JP7647288B2 - MODEL GENERATION DEVICE, MODEL GENERATION METHOD, MODEL GENERATION PROGRAM, QUALITY PREDICTION MODEL, AND PRODUCT QUALITY PREDICTION DEVICE - Google Patents

Description

One aspect of the present disclosure relates to a model generation device, a model generation method, a model generation program, a quality prediction model, and a product quality prediction device.

Product quality in chemical plants and the like has traditionally been measured by, for example, sampling inspection. Sampling inspection has drawbacks, such as being costly, time-consuming, and not immediate. To address these drawbacks, product quality has been estimated by analyzing the correlation between product quality and sensor data such as temperature and pressure detected during the product manufacturing process. This estimation has been performed, for example, using a model including a neural network. As a method for predicting product quality using a model built by machine learning including such a neural network, a technology is known that predicts quality by applying LSTM (Long short-term memory) to a manufacturing process consisting of multiple phases (see, for example, non-patent document 1).

Kai Wang, Bhushan Gopaluni, Junghui Chen andZhihuan Song, “DeepLearning of Complex Batch Process Data and Its Application on QualityPrediction”, [online], IEEETransactions on Industrial Informatics, vol.16, no.12, pp.7233-7242, Dec. 2020,doi: 10.1109/TII.2018.2880968, [Retrieved December 11, 2020], Internet <https://dais.chbe.ubc.ca/assets/preprints/2019J2_Wang_IEEEII.pdf>

A method of predicting product quality based on sensor data using a model constructed by machine learning is effective in continuous plants that mass-produce one type of product. However, in so-called batch plants that produce a wide variety of products in small quantities, it is difficult to apply the above method for reasons such as the small amount of data acquired for each product, the fact that the plant is a non-steady process in which the state of the plant changes over time, and the fact that product quality is affected by the entire process history. Although the technology described in Non-Patent Document 1 mentions batch processes, it does not anticipate the production of a wide variety of products or the prediction of their product quality.

The present invention was made in consideration of the above problems, and aims to make it possible to predict the quality of products produced in plants that produce a wide variety of products in small quantities, such as batch plants.

A model generation device according to one aspect of the present disclosure is a model generation device that generates a quality prediction model by machine learning to predict the product quality of multiple types of products to be manufactured, the quality prediction model including a neural network, sensor data and physical property data as input, and product quality prediction data representing predicted product quality as output, the learning data used in the machine learning of the quality prediction model is composed of pairs of sensor data and product quality data, the sensor data is data including inspection values of one or more detection items that indicate the state of manufacture detected in the manufacturing process of a unit quantity of a product, the product quality data is data indicating the product quality of a product manufactured in the manufacturing process in which the corresponding sensor data was detected, and the physical property data is data indicating the characteristics of the product associated with each product, the model generation device includes a model input/output unit that inputs the sensor data included in the learning data and the physical property data of the product corresponding to the learning data into the quality prediction model and acquires the product quality prediction data output from the quality prediction model, a model learning unit that updates parameters included in the quality prediction model based on the error between the product quality prediction data and the product quality data included in the learning data, and a model output unit that outputs the learned quality prediction model whose parameters have been updated by the model learning unit.

A model generation method according to one aspect of the present disclosure is a model generation method in a model generation device that generates a quality prediction model for predicting the product quality of multiple types of products to be manufactured by machine learning, the quality prediction model includes a neural network, receives sensor data and physical property data as input, and outputs product quality prediction data representing predicted product quality, the learning data used in the machine learning of the quality prediction model consists of pairs of sensor data and product quality data, the sensor data is data including inspection values of one or more detection items that indicate the state of production detected in the production process of a unit quantity of products, the product quality data is data indicating the product quality of the product manufactured in the production process in which the corresponding sensor data was detected, and the physical property data is data indicating the characteristics of the product associated with each product, the model generation method includes a model input/output step of inputting the sensor data included in the learning data and the physical property data of the product corresponding to the learning data into the quality prediction model and acquiring product quality prediction data output from the quality prediction model, a model learning step of updating parameters included in the quality prediction model based on the error between the product quality prediction data and the product quality data included in the learning data, and a model output step of outputting the learned quality prediction model whose parameters have been updated in the model learning step.

A model generation program according to one aspect of the present disclosure is a model generation program for causing a computer to function as a model generation device that generates, by machine learning, a quality prediction model for predicting the product quality of multiple types of products to be manufactured, wherein the quality prediction model includes a neural network, receives sensor data and physical property data as input, and outputs product quality prediction data representing predicted product quality, and the learning data used in the machine learning of the quality prediction model consists of pairs of sensor data and product quality data, the sensor data is data including inspection values of one or more detection items that indicate the state of manufacture detected in the manufacturing process of a unit quantity of a product, the product quality data is data indicating the product quality of a product manufactured in the manufacturing process in which the corresponding sensor data was detected, and the physical property data is data associated with each product and indicating the characteristics of the product, and the model generation program causes the computer to realize a model input/output function that inputs the sensor data included in the learning data and the physical property data of the product corresponding to the learning data into the quality prediction model and acquires product quality prediction data output from the quality prediction model, a model learning function that updates parameters included in the quality prediction model based on the error between the product quality prediction data and the product quality data included in the learning data, and a model output function that outputs the learned quality prediction model whose parameters have been updated by the model learning function.

