JP7040506B2 - Factory line load prediction model creation method, factory line load prediction model, factory line production plan creation method, factory line load prediction model creation program, factory line load prediction model creation device, and factory line production plan creation device - Google Patents

Description

In the present invention, for example, a method for creating an in-factory line load prediction model for predicting a line load in a factory, a method for creating an in-factory line load prediction model, a method for creating an in-factory line production plan, and a factory in which a material is manufactured through a plurality of processes. In-line load prediction model creation program, factory line load prediction model creation device, and factory line production plan creation device.

Conventionally, a technique for predicting a load on a production line has been known in a factory.
For example, in Patent Document 1, products having the same product attributes such as product order information and manufacturing specification information or within a certain range are grouped as a product type group, and the manufacturing load for each product type group created is calculated in the past. We are building a manufacturing load prediction model that is predicted based on actual data.
Further, in Patent Document 2, a process-specific product classification logic for sorting products having the same product attributes or within a predetermined range as the same product based on the manufacturing record order information and the manufacturing record information is provided in the manufacturing process. In addition to creating each product, the manufacturing load prediction model is calculated based on the manufacturing record order information and the manufacturing record information using the product type classification logic for each process. Then, in Patent Document 2, the manufacturing load of each manufacturing process is predicted for the production plan for ordering a new product by using the process-specific product classification logic and the manufacturing load prediction model.

Japanese Patent No. 4757729 Japanese Patent No. 5402621

By the way, in a factory that manufactures products through multiple processes, in a line that produces a variety of small lots of products, the load on the factory line changes depending on the change in the product mix, the process that becomes the bottleneck, and the manufacturing line that passes through. .. In the above-mentioned Patent Documents 1 and 2, statistical processing is performed from the manufacturing record data, and the logic and the model are created based on the result of the statistical processing, but the change in the product mix is not considered. If the techniques of Patent Documents 1 and 2 are used to make predictions in consideration of changes in the product mix, statistical processing becomes complicated. Further, in Patent Documents 1 and 2, a decision tree, which is a kind of machine learning method, is used as a variety classification logic, but time-series elements are not considered and information on time-series changes is reflected. I can't.

The present invention has been made in view of the above problems, and an object thereof is an in-factory line load prediction model capable of predicting the load of each process corresponding to a product mix without requiring complicated statistical processing. To provide a creation method, an in-factory line load prediction model, an in-factory line production plan creation method, an in-factory line load prediction model creation program, an in-factory line load prediction model creation device, and an in-factory line production plan creation device. ..

In the method for creating an in-factory line load prediction model according to the present invention, a production data accumulation step for accumulating production data in the factory including past production results and data for learning are created based on the accumulated production data. It is characterized by including a training data creation step to be performed, and a prediction model creation step of creating a prediction model of a load related to production by machine learning based on the training data created in the training data creation step.

Further, the in-factory line load prediction model according to the present invention is characterized by being created by the in-factory line load prediction model creation method according to the above invention.

Further, in the in-factory line load prediction model according to the present invention, in the above invention, the load prediction model is based on either a locally constructed computer or a machine learning server existing on an external network including a cloud. It is characterized by being created.

Further, the method for creating an in-factory line production plan according to the present invention includes an input step for inputting a prediction start time, a prediction end time, and a production plan after the prediction start time, and the prediction start input in the input step. Machine learning of the load of the line in the factory up to the predicted end time based on the time, the production plan after the predicted start time, and the information about the state of the past factory line including at least the predicted start time. It is characterized by including a load prediction step for predicting using the in-factory line load prediction model, and a production plan correction step for modifying the production plan based on the result obtained in the load prediction step.

Further, in the method for creating an in-factory line load prediction model according to the present invention, in the above invention, the learning data created in the learning data creation step includes at least information on the size and variety of manufactured products and the equipment for predicting. It is characterized by being time-series data including inventory information.

Further, the method for creating an in-factory line load prediction model according to the present invention is characterized in that, in the above invention, there are a plurality of prediction models created by machine learning in the prediction model creation step.

Further, the method for creating an in-factory line production plan according to the present invention has a plurality of in-factory line load prediction models in the above invention, and the load predicted in the load prediction step is a plurality of in-factory line load prediction models. It is characterized in that it is at least one predicted value, or an average value, a maximum value, and a minimum value among each predicted value of the model.

Further, the in-factory line load prediction model creation program according to the present invention is for learning based on the production data accumulation step for accumulating the in-factory production record data including the past production record and the accumulated production record data. A computer is made to execute a training data creation step for creating data of the above, and a prediction model creation step for creating a prediction model of a load related to production by machine learning based on the training data created in the training data creation step. It is characterized by that.

Further, in the in-factory line load prediction model creation program according to the present invention, in the above invention, the prediction model creation step is either a machine learning server existing on a locally constructed computer or an external network including a cloud. It is characterized in that the prediction model of the load is created by the method.

Further, the in-factory line load prediction model creating device according to the present invention is for learning based on the production data storage unit that stores the in-factory production record data including the past production record and the accumulated production record data. It is characterized by having a training data creation unit that creates the data of the above, and a prediction model creation unit that creates a prediction model of the load related to production by machine learning based on the training data created by the training data creation unit. And.

