Knowledge Model Execution Device

Abstract

To provide a knowledge model creation support system making it possible to efficiently create a knowledge model.SOLUTION: A knowledge model creation support system 30 includes: an object factor acquisition unit 31 that acquires parent factors A and B and a child factor C as an object factor 31a for use in defining relationship information 24; an observation data acquisition unit 32 that acquires observation data 32a concerning factor values of the parent factors A and B and a factor value of the child factor C; a generation unit 36 that generates a conditional probability table 36b, in which a probability equivalent to the factor value of the child factor C is defined in a case where the factor values of the parent factors A and B are regarded as requirements, on the basis of the observation data 32a; and a knowledge model storage unit 38 that regards the generated conditional probability table 36b as the relationship information 24 among the parent factors A and B and the child factor C, and stores a knowledge model 4a containing the relationship information 24. When acquiring new observation data 32a, the generation unit 36 generates a new conditional probability table 36b and updates the knowledge model 4a stored in the knowledge model storage unit 38.SELECTED DRAWING: Figure 6

Description

[Technical Field]

[0001] The present invention provides Knowledge Model Execution Device Regarding. [Background technology]

[0002] For example, it is known that in order to determine machining conditions for industrial machinery, the knowledge of experts is compiled into a database and the database is utilized. For example, Patent Document 1 describes a knowledge model compiled into a database. The knowledge model is a model composed of multiple factors defined by various technical terms and the relationships between the factors. It is said that the knowledge model can be used to optimize machining conditions.

[0003] Patent Document 2 also describes the application of machine learning when defining the relationships between factors in a knowledge model. In particular, as means for defining the relationships between factors, it describes a means for defining the correspondence relationship (rank-rank relationship) between the rank value of a reference factor and the rank value of a factor to be connected, and a means for defining the correspondence relationship (rank-range relationship) between the rank value of a reference factor and the range value of a factor to be connected. Furthermore, when a reference factor has a relationship with multiple factors, the influence of multiple factors on the reference factor can be determined by setting contribution degrees. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2020-177547 [Patent Document 2] Patent Publication No. 2021-071856 Summary of the Invention [Problem to be solved by the invention]

[0005] The relationships between factors described in Patent Document 2 directly correspond to factor values ​​of each factor, such as rank values ​​and range values. It is not easy for a person to determine the correspondence between factor values. Although it is possible to determine the correspondence between factor values ​​by applying machine learning, it may not be possible to properly define the relationships between factors, especially when the amount of data in the training dataset is small.

[0006] The present invention has been made in view of the above problem, and is capable of efficiently creating a knowledge model by a method different from the conventional method. Knowledge Model Execution Device The aim is to provide the following. [Means for solving the problem]

[0007] One aspect of the present invention is a device for supporting creation of a knowledge model that is configured by a plurality of factors defined by industrial technical terms and information on relationships between the factors. and, an output value determination unit that determines a factor value of an output factor based on a factor value of an input factor and the knowledge model, The creation support device includes: a target factor acquisition unit that acquires parent factors and child factors as target factors for defining the relationship information; an observation data acquisition unit that acquires observation data regarding the factor values ​​of the parent factors and the factor values ​​of the child factors; a generation unit that generates a conditional probability table that defines the probability of a factor value of the child factor being a condition for a factor value of the parent factor, based on the observation data; a knowledge model storage unit that stores the knowledge model including the relationship information, using the generated conditional probability table as the relationship information between the parent factor and the child factor, The generation unit generates a new conditional probability table when the observation data acquisition unit acquires new observation data, and updates the knowledge model stored in the knowledge model storage unit. death, In the conditional probability table, the parent factor and the child factor have their respective factor values ​​divided into a plurality of intervals, The output value determination unit The input factor is the parent factor, and the output factor is the child factor; In the conditional probability table, a probability in each interval of the child factor corresponding to the input factor is obtained; determining a factor value of the output factor based on a value corresponding to each interval of the child factor and a probability in each interval of the child factor; knowledge model Execution device is located. [Effects of the Invention]

[0008] Knowledge Model Execution Device According to [the text], a conditional probability table is used as relationship information between factors that make up a knowledge model. A conditional probability table is defined by the probability that a child factor value corresponds to a given parent factor value, given the parent factor value. In other words, a conditional probability table does not directly relate the parent factor value to the child factor value, but instead uses probabilities to represent the relationship. This allows for flexible definition of the relationship between factors.

