Feedback system based on causal inference and feedback method based on causal inference

Abstract

To provide feedback for process improvement to increase the efficiency of on-site processes.SOLUTION: A feedback system using causal inference of the present invention includes: a data acquisition unit 121 that acquires sensor data in a manufacturing process 300; a causal inference unit 122 that performs causal inference using a causal structure in which the sensor data is a variable and a conditional probability between variables; an uncontrollable variable avoidance unit 123 that constructs an inference model by replacing an uncontrollable variable with a controllable variable when the result of the causal inference includes an uncontrollable variable; an optimization unit 125 that optimizes each controllable variable in the inference model; and an output unit 126 that feeds back the optimal value of each optimized controllable variable to the manufacturing process.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present invention relates to a feedback system and a feedback method based on causal inference. [Background technology]

[0002] The use of big data using AI (artificial intelligence) is becoming more common across industries. In particular, in the manufacturing industry, there is a growing need to mathematically analyze operational process data and product quality data acquired by sensors, etc., to improve processes with the aim of stabilizing and improving quality. In addition, there is a demand from manufacturing customers to streamline on-site processes by understanding the factors that affect quality.

[0003] Conventionally, in manufacturing plants for manufacturing products, there is a technology in which performance, functional design, production technology, etc. are reviewed, and further operational methods are improved, and new proposals are shared as feedback information among related companies (for example, see Patent Document 1). Patent Document 1 discloses a technology that "optimizes operations throughout the entire life cycle of a plant that manufactures products by taking appropriate measures for periodic plant shutdown maintenance against degradation events that occur over time, searching for and determining optimal measures for specific degradation events that become apparent over time, and implementing the optimal measures." [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2022-177561 A Summary of the Invention [Problem to be solved by the invention]

[0005] Incidentally, the sensor data acquired from related companies and plants may contain variables (factors) that an operator can directly control regarding control parameters, as well as variables (factors) that an operator cannot directly control because the numerical values ​​of the control parameters are determined dependently or indirectly due to the influence of other variables (factors). However, these controllable / uncontrollable variables are not taken into consideration in the conventional technology described in Patent Document 1.

[0006] In related companies and plants, when process improvements are made to improve the efficiency of on-site processes, that is, when feedback is provided for process improvements aimed at stabilizing and improving quality, it is difficult for operators to provide feedback when uncontrollable variables (factors) are involved.

[0007] The present invention has been made in consideration of the above circumstances, and aims to provide a feedback system using causal inference and a feedback method using causal inference that can realize feedback aimed at process improvement to increase the efficiency of on-site processes. [Means for solving the problem]

[0008] The feedback system using causal inference of the present invention for solving the above problems includes a data acquisition unit that acquires sensor data in a manufacturing process, a causal inference unit that performs causal inference using a causal structure and conditional probability in which the sensor data is a variable, an uncontrollable variable avoidance unit that identifies important factors included in the result of the causal inference based on importance, which is an index of the magnitude of influence on the target variable, using the conditional probability, and if the important factor included in the result of the causal inference is an uncontrollable variable, constructs an inference model by replacing the uncontrollable variable with a controllable variable, an optimization unit that optimizes each controllable variable in the inference model, and an output unit that feeds back the optimal value of each optimized controllable variable to the manufacturing process.

[0009] In addition, the feedback method using causal inference of the present invention for solving the above-mentioned problems executes the following steps: acquiring sensor data in a manufacturing process; performing causal inference using a causal structure and conditional probability in which the sensor data is a variable; identifying important factors included in the result of the causal inference based on importance, which is an index of the magnitude of influence on the target variable, using the conditional probability; and, if an important factor included in the result of the causal inference is an uncontrollable variable, constructing an inference model by replacing the uncontrollable variable with a controllable variable; optimizing each controllable variable in the inference model; and feeding back the optimal values ​​of each optimized controllable variable to the manufacturing process. Effect of the Invention

[0010] According to the present invention, based on the causal structure calculated in the process of causal inference and the conditional probability between each variable, it is possible to realize feedback composed of variables (factors) that can be controlled by the operator, aimed at process improvement to make on-site processes more efficient, thereby making it possible to study quality improvement and achieve stable production at less cost.