According to this aspect, the physical property data associated with the product is input into the quality prediction model together with the sensor data detected during the product manufacturing process, so that the characteristics of the sensor data are learned in association with the type of product. As a result, sensor data from different types of products can be used to train the quality prediction model, so that even if the amount of sensor data acquired from the manufacturing results of each product type is not sufficient for training the quality prediction model, the sensor data from multiple types of products can be used to train the quality prediction model. Therefore, it becomes possible to predict the quality of products manufactured in plants that manufacture a wide variety of products in small quantities, such as batch plants.

In a model generation device according to another aspect, the quality prediction model may include a first model and a second model, and the model input/output unit may input sensor data to the first model, obtain latent variables expressing characteristics of the sensor data from the first model, input the latent variables and physical property data to the second model, and obtain product quality prediction data output from the second model.

According to this aspect, the characteristics of the sensor data are learned by the first model. Then, the latent variables and physical property data output from the first model are input to the second model, and the second model learns the correlation between the characteristics of the sensor data and the product. Therefore, the characteristics of the sensor data are learned in association with the type of product, and a quality prediction model applicable to multiple product varieties can be constructed by the first model and the second model.

In a model generation device according to another aspect, the first model may be an encoder-decoder model including a neural network and configured with an encoder and a decoder, the encoder generates latent variables based on sensor data, the decoder generates output data based on the latent variables, the model input/output unit inputs the sensor data to the encoder and acquires output data output from the decoder, and the model learning unit updates parameters of the first model based on an error between the output data and the sensor data input to the encoder, and updates parameters of the second model based on an error between product quality prediction data output from the second model and product quality data.

According to this aspect, latent variables that compress and express the features of the sensor data can be obtained.
The latent variables and the physical property data are input together into the second model, and the parameters of the second model are updated based on the error between the product quality prediction data output from the second model and the product quality data in the training data, thereby constructing the second model, thereby making it possible to construct a quality prediction model that can be applied to multiple product varieties.

In another aspect of the model generation device, the sensor data includes data acquired in a time series for each item in the product manufacturing process, the encoder-decoder model is configured with any of a recurrent neural network (RNN), a long short-term memory (LSTM), a transformer, and a gated recurrent unit (GRU), and the model input/output unit may sequentially input the sensor data to the encoder in a time series.

In accordance with this aspect, in view of the fact that the sensor data is time-series data indicating the state of the manufacturing process, the encoder-decoder model constituting the first model is composed of any one of an RNN, an LSTM, a Transformer, and a GRU, which are suitable for learning the characteristics of the time-series data. Therefore, the characteristics of the sensor data are appropriately learned by the first model.

In another aspect of the model generation device, the sensor data may include at least one of the temperature and pressure during the manufacturing process, and the physical property data may be data indicating at least one of the physical properties and characteristics of the raw materials associated with the raw materials of the product.

According to this aspect, the sensor data includes at least either the temperature or the pressure in the manufacturing process, and physical property data including either the physical properties or the characteristics of the raw materials of the product is input into the quality prediction model together with the sensor data, so that the characteristics of the state of the manufacturing process are learned in the quality prediction model while being appropriately associated with the type of product.

A quality prediction model according to one aspect of the present disclosure is a quality prediction model trained by machine learning for making a computer function and predicting the product quality of multiple types of products to be manufactured, the quality prediction model includes a neural network, receives sensor data and physical property data as input, and outputs product quality prediction data representing predicted product quality, the learning data used in the machine learning of the quality prediction model consists of pairs of sensor data and product quality data, the sensor data is data including inspection values of one or more detection items that indicate the state of manufacture detected in the manufacturing process of a unit quantity of a product, the product quality data is data indicating the product quality of a product manufactured in the manufacturing process in which the corresponding sensor data was detected, the physical property data is data corresponding to each product and indicating the characteristics of the product, and the quality prediction model is constructed by machine learning that updates parameters based on the error between the product quality prediction data output in response to the input of the sensor data included in the learning data and the physical property data corresponding to the sensor data, and the product quality data included in the learning data.

According to this aspect, physical property data associated with the product is input together with sensor data detected during the product manufacturing process, so that a quality prediction model can be obtained in which the characteristics of the sensor data are learned in association with the type of product. As a result, sensor data from different types of products can be used to train the quality prediction model, so that even if the amount of sensor data acquired from the manufacturing results of each product type is not sufficient for training the quality prediction model, the sensor data from multiple types of products can be used to train the quality prediction model. Therefore, a quality prediction model that can be applied to predicting the quality of products manufactured in plants that manufacture a wide variety of products in small quantities, such as batch plants, can be obtained.

