.

The values in the "unit length" are treated as the average value of the observed data. In addition, some of the items may include items that take the maximum value or items that take the minimum value. A feature of the "unit length" is that the unit length of the steel plate manufactured in the final process has a minimum unit limit that corresponds to the data collection period in the length direction of the quality data required to build a quality prediction model. In addition, the "unit length" is preferably a unit longer than the data collection period corresponding to the unit length of the final process for the operating condition data in the length direction obtained in the intermediate process, and is set individually according to the data collection period of the target of prediction.

The "predetermined length" is preferably a small value in order to specify the position within the coil as precisely as possible. On the other hand, to achieve high quality prediction performance, a minimum length is required that reveals the characteristics of fluctuations in the operating condition data. Therefore, the "predetermined length" has an appropriate value according to the requirements for position specification and the characteristics of the input data, and is set individually according to the target and purpose of the prediction.

In this way, the data linking unit 31 links the operating condition data and quality data for each process for each predetermined unit length (e.g., 1 m) in the length direction of the final product, for the quality for each specified length (e.g., 100 m), to create two-dimensional data.

In other words, when the quality of a product is evaluated for each specific position of the product, the data linking unit 31 creates performance data linking the operating condition data and the quality data for each specific position of the product. In addition, when the quality of a product is evaluated for each specific position of the product and the product is manufactured through multiple manufacturing processes, the data linking unit 31 creates performance data linking the operating condition data and the quality data for each manufacturing process for each specific position of the product.

The model construction unit 32 uses the performance data created by the data linking unit 31 to construct a quality prediction model (mathematical model) that predicts product quality for any operating conditions. This quality prediction model is a model that predicts product quality using operating condition data as input data, and is a model that learns the structure of the input data as a feature pattern. Specifically, the quality prediction model is composed of an input part, a feature discrimination part, a prediction judgment part, and an output part.

In the input section, as shown in Figure 5 above, input data (two-dimensional data) is input that has a two-dimensional matrix structure with the position coordinates of the product in a specific direction (e.g., the length direction of the product) in the row direction and each operating condition in the column direction.

In the feature discrimination section, the model is trained to capture patterns of two-dimensional features in the input data and discriminate features. This forms a feature discrimination layer that reacts to specific two-dimensional patterns and transmits information. This feature discrimination layer is formed in multiple layers, and information is transmitted from one layer to the next in sequence. Information is also transmitted from the layer connected to the input part that passes on information to the layer connected to the prediction and judgment part that receives information.

The prediction and judgment part outputs quality judgment prediction results for the input of two-dimensional data, based on the complex patterns of multiple operating conditions that affect quality.

In this embodiment, as in the first embodiment, a quality prediction model is constructed using CNN, which is a type of deep learning algorithm that has the function of extracting features of input data. The CNN is composed of an input layer, a convolution layer, a pooling layer, a fully connected layer, and an output layer, and the processing content in each layer is the same as in the first embodiment.

By applying CNN, it is possible to extract characteristic patterns from input data and obtain classifications that are identified with high accuracy based on those patterns as output data. In this embodiment, complex patterns of multiple operating conditions that affect quality are identified by a CNN (two-dimensional CNN) that outputs the classification probability of the occurrence or absence of quality defects for a two-dimensional input of each operating condition. In this case, a threshold for the classification probability is set so that if the classification probability of the occurrence of quality defects is less than an arbitrary value p (0≦p≦1), it is classified as "no quality defect," and if the classification probability is equal to or greater than an arbitrary value p, it is classified as "quality defect."

The model construction unit 32 uses the operating conditions in multiple processes when manufacturing a product as input data and the defect occurrence probability score for each quality category of the manufactured product as output data to train the CNN. Note that the "defect occurrence probability score for each quality category" refers to the classification probability regarding the presence or absence of multiple types of defects, such as "the probability of occurrence of defect A is 70%" and "the probability of occurrence of defect B is 30%."

The data created by the data linking unit 31, i.e., the two-dimensional matrix data input when constructing a quality prediction model, can be considered to contain, as latent information, information on fluctuations in operational conditions for a specific direction of the product, and information on the relationship between adjacent operational conditions in the data structure of the input data. Therefore, by making the input data two-dimensional and increasing the amount of information on operational information for a specific direction of the product, it becomes easier to capture patterns of operational fluctuations related to the quality of the product.

Now consider a case where multiple operating condition sequences, i.e., the order of each element of the input data (two-dimensional data), have a relationship to the quality of the product. In this case, a situation may arise where the order of the operating condition sequences makes it easier for the feature discrimination layer, which reacts to specific two-dimensional patterns and transmits information, to discriminate patterns of information relating to the relationships between operating condition data.

The model construction unit 32 then uses the evaluation data to evaluate the prediction accuracy of the quality prediction model. Based on the evaluation results of the prediction accuracy obtained at that time, the structure of the input data is changed to optimize the prediction accuracy of the quality prediction model. Here, the above-mentioned "input data structure" specifically refers to the order of the input data that corresponds to the order of the operating conditions.