[0011] Problems, configurations, and effects other than those described above will become apparent from the following description of the mode for carrying out the invention (hereinafter, referred to as an embodiment). [Brief description of the drawings]

[0012] [Figure 1] FIG. 1 is an image diagram illustrating an example of a manufacturing process of a chemical plant to which the present invention is applied. [Diagram 2] 1A and 1B are diagrams illustrating an example of sensor data (manufacturing data) of a manufacturing process and an example of a causal structure. [Diagram 3] 1 is a block diagram showing an example of a system configuration of a feedback system using causal inference according to an embodiment of the present invention; [Figure 4] 1 is a flowchart illustrating an example of a process of a feedback method using causal inference according to an embodiment of the present invention. [Diagram 5] FIG. 13 is a diagram showing an example of a display screen for changes in model accuracy in a process for confirming model accuracy. [Figure 6] FIG. 13 is a diagram showing an example of a model selection screen in the process of model selection and model accuracy confirmation. [Figure 7] FIG. 13 is a diagram showing an example of a display screen of a feedback value to be fed back to a manufacturing process. [Figure 8] 13 is a flowchart illustrating an example of a process of causal inference. [Figure 9] 13A and 13B are diagrams illustrating an example of a confirmation screen for a causal structure and a correction screen for the causal structure. [Figure 10] FIG. 13 is a diagram showing an example of a screen for creating a controllable / uncontrollable variable list. [Figure 11] 13 is a flowchart showing an example of a process for avoiding an uncontrollable variable. [Figure 12] 3A to 3C are diagrams illustrating examples of data structures of manufacturing data, a hierarchical information list, and a causal structure table. [Figure 13] 13A and 13B are diagrams illustrating an example of the data structure of a controllable / uncontrollable variable list and a White / Black list. [Figure 14] 13A and 13B are diagrams illustrating an example of data structures of model accuracy, feedback values, and an important factor list. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] Hereinafter, an embodiment of the present invention will be described with reference to the accompanying drawings. In this specification and the drawings, components having substantially the same functions or configurations are designated by the same reference numerals, and redundant explanations will be omitted.

[0014] <Manufacturing process example> 1 is a schematic image diagram of an example of a production process of a chemical plant to which the present invention is applied. The production process of the chemical plant illustrated here is just one example.

[0015] The manufacturing process of the chemical plant shown in Figure 1 is configured to have, for example, two reaction tanks (tank A and tank B) 11A, 11B, with material 12 being fed into reaction tank (tank A) 11A and product 13 being discharged from reaction tank (tank B) 11B.

[0016] In this manufacturing process, the two reaction vessels (vessel A and vessel B) 11A and 11B have various parameters such as the rotation speed of the motor that drives the stirring blade in the reaction vessel, the pressure in the reaction vessel, the temperature, the liquid level, the reaction time of the reaction vessel, etc. These parameters are detected by sensors such as a rotation speed gauge, a pressure gauge, a thermometer, a liquid level gauge, etc., and can be acquired as sensor data.

[0017] Among the various parameters in the above manufacturing process, in addition to factors (parameters) whose numerical values ​​can be controlled based on the control of the processing conditions in the process, there may be factors whose numerical values ​​cannot be controlled based on the control of the processing conditions in the process. In the following, controllable factors will be referred to as controllable variables, and uncontrollable factors will be referred to as uncontrollable variables.

[0018] In the above manufacturing process, for example, if the temperature in the reaction tank is increased by control, the pressure in the reaction tank will increase accordingly. In this case, the temperature, which is a controlled value, is a controllable variable, and the pressure, which is not a controlled value, is an uncontrollable variable. In the example of the manufacturing process shown in Figure 1, the reaction time of the reaction tank is also an uncontrollable variable because it is the time required for the input material to reach the desired state in the reaction tank according to the environment of the reaction tank and is a time that is determined independently of external control.