A product quality prediction device according to one aspect of the present disclosure is a product quality prediction device that predicts the product quality of multiple types of products to be manufactured using a quality prediction model constructed by machine learning, the quality prediction model including a neural network, receiving sensor data and physical property data as input, and outputting product quality prediction data representing predicted product quality, the learning data used in the machine learning of the quality prediction model consisting of pairs of sensor data and product quality data, the sensor data being data including inspection values of one or more detection items that indicate the state of manufacture, detected in the manufacturing process of a unit quantity of a product, and the product quality data being data indicating the product quality of a product manufactured in the manufacturing process in which the corresponding sensor data was detected. The physical property data is data associated with each product and indicates the characteristics of the product. The quality prediction model is constructed by machine learning that updates parameters based on the error between the product quality prediction data output in response to input of sensor data included in the learning data and physical property data corresponding to the sensor data, and the product quality data included in the learning data. The product quality prediction device inputs sensor data detected in the manufacturing process of the product whose product quality is to be predicted and the physical property data of the product corresponding to the sensor data into the quality prediction model, and includes a product quality prediction unit that acquires the product quality prediction data output from the quality prediction model, and a product quality prediction data output unit that outputs the product quality prediction data acquired by the product quality prediction unit.

According to this aspect, by inputting physical property data associated with the product along with sensor data detected during the product manufacturing process, a quality prediction model that has been learned by associating the characteristics of the sensor data with the type of product is used to predict the quality of the product. Therefore, it is possible to predict the quality of products manufactured in plants that manufacture a wide variety of products in small quantities, such as batch plants.

One aspect of the present disclosure makes it possible to predict the quality of products produced in plants that produce a wide variety of products in small quantities, such as batch plants.

1 is a block diagram showing an example of a functional configuration of a model generating device according to an embodiment. 1 is a block diagram showing an example of a functional configuration of a product quality prediction device according to an embodiment. FIG. 2 is a hardware block diagram of a model generating device and a product quality predicting device according to the embodiment. FIG. 2 is a diagram showing a schematic configuration of a quality prediction model. FIG. 10 is a diagram illustrating an example of learning data including sensor data and product quality data. FIG. 11 is a diagram showing an example of physical property data. FIG. 2 is a diagram illustrating an example of a configuration of a quality prediction model. FIG. 11 is a diagram illustrating an example of a first model configured by an encoder-decoder model. 4 is a flowchart showing the process of a model generation method in the model generation device according to the embodiment. 4 is a flowchart showing the process of a product quality prediction method in the product quality prediction device according to the embodiment. FIG. 13 is a diagram illustrating a configuration of a model generation program. FIG. 13 is a diagram showing a configuration of a product quality prediction program.

Hereinafter, an embodiment of the present invention will be described in detail with reference to the attached drawings. In the description of the drawings, the same or equivalent elements are given the same reference numerals, and duplicate descriptions will be omitted.

FIG. 1 is a block diagram showing an example of the functional configuration of a model generation device according to an embodiment. The model generation device 10 is a device that generates a quality prediction model by machine learning to predict the product quality of multiple types of products produced in a plant. In this embodiment, the term "product" is a concept that can include finished products, work in progress, semi-finished products, parts, and prototypes.

As shown in FIG. 1, the model generating device 10 may include functional units configured in a processor 101, a learning data storage unit 30, a physical property data storage unit 40, and a quality prediction model storage unit 50. The model generating device 10 includes a learning data acquisition unit 11, a first physical property data acquisition unit 12, a model input/output unit 13, a model learning unit 14, and a model output unit 15. Each of these functional units 11 to 15 may be configured in a single device, or may be distributed across multiple devices.

Each of the functional units 11 to 15 is configured to be able to access the learning data storage unit 30, the physical property data storage unit 40, and the quality prediction model storage unit 50. The learning data storage unit 30, the physical property data storage unit 40, and the quality prediction model storage unit 50 may be configured inside the model generation device 10 as shown in FIG. 1, or may be configured outside the model generation device 10 in a separate device accessible from the model generation device 10. Each of the functional units 11 to 15 and each of the storage units 30, 40, 50 will be described in detail later.

Figure 2 is a block diagram showing an example of the functional configuration of a product quality prediction device according to an embodiment. The product quality prediction device 20 is a device that predicts the product quality of multiple types of products manufactured in a plant using a quality prediction model constructed by machine learning.

As shown in FIG. 2, the product quality prediction device 20 may include functional units configured in the processor 101, a sensor data storage unit 60, a physical property data storage unit 40, and a quality prediction model storage unit 50. The product quality prediction device 20 includes a sensor data acquisition unit 21, a second physical property data acquisition unit 22, a product quality prediction unit 23, and a product quality prediction data output unit 24. Each of these functional units 21 to 24 may be configured in a single device, or may be distributed across multiple devices.

Each of the functional units 21 to 24 is configured to be able to access the sensor data storage unit 60, the physical property data storage unit 40, and the quality prediction model storage unit 50. The sensor data storage unit 60, the physical property data storage unit 40, and the quality prediction model storage unit 50 may be configured inside the product quality prediction device 20 as shown in FIG. 2, or may be configured outside the product quality prediction device 20 as a separate device accessible from the product quality prediction device 20. The physical property data storage unit 40 and the quality prediction model storage unit 50 shown in FIG. 2 may be configured as the same storage unit as the same storage unit shown in FIG. 1. Each of the functional units 21 to 24 and each of the storage units 60, 40, 50 will be described in detail later.