[0009] Furthermore, the generation unit generates a conditional probability table as relationship information between factors based on the observation data. If new observation data is acquired after the generation of the conditional probability table, the generation unit generates a new conditional probability table based on the new observation data. Then, the generation unit updates the knowledge model stored in the knowledge model storage unit to a knowledge model that uses the newly generated conditional probability table as relationship information. Therefore, because the knowledge model is updated, even if the accuracy of the knowledge model is low initially, it gradually improves.

[0010] As described above, according to the above aspect, a knowledge model can be efficiently created using a method different from the conventional method. Knowledge Model Execution Device can be provided. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 1 is a diagram illustrating a configuration of a knowledge model utilization system. [Figure 2] FIG. 1 is a diagram showing a knowledge network diagram representing a knowledge model. [Figure 3] FIG. 1 is a diagram illustrating a part of a knowledge model. [Figure 4] FIG. 4 is a diagram showing factor values ​​of each factor constituting the knowledge model of FIG. 3. [Figure 5] FIG. 4 is a diagram showing relationship information that constitutes the knowledge model of FIG. 3. [Figure 6] FIG. 1 is a diagram illustrating a configuration of a knowledge model creation support device. [Figure 7] The target factor is acquired by the target factor acquisition unit. [Figure 8] This is the observation data acquired by the observation data acquisition unit. [Figure 9] 1 shows discrete intervals determined by the interval determination unit, where (a) shows discrete intervals set at equal intervals, (b) shows discrete intervals determined based on mutual information, and (c) shows discrete intervals set at equal frequency. [Figure 10] FIG. 10 is a diagram for explaining mutual information. [Figure 11] 10 is a conditional probability table generated based on the parameters of the prior distribution stored in the prior distribution storage unit. [Figure 12] 10 is a conditional probability table generated based on the posterior distribution generated by the generation unit. [Figure 13] 1 shows the procedure for updating a conditional probability table, in which the initial state of the conditional probability table is generated using only the parameters of the prior distribution. [Figure 14] 10 shows a procedure for updating a conditional probability table, in which an updated conditional probability table is generated based on a posterior distribution. [Figure 15] This shows the procedure for updating the conditional probability table, in which a further updated conditional probability table is generated based on a new posterior distribution. [Figure 16] 1 shows a drawing GUI window for creating a knowledge model. [Figure 17] 1 shows a drawing GUI window for creating a knowledge model. [Figure 18] 1 shows a drawing GUI window for creating a knowledge model. DETAILED DESCRIPTION OF THE INVENTION

[0012] (Embodiment) 1. Configuration of knowledge model utilization system 1 The configuration of the knowledge model utilization system 1 will be described with reference to FIG. 1. The knowledge model utilization system 1 can be targeted at industrial technology fields such as engineering, medicine, pharmacy, agriculture, and biology. In particular, the knowledge model utilization system 1 can be targeted at the mechanical processing field and material forming field, both of which are included in the engineering field. The mechanical processing field includes, for example, cutting, grinding, electrical discharge machining, and press working. The material forming field includes, for example, injection molding and casting. The knowledge model utilization system 1 is a system that utilizes a knowledge model 4a that describes knowledge about technical information by an expert in the above industrial technology fields.

[0013] As shown in Fig. 1, the knowledge model utilization system 1 includes a plurality of industrial machines 2 and 3 in the aforementioned industrial technology field, and a knowledge model execution device 4. The plurality of industrial machines 2 and 3 and the knowledge model execution device 4 form a network. In other words, the plurality of industrial machines 2 and 3 and the knowledge model execution device 4 are configured to be able to communicate with each other.

[0014] The multiple industrial machines 2, 3 are, for example, machine tools, various molding machines, conveying devices, construction machines, agricultural machines, chemical industry machines, etc. The knowledge model execution device 4 stores a knowledge model 4a. Using the knowledge model 4a, the knowledge model execution device 4 performs, for example, support for the operation of the industrial machines 2, 3 in the aforementioned industrial technology field, determination of operating conditions, judgment of machine conditions, judgment of workpiece conditions, etc.

[0015] Generally, in the field of machining, an operator determines machining conditions such as cutting speed and cutting depth per unit time, taking into consideration various information such as the material of the workpiece, the material of the tool, the quality of the workpiece, and the machining cycle time. In this case, the knowledge model 4a is a model of the operator's thought process when the operator acquires various input information and determines the machining conditions.