[0019] In this type of manufacturing process, the present invention aims to optimize the process, i.e., to set each parameter to an optimal value. To achieve this goal, the present invention acquires sensor data in the manufacturing process and performs causal inference on the causal structure based on the acquired sensor data. An example of the sensor data (manufacturing data) is shown in FIG. 2A, and an example of the causal structure is shown in FIG. 2B.

[0020] In the present invention, after performing causal inference, if an important factor included in the result of the causal inference is an uncontrollable variable, the uncontrollable variable is replaced with a controllable variable to construct an inference model, and each controllable variable of the inference model is optimized.Then, the optimal value of each optimized controllable variable is fed back to the manufacturing process.

[0021] In the present invention, after performing causal inference, if an important factor included in the result of the causal inference is an uncontrollable variable, the uncontrollable variable is replaced with a controllable variable to construct an inference model, and each controllable variable of the inference model is optimized.Then, the optimal value of each optimized controllable variable is fed back to the manufacturing process.

[0022] The feedback system and the feedback method based on causal inference of the present invention will be described in detail below.

[0023] <Feedback system> A feedback system using causal inference according to one embodiment of the present invention is a feedback system using causal inference aimed at process improvement for improving the efficiency of on-site manufacturing processes, i.e., process improvement aimed at stabilizing and improving quality.

[0024] [System configuration example] FIG. 3 is a block diagram showing an example of the system configuration of a feedback system using causal inference according to one embodiment of the present invention.

[0025] As shown in FIG. 3, the feedback system 100 using causal inference according to this embodiment includes a display unit 110, a control unit 120, a memory unit 130, and a communication unit 140, and is configured to be controllable by a user 200 as necessary.

[0026] (Display) The display unit 110 consists of a liquid crystal display, an organic EL (electroluminescence) display, etc., and has the following display function units: a causal structure unit 111, a controllable / uncontrollable variable list unit 112, an algorithm selection unit 113, a model accuracy unit 114, and a feedback unit 115.

[0027] The causal structure unit 111 has a display function of displaying a causal structure, which is an output result of the causal inference, on a screen. The controllable / uncontrollable variable list unit 112 has a display function of displaying a list of controllable variables and uncontrollable variables on a screen. The algorithm selection unit 113 has a display function of displaying a screen for selecting an algorithm to be used based on an accuracy list when constructing an inference model (hereinafter, may be simply described as a "model") using a plurality of algorithms (DL: Deep Learning, ML: Machine Learning, etc.) and displaying a list of accuracy results. The model accuracy unit 114 has a display function of displaying a screen for analyzing anomaly factors, for example, when the accuracy deteriorates without changing the causal structure, or when the accuracy deteriorates due to a change in the causal structure. The feedback unit 115 has a display function of displaying a feedback value to the manufacturing process 300 on a screen.

[0028] (Control unit) The control unit 120 is composed of a CPU (Central Processing Unit) and the like, and has the following functional units: a data acquisition unit 121, a causal inference unit 122, an uncontrollable variable avoidance unit 123, a model accuracy analysis unit 124, an optimization unit 125, and an output unit 126.

[0029] The data acquisition unit 121 has a function of acquiring sensor data in the manufacturing process 300. The acquisition of this sensor data is performed for the purpose of setting the parameters of the manufacturing process 300 to optimal values, that is, for the purpose of optimizing the process.

[0030] The causal inference unit 122 has a function of performing causal inference based on confirmation by the user 200 as to whether or not the causal structure, which is the output result of the causal inference based on the sensor data acquired by the data acquisition unit 121, matches with a causal structure based on on-site know-how of the manufacturing process 300 (so-called on-site intuition). This causal inference unit 122 allows the user 200 to modify the causal structure based on on-site know-how.