FIG. 3 is a diagram showing an example of the hardware configuration of a computer 100 that constitutes the model generating device 10 and the product quality predicting device 20 according to an embodiment. That is, the computer 100 can constitute the model generating device 10 and the product quality predicting device 20.

As an example, the computer 100 includes, as hardware components, a processor 101, a main memory device 102, an auxiliary memory device 103, and a communication control device 104. The computer 100 constituting the model generating device 10 and the product quality prediction device 20 may further include an input device 105 such as a keyboard, a touch panel, a mouse, etc., which are input devices, and an output device 106 such as a display.

The processor 101 is a computing device that executes an operating system and application programs. Examples of processors include a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit), but the type of processor 101 is not limited to these. For example, the processor 101 may be a combination of a sensor and a dedicated circuit. The dedicated circuit may be a programmable circuit such as an FPGA (Field-Programmable Gate Array), or may be another type of circuit.

The main memory device 102 is a device that stores programs for implementing the model generating device 10, etc., and the results of calculations output from the processor 101. The main memory device 102 is composed of at least one of a ROM (Read Only Memory) and a RAM (Random Access Memory), for example.

The auxiliary storage device 103 is generally a device capable of storing larger amounts of data than the main storage device 102. The auxiliary storage device 103 is composed of a non-volatile storage medium such as a hard disk or a flash memory. The auxiliary storage device 103 stores the model generation program P1 and/or the product quality prediction program P2 for causing the computer 100 to function as the model generation device 10 or the product quality prediction device 20, as well as various types of data.

The communication control device 104 is a device that executes data communication with other computers via a communication network. The communication control device 104 is configured, for example, by a network card or a wireless communication module.

Each functional element of the model generating device 10 and the product quality predicting device 20 is realized by loading the corresponding programs P1, P2 onto the processor 101 or the main memory device 102 and having the processor 101 execute the programs. The programs P1, P2 contain code for realizing each functional element of the corresponding server. The processor 101 operates the communication control device 104 in accordance with the programs P1, P2, and executes reading and writing of data in the main memory device 102 or the auxiliary memory device 103. Through such processing, each functional element of the corresponding server is realized.

The programs P1 and P2 may be provided in a form that is fixedly recorded on a tangible recording medium such as a CD-ROM, a DVD-ROM, or a semiconductor memory. Alternatively, at least one of these programs may be provided via a communications network as a data signal superimposed on a carrier wave.

Next, the quality prediction model generated by the model generation device 10 will be described with reference to FIG. 4. FIG. 4 is a diagram showing a schematic configuration of the quality prediction model. The quality prediction model MD includes a neural network and is constructed by machine learning in the model generation device 10. The quality prediction model MD receives sensor data sd and physical property data pd as input, and outputs product quality prediction data qd that represents the predicted product quality.

The sensor data sd is data including inspection values of one or more detection items that indicate the state of production, detected during the manufacturing process of one unit quantity (e.g., one lot). The sensor data sd may include, for example, at least one of the temperature and pressure during the product manufacturing process as a detection item. The sensor data sd is not limited to the above example, as long as it is data detected from the product manufacturing system. In addition, the sensor data sd may be data (inspection values) detected at a certain time during the manufacturing process, or may include time-series data. The sensor data sd is converted into a vector representation and input to the quality prediction model MD. Details of the sensor data sd will be described later with reference to FIG. 5.

The physical property data pd is data associated with each product produced in the plant and indicates the characteristics of each product. That is, the physical property data pd constitutes product-specific information that can identify each product. More specifically, the physical property data pd is data associated with the raw materials of a product and indicates at least one of the physical properties and characteristics of the raw materials. The physical property data pd is converted into a vector representation and input to the quality prediction model MD. Details of the physical property data pd will be described later with reference to FIG. 6.

The product quality prediction data qd is data indicating the product quality predicted by the quality prediction model MD. The quality prediction model MD outputs a vector representing a predetermined product quality. The product quality may be, for example, viscosity and the concentration of specific impurities, but is not limited to these examples.

The quality prediction model MD can be configured with a multi-layer neural network, and the network configuration is not limited and can be any network configuration. The quality prediction model MD can also be configured with a known neural network, and the type is not limited and can be configured with, for example, a multi-layer perceptron. As described later, the quality prediction model MD can also be configured with multiple models including a neural network.

Referring again to FIG. 1, the functional units of the model generating device 10 will be described. The learning data acquisition unit 11 acquires learning data used for machine learning of the quality prediction model MD. In this embodiment, the learning data acquisition unit 11 acquires learning data from the learning data storage unit 30.

The learning data consists of pairs of sensor data sd and product quality data obtained based on the track record of product production. Figure 5 is a diagram showing an example of learning data consisting of sensor data sd and product quality data pq. The learning data ld consists of pairs of sensor data sd obtained by producing one unit quantity (one lot) of a product and product quality data pq indicating a predetermined quality of the product involved in that production. Figure 5 shows learning data ld obtained by producing L lots of products.

The sensor data sd for one lot is composed of time-series sensor data (1 to D) acquired by D types of sensors at each time step from time t1 to tn. The inspection values acquired by each of the D types of sensors may constitute the inspection values of the D types of detection items. The number of time steps from time t1 to tn is variable. By manufacturing one lot of products, product quality data pq indicating the quality of the manufactured products is acquired. The product quality data pq (q1 to q _L ) of each lot is associated with the corresponding sensor data sd1 to sd _L , respectively, and constitutes the learning data ld for each lot. Note that the products of each lot may be different types of products.