Further, in the in-factory line load prediction model creation device according to the present invention, in the above invention, the prediction model creation unit is in the same system as the production data storage unit and the learning data creation unit, or an external network including a cloud. It is characterized by being provided on the top.

Further, the in-factory line production plan creating device according to the present invention has an input unit for inputting a prediction start time, a prediction end time, a production plan after the prediction start time, and the prediction start input in the input unit. Machine learning of the load of the line in the factory up to the predicted end time based on the time, the production plan after the predicted start time, and the information about the state of the past factory line including at least the predicted start time. It is characterized by including a load prediction unit that predicts using the in-factory line load prediction model, and a production plan correction unit that corrects the production plan based on the results obtained by the load prediction unit.

Factory line load prediction model creation method, factory line load prediction model, factory line production plan creation method, factory line load prediction model creation program, factory line load prediction model creation device, and factory line according to the present invention. According to the production planning device, it is possible to predict the load of each process corresponding to the product mix without requiring complicated statistical processing.

FIG. 1 is a block diagram showing a configuration of a management system including a production system according to an embodiment of the present invention. FIG. 2 is a diagram illustrating an example of a production line in a production system according to an embodiment of the present invention. FIG. 3 is a diagram for explaining the feature amount of the input material used when the production system according to the embodiment of the present invention predicts the state amount. FIG. 4 is a diagram illustrating a feature amount of a line used when a production system according to an embodiment of the present invention predicts a state amount. FIG. 5 is a flowchart illustrating an example of learning data creation processing performed by the production system according to the embodiment of the present invention. FIG. 6 is a diagram illustrating a state quantity performed by the production system according to the embodiment of the present invention. FIG. 7 is a diagram illustrating the creation of a predictive model by learning using a neural network. FIG. 8 is a diagram illustrating a neural network having three layers of weights. FIG. 9 is a diagram illustrating the generation of load prediction values using a neural network. FIG. 10 is a diagram illustrating evaluation of a model by cross-validation. FIG. 11 is a diagram illustrating an example of state quantity prediction performed by the production system according to the embodiment of the present invention. FIG. 12 is a diagram illustrating an example of state quantity prediction performed by the production system according to the embodiment of the present invention. FIG. 13 is a diagram illustrating an example of state quantity prediction performed by the production system according to the embodiment of the present invention. FIG. 14 is a diagram showing an example of the result of prediction created by the production system according to the embodiment of the present invention. FIG. 15 is a diagram showing an example of the result of prediction created by the production system according to the embodiment of the present invention. FIG. 16 is a flowchart illustrating an example of a predictive model creation process performed by the production system according to the embodiment of the present invention. FIG. 17 is a flowchart illustrating an example of load prediction processing performed by the production system according to the embodiment of the present invention. FIG. 18 is a block diagram showing a configuration of a management system including a production system according to a modification of the embodiment of the present invention.

Hereinafter, a production system including an in-factory line load prediction model creating device and an in-factory line production plan creating device, which is an embodiment of the present invention, will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing a configuration of a management system including a production system according to an embodiment of the present invention. The production system 1 shown in FIG. 1 is a system provided in a factory that produces products and manages the production of products. The production system 1 includes an internal memory such as a ROM (Read Only Memory) in which various programs are pre-installed, a RAM (Random Access Memory) for recording calculation parameters of each process, and a CPU (Central Processing Unit). It is realized by using a computer equipped with a general-purpose processor. The internal memory stores various programs including a program for the production system 1 to execute a factory line load prediction model creation method, a factory line production plan creation method, and a factory line load prediction model creation program. It is also possible to record various programs on a computer-readable recording medium such as a hard disk, a flash memory, a CD-ROM, or a DVD-ROM and distribute them widely. In addition, various programs can also be acquired by downloading via a telecommunication line (network). The telecommunication line (network) referred to here is realized by, for example, an existing public line network, LAN (Local Area Network), WAN (Wide Area Network), etc., and may be wired or wireless. The production system 1 is composed of one or a plurality of computers.

Here, the production line in the production system 1 will be described with reference to FIG. FIG. 2 is a diagram illustrating an example of a production line in a production system according to an embodiment of the present invention. In the production line shown in FIG. 2, the first treatment step P1, the second treatment step P2, the third treatment step P3, the fourth treatment step P4, the fifth treatment step P5, and the warehouse delivery step WH are carried out. Will be done. In the second processing step P2 to the fifth processing step P5, the work-in-process produced in the previous process is set aside in the inventory storage area (inventory storage area SY2 to SY5), and the manufacturing process is performed using the work-in-process items stocked in the inventory storage area. Will be.
Each treatment step is a step of manufacturing a product or carrying out a certain treatment by a manufacturing facility or a treatment facility. For example, in a steel mill, in addition to heat treatment of slabs by a heating furnace and rolling treatment by a rolling mill, heat treatment and surface treatment. , Annealing, cooling, shearing, etc. are carried out.