[0031] The uncontrollable variable avoidance unit 123 has a function of avoiding the inclusion of an uncontrollable variable in the result of causal inference. Specifically, when an important factor included in the result of causal inference is an uncontrollable variable, the uncontrollable variable avoidance unit 123 extracts important factors for model construction based on the intended importance of the target uncontrollable variable, and replaces the uncontrollable variable with a controllable variable. Here, model construction means constructing an inference model that predicts quality. Also, the importance means the degree of influence on the objective variable when the value of each variable changes, by using the conditional probability between each variable.

[0032] The model accuracy analysis unit 124 has a function of analyzing anomaly factors, for example, when the accuracy deteriorates without changing the causal structure, or when the accuracy deteriorates due to a change in the causal structure. The analysis result of the anomaly factors can be used to consider the next processing content. For example, the inference model can be updated when the process itself changes, or the inference model can be maintained in the event of temporary data drift due to degradation, and the like, to consider the next processing content.

[0033] The optimization unit 125 has the function of using the inference model created in model construction to perform a process of adjusting the sensor data to a target value (target quality), i.e., an optimization process of setting each controllable variable of the inference model to the optimal value.

[0034] The output unit 126 outputs and displays the optimal value of each controllable variable derived by the optimization process in the optimization unit 125 on the display unit 110. Furthermore, the output unit 126 calculates a predicted value of the objective variable based on the optimal value of each controllable variable, and outputs and displays it on the display unit 110. This allows the numerical value of each controllable variable to be confirmed on the screen of the display unit 110. Then, if the numerical value of each controllable variable is determined to be appropriate, the output unit 126 feeds it back to the manufacturing process 300.

[0035] (Storage part) The memory unit 130 is composed of various storage media for storing data used in the control in the feedback system 100 according to this embodiment, and has a manufacturing data storage unit 131, a causal structure storage unit 132, a controllable / uncontrollable variable list storage unit 133, a model storage unit 134, a model accuracy storage unit 135, and a feedback value storage unit 136.

[0036] The manufacturing data storage unit 131 stores manufacturing data. The causal structure storage unit 132 stores a causal structure. The controllable / uncontrollable variable list storage unit 133 stores a list of controllable / uncontrollable variables. The model storage unit 134 stores an inference model. The model accuracy storage unit 135 stores model accuracy. The feedback value storage unit 136 stores a feedback value fed back to the manufacturing process 300 as a previous feedback value.

[0037] (Communications Department) The communication unit 140 imports sensor data from the manufacturing process 300 and provides it to the data acquisition unit 121 of the control unit 120, and also feeds back to the manufacturing process 300 the optimal values ​​of each controllable variable that are optimized by the optimization unit 125 and output from the output unit 126.

[0038] In the feedback system 100 according to this embodiment having the above configuration, feedback is provided for process improvement in order to improve the efficiency of on-site processes in a manufacturing process such as a chemical plant.

[0039] Specifically, in the feedback system 100 according to this embodiment, the function of the causal inference unit 122 performs causal inference on a causal structure that is an output result of causal inference based on sensor data acquired from the manufacturing process 300. The result of this causal inference makes it possible to correct the causal structure based on on-site know-how. It is also possible to check whether or not the causal structure that is the inference result includes an uncontrollable variable. If an uncontrollable variable is included in the feedback aimed at process improvement to improve the efficiency of on-site processes, it is difficult to realize the feedback.

[0040] Therefore, in the feedback system 100 according to the present embodiment, when an important factor included in the result of causal inference is an uncontrollable variable, a process of replacing the uncontrollable variable with a controllable variable is performed by the function of the uncontrollable variable avoidance unit 123. Specifically, a process of extracting important factors for model construction based on the intended importance of the target uncontrollable variable is repeated, and a process of tracing causality back until the uncontrollable variable is replaced with a controllable variable is performed. This process makes it possible to avoid the inclusion of uncontrollable variables in the result of causal inference and to construct an inference model in which all factors are controllable variables, and therefore it is possible to construct an inference model aimed at process improvement for improving the efficiency of on-site processes.