[0016] In other words, in knowledge model 4a, in addition to the material of the workpiece, the material of the tool, the quality of the workpiece, the processing cycle time, the cutting speed, the cutting depth, etc., each of the industrial technology elements that appear in the worker's thought process is defined as a factor, and the relationship between the factors is defined.

[0017] For example, when determining the machining conditions for a machine tool that is one of the industrial machines 2 and 3, the knowledge model execution device 4 can use the knowledge model 4a to output machining conditions such as cutting speed and depth of cut as output factors if it acquires the material and quality of the workpiece as input factors. Furthermore, the knowledge model execution device 4 can acquire the observation data 32a of the industrial machines 2 and 3 as input factors and use the knowledge model 4a to output the machine state such as an abnormal state or a deteriorated state of the industrial machines 2 and 3, or the presence or absence of an abnormality in the quality of the workpiece caused by the industrial machines 2 and 3 as output factors.

[0018] 2. Knowledge Network Diagram 10 The knowledge model 4a is conceptually expressed in the form of a network. An example of a knowledge network diagram 10 that expresses the knowledge model 4a in the form of a network will be described with reference to FIG. 2. In other words, the knowledge network diagram 10 is a graphical representation of the knowledge model 4a. In this embodiment, a knowledge network diagram 10 relating to a knowledge model 4a in the machining field will be taken as an example.

[0019] As shown in FIG. 2, the knowledge network diagram 10 comprises a plurality of node figures 11 and link figures 12 that connect the node figures 11 together. The node figures 11 are expressed, for example, by high quality, high efficiency, cutting speed, and rotational feed. The node figures 11 represent factors defined in industrial technology terms in the knowledge model 4a. Multiple factors may have a technical inclusion relationship (also called a hierarchical relationship or a master-slave relationship), or may have a technical heterogeneous dependency relationship. In other words, the relationships between factors are classified into the above two types.

[0020] For example, the workpiece material description has a relationship of inclusion with workpiece material thermal properties, workpiece material hardness, workpiece material elongation, etc. In other words, as factors having a technical inclusion relationship, the workpiece material description is a higher-level conceptual factor, and workpiece material thermal properties, workpiece material hardness, workpiece material elongation, etc. are lower-level conceptual factors. Furthermore, factors having a technical heterogeneous dependency relationship include workpiece material thermal properties and required tool heat resistance, etc. In what follows, the relationship between two factors having a technical inclusion relationship will be simply referred to as an inclusion relationship, and the relationship between two factors having a technical heterogeneous dependency relationship will be simply referred to as a heterogeneous dependency relationship.

[0021] The link graphics 12 represent the relationships between the node graphics 11. The link graphics 12 represent the relationships between the factors in the knowledge model 4a. The link graphics 12 are displayed so that a first link graphic 12a representing an inclusion relationship and a second link graphic 12b representing a heterogeneous dependency relationship can be distinguished from each other.

[0022] In Figure 2, the first link graphic 12a, which represents an inclusion relationship, is represented by a frame that indicates the area of ​​the superordinate concept factor, and the subordinate concept factors are placed within the frame that represents the first link graphic 12a. The second link graphic 12b is represented by an arrow line that indicates the direction of the definition of the factors. The origin of the arrow of the second link graphic 12b is the parent factor, and the destination of the arrow of the second link graphic 12b is the child factor. However, the representation method of the first link graphic 12a and the second link graphic 12b is not limited to the above, and other forms can also be used.

[0023] 3. Construction of Knowledge Model 4a The knowledge model 4a that defines the knowledge network diagram 10 shown in Fig. 2 will be described with reference to Fig. 3 to Fig. 5. The knowledge model 4a is composed of data and processing programs, and defines the knowledge network diagram 10 shown in Fig. 2 as data.

[0024] Here, knowledge model 4a is a representation of knowledge network diagram 10 as shown in Figure 2, and is therefore a model that has many factors and defines the relationships between these factors. However, for ease of explanation, the smallest unit that constitutes part of knowledge model 4a, that is, one unit defined by a parent factor and a child factor, will be taken as an example below.

[0025] As shown in FIG. 3, the minimum unit constituting the knowledge model 4a includes a plurality of factors 21, 22, and 23 defined in industrial technical terms, relationship information 24 between the factors, and an output value determination unit 25.