Specifically, the material is charged and the work-in-process produced in the first processing step P1 is sent to the second processing step P2 and / or the third processing step P3.
The work-in-process produced in the second treatment step P2 and / or the third treatment step P3 is sent to the fourth treatment step P4, the fifth treatment step P5 and / or the warehouse delivery step WH.
The work-in-process produced in the fourth processing step P4 is sent to the fifth processing step P5 and / or the warehouse delivery process WH.
The work-in-process produced in the fifth processing step P5 is sent to the warehouse delivery step WH.
That is, in the production line shown in FIG. 2, the product is sent to the warehouse delivery process WH by any one of the following patterns 1 to 6.
1. 1. 1st processing process P1 → 2nd processing process P2 → Warehouse delivery process WH
2. 2. 1st processing process P1 → 2nd processing process P2 → 4th processing process P4 → Warehouse delivery process WH
3. 3. 1st processing process P1 → 2nd processing process P2 → 4th processing process P4 → 5th processing process P5 → Warehouse delivery process WH
4. 1st processing process P1 → 3rd processing process P3 → Warehouse delivery process WH
5. 1st processing process P1 → 3rd processing process P3 → 4th processing process P4 → Warehouse delivery process WH
6. 1st processing process P1 → 3rd processing process P3 → 4th processing process P4 → 5th processing process P5 → Warehouse delivery process WH
In this way, the product produced via the predetermined processing process is sent to the warehouse carry-in process WH and stored in the warehouse.

Returning to FIG. 1, the production system 1 includes a production data storage unit 11, a learning data creation unit 12, a prediction model creation unit 13, a learning model storage unit 14, an input unit 15, and a load prediction unit 16. A result output unit 17 and a production plan correction unit 18 are provided. Further, the production system 1 (production plan correction unit 18) outputs information to the host system 19. In the production system 1, some configurations may be electrically connected to each other via a network. A management system for performing production control is configured by the production system 1 and the host system 19.

The production data storage unit 11 stores production data in the factory including past production results. Specifically, it stores the production volume of the product, data on the material for manufacturing the product, and data on the manufacturing process and the manufacturing line.

The learning data creation unit 12 creates learning data for predicting the line load by using the production data stored in the production data storage unit 11. The training data is data in which a state quantity and a target quantity for predicting a line load are combined. In the present embodiment, the state quantity is from an arbitrary date and time (time t) set by the user in the past or present to an arbitrarily specified date and time (time t—T: T is a retroactive unit time). The feature amount of the input material for each unit time (hereinafter, also referred to as the input material feature amount) and the feature amount in the line (hereinafter, also referred to as the line feature amount). The target amount is a feature amount on the line at time t + 1. The target amount is a feature amount at a time (future) before the current time, or a feature amount at a time when the production record is not accumulated in the production data storage unit 11.
Here, the unit time is the minimum time for which the state change of the production line is to be predicted, and the retroactive unit time T is determined in a unit that makes it easy to grasp before and after the change in the product mix. For example, it is desirable that the retroactive time T is set to be at least twice the predicted time unit and at least half the change time of the lot unit. When predicting the state of the production line in 1-hour units, the retroactive unit time T is set to 2 hours or more, and when the lot unit of the input material changes every 8 hours, the retroactive unit time T is set to 4 hours or more. Set. By determining the retroactive time in this way, a predictive model that takes into account changes in the product mix is constructed.

The feature amount of the material to be charged is the amount (weight) of the material to be charged between the time determined based on the set time and the time when the unit time has elapsed. FIG. 3 is a diagram illustrating the feature amount of the input material used when the in-factory line load prediction model creating apparatus according to the embodiment of the present invention predicts the state amount. As shown in FIG. 3, the characteristic amount of the input material is an amount defined for the size, variety, manufacturing conditions, and state of the material before processing or the material in the middle of processing (the product produced in each processing step). be. For example, when the material is a plate, the size is a classification of the plate thickness and the plate length according to the size. The varieties are defined by grouping materials with similar attributes in advance, or classified using the classification of varieties used in the factory. The manufacturing conditions are classified according to the heating temperature and the rolling method. The state is a classification of states such as hot pieces and cold pieces.
In the present embodiment, a value calculated based on this information and the number and weight of the charged material within each time is used as the feature amount of the charged material. In the present embodiment, the number or weight of preset items at each time is used as the feature amount of the input material. The sum of the number of sheets in each size, each type, each manufacturing condition, each state, or the sum of the weights classified by each feature amount is equal to the total number of input materials or the total weight scheduled within each time. It shall be.

FIG. 4 is a diagram illustrating a feature amount of a line used when a production system according to an embodiment of the present invention predicts a state amount. As shown in FIG. 4, the feature amount in the line is a value calculated based on the processing amount (number or weight) of each processing step and the amount of inventory (number or weight). In the present embodiment, the number or weight of the processed amount and the number or weight of the stock amount at each time are output as the feature amount in the line.
In the present embodiment, the feature amount treated as learning data is time-series data including the input material feature amount and the line feature amount at each time.

FIG. 5 is a flowchart illustrating an example of learning data creation processing performed by the production system according to the embodiment of the present invention. The learning data creation unit 12 determines the conditions for creating the learning data input via the input unit 15 (steps S101 to S103).

The learning data creation unit 12 first determines the data creation start time t (step S101). The learning data creation unit 12 determines the data creation start time t based on the time input via the input unit 15.