[0041] Then, in the feedback system 100 according to the present embodiment, an optimization process is performed for each controllable variable of the inference model constructed by replacing uncontrollable variables with controllable variables, and then the optimal value of each controllable variable is fed back to the manufacturing process 300. This makes it possible to realize feedback for process improvement to improve the efficiency of on-site processes, thereby enabling consideration for quality improvement and realization of stable production.

[0042] [Example of feedback processing method] Next, a feedback method executed in the feedback system 100 using causal inference configured as above will be described.

[0043] 4 is a flowchart showing an example of a process of the feedback method using causal inference according to an embodiment of the present invention. A series of processes of the feedback method using causal inference according to the present embodiment is executed under the control of a control unit 120 constituted by a CPU or the like.

[0044] First, the control unit 120 acquires the latest sensor data for each unit, such as a lot, in the manufacturing process 300 (step S11) for the purpose of optimizing the process, and then performs preprocessing (step S12). In this preprocessing, the sensor data is processed so that causal inference and model creation are possible, specifically, processing such as interpolation of missing values ​​in the sensor data is performed.

[0045] Next, the control unit 120 determines whether or not a new model needs to be created (step S13), and if a new model needs not to be created (No in S13), it reads the model created last time (step S14), and then checks the accuracy of the model, i.e., checks whether the accuracy of the model has deteriorated below a predetermined threshold (step S15).

[0046] In the model accuracy confirmation process in step S15, a model accuracy screen for the most recent sensor data is displayed on the display unit 110 by the display function of the model accuracy unit 114. An example of the model accuracy change display screen is shown in Fig. 5. The model accuracy change display screen shown in Fig. 5 displays, for example, a change in model accuracy from previous accuracy AUC: 0.75 to current accuracy AUC: 0.55 in the causal structure of temperature, reaction time, pressure, and quality.

[0047] In the model accuracy change display screen shown in Figure 5, AUC (Area Under the ROC Curve: a machine learning evaluation index) is one of the accuracy indices for binary classification tasks, and takes a value in the range of 0 to 1, with the closer to 1 the better.

[0048] If the control unit 120 confirms in the confirmation process of step S15 that the accuracy of the model has not deteriorated (No in S15), it feeds back each controllable variable of the model, i.e., the same sensor data (control value) as the previous time, as feedback values ​​to the manufacturing process 300 (step S16).

[0049] If the control unit 120 determines in step S13 that a new model needs to be created (Yes in S13), or if it confirms in step S15 that the accuracy of the model has deteriorated (Yes in S15), it proceeds to step S17.

[0050] In step S17, the control unit 120 performs a process of acquiring past data in order to create a new model. Next, the control unit 120 performs causal inference to infer important factors related to quality using the past data (step S18). In this causal inference, a process is performed to confirm whether the causal structure, which is the output result of the causal inference, matches the causal structure based on on-site know-how. The details of the causal inference process will be described later.

[0051] Next, when an uncontrollable variable is included in the important factors that are the inference results of the causal inference, the control unit 120 performs a process of avoiding the uncontrollable variable (step S19). In this impossible variable avoidance process, a process of replacing the uncontrollable variable included in the important factors with a controllable variable is performed. The details of the impossible variable avoidance process of replacing the uncontrollable variable with the controllable variable will be described later.

[0052] Next, the control unit 120 constructs an inference model in which all factors are made up of controllable variables by replacing uncontrollable variables with controllable variables (step S20), and then performs model selection and model accuracy confirmation processing (step S21). In the processing of step S21, a model selection screen is displayed on the display unit 110 by the display function of the model accuracy unit 114. An example of the model selection screen is shown in FIG. 6. On the model selection screen shown in FIG. 6, the accuracy of quality prediction models based on multiple algorithms is displayed in order to select a priority accuracy index. On this display screen, the model (algorithm) to be adopted is selected by operation by the user 200.