[0026] In FIG. 3, the multiple factors 21, 22, 23 include parent factors A and B and a child factor C. Parent factors A and B are factors that influence child factor C. Conversely, child factor C is a factor that is influenced by parent factors A and B. In other words, the factor value of child factor C is determined based on the factor values ​​of parent factors A and B. Note that while FIG. 3 illustrates two parent factors A and B, there are also cases where there is one parent factor or three or more parent factors.

[0027] As shown in FIG. 4, parent factor A defines, for example, factor values ​​by range values. For example, parent factor A is defined as a range value, which is a factor value, by a plurality of discretized intervals, such as "10 to 20" and "20 to 30." In FIG. 4, the range values ​​of parent factor A are defined by equally discretized intervals, but they can also be defined by unevenly discretized intervals. Furthermore, parent factor A includes, for example, information on the median of the range values ​​as the nominal range value. For example, the nominal range values ​​of parent factor A are "15" and "25." Note that the nominal range value can also be the minimum or maximum value of the range value.

[0028] Like parent factor A, parent factor B defines its factor values ​​as range values. For example, parent factor B is defined as a range value, which is a factor value, using multiple discretized intervals such as "20 to 40" and "40 to 60." In FIG. 4, the range values ​​of parent factor B are defined as evenly discretized intervals, but they can also be defined as unevenly discretized intervals. Furthermore, parent factor B includes information on the median of the range values ​​as the nominal range value. For example, the nominal range values ​​of parent factor B are "30" and "50."

[0029] Like parent factors A and B, child factor C defines its factor values ​​using range values. For example, child factor C is defined as a range value, which is a factor value, using a plurality of discretized intervals, "100-200," "200-300," and "300-400." In FIG. 4, the range values ​​of child factor C are defined using evenly discretized intervals, but they can also be defined using unevenly discretized intervals. Furthermore, child factor C includes, for example, information on the median of the range values ​​as the nominal range value. For example, the nominal range values ​​of child factor C are "150," "250," and "350."

[0030] Returning to Fig. 3, the explanation will be made. The relationship information 24 is a database that defines the relationship between parent factors A and B and child factor C. In this embodiment, a conditional probability table is used as the relationship information 24. In detail, the relationship information 24 is a conditional probability table that defines the probability of the factor value of child factor C occurring when the factor values ​​of parent factors A and B are used as conditions.

[0031] 5, the relationship information 24 is a conditional probability table that defines the relationship between parent factors A and B and child factor C. If the factor value of parent factor A is "10 to 20" and the factor value of parent factor B is "20 to 40," the probability that child factor C will be "100 to 200" is "0.5," the probability that the factor value of child factor C will be "200 to 300" is "0," and the probability that the factor value of child factor C will be "300 to 400" is "0.5."

[0032] If the factor value of parent factor A is "20-30" and the factor value of parent factor B is "20-40," the probability that the factor value of child factor C will be "100-200" is "0.3333," the probability that the factor value of child factor C will be "200-300" is "0.6667," and the probability that the factor value of child factor C will be "300-400" is "0." When the factor value of parent factor A is "10-20" and the factor value of parent factor B is "40-60," or when the factor value of parent factor A is "20-30" and the factor value of parent factor B is "40-60," the results are as shown in Figure 5. Note that if the discretized range values ​​of factors A-C are set to more intervals or if the number of parent factors is increased, the number of cells in the conditional probability table shown in Figure 5 will increase.

[0033] When the factor values ​​of parent factors A and B are input, the output value determination unit 25 obtains the probability of each range value of child factor C from the conditional probability table of the relationship information 24. Then, the output value determination unit 25 determines the output value as the factor value of child factor C based on the obtained probability of each range value of child factor C.

[0034] For example, in Figure 5, assume that the factor value of parent factor A is "20-30" and the factor value of parent factor B is "20-40". In this case, the probability that the factor value of child factor C will be "100-200" is "0.3333", the probability that the factor value of child factor C will be "200-300" is "0.6667", and the probability that the factor value of child factor C will be "300-400" is "0".

[0035] At this time, the output value determination unit 25 can output, for example, the factor value with the highest probability. In the above example, since the probability that the factor value of child factor C will be "200 to 300" is the highest, the output value determination unit 25 sets the factor value "200 to 300" of child factor C as the output value. Note that the output value of child factor C can also be the range nominal value "250."