Further, the learning data creation unit 12 determines a unit time for extracting the actual value of the input material or the like (step S102). The learning data creation unit 12 determines the unit time based on the time input via the input unit 15. Here, the unit time will be described as 1.

Further, the learning data creation unit 12 determines the retroactive unit time T from the data creation start time t (step S103). The learning data creation unit 12 determines the retroactive unit time T based on the time input via the input unit 15.
The above-mentioned steps S101 to S103 may be executed at the same time, or may be executed in an order different from the above-mentioned order. Further, a preset value may be used for the unit time and the retroactive unit time T.

The learning data creation unit 12 creates learning data from the time (t-T) to the time t after determining the learning data creation conditions (step S104). The learning data creation unit 12 creates learning data at time t'= (t-T) with reference to the production data storage unit 11.

When the learning data at time t'is created, the created learning data is stored in a buffer or the like (not shown) (step S105).

After that, the learning data creation unit 12 updates the time t'= t'+ 1 (step S106).

The learning data creation unit 12 determines whether or not actual data such as production results exist at the updated time t'with reference to the production data storage unit 11 (step S107). When the learning data creation unit 12 determines that there is actual data such as production results at the updated time t'(step S107: Yes), the process proceeds to step S104, and the updated time t'is described above. Repeat the process. On the other hand, when the learning data creation unit 12 determines that the actual data such as the production actual does not exist at the updated time t'(step S107: No), the learning data creation process ends.

By the training data creation process described above, for example, when there are N times separated by a unit time between the time (tT) where the actual data exists and the time t, N sets of state quantities (input material). Feature amount and line feature amount) are created.
Further, the learning data creation unit 12 sets the learning data at the time t + 1 in which the unit time is added to the set time t and the actual data does not exist as the target amount generated by the prediction described later. ..

FIG. 6 is a diagram illustrating a state quantity performed by the production system according to the embodiment of the present invention. As shown in FIG. 6, the actual value of the state amount of the line at each time from time (tT) to time t (actual state amount of the line) and the actual value of the characteristic amount of the input material for each unit time (input). A state quantity that is a set of material feature quantities (actual results) is created as training data. Further, when predicting the time t + 2, the predicted value of the state quantity of the line at the predicted time t + 1 (predicted value of the state quantity of the line: corresponding to the target quantity) is added as learning data.

The prediction model creation unit 13 creates a prediction model using the training data created by the training data creation unit 12. In this embodiment, a prediction model is created by learning using a neural network. Neural network learning is realized by using development environments such as Chainer and Tensorflow.
The learning model storage unit 14 stores K models after learning by the prediction model creation unit 13.

First, a neuron model used in the modeled learning device will be described with reference to FIG. 7. FIG. 7 is a diagram illustrating the creation of a predictive model by learning using a neural network. As shown in FIG. 7, the neuron outputs an output Y for a plurality of inputs X (here, as an example, inputs X1 to X3). Each input X is multiplied by a weight w corresponding to this input X. As a result, the neuron outputs the output Y expressed by the following equation (1). The input X, the output Y, and the weight w are all vectors. Here, B is a bias and F is an activation function.
Y = F (Σ _i Xi * wi + B) ・ ・ ・ (1)

Next, a neural network having three layers of weights in which the above-mentioned neurons are combined will be described with reference to FIG. FIG. 8 is a diagram illustrating a neural network having three layers of weights. As shown in FIG. 8, a plurality of inputs X (here, inputs X11 to X13 as an example) are input from the left side of the neural network, and a plurality of outputs Y (here, outputs Y11 to outputs Y13 as an example) are output from the right side. Will be done. In the neural network shown in FIG. 8, a plurality of outputs Y are output via the input layer D1, the intermediate layer D2, and the output layer D3.

Specifically, the inputs X11 to X13 are input with corresponding weights applied to each of the three neurons N11 to N13 in the input layer D1. The weights applied to these inputs are collectively referred to as w11.

The neurons N11 to N13 output the feature vector Z11 to the feature vector Z13, respectively. These feature vectors Z11 to feature vectors Z13 are collectively labeled as Z1. This feature vector Z1 is a feature vector between the weight w11 and the weight w12.
The feature vector Z11 to the feature vector Z13 are input with corresponding weights applied to each of the two neurons (N21, N22) in the intermediate layer D2. The weights applied to these feature vectors are collectively marked as w12.

The neurons N21 and 22 output the feature vectors Z21 and Z22, respectively. These feature vectors Z21 and Z22 are collectively labeled as Z2. This feature vector Z2 is a feature vector between the weight w12 and the weight w13.
The feature vectors Z21 and Z22 are input with corresponding weights applied to each of the three neurons (N31 to N33) in the output layer D3. The weights applied to these feature vectors are collectively referred to as w13.

Finally, the neurons N31 to N33 output outputs Y11 to Y13, respectively.
Here, the weights w11 to w13 can be learned by the error back propagation method (backpropagation). Error information is input from the right side and flows to the left side. The error back propagation method is a method of adjusting (learning) the weights of each neuron so as to reduce the difference between the output Y when the input X is input and the true output Y (teacher). ..