The first physical property data acquisition unit 12 acquires the physical property data pd of the product corresponding to the learning data ld acquired by the learning data acquisition unit 11. In this embodiment, the first physical property data acquisition unit 12 acquires the physical property data pd by referring to the physical property data storage unit 40. FIG. 6 is a diagram showing the data stored in the physical property data storage unit 40. As shown in FIG. 6, the physical property data storage unit 40 stores the feature amount for each raw material used in the production of the product. The feature amounts 1 to d may be, for example, information on the physical properties such as the boiling point and vapor pressure of the compound constituting the raw material. The feature amounts 1 to d may also be latent variables generated based on the molecular descriptors and physical property values of the compound constituting the raw material. The feature amounts 1 to d may be any data that represents the characteristics of the raw material, and are not limited to the above examples.

The physical property data pd of this embodiment can be composed of the feature quantities of the raw materials of the product, as an example. Specifically, the first physical property data acquisition unit 12 determines the type of product of one lot of learning data among the learning data acquired by the learning data acquisition unit 11 based on the attribute information (not shown) of the learning data, and determines the raw materials used in the determined product based on the attribute information (not shown) that associates the raw materials used in the production of the product with the product. Furthermore, the first physical property data acquisition unit 12 refers to the physical property data storage unit 40 to acquire information on the feature quantities associated with the raw materials of the product, and composes the physical property data pd consisting of the acquired information on the feature quantities. The physical property data pd thus configured is data having a number of dimensions obtained by multiplying (the number of types of raw materials used in the production of the product) by the number d of the feature quantities associated with each raw material. Note that the physical property data pd of the product corresponding to the learning data ld may be acquired in advance and constitute a part of the learning data ld.

The model input/output unit 13 inputs the sensor data sd included in the learning data ld and the product physical property data pd corresponding to the learning data ld (sensor data sd) into the quality prediction model MD. Then, the model input/output unit 13 acquires the product quality prediction data qd output from the quality prediction model MD.

The model learning unit 14 performs machine learning of the quality prediction model MD by updating the parameters (weighting coefficients) included in the quality prediction model MD based on the error between the product quality prediction data qd output from the quality prediction model MD and acquired by the model input/output unit 13, and the product quality data pq included in the learning data ld.

The model output unit 15 outputs the trained quality prediction model MD whose parameters have been updated by the model learning unit 14. In this embodiment, the model output unit 15 stores the trained quality prediction model MD in the quality prediction model storage unit 50. When machine learning based on all previously prepared training data ld is performed by the model learning unit 14, the model output unit 15 may determine that the quality prediction model MD has been trained and output the quality prediction model MD.

Next, a specific example of the configuration of the quality prediction model MD will be described with reference to FIG. 7. FIG. 7 is a diagram showing an example of the configuration of the quality prediction model MD. As shown in FIG. 7, the quality prediction model MD may include a first model md1 and a second model md2.

The first model md1 is a model that includes a neural network, has sensor data sd as input, and outputs a latent variable h. The latent variable h is a vector that expresses the characteristics of the sensor data sd. The second model md2 is a model that includes a neural network, has latent variable h and physical property data pd as input, and outputs product quality prediction data qd.

As shown in FIG. 7, when the quality prediction model MD includes a first model md1 and a second model md2, the model input/output unit 13 inputs the sensor data sd to the first model md1 and acquires a latent variable h that expresses the characteristics of the sensor data sd from the first model md1. Then, the model input/output unit 13 inputs the latent variable h and the physical property data pd to the second model md2 and acquires the product quality prediction data qd output from the second model md2.

The first model md1 is required only to generate a latent variable h that is a compressed representation of the features of the input sensor data sd. The configuration of the network and layers is not limited, but may be configured, for example, by an encoder-decoder model.

The network configuration of the neural network that constitutes the second model md2 is not limited, but may be, for example, a multi-layer perceptron. The second model md2 may also include a fully connected layer that receives inputs of the latent variables h and the physical property data pd.

Figure 8 is a diagram showing an example of a first model md1 configured by an encoder-decoder model. As shown in Figure 8, the first model md1 includes an encoder en1 and a decoder de1.

Encoder en1 receives time-series sensor data sd as input and outputs latent variables h. That is, the model input/output unit 13 inputs vectors of sensor data acquired at each time step from time t1 to tn as input data x1 to xn to encoder en1 in a chronological order.

The decoder de1 sequentially outputs output data y1 to yn based on the input of the latent variable h output from the encoder en1. The model input/output unit 13 acquires the output data y1 to yn output from the decoder de1.

The model learning unit 14 updates (learns) the parameters of the first model md1, which is an encoder-decoder model, based on the error between the output data y1 to yn and the input data x1 to xn. In other words, the first model md1 is an autoencoder that is machine-trained so that the output data y1 to yn and the input data x1 to xn are equal.