[0036] Furthermore, the output value determination unit 25 can also determine the output value of the child factor C, for example, using each factor value and its probability. For example, the output value determination unit 25 can use the expected value as the output value. In the above example, the output value determination unit 25 determines that the probability of the range nominal value "150" is "0.3333," the probability of the range nominal value "250" is "0.6667," and the probability of the range nominal value "350" is "0." Therefore, the output value determination unit 25 may multiply the range nominal value by the probability and use the sum as the output value, as shown in equation (1).

[0037]

number

[0038] In the above, when the factor values ​​of parent factors A and B are input, the output value determination unit 25 outputs the factor value of child factor C. In other words, the output value determination unit 25 outputs the factor value of child factor C as a result of forward probabilistic inference. In addition, when the factor value of child factor C is input, the output value determination unit 25 can output the factor values ​​of parent factors A and B. In other words, the output value determination unit 25 can output the factor values ​​of parent factors A and B as a result of backward probabilistic inference.

[0039] When the factor value of child factor C is input, the output value determination unit 25 can acquire probability values ​​of the range values ​​of parent factor A and factor B based on the conditional probability table shown in Fig. 5 and the probability tables for factors A and B (not shown). Then, based on the acquired probabilities of the range values ​​of parent factor A and parent factor B, the output value determination unit 25 can determine the factor value of parent factor A and the factor value of parent factor B as output values.

[0040] 4. Configuration of the knowledge model creation support device 30 The configuration of the knowledge model creation support device 30 for creating the above-mentioned knowledge model 4a will be described with reference to Figs. 6 to 15. The knowledge model creation support device 30 is a device for creating the knowledge model 4a using a conditional probability table. The knowledge model creation support device 30 is mounted on the knowledge model execution device 4. Therefore, the knowledge model creation support device 30 forms a network with the industrial machines 2 and 3, and is configured to be able to communicate with the industrial machines 2 and 3.

[0041] 6, the knowledge model creation support device 30 includes a target factor acquisition unit 31, an observation data acquisition unit 32, an interval determination unit 33, an interval edit processing unit 34, a prior distribution storage unit 35, a generation unit 36, a weight storage unit 37, a knowledge model storage unit 38, and a model edit processing unit 39. The knowledge model creation support device 30 includes at least a processor and a storage device. The target factor acquisition unit 31, the observation data acquisition unit 32, the interval determination unit 33, the interval edit processing unit 34, the generation unit 36, and the model edit processing unit 39 are configured by a processor, and the prior distribution storage unit 35, the weight storage unit 37, and the knowledge model storage unit 38 are configured by a storage device.

[0042] The target factor acquisition unit 31 acquires a target factor 31a for defining the relationship information 24. In a factor network such as that shown in Fig. 7, the target factor acquisition unit 31 acquires parent factors A and B and child factor C as the target factor 31a. In Fig. 7, in forward probabilistic inference, parent factors A and B are conditions in the conditional probability table, and child factor C is an output factor. Furthermore, in probabilistic inference in the opposite direction to the arrow shown in Fig. 7, child factor C is an input factor, and parent factors A and B are output factors.

[0043] The observation data acquisition unit 32 acquires observation data 32a about the factor values ​​of parent factors A and B and the factor value of child factor C acquired by the target factor acquisition unit 31. The observation data acquisition unit 32 is configured to be able to communicate with the industrial machines 2 and 3, and is therefore able to acquire the observation data 32a of the industrial machines 2 and 3 in real time. Furthermore, the observation data acquisition unit 32 can centrally acquire the observation data 32a of multiple industrial machines 2 and 3. However, the observation data acquisition unit 32 can also acquire a set of observation data 32a for a predetermined period of time, rather than in real time.

[0044] The observation data acquisition unit 32 acquires a large amount of observation data 32a, as shown in Fig. 8. In Fig. 8, numerical values ​​in the same row indicate the factor values ​​of factors A, B, and C at the same time. In other words, the numerical value in the top row indicates that when the factor value of parent factor A was "10" and the factor value of parent factor B was "20," the factor value of child factor C was "115."

[0045] The interval determination unit 33 determines discrete intervals by discretizing each of the factor values ​​of the target factor 31a, i.e., the factor values ​​of parent factors A and B and the factor value of child factor C, in the conditional probability table shown in Figure 5. The interval determination unit 33 can select one from a plurality of discrete interval determination methods, for example, as shown in Figures 9(a), (b), and (c). However, for convenience of illustration, Figures 9(a), (b), and (c) illustrate discrete intervals related to parent factors A and B. For child factor C, a discrete interval can be determined in the same way as for parent factors A and B.