FIG. 9 is a diagram illustrating the generation of load prediction values using a neural network. In the present embodiment, the learning data (input material feature amount and line feature amount) from the created time t to time t as shown in FIG. 9 is input to the input layer and output via the intermediate layer. A neural network that outputs load prediction values (processing amount and inventory amount) at time t + 1 is constructed and trained in the layer.

Further, in the present embodiment, in order to avoid overfitting, which is a weak point of the neural network, a model is created by evaluating the model by cross-validation. FIG. 10 is a diagram illustrating evaluation of a model by cross-validation. In FIG. 10, the unhatched squares show the training data and the hatched squares show the verification data.

The prediction model creation unit 13 divides L training data into K data, and arranges verification data while changing the arrangement of K-1 training data and one verification data as shown in FIG. Create K models (model 1 to model K) that are different from each other. In this method, K models are created and evaluated using the average value of K model outputs. The prediction model creation unit 13 stores the K models for which learning has been completed in the learning model storage unit 14.
A prediction model is created by executing the process described above.
The predictive model is created by the machine learning described above, and is provided for each production line, for example. That is, in a factory having a plurality of production lines, a plurality of prediction models are generated.

The input unit 15 is realized by using a user interface such as a keyboard, a mouse, a switch, and a touch panel, and receives inputs of various signals such as a signal instructing the operation of the production system 1.

The load prediction unit 16 predicts the load of the line using the prediction model created by the prediction model creation unit 13. When the load prediction unit 16 performs load prediction from time t + 1 to time t + 3 using the prediction model based on the above-mentioned training data from time t to time t, for example, the load prediction unit 16 is a scheduled input material from time t to time t + 1. Time t + 1 by inputting the feature amount, the feature amount of the scheduled input material from time t + 1 to time t + 2, and the feature amount of the scheduled input material from time t + 2 to time t + 3 from an externally created production plan, etc. Predict the load up to t + 3. Further, the execution of the prediction by the load prediction unit 16 is realized by using a development environment such as Chainer or Tensorflow.

Subsequently, a method of load prediction using a prediction model by the load prediction unit 16 will be described. 11 to 13 are diagrams illustrating an example of state quantity prediction performed by the production system according to the embodiment of the present invention. FIG. 11 is a diagram illustrating an outline in the case of performing load prediction at time t + 3 three unit times ahead of the prediction start time t. FIG. 12 is a diagram illustrating an outline in the case of performing load prediction at time t + 1 one unit time ahead of the prediction start time t. FIG. 13 is a diagram illustrating an outline in the case of performing load prediction at time t + 2 two unit times ahead of the prediction start time t.

(1) Prediction from time t to time t + 1 State quantity (actual value) of the line from time t-T to time t and feature quantity (actual value) of input material from time t-T to time t-1 ) Is read from the production data storage unit 11.
As the feature amount of the input material from the time t to the time t + 1, the feature amount of the scheduled input material is used.
These data are state quantities that are input when making predictions from time t to time t + 1 (see FIG. 12).

The load prediction unit 16 reads out K prediction models from the learning model storage unit 14, inputs a state quantity to be input when making predictions up to time t + 1 for each prediction model, and inputs K prediction values. Let the average value of be the load prediction value at time t + 1. As the load prediction value, in addition to the average value, one of K predicted values may be selected, or the maximum value, the minimum value, and the mode value may be used as the load prediction value.

(2) Prediction from time t + 1 to time t + 2 State quantity (actual value) of the line from time t- (T-1) to time t and input material from time t- (T-1) to time t The feature amount (actual value) is read from the production data storage unit 11.
As the feature amount of the input material from the time t to the time t + 1, the feature amount of the scheduled input material is used.
As the state quantity of the line at time t + 1, the predicted value of the predicted line state quantity at time t + 1 is used.
These data are state quantities that are input when making predictions from time t to time t + 1 (see FIG. 13).

The load prediction unit 16 reads out K prediction models from the learning model storage unit 14, inputs a state quantity to be input when making predictions up to time t + 1 for each prediction model, and inputs K prediction values. Let the average value of be the load prediction value at time t + 2.

The load prediction unit 16 predicts from time t + 2 to time t + 3 in the same manner as above, and determines the feature amount of each scheduled input material from time t, t + 1, t + 2 and the predicted value of the state amount of the line at time t + 1, t + 2. It is used to generate a predicted value at time t + 3.
Even if the prediction period is arbitrarily input, the prediction can be performed in the same manner as above.

The load prediction unit 16 outputs the generated predicted value to the prediction result output unit 17 and / or the host system 19.

The prediction result output unit 17 outputs a prediction result based on the prediction value generated by the load prediction unit 16. The prediction result output unit 17 outputs the data to the outside and displays the prediction result on the monitor. When displaying the prediction result on the monitor, the prediction result output unit 17 is configured by using a monitor such as a liquid crystal display or an organic EL (Electro Luminescence).

14 and 15 are diagrams showing an example of the result of prediction made by the production system according to the embodiment of the present invention. FIG. 14 shows the relationship between the time and the inventory amount (time t + 1 to t + 3 are predicted values) in each processing step (second processing step P2 to fifth processing step P5). FIG. 15 shows each processing step (first processing step). The relationship between the time and the processing amount (time t + 1 to t + 3 are predicted values) in the 2 processing steps P2 to the 5th processing step P5) is shown.