The model construction unit 32 rearranges the sequence of operating conditions in the input data, for example as shown in FIG. 6, and searches for the order of the sequence of operating conditions that maximizes the prediction accuracy.

For example, in the first search, the model construction unit 32 uses input data (two-dimensional data) arranged in the order of "operation condition data _x11-M1 ... _x12-M2 ... _x1n-Mn ... x1N _{- MN" to train the CNN and evaluate the prediction accuracy (see (a) in the figure). Also, in the second search, the model construction unit 32 uses input data (two-dimensional data) arranged in the order of "operation condition data x12-M2} ... _x11-M1 ... _x1n-Mn ... _x1N-MN " to train the CNN and evaluate the prediction accuracy (see (b) in the figure). Then, in the i-th search, the model construction unit 32 trains the CNN using input data (two-dimensional data) consisting of the arrangement of “operating condition data x _{1N to MN} ... x _{11 to M1 ...} x 1n to Mn ... x _{1N-1 to MN-1} ” and evaluates the prediction accuracy (see (c) in the same figure).

For example, in FIG. 6(b), for data D1 (front side of the page), the order of the operating condition column in the first column from the left and the operating condition column in the second column from the left are swapped with respect to FIG. 6(a), and a similar swap is performed for data D2 (rear side of the page). That is, for data D2, the order of the operating condition column in the first column from the left and the operating condition column in the second column from the left are swapped, just like for data D1. When swapping the order of the input data, the model construction unit 32 also performs a similar swap for each of the two-dimensional data that make up the input data.

In this way, the model construction unit 32 evaluates the prediction accuracy for each order while rearranging the order of the operation condition columns in the input data. In addition, when rearranging the order of the operation condition columns, the model construction unit 32 rearranges the operation condition data in the column direction of the performance data shown in FIG. 5, that is, in column units (process units).

In this embodiment, we deal with a model that outputs the classification of the input as a prediction result. Therefore, specifically, "Accuracy" is used as an index for evaluating the prediction accuracy in the model construction unit 32. In other words, the model construction unit 32 searches for the order of the operating condition string in the input data that maximizes the value of the evaluation formula shown in the following formula (1).

However, in the above formula (1), TP (True Positive) is the total number of prediction results where the quality prediction model predicts "Positive" and the correct answer is "Positive". Also, TN (True Negative) is the total number of prediction results where the quality prediction model predicts "Negative" and the correct answer is "Negative". Also, FP (False Positive) is the total number of prediction results where the quality prediction model predicts "Positive" but the correct answer is "Negative". Also, FN (False Negative) is the total number of prediction results where the quality prediction model predicts "Negative" but the correct answer is "Positive".

In addition, the above indicators can be evaluated each time the operating condition string in the input data is rearranged using a method called hold-out validation or a method called cross-validation.

In hold-out validation, the performance data used for learning is divided into training data for training a quality prediction model and evaluation data for evaluating the quality prediction model. An index is then calculated based on the results for the evaluation data. In cross-validation, the performance data used for learning is divided into K groups, and in the kth round (1≦k≦K), the kth group is used as the evaluation data, and all the other groups are used as the learning data to repeatedly perform learning and evaluation. The indexes calculated each time are then averaged to calculate the final index.

The number of possible input data structures, i.e., possible orderings of the operating condition strings in the input data, is equal to the factorial of the total number of items in the operating condition data, so it is not realistic to calculate the value of the evaluation function for all input data structures due to the combinatorial explosion problem. Methods for solving these optimization problems include, for example, metaheuristics, genetic algorithms, and mathematical programming, and any of these may be used.

The order of the sequence of operating conditions in the input data obtained here does not need to be the optimal solution; conditions may be set for the calculation time, and the best solution obtained within that time may be used. Alternatively, all solutions obtained within a preset calculation time may be saved and then all output at the end.

In this way, the model construction unit 32 functions as a means for constructing a quality prediction model, and also functions as a prediction accuracy evaluation means for evaluating the prediction accuracy of the quality prediction model, and a prediction accuracy optimization means for optimizing the prediction accuracy of the quality prediction model based on the evaluation of the prediction accuracy.

The quality prediction unit 33 outputs quality prediction results by inputting two-dimensional input data consisting of multiple operating condition sequences for each specific position in a specific direction of the product whose quality is to be predicted into the quality prediction model constructed by the model construction unit 32. "Output of quality prediction results" specifically refers to calculating a defect occurrence probability score for the target product.

Then, the quality prediction unit 33 displays the defect occurrence probability score, which is the quality prediction result, on the display unit 40. In this case, the quality prediction unit 33 may display a circle mark in the case of a prediction result of no quality defect, and a cross mark in the case of a prediction result of a quality defect, on the display unit 40 where information on the target product can be viewed, or may display a statement such as "The target product is predicted to have a quality defect" only in the case of a prediction result of a quality defect.

[Model generation method]
A model generation method by the model generation device according to the second embodiment will be described with reference to Fig. 7. The quality prediction method includes a data collection step (step S11), a data linking step (step S12), and a model construction step (step S13).