[0053] In the model selection screen shown in Figure 6, RMSE (Root Mean Squared Error) is one of the common indicators for regression tasks, with the minimum value being 0, and the closer to 0 the better.

[0054] Next, the control unit 120 performs an optimization process to optimize each controllable variable of the inference model (step S22), and then proceeds to step S16 to feed back the optimized optimal values ​​of each controllable variable to the manufacturing process 300 as feedback values.

[0055] At this time, in the feedback process of step S16, a feedback value screen for each important factor is displayed on the display unit 110 by the display function of the feedback unit 115. An example of the feedback value display screen is shown in Fig. 7. On the feedback value display screen shown in Fig. 7, the predicted value of quality and the set values ​​of the controllable variables (feedback values) are displayed. Among the set values ​​of the controllable variables, values ​​that have changed from the previous input value by more than a threshold value are highlighted (highlighted).

[0056] (Example of causal inference processing) Next, an example of the process of step S18 in Fig. 4, that is, the process of causal inference, will be described. Fig. 8 is a flowchart showing an example of the process of causal inference.

[0057] The control unit 120 performs causal inference using the past data acquired in step S17 (step S181). As a result of this causal inference, a causal structure, that is, a conditional probability for determining the presence or absence of causality between each variable, is output.

[0058] The user 200 judges whether this causal structure is correct or not (step S182). The user 200 judges whether the causal structure derived by the causal inference is correct or not depending on whether it matches a causal structure based on on-site know-how (sense). In step S182, a confirmation screen for the causal inference is displayed on the display unit 110 by the display function of the causal structure unit 111. An example of the confirmation screen for the causal inference is shown in FIG. 9.

[0059] If user 200 determines that the causal structure derived by causal inference does not match the causal structure based on on-site know-how, the causal structure needs to be corrected, and so selects "correct" on the causal inference confirmation screen shown in Figure 9. If user 200 confirms that the causal structure matches, he or she selects "OK".

[0060] If the causal structure derived by the causal inference is incorrect (No in S182), the user 200 corrects the parts of the causal structure derived by the causal inference that differ from the causal structure based on on-site know-how on the causal inference correction screen shown in Fig. 9 (step S183). In this correction work, for example, it is assumed that the user performs an interactive operation on the screen to change the connection of arrows.

[0061] After correcting the causal structure derived by the causal inference, the process returns to step S181 and performs causal inference again.

[0062] If the causal structure derived by causal inference is correct (Yes in S182), the control unit 120 uses the causal structure to extract important factors for model construction based on their importance (degree of influence on the target variable) (step S184), and then creates a list of controllable variables / uncontrollable variables (step S185).

[0063] In step S185, a screen for creating a controllable / uncontrollable variable list is displayed on the display unit 110 by the display function of the controllable / uncontrollable variable list unit 112. An example of the screen for creating a controllable / uncontrollable variable list is shown in Fig. 10. On this screen, the user 200 flags the factors of the causal structure derived by causal inference as to whether they are controllable variables or not.

[0064] As described above, the process of causal inference involves a process of checking whether the causal structure, which is the output result of the causal inference, matches the causal structure based on on-site know-how (intuition). This process makes it possible to modify the causal structure based on on-site know-how.

[0065] (Example of avoiding uncontrollable variables) Next, an example of the process of step S19 in Fig. 4, that is, the process of avoiding an uncontrollable variable, will be described. Fig. 11 is a flowchart showing an example of the process of avoiding an uncontrollable variable.

[0066] The control unit 120 first determines whether or not the important factors extracted in step S184 in FIG. 8 include an uncontrollable variable (step S191), and if an uncontrollable variable is not included (No in S191), ends this process.