Here, the point Q shown in FIG. 14 indicates that the work-in-process inventory of the second processing step P2 exceeds the maximum at time t + 1. As in the example shown in FIG. 14, when the work-in-process inventory is exceeded in the processing process, the production plan correction unit 18 corrects the input production plan.

The production plan correction unit 18 generates information for correcting the production plan, for example, when the work-in-process inventory is exceeded in a certain processing process.

Here, in the prediction model (in-process line load prediction model) according to the present embodiment, the characteristic quantities of the input materials at the times τ, τ-1, ..., τ-T are X (τ), X (τ). -1) ..., X (τ-T), time τ, τ-1, ..., Processing amount P (m, τ), P (m, τ-, which is the state quantity of the processing step m in τ-T. 1), ..., P (m, τ-T), time τ, τ-1, ..., In-process inventory amount S (n, τ), S (n) which is the state quantity of the processing step n at time τ, τ-1, ..., τ-T , Τ-1), ..., S (n, τ-T) are input, the processing amount P (m, τ + 1) of the processing step m at the time τ + 1, and the in-process inventory amount S (n, It can be regarded as a function that outputs τ + 1). When the prediction model is NN (), this prediction model NN () is expressed as the following equation (2). For example, when there are five processing steps as described above, m = 1,2,3,4,5, n = 2,3,4,5.
NN (X (τ), ..., X (τ-T), P (m, τ), ...,
P (m, τ-T)), S (n, τ), ..., S (n, (τ-T))
= (P (m, τ + 1), S (n, τ + 1)) ... (2)

Next, specifically, assuming that the maximum processing amount in each processing process is MaxP (m) and the maximum amount of work-in-process inventory is MaxS (n), the prediction from time t + 1 to time t + 3 is performed in each manufacturing process. The constraint conditions of the processing amount and the work-in-process inventory amount can be expressed by the following equations (3) and (4).
For τ = t + 1, ..., t + 3;
For all m;
P'(m, τ) -MaxP (m) <= 0 ... (3)
Here, P'(m, τ) is the processing amount of each processing step obtained by substituting the feature amount X'(τ) of the input material corrected at each time into the above equation (2).
For τ = t + 1, ..., t + 3;
For all n;
S'(n, τ) -MaxS (n) <= 0 ... (4)
Here, S'(n, τ) is a work-in-process inventory amount in each processing step obtained by substituting the feature amount X'(τ) of the input material corrected at each time into the above equation (2). be.

The production plan correction unit 18 obtains X'(τ) that maximizes the lower equation (5) under the constraints of the above equations (3) and (4). As a result, it can be formulated as a feature optimization problem.
For τ = t + 1, ..., t + 3;
Σ _τ G (X'(τ)) → Max ・ ・ ・ (5)

In this embodiment, since the prediction model NN () that makes a prediction is configured by a neural network, it becomes a kind of black box function optimization problem. Regarding the black box function optimization problem, it is known that evolutionary calculations such as genetic algorithms and methods such as Bayesian optimization are applied, and X'(5) that maximizes the above equation (5) by using these. τ) can be obtained.
Here, in this embodiment, an example of predicting and correcting in the range of τ = t + 1, ..., T + 3 is shown, but for any Tm, τ = t + 1, ..., T + Tm. Prediction intervals can be set appropriately and the plan can be revised.

The production plan correction unit 18 outputs the corrected feature amount X'(τ) at each time to the host system 19.

The host system 19 is a system that is realized by using, for example, a computer equipped with a general-purpose processor such as a CPU and a memory, and manages a production plan and the like. The host system 19 generates a production plan based on the correction information based on the predicted value received from the load prediction unit 16 and the corrected feature amount X'(τ) received from the production plan correction unit 18, and performs each processing process. Let it be carried out.

An example of the flow of processing executed by the production system 1 described above will be described with reference to FIGS. 16 and 17. FIG. 16 is a flowchart illustrating an example of a predictive model creation process performed by the production system according to the embodiment of the present invention.

First, when the production data in the factory including the past production results is input via the input unit 15 or the like, the input production data is accumulated in the production data storage unit 11 (step S201: production data storage step). ..

The learning data creation unit 12 creates learning data based on the accumulated production data (step S202: learning data creation step). The learning data creation unit 12 creates learning data as described with reference to FIGS. 3 to 5. The created learning data is stored in the learning model storage unit 14.

As described above, the prediction model creation unit 13 creates a prediction model of the load related to production by machine learning based on the learning data created by the training data creation unit 12 (step S203: prediction model creation step).

After that, load prediction processing using the created prediction model is performed. In the load prediction process, the load is predicted from the prediction model and the production plan is created. FIG. 17 is a flowchart illustrating an example of load prediction processing performed by the production system according to the embodiment of the present invention. In the load prediction process, first, the prediction start time (corresponding to the above-mentioned time t) and the prediction end time (for example, corresponding to the above-mentioned time t + 3) are input via the input unit 15, and the production data storage unit 11 inputs. The production plan (feature amount of input material, etc.) after the prediction start time is input (step S301: input step).