[0046] As shown in Fig. 9(a), the interval determination unit 33 can determine intervals in which the factor values ​​of the target factor 31a are evenly discretized. The interval determination unit 33 determines intervals in which the factor values ​​of parent factors A and B are evenly discretized. In Fig. 9(a), vertical dashed lines represent boundaries of the discrete intervals of parent factor A. In Fig. 9(a), horizontal dashed lines represent boundaries of the discrete intervals of parent factor B.

[0047] The interval determination unit 33 can determine intervals that are unevenly discretized, as shown in Fig. 9(b). The interval determination unit 33 can determine intervals that are unevenly discretized for the factor values ​​of parent factors A and B. Furthermore, the interval determination unit 33 can determine intervals that are unevenly discretized for the factor value of child factor C.

[0048] The interval determination method shown in Figure 9(b) determines discrete intervals based on the mutual information I(X;Y) between each random variable in factors A, B, and C, which are the target factors 31a. The mutual information I(X;Y) is a quantity that represents a measure of the interdependence between two random variables. The mutual information I(X;Y) indicates how much one variable can be predicted by understanding the other variable.

[0049] The mutual information I(X;Y) is expressed by equation (2). Furthermore, the mutual information I(X;Y) is expressed as shown in Figure 10. In equation (2), H(X) and H(Y) are entropy (also called information entropy) and represent the uncertainty of the random variable. H(X) is expressed by equation (3). H(X) is the average amount of information obtained when the results of individual events are known when an event is represented by random variable X. From equation (3), H(X) is a weighted average of the probability that each event will occur. H(Y) is similarly expressed by equation (4).

[0050]

number

[0051]

number

[0052]

number

[0053] In equation (2), H(X|Y) is the conditional entropy, which is expressed by equation (5). Also, in equation (2), H(X, Y) is the joint entropy, which is expressed by equation (6).

[0054]

number

[0055]

number

[0056] For example, a case will be described in which the interval determination unit 33 determines discrete intervals for factor A based on the mutual information I(X;Y) between factors A and B and the mutual information I(X;Y) between factors A and C. First, the interval determination unit 33 acquires the final number of discretizations for factor A and data including continuous variables as input information. Next, as initialization, the unit 33 discretizes factor A so that each interval contains only one data point. However, discretization may be performed using a specified initial number of discretizations, and in this case, any algorithm may be used.

[0057] Next, the current number of discretizations is reduced by one until it reaches the final number of discretizations. For example, if the total number of discretizations is 10 in the first iteration of the loop, one adjacent discrete interval is combined into one discrete interval, bringing the total number of discretizations to 9. In the second iteration of the loop, one adjacent discrete interval is combined into one discrete interval, bringing the total number of discretizations from 9 to 8. This process is repeated.

[0058] When merging adjacent discrete intervals, the adjacent discrete intervals are merged so that the reduction in the sum of the mutual information between factor A and other factors B and C is minimized. For example, when reducing the total number of discretizations from 10 to 9, the adjacent discrete intervals that result in the smallest reduction in the mutual information I(X;Y) when merging are calculated for all adjacent discrete intervals. The adjacent discrete intervals that result in the smallest reduction in the mutual information I(X;Y) are then merged.

[0059] More specifically, first, for factor A and factor B, the mutual information I(A;B) of factors A and B is calculated when adjacent discrete intervals of factor A are combined. Next, for factor A and factor C, the mutual information I(A;C) of factors A and C is calculated when adjacent discrete intervals of factor A are combined. Then, the sum of the mutual information I(A;B) of factors A and B and the mutual information I(A;C) of factors A and C is calculated, and it is determined which adjacent discrete intervals of factor A, when combined, will result in the least decrease in the sum of the mutual information I of both, and the corresponding adjacent discrete intervals are combined.

[0060] 9(c), the interval determination unit 33 can determine intervals in which the factor values ​​are discretized at equal frequencies. In this case, the interval determination unit 33 also determines intervals in which the factor values ​​are discretized at uneven frequencies.