The encoder-decoder model constituting the first model md1 may be composed of any of the following: RNN (Recurrent Neural Network), LSTM (Long short-term memory), Transformer, and GRU (Gated Recurrent Unit). These exemplified models are suitable for learning time-series data.

The model input/output unit 13 acquires the latent variable h output from the encoder en1 as the output of the first model md1. The model input/output unit 13 inputs the latent variable h and the physical property data pd to the second model md2, and acquires the product quality prediction data qd output from the second model md2. The model learning unit 14 updates (learns) the parameters of the second model md2 based on the error between the product quality prediction data qd and the product quality data pq included in the learning data ld.

The quality prediction model MD, which is a model that includes a trained neural network, can be considered as a program that is loaded or referenced by a computer and causes the computer to execute specified processes and realize specified functions.

That is, the trained quality prediction model MD of this embodiment is used in a computer equipped with a CPU and memory. Specifically, the computer's CPU operates in accordance with instructions from the trained quality prediction model MD stored in the memory to perform calculations on input data input to the input layer of the neural network based on trained weight coefficients (parameters) and functions corresponding to each layer, and to output the results of the calculations from the output layer.

Referring again to FIG. 2, the functional units of the product quality prediction device 20 will be described. The sensor data acquisition unit 21 acquires sensor data sd detected in the manufacturing process of the product whose product quality is to be predicted. In this embodiment, the sensor data acquisition unit 21 acquires sensor data sd stored in the sensor data storage unit 60. The sensor data sd acquired here is data indicating the manufacturing state detected in the manufacturing process of one lot of products, and has the same configuration as the sensor data sd included in the learning data ld described with reference to FIG. 5.

The second physical property data acquisition unit 22 acquires product physical property data pd corresponding to the sensor data sd acquired by the sensor data acquisition unit 21. Specifically, the second physical property data acquisition unit 22 acquires the physical property data pd by referring to the physical property data storage unit 40, similar to the first physical property data acquisition unit 12.

The product quality prediction unit 23 inputs the sensor data sd acquired by the sensor data acquisition unit 21 and the physical property data pd acquired by the second physical property data acquisition unit 22 into the quality prediction model MD, and acquires the product quality prediction data qd output from the quality prediction model MD.

The product quality prediction data output unit 24 outputs the product quality prediction data qd acquired by the product quality prediction unit 23 as information predicting the product quality of the product to be predicted.

Figure 9 is a flowchart showing an example of the processing contents of a model generation method in the model generation device 10 according to an embodiment. The model generation method is executed by loading a model generation program P1 into the processor 101 and executing the program to realize each of the functional units 11 to 15.

In step S1, the learning data acquisition unit 11 acquires learning data ld consisting of pairs of sensor data sd and product quality data pq acquired from the manufacturing process of the product and lot to be learned.

In step S2, the first physical property data acquisition unit 12 acquires product physical property data pd corresponding to the learning data ld acquired by the learning data acquisition unit 11 in step S1. Note that the product physical property data pd corresponding to the learning data ld may be acquired in advance and may form part of the learning data ld.

In step S3, the model input/output unit 13 inputs the sensor data sd contained in the learning data ld acquired in step S1 and the physical property data pd acquired in step S2 into the quality prediction model MD.

In step S4, the model input/output unit 13 acquires the product quality prediction data qd output from the quality prediction model MD.

In step S5, the model learning unit 14 performs machine learning of the quality prediction model MD by updating the parameters (weighting coefficients) included in the quality prediction model MD based on the error between the product quality data pq included in the learning data ld and the product quality prediction data qd acquired in step S4.

In step S6, the model output unit 15 determines whether or not to end the learning of the quality prediction model MD. For example, when machine learning based on all previously prepared learning data ld is performed by the model learning unit 14, it may be determined to end the learning. If it is determined to end the learning, the process proceeds to step S7. On the other hand, if it is not determined to end the learning, the process returns to step S1.

In step S7, the model output unit 15 outputs the learned quality prediction model MD whose parameters have been updated by the model learning unit 14.

Figure 10 is a flowchart showing an example of the processing contents of the product quality prediction method in the product quality prediction device 20 according to the embodiment. The product quality prediction method is executed by loading the product quality prediction program P2 into the processor 101 and executing the program to realize each of the functional units 21 to 24.

In step S11, the sensor data acquisition unit 21 acquires sensor data sd detected during the manufacturing process of the product whose product quality is to be predicted.

In step S12, the second physical property data acquisition unit 22 acquires product physical property data pd corresponding to the sensor data sd acquired by the sensor data acquisition unit 21.

In step S13, the product quality prediction unit 23 inputs the sensor data sd acquired by the sensor data acquisition unit 21 and the physical property data pd acquired by the second physical property data acquisition unit 22 into the trained quality prediction model MD.

In step S14, the product quality prediction unit 23 acquires the product quality prediction data qd output from the quality prediction model MD.

In step S15, the product quality prediction data output unit 24 outputs the product quality prediction data qd acquired by the product quality prediction unit 23 in step S14 as information predicting the product quality of the product to be predicted.

Next, we will explain the model generation program for causing a computer to function as the model generation device 10 of this embodiment. Figure 11 is a diagram showing the configuration of the model generation program.