After the prediction start time and the production plan after the prediction start time are input, the load prediction unit 16 performs information on the state of the past factory line including the prediction start time (here, input material feature amount and line feature amount). ), The load of the line in the factory up to the prediction end time is predicted using the line load prediction model in the factory (step S302: load prediction step). At this time, the prediction result output unit 17 may display the prediction result on a monitor or the like based on the generated prediction value.

Upon receiving the prediction result of the load prediction unit 16 via the prediction result output unit 17, the production plan correction unit 18 determines whether or not it is necessary to correct the production plan (step S303). The production plan correction unit 18 compares, for example, the predicted value (predicted value of the processed amount or the inventory amount) at each time with a preset threshold value. Here, when the production plan correction unit 18 determines that the predicted value exceeds the threshold value (step S303: Yes), the production plan correction unit 18 proceeds to step S304. On the other hand, when the production plan correction unit 18 determines that the predicted value does not exceed the threshold value (step S303: No), the production plan correction unit 18 proceeds to step S306.

In step S304, the production plan correction unit 18 obtains X'(τ) that maximizes the lower equation (5) under the constraints of the upper equations (3) and (4). The production plan correction unit 18 outputs the corrected feature amount X'(τ) at each time to the host system 19 as production plan correction information.
In addition to the correction necessity judgment by the production plan correction unit 18, it is determined that the user who confirmed the prediction result output by the prediction result output unit 17 has input the correction information of the production plan via the input unit 15 or the like. In this case, the input information may be output to the host system 19 as production plan correction information. In this case, the prediction result output unit 17 may perform determination processing and output processing, or the control unit (not shown) that collectively controls the production system 1 may perform the above-mentioned processing.

In step S305 following step S304, the host system 19 corrects the production plan based on the acquired production plan correction information. The host system 19 may automatically correct the production plan based on the acquired production plan correction information, or may correct the production plan by reflecting the information input by the user based on the production plan correction information. You may do it.
The processing of step S304 and step S305 corresponds to the production plan modification step.

The host system 19 causes each processing step to be executed based on the production plan (step S306). At this time, when the production plan is revised, the host system 19 causes the processing to be carried out in the revised production plan.

In the embodiment described above, the load predicted value is generated based on the training data created using the actual data and the predicted value at the time before the predicted time. Therefore, for example, in a certain production line. The load (load related to processing amount and inventory amount) can be predicted and reflected in the production plan. According to this embodiment, the load of each process corresponding to the product mix can be predicted by simple statistical processing. By predicting the load on the production line of the factory, it is possible to grasp the equipment that exceeds the capacity in advance, and by taking measures such as changing the production plan, it becomes possible to perform appropriate production such as reduction of intermediate inventory.

Further, in the present embodiment, the production plan correction unit 18 generates information for correcting the production plan when, for example, the in-process inventory is exceeded in a certain processing process, so that the process has a large load. If there is, the production plan can be modified to carry out appropriate production.

(Modification example)
In the above-described embodiment, a neural network on a computer constructed in the production system 1 (locally) is used for each process of the prediction model creation unit 13 and the load prediction unit 16. In recent years, a technique for constructing and executing a machine learning model on the cloud has been known, and this technique can also be applied to the processing of the above-described embodiment. When the technique of executing the process on the cloud described above is applied to the above-described embodiment, it is not necessary to have the programs and storage areas of the prediction model creation unit 13, the learning model storage unit 14, and the load prediction unit 16 in FIG. Therefore, there is an advantage that the system configuration as a production system is simplified and maintenance is easy.

FIG. 18 is a block diagram showing a configuration of a management system including a production system according to a modification of the embodiment of the present invention. The production system shown in FIG. 18 includes a production data storage unit 11, a learning data creation unit 12, a prediction model creation unit 13, an input unit 15, a prediction result output unit 17, a production plan correction unit 18, and prediction data. It is provided with a creating unit 20. Further, the production plan correction unit 18 outputs information to the host system 19. The management system is configured by the above-mentioned production system, higher-level system 19, and machine learning server 21.

The input unit 15 inputs data for performing load prediction to the prediction data creation unit 20. The prediction data creation unit 20 creates prediction data to be used for load prediction based on the data input via the input unit 15. The prediction data includes, for example, the time when the prediction is made.

The machine learning server 21 creates a prediction model based on the learning data. Further, the machine learning server 21 performs load prediction using the input prediction data and the created prediction model. The machine learning server 21 has the functions of the prediction model creation unit 13 and the load prediction unit 16 described above.

The learning data creation unit 12 transmits the above-mentioned learning data to the machine learning server 21 that performs machine learning via the external network environment (network Nt1). Further, the machine learning server 21 builds a prediction model in the server using the acquired learning data. Further, when performing load prediction, the prediction data creation unit 20 transmits the prediction data to the machine learning server 21 via the external network environment (network Nt2). The machine learning server 21 performs load prediction using the acquired prediction data and the constructed prediction model. The prediction result output unit 17 receives the load prediction result from the machine learning server 21 via the external network environment (network Nt3). The production plan correction unit 18 is processed according to the flowchart of FIG. 17, and the modified production plan is executed by the host system 19. The production plan corrected by the production plan correction unit 18 is output to the machine learning server 21 via the network Nt3.