[0061] The section edit processor 34 is configured to receive input from an operator and edit discrete sections in accordance with the operator's input. The section edit processor 34 makes it possible to adjust discrete sections determined by any of Fig. 9(a), (b), and (c) based on the operator's input. The section edit processor 34 can also create discrete sections from scratch based on the operator's input.

[0062] The prior distribution storage unit 35 stores a parameter α of a preset prior distribution 35a. jk The prior distribution 35a uses the Dirichlet distribution as expressed in equation (7). j is the discrete interval of the parent factor A, and k is the discrete interval of the child factor C.

[0063]

number

[0064] Furthermore, the prior distribution storage unit 35 stores the parameter α jk The expected value θa of the prior distribution 35a generated based on jk 11 shows a conditional probability table 35b of parent factor A and child factor C for ease of explanation. Here, the conditional probability table 35b stored in the prior distribution storage unit 35 is generated based on the discrete intervals determined by the interval determination unit 33.

[0065] Expected value θa of prior distribution 35a jk is expressed by the formula (8). In the formula (8), Na j is the parameter α of the prior distribution 35a in the discrete interval j. jk is the sum of Na jk is the parameter α of the prior distribution 35a in the discrete interval j, k. jk is.

[0066]

number

[0067] The conditional probability table 35b shown in FIG. 11 shows the corresponding expected value θa jk Here, the parameter α of the prior distribution 35a is jk is set based on, for example, the worker's experience, information on similar fields, etc. In this case, the conditional probability table 35b based on the prior distribution 35a is a conditional probability table generated based on the worker's experience, information on similar fields, etc.

[0068] The generation unit 36 ​​acquires the observation data 32a acquired by the observation data acquisition unit 32. The data frequency β of the observation data 32a jk is expressed by the formula (9).jk is expressed by a multinomial distribution. jk The maximum likelihood estimator φ for jk is expressed by equation (10). N j is the total number of observation data 32a in the discrete interval j, and N jk is the data frequency β of the observation data 32a in the discrete interval j, k. jk is.

[0069]

number

[0070]

number

[0071] The generation unit 36 ​​generates the observation data 32 a acquired by the observation data acquisition unit 32 and the parameter α of the prior distribution 35 a stored in the prior distribution storage unit 35. jk The posterior distribution 36a is calculated using Bayesian estimation based on the above. The posterior distribution 36a is a Dirichlet distribution that is configured with the same discrete intervals as the prior distribution 35a.

[0072] Furthermore, the generation unit 36 ​​calculates the expected value θb of the posterior distribution 36a based on the calculated posterior distribution 36a. jk The conditional probability table 36b is generated based on the discrete intervals determined by the interval determination unit 33. In other words, the conditional probability table 36b generated by the generation unit 36 ​​is configured with the same discrete intervals as the conditional probability table 35b generated based on the prior distribution 35a. The generated conditional probability table 36b is shown in the upper part of Figure 12. The conditional probability table 36b as the relationship information 24 has the value θb as the element value of each discrete interval. jk is stored.

[0073] The lower part of Fig. 12 shows a diagram for explaining in detail the discrete interval j of the parent factor A in the upper part of Fig. 12. Each element value θb of the conditional probability table 36b in the upper part of Fig. 12 jk is the expected value θb of the posterior distribution 36a in the lower part of FIG. jk Matches.

[0074] Expected value θb of posterior distribution 36a jk is expressed by equation (11). In equation (11), W j is the weight. Na jk is the parameter α of the prior distribution 35a jk is the value of the discrete interval j in j is the parameter α of the prior distribution 35a in the discrete interval j. jk is the sum of N jk is the data frequency β of the observation data 32a jk is the value of the discrete interval j in N j is the total number of data in the observation data 32a in the discrete interval j.

[0075]

number

[0076] That is, the generator 36 generates the parameter α of the observation data 32a and the prior distribution 35a. jk Based on this, the posterior distribution 36a is calculated using Bayesian estimation, and the expected value θb jk A conditional probability table 36b is generated with the following element values:

[0077] Here, as shown in equation (11), the generation unit 36 ​​does not simply use Bayesian estimation but also uses the weight W j In other words, the posterior distribution 36a is based on the observed data 32a and the weights W j Parameter α of the prior distribution 35a taking into account jk Furthermore, the element values ​​Nb obtained in the process of generating the posterior distribution 36a by the generating unit 36 ​​are calculated using Bayesian estimation. jk is expressed by equation (12). jkcorresponds to the numerator of equation (11).