[0067] If the important factors include an uncontrollable variable (Yes in S191), the control unit 120 extracts one of the uncontrollable variables (step S192), and then determines whether the extracted uncontrollable variable has a direct cause (step S193). If there is no direct cause (No in S193), the control unit 120 deletes the extracted uncontrollable variable from the important factors (step S194), and then returns to step S191. Here, a direct cause is a cause that is in a higher hierarchy (a hierarchy with a smaller number of layers) than the extracted uncontrollable variable in the causal structure of the extracted uncontrollable variable and has a causal relationship. For example, if quality in FIG. 12 described later is the extracted uncontrollable variable, the layer of quality is 3 according to the hierarchical information list, so that A tank_temperature, B tank_temperature, and B tank_reaction time, which are in layer 1 or layer 2 and have a causal relationship in the causal structure table, correspond to direct causes. Note that the direct cause may be a cause that is in a higher hierarchy than the extracted uncontrollable variable and has a causal relationship.

[0068] If the extracted uncontrollable variable has a direct factor (Yes in S193), the control unit 120 calculates the importance (degree of influence on the objective variable) of the uncontrollable variable based on the conditional probability (step S195), and then extracts one variable with the highest importance from the direct factors (step S196).

[0069] Next, the control unit 120 judges whether or not the extracted variable is a controllable variable (step S197), and if the variable is not a controllable variable (No in S197), the process returns to step S193 and the same process is performed.

[0070] If it is a controllable variable (Yes in S197), the control unit 120 replaces the uncontrollable variable with the controllable variable extracted in step S195 (step S198), then updates the important factors (step S199), and then returns to step S191 to repeat the same processing.

[0071] As described above, in the process of avoiding uncontrollable variables, the process of extracting important factors for model construction based on the intended importance of the target uncontrollable variable is repeated, and causation is traced back until the uncontrollable variable is replaced with a controllable variable. This series of processes makes it possible to avoid the inclusion of uncontrollable variables in the results of causal inference, and therefore makes it possible to build an inference model aimed at process improvement to improve the efficiency of on-site processes.

[0072] (Data Structure) Next, various data structures used in feedback processing based on causal inference will be described.

[0073] FIG. 12 is a diagram showing an example of the data structure of the manufacturing data, the hierarchical information list, and the causal structure table. The manufacturing data is, for example, data such as A tank_temperature, A tank_reaction time, ..., B tank_temperature, B tank_reaction time, ..., and quality in the manufacturing process shown in FIG. 1. The numerical values ​​shown here are examples. The hierarchical information list is a list in which, for each variable when performing causal inference, A tank_temperature, A tank_reaction time, ... is layer 1, B tank_temperature, B tank_reaction time, ... is layer 2, and quality is layer 3. The causal structure table is a structure table showing the causal relationship between each variable, which is the result of causal inference, with 0 / 1, where 0 indicates no causal relationship and 1 indicates a causal relationship, and the variables are connected by arrows.

[0074] Each data of the manufacturing data, the hierarchical information list, and the causal structure table is stored in a predetermined storage unit of the memory unit 130 shown in Fig. 3. For example, each data of the manufacturing data and the hierarchical information list is stored in the manufacturing data storage unit 131, and the data of the causal structure table is stored in the causal structure storage unit 132.

[0075] Fig. 13 is a diagram showing an example of the data structure of the controllable / uncontrollable variable list and the White / Black list. The controllable / uncontrollable variable list is a list indicating whether each variable, which is the result of causal inference, is a controllable variable or an uncontrollable variable, with 1 indicating a controllable variable and 0 indicating an uncontrollable variable. This controllable / uncontrollable variable list corresponds to the controllable / uncontrollable variable list creation screen shown in Fig. 10.

[0076] The White / Black list is a list necessary for causal inference, and places restrictions on the modification of the causal structure. Here, the White list includes factors for which you want to actively connect causal relationships, while the Black list includes factors that you want to cut off the network because they have no causal relationship based on physical phenomena or past experience, such as factors that cause the analysis results to be incorrect due to factors such as a lack of data or the inclusion of outliers. Specifically, this is a list for performing causal inference by entering a "1" for causality when causality is definitely present, and a "0" for no causality when causality is absolutely not present, so that no causality can be found there.