Even when the production plan is modified using an external machine learning service, the feature quantity X'(τ) that optimizes equation (5) is used as a black box function optimization problem regardless of the form of the prediction model. You can ask.

Although the embodiment to which the invention made by the present inventors has been applied has been described above, the present invention is not limited by the description and the drawings which form a part of the disclosure of the present invention according to the present embodiment. That is, other embodiments, examples, operational techniques, and the like made by those skilled in the art based on the present embodiment are all included in the scope of the present invention.

1 Production system 11 Production data storage unit 12 Learning data creation unit 13 Prediction model creation unit 14 Learning model storage unit 15 Input unit 16 Load prediction unit 17 Prediction result output unit 18 Production plan correction unit 19 Upper system 20 Prediction data creation unit 21 Machine Learning server

Claims (13)

A production data accumulation step that accumulates production data in the factory including past production results,
Learning to create learning data for learning based on the production data up to the retroactive unit time set by more than twice the predicted time unit and more than half of the change time of the lot unit among the accumulated production data. Data creation steps and
Based on the training data created in the training data creation step, a prediction model creation step that creates a prediction model of the load related to production by machine learning, and
A method for creating an in-factory line load prediction model, which comprises. The learning data is data in which a state quantity and a target quantity for predicting a line load are combined.
The state quantity is a feature quantity of the input material for each unit time up to the time retroactive by the retroactive unit time, and a feature quantity of the production line based on the treatment characteristics for each treatment step.
The target amount is a characteristic amount of the production line for each processing process at a time earlier than the current time, or a characteristic amount of the production line for each processing process at a time when the production record is not accumulated.
The method for creating an in-factory line load prediction model according to claim 1. An in-factory line load prediction model, characterized in that it was created by the method for creating an in-factory line load prediction model according to claim 1. The in-factory line load according to claim 3, wherein the load prediction model is created by either a locally constructed computer or a machine learning server existing on an external network including a cloud. Predictive model. An input step for inputting the forecast start time, the forecast end time, and the production plan after the forecast start time ,
Up to the forecast end time based on the forecast start time, the production plan after the forecast start time , and the information about the state of the past factory line including at least the forecast start time, which was input in the input step. , A load prediction step that predicts the load of a line in a factory using a machine-learned line load prediction model in a factory.
A production plan correction step that corrects the production plan based on the results obtained in the load prediction step, and
Including
The in-factory line load prediction model is created based on the production data in the factory up to the retroactive unit time set by more than twice the predicted time unit and more than half of the change time in the lot unit.
A method of creating an in-factory line production plan, which is characterized by this. The training data created in the training data creation step is characterized by being time-series data including at least information on the size and variety of manufactured products, and predicted equipment processing amount and inventory information. The method for creating an in-factory line load prediction model according to claim 1. The method for creating an in-factory line load prediction model according to claim 1, wherein the method has a plurality of prediction models created by machine learning in the prediction model creation step. It has multiple in-factory line load prediction models and has
5. The load predicted in the load prediction step is at least one predicted value, or an average value, a maximum value, and a minimum value among the predicted values of each of the plurality of factory line load prediction models. How to create an in-factory line production plan as described in. The production data accumulation step that accumulates the production record data in the factory including the past production record,
Learning to create learning data based on the production data up to the retroactive unit time set by more than twice the predicted time unit and more than half of the change time of the lot unit among the accumulated production performance data. Data creation steps and
Based on the training data created in the training data creation step, a prediction model creation step that creates a prediction model of the load related to production by machine learning, and
An in-factory line load prediction model creation program characterized by having a computer execute. The ninth aspect of claim 9, wherein the predictive model creation step creates a predictive model of the load by either a locally constructed computer or a machine learning server existing on an external network including a cloud. In-factory line load prediction model creation program. A production data storage unit that stores production performance data in the factory, including past production performance,
Learning to create learning data based on the production data up to the retroactive unit time set by more than twice the predicted time unit and more than half of the change time of the lot unit among the accumulated production performance data. Data creation department and
Based on the learning data created by the training data creation unit, a prediction model creation unit that creates a prediction model of the load related to production by machine learning, and a prediction model creation unit.
An in-factory line load prediction model creation device characterized by being equipped with. The in-factory line load prediction according to claim 11, wherein the prediction model creation unit is provided in the same system as the production data storage unit and the learning data creation unit, or on an external network including a cloud. Modeling device. An input unit for inputting the forecast start time, the forecast end time, and the production plan after the forecast start time ,
Up to the forecast end time based on the forecast start time, the production plan after the forecast start time , and the information about the state of the past factory line including at least the forecast start time, which is input by the input unit. , A load prediction unit that predicts the load of the line in the factory using a machine-learned line load prediction model in the factory.
The production plan correction unit that corrects the production plan based on the results obtained by the load prediction unit, and the production plan correction unit.
Equipped with
The in-factory line load prediction model is created based on the production data in the factory up to the retroactive unit time set by more than twice the predicted time unit and more than half of the change time in the lot unit.
An in-factory line production planning device that is characterized by this.

Images