The model generation program P1 is configured to include a main module m10 that provides overall control of the model generation process in the model generation device 10, a learning data acquisition module m11, a first physical property data acquisition module m12, a model input/output module m13, a model learning module m14, and a model output module m15. The modules m11 to m15 realize the functions of the learning data acquisition unit 11, the first physical property data acquisition unit 12, the model input/output unit 13, the model learning unit 14, and the model output unit 15.

The model generation program P1 may be transmitted via a transmission medium such as a communication line, or may be stored in a recording medium M1 as shown in FIG. 11.

Next, we will explain the product quality prediction program for causing a computer to function as the product quality prediction device 20 of this embodiment. Figure 12 is a diagram showing the configuration of the product quality prediction program.

The product quality prediction program P2 is configured to include a main module m20 that controls the overall product quality prediction process in the product quality prediction device 20, a sensor data acquisition module m21, a second physical property data acquisition module m22, a product quality prediction module m23, and a product quality prediction data output module m24. Each of the modules m21 to m24 realizes the functions of the sensor data acquisition unit 21, the second physical property data acquisition unit 22, the product quality prediction unit 23, and the product quality prediction data output unit 24.

The product quality prediction program P2 may be transmitted via a transmission medium such as a communication line, or may be stored in a recording medium M2 as shown in FIG. 12.

According to the model generation device 10, model generation method, and model generation program P1 of this embodiment described above, physical property data associated with a product is input to the quality prediction model together with sensor data detected during the product manufacturing process, so that the characteristics of the sensor data are learned in association with the type of product. As a result, sensor data of different types of products can be used to train the quality prediction model, so that even if the amount of sensor data acquired from the manufacturing results of each type is not sufficient for training the quality prediction model, sensor data of multiple types of products can be used to train the quality prediction model. Therefore, it becomes possible to predict the quality of products manufactured in plants that manufacture a wide variety of products in small quantities, such as batch plants.

The present invention has been described in detail above based on the embodiments. However, the present invention is not limited to the above embodiments. The present invention can be modified in various ways without departing from the gist of the invention.

P1...model generation program, P2...product quality prediction program, m10...main module, m11...learning data acquisition module, m12...first physical property data acquisition module, m13...model input/output module, m14...model learning module, m15...model output module, m20...main module, m21...sensor data acquisition module, m22...second physical property data acquisition module, m23...product quality prediction module, m24...product quality prediction data output module 10...model generation device, 11...learning data acquisition unit, 12...first physical property data acquisition unit, 13...model input/output unit, 14...model learning unit, 15...model output unit, 20...product quality prediction device, 21...sensor data acquisition unit, 22...second physical property data acquisition unit, 23...product quality prediction unit, 24...product quality prediction data output unit, 30...learning data storage unit, 40...physical property data storage unit, 50...quality prediction model storage unit, 60...sensor data storage unit, M1, M2...recording medium, MD...quality prediction model.

Claims (8)