In the data collection step, the data linking unit 31 collects operation condition data and quality data for multiple processes from the operation DB 21 (step S11). Next, in the data linking step, the data linking unit 31 creates performance data by linking the operation condition data and the quality data as two-dimensional data (step S12). Next, in the model construction step, the model construction unit 32 constructs a quality prediction model in which the operation conditions for multiple processes when manufacturing a product are used as input data and the defect occurrence probability score for each quality category of the manufactured product is used as output data (step S13).

In the model construction step, when constructing a quality prediction model, a prediction accuracy evaluation step and a prediction accuracy optimization step are carried out. In the prediction accuracy evaluation step, the model construction unit 32 evaluates the prediction accuracy of the quality prediction model using evaluation data. In the prediction accuracy optimization step, the model construction unit 32 optimizes the prediction accuracy of the quality prediction model by changing the structure of the input data (specifically, the order of the input data corresponding to the order of the operating conditions) based on the evaluation results of the prediction accuracy obtained in the prediction accuracy evaluation step.

[Quality prediction method]
In the quality prediction method by the quality prediction device according to the second embodiment, the quality prediction unit 33 performs a quality prediction step of predicting the quality of a manufactured product based on the quality prediction model generated by the above-mentioned model generation method. After this quality prediction step, the quality prediction unit 33 may perform a prediction result display step of displaying the quality prediction result on the display unit 40.

In the quality prediction model generation device, quality prediction device, quality prediction model generation method, and quality prediction method according to the embodiments described above, when constructing a quality prediction model that learns the structure of input data as a feature pattern, the structure of the input data is changed to optimize prediction accuracy. This makes it possible to generate a quality prediction model that can predict product quality with high accuracy, and to predict product quality with high accuracy using the quality prediction model.

The quality prediction model generating device, quality prediction device, quality prediction model generating method, and quality prediction method according to the present invention have been specifically described above using the form and examples for implementing the invention, but the spirit of the present invention is not limited to these descriptions and must be interpreted broadly based on the claims. Furthermore, it goes without saying that various changes, modifications, etc. based on these descriptions are also included in the spirit of the present invention.

Reference Signs List 1 Information processing device 10 Input unit 20 Storage unit 21 Operation DB
30 Calculation unit 31 Data linking unit 32 Model construction unit 33 Quality prediction unit 40 Display unit

Claims (6)

A performance data storage means for storing performance data in a database that links operational conditions when a product is manufactured with the quality of the manufactured product;
a quality prediction model construction means for constructing a quality prediction model for predicting the quality using the operational conditions as input data, the quality prediction model learning a structure of the input data as a feature pattern, by using the performance data;
a prediction accuracy evaluation means for evaluating the prediction accuracy of the quality prediction model using evaluation data;
a prediction accuracy optimization means for optimizing the prediction accuracy of the quality prediction model by changing a structure of the input data based on the evaluation result of the prediction accuracy obtained by the prediction accuracy evaluation means;
Equipped with
The performance data storage means includes:
When the quality of the product is evaluated for each specific position of the product, performance data is created linking the operating conditions and the quality of the product for each specific position of the product;
When the quality of the product is evaluated for each specific position of the product and the product is manufactured through a plurality of manufacturing processes, performance data is created for each specific position of the product, linking the operating conditions for each manufacturing process with the quality of the product.
A device for generating quality prediction models. The structure of the input data changed by the prediction accuracy optimization means is an arrangement order of the input data corresponding to an order of the operation conditions.
The quality prediction model generating device according to claim 1 . The quality prediction model is a neural network having a function of extracting features of the input data.
A quality prediction model generating device according to claim 1 or 2 . A quality prediction device comprising: a quality prediction means for predicting the quality of a manufactured product based on a quality prediction model generated by the quality prediction model generation device according to any one of claims 1 to 3 . A method for generating a quality prediction model executed by a computer-implemented device, comprising:
a performance data storage step in which performance data storage means included in the computer stores performance data in a database, the performance data linking the operating conditions when a product is manufactured with the quality of the manufactured product;
a quality prediction model construction step in which a quality prediction model construction means included in the computer constructs, using the performance data, a quality prediction model that predicts the quality using the operation conditions as input data, the quality prediction model learning a structure of the input data as a feature pattern;
a prediction accuracy evaluation step in which a prediction accuracy evaluation means included in the computer evaluates the prediction accuracy of the quality prediction model using evaluation data;
a prediction accuracy optimization step in which a prediction accuracy optimization means included in the computer optimizes the prediction accuracy of the quality prediction model by changing a structure of the input data based on the evaluation result of the prediction accuracy obtained in the prediction accuracy evaluation step;
Including,
The performance data storage step includes:
When the quality of the product is evaluated for each specific position of the product, performance data is created linking the operating conditions and the quality of the product for each specific position of the product;
When the quality of the product is evaluated for each specific position of the product and the product is manufactured through a plurality of manufacturing processes, performance data is created for each specific position of the product, linking the operating conditions for each manufacturing process with the quality of the product.
How to generate quality prediction models. 1. A quality prediction method implemented by a computer-implemented apparatus, comprising:
A quality prediction method comprising a quality prediction step of predicting quality of a manufactured product based on the quality prediction model generated by the quality prediction model generating method according to claim 5 .

Images