[0077] Each data of the controllable / uncontrollable variable list and the White / Black list is stored in a predetermined storage section of the memory unit 130 shown in Fig. 3. For example, each data of the controllable / uncontrollable variable list and the White / Black list is stored in the controllable / uncontrollable variable list storage section 133.

[0078] Fig. 14 is a diagram showing an example of the data structure of model accuracy, feedback value, and important factor list. For model accuracy, accuracy indexes AUC, RMSE, etc. of each model are stored for each Model_No. 1, 2, .... The numerical values ​​exemplified here are merely examples. For feedback values, numerical values ​​to be fed back for each variable are stored for each Lot_No. 1, 2, .... For the important factor list, variables determined as important factors for each variable are set to "1" and variables other than important factors are set to "0".

[0079] The data on the model accuracy, the feedback value, and the important factor list are stored in a predetermined storage unit of the memory unit 130 shown in Fig. 3. For example, the data on the model accuracy is stored in the model accuracy storage unit 135, and the data on the feedback value and the important factor list are stored in the feedback value storage unit 136. [Explanation of symbols]

[0080] 11A, 11B... reaction tank (tank A, tank B), 12... material, 13... product, 100... feedback system, 110... display unit, 111... causal structure unit, 112... controllable / uncontrollable variable list unit, 113... algorithm selection unit, 114... model accuracy unit, 115... feedback unit, 120... control unit, 121... data acquisition unit, 122... causal inference unit, 123... uncontrollable variable avoidance unit, 124... model accuracy analysis unit, 125... optimization unit, 126... output unit, 130... memory unit, 131... manufacturing data storage unit, 132... causal structure storage unit, 133... controllable / uncontrollable variable list storage unit, 134... model storage unit, 135... model accuracy storage unit, 136... feedback value storage unit, 140... communication unit, 200... user, 300... manufacturing process

Claims

1. a data acquisition unit for acquiring sensor data in a manufacturing process; a causal inference unit that performs causal inference using a causal structure and a conditional probability with the sensor data as a variable; an uncontrollable variable avoidance unit that uses the conditional probability to identify an important factor included in the result of the causal inference based on an importance that is an index of the magnitude of an influence on the target variable, and when the important factor included in the result of the causal inference is an uncontrollable variable, constructs an inference model by replacing the uncontrollable variable with a controllable variable; an optimization unit that optimizes each controllable variable of the inference model; an output unit that feeds back the optimized values ​​of each of the controllable variables to the manufacturing process; A feedback system based on causal inference.

2. Further equipped with a model accuracy analysis unit that analyzes the cause of abnormalities 2. The causal inference feedback system of claim 1.

3. The causal inference unit performs causal inference by confirming whether or not a causal structure based on the sensor data coincides with a causal structure based on on-site know-how of the manufacturing process.

2. The causal inference feedback system of claim 1.

4. The uncontrollable variable avoidance unit extracts important factors based on a desired importance of a target uncontrollable variable when an important factor included in a result of the causal inference is an uncontrollable variable, and constructs an inference model by replacing the uncontrollable variable with the extracted important factor.

2. The causal inference feedback system of claim 1.

5. The uncontrollable variable avoidance unit constructs an inference model by replacing an important factor included in the result of the causal inference with a controllable variable when the important factor is an uncontrollable variable and the uncontrollable variable has a direct cause. The feedback system for causal inference according to claim 1 .

6. acquiring sensor data from a manufacturing process; A step of performing causal inference using a causal structure and conditional probability with the sensor data as variables; a step of identifying an important factor included in the result of the causal inference based on the importance, which is an index of the magnitude of the influence on the objective variable, using the conditional probability, and constructing an inference model by replacing the uncontrollable variable with a controllable variable when the important factor included in the result of the causal inference is an uncontrollable variable; optimizing each controllable variable of the inference model; feeding back the optimized values ​​of each of the controllable variables to the manufacturing process; A feedback method using causal inference to execute each process.

Images