A model generation device that generates a quality prediction model for predicting product quality of a plurality of types of products to be manufactured by machine learning,
the quality prediction model includes a neural network, receives sensor data and physical property data as input, and outputs product quality prediction data representing a predicted product quality; and includes a first model and a second model;
learning data used in machine learning of the quality prediction model includes pairs of sensor data and product quality data;
the sensor data is data including inspection values of one or more detection items that are detected in a manufacturing process of one unit quantity of the product and indicate a manufacturing state,
the product quality data is data indicating a product quality of the product manufactured in a manufacturing process in which the corresponding sensor data is detected,
The physical property data is data associated with each product and indicates characteristics of the product;
The model generating device comprises:
a model input/output unit that inputs the sensor data included in the learning data and the physical property data of the product corresponding to the learning data into the quality prediction model, and acquires the product quality prediction data output from the quality prediction model;
a model learning unit that updates parameters included in the quality prediction model based on an error between the product quality prediction data and the product quality data included in the learning data;
a model output unit that outputs the trained quality prediction model in which the parameters have been updated by the model training unit ,
the model input/output unit inputs the sensor data to the first model, acquires latent variables expressing features of the sensor data from the first model, inputs the latent variables and the physical property data to the second model, and acquires the product quality prediction data output from the second model.
Model generation device. The first model is an encoder-decoder model including a neural network and configured with an encoder and a decoder, the encoder generates the latent variables based on the sensor data, and the decoder generates output data based on the latent variables;
the model input/output unit inputs the sensor data to the encoder and acquires the output data output from the decoder;
the model learning unit updates parameters of the first model based on an error between the output data and the sensor data input to the encoder, and updates parameters of the second model based on an error between the product quality prediction data output from the second model and the product quality data.
The model generating device according to claim 1 . the sensor data includes data acquired in a time series for each item in a manufacturing process of the product,
The encoder-decoder model is configured by any one of a recurrent neural network (RNN), a long short-term memory (LSTM), a transformer, and a gated recurrent unit (GRU),
The model input/output unit sequentially inputs the sensor data to the encoder in time series.
The model generating device according to claim 2 . The sensor data includes at least one of a temperature and a pressure in a manufacturing process,
The physical property data is data indicative of at least one of physical properties and characteristics of a raw material associated with the raw material of a product.
The model generating device according to any one of claims 1 to 3 . A model generation method for a model generation device that generates a quality prediction model for predicting product quality of a plurality of types of products to be manufactured by machine learning, comprising:
the quality prediction model includes a neural network, receives sensor data and physical property data as input, and outputs product quality prediction data representing a predicted product quality; and includes a first model and a second model;
learning data used in machine learning of the quality prediction model includes pairs of sensor data and product quality data;
the sensor data is data including inspection values of one or more detection items that are detected in a manufacturing process of one unit quantity of the product and indicate a manufacturing state,
the product quality data is data indicating a product quality of the product manufactured in a manufacturing process in which the corresponding sensor data is detected,
The physical property data is data associated with each product and indicates characteristics of the product;
The model generation method includes:
a model input/output step of inputting the sensor data included in the learning data and the physical property data of the product corresponding to the learning data into the quality prediction model, and acquiring the product quality prediction data output from the quality prediction model;
a model learning step of updating parameters included in the quality prediction model based on an error between the product quality prediction data and the product quality data included in the learning data;
a model output step of outputting the trained quality prediction model in which the parameters have been updated in the model training step ,
In the model input/output step, the sensor data is input to the first model, latent variables expressing features of the sensor data are obtained from the first model, the latent variables and the physical property data are input to the second model, and the product quality prediction data output from the second model is obtained.
Model generation method. A model generation program for causing a computer to function as a model generation device that generates, by machine learning, a quality prediction model for predicting product quality of a plurality of types of products to be manufactured,
the quality prediction model includes a neural network, receives sensor data and physical property data as input, and outputs product quality prediction data representing a predicted product quality; and includes a first model and a second model;
learning data used in machine learning of the quality prediction model includes pairs of sensor data and product quality data;
the sensor data is data including inspection values of one or more detection items that are detected in a manufacturing process of one unit quantity of the product and indicate a manufacturing state,
the product quality data is data indicating a product quality of the product manufactured in a manufacturing process in which the corresponding sensor data is detected,
The physical property data is data associated with each product and indicates characteristics of the product;
The model generation program is configured to:
a model input/output function for inputting the sensor data included in the learning data and the physical property data of the product corresponding to the learning data into the quality prediction model, and acquiring the product quality prediction data output from the quality prediction model;
a model learning function for updating parameters included in the quality prediction model based on an error between the product quality prediction data and the product quality data included in the learning data;
A model output function that outputs the trained quality prediction model in which the parameters have been updated by the model training function ,
The model input/output function inputs the sensor data into the first model, acquires latent variables expressing features of the sensor data from the first model, inputs the latent variables and the physical property data into the second model, and acquires the product quality prediction data output from the second model.
Model generator. A quality prediction model trained by machine learning for causing a computer to function and predicting product quality of multiple types of products to be manufactured,
The method includes a neural network, receives sensor data and physical property data as input, outputs product quality prediction data representing predicted product quality, and includes a first model and a second model;
learning data used in machine learning of the quality prediction model includes pairs of sensor data and product quality data;
the sensor data is data including inspection values of one or more detection items that are detected in a manufacturing process of one unit quantity of the product and indicate a manufacturing state,
the product quality data is data indicating a product quality of the product manufactured in a manufacturing process in which the corresponding sensor data is detected,
The physical property data is data associated with each product and indicates characteristics of the product,
The quality prediction model is
The system is constructed by machine learning that updates parameters based on an error between the product quality prediction data output in response to input of the sensor data and the physical property data corresponding to the sensor data, and the product quality data included in the learning data;
a latent variable expressing a feature of the sensor data is output from the first model in response to the sensor data being input to the first model, and the product quality prediction data is output from the second model in response to the latent variable and the physical property data being input to the second model.
A trained quality prediction model. A product quality prediction device that predicts product quality of a plurality of types of products to be manufactured using a quality prediction model constructed by machine learning,
the quality prediction model includes a neural network, receives sensor data and physical property data as input, and outputs product quality prediction data representing a predicted product quality; and includes a first model and a second model;
learning data used in machine learning of the quality prediction model includes pairs of sensor data and product quality data;
the sensor data is data including inspection values of one or more detection items that are detected in a manufacturing process of one unit quantity of the product and indicate a manufacturing state,
the product quality data is data indicating a product quality of the product manufactured in a manufacturing process in which the corresponding sensor data is detected,
The physical property data is data associated with each product and indicates characteristics of the product,
The quality prediction model is
The system is constructed by machine learning that updates parameters based on an error between the product quality prediction data output in response to input of the sensor data and the physical property data corresponding to the sensor data, and the product quality data included in the learning data;
a latent variable expressing a feature of the sensor data is output from the first model in response to the sensor data being input to the first model, and the product quality prediction data is output from the second model in response to the latent variable and the physical property data being input to the second model.
The product quality prediction device includes:
a product quality prediction unit that inputs sensor data detected in a manufacturing process of a product whose product quality is to be predicted and the physical property data of the product corresponding to the sensor data into the quality prediction model, and obtains the product quality prediction data output from the quality prediction model;
a product quality prediction data output unit that outputs the product quality prediction data acquired by the product quality prediction unit;
A product quality prediction device comprising:

Images