[0078]

number

[0079] Here, for comparison, the weight W j The expected value θb of the posterior distribution 36a when not applying jk is expressed by equation (13).

[0080]

number

[0081] In this embodiment, as shown in equation (11), the weight W j Using this, the expected value θb of the posterior distribution 36a jk The expected value θb of the posterior distribution 36a in this embodiment is calculated. jk represents the prior distribution 35a jk , Na j Weight W j In other words, the weight W j is the parameter α of the prior distribution 35a for the observed data 32a jk From equation (11), the weight W j The larger the parameter α of the prior distribution 35a for the observed data 32a, the jk The influence ratio of is high, and the weight W j The smaller the parameter α of the prior distribution 35a for the observed data 32a, jk The impact rate of

[0082] Weight W j is expressed by equation (14). In equation (14), K j is the weight W j It is a coefficient to represent Na j is the parameter α of the prior distribution 35a in the discrete interval j. jk is the sum of N j is the total number of data in the observation data 32a in the discrete interval j. In other words, the weight Wj Na j , N j It is expressed by:

[0083]

number

[0084] When the amount of observed data 32a is small, the generation unit 36 ​​calculates the observed data 32a and the parameter α jk Based on this, we use Bayesian estimation to calculate the expected value θb of the posterior distribution 36a. jk It is preferable to generate a conditional probability table 36b in which each element value is as the relationship information 24. If a large amount of observation data 32a is acquired, the generation unit 36 ​​may generate the conditional probability table 36b based only on the observation data 32a. Therefore, the generation unit 36 ​​is configured to be able to select between generating the conditional probability table 36b using Bayesian estimation and generating the conditional probability table 36b using maximum likelihood estimation based on the observation data 32a.

[0085] The weight storage unit 37 stores the weight W j Store the weight W j can be set to different values ​​for each discrete interval of the factor value of the parent factor A, or can be set to the same value. As mentioned above, from equation (11), the weight W j The larger the parameter α of the prior distribution 35a for the observed data 32a, the jk The influence ratio of is high, and the weight W j The smaller the parameter α of the prior distribution 35a for the observed data 32a, jk For example, the weight storage unit 37 stores a plurality of weights W as shown in equation (15). j_L ,W j_M ,W j_S Store.

[0086]

number

[0087] Weight W j The meaning of the weight W will be explained in detail using specific values. As shown in the above formula (13), j The expected value θb of the posterior distribution 36a without using jk Here, N j4 is "80" and Na j When is "100", the parameter α of the prior distribution 35a jk Compare the case of {1,2,2,5} with the case of {100,200,200,500}. The parameter α of the prior distribution 35a jk The ratios of these are all {1:2:2:5}.

[0088] Parameter α of prior distribution 35a jk The expected value θb of the posterior distribution 36a when j4 is "0.7727" as shown in equation (16). The calculation formula is "(5 + 80) / (10 + 100)". The parameter α of the prior distribution 35a jk The expected value θb of the posterior distribution 36a when j4 As shown in equation (17), the result is "0.5272". The calculation formula is "(500+80) / (1000+100)".

[0089]

number

[0090]

number

[0091] Thus, the parameter α of the prior distribution 35a jk Even if the ratio of the parameters is the same {1:2:2:5}, the parameter α of the prior distribution 35a jk Depending on the magnitude of the absolute value of j4 As shown in equation (16), the parameter α of the prior distribution 35a jkThe smaller the absolute value of , the greater the expected value θb of the posterior distribution 36a. j4 is the data frequency β of the observation data 32a jk The maximum likelihood estimator φ for j4 On the other hand, as shown in equation (17), the parameter α jk The larger the absolute value of , the greater the expected value θb of the posterior distribution 36a. j4 is the expected value θa of the prior distribution 35a j4 The value is close to 0.5 (= 5 / 10).

[0092] And the weight W j can be regarded as the confidence of the prior distribution 35a. In other words, when the confidence of the prior distribution 35a is high, the weight W j In this case, the expected value θb of the posterior distribution 36a j4 is the expected value θa of the prior distribution 35a j4 On the other hand, when the confidence level of the prior distribution 35a is low, the weight W j In this case, the expected value θb of the posterior distribution 36a j4 is the data frequency β of the observation data 32a jk The maximum likelihood estimator φ for j4 In this way, the weight W j By adjusting the expected value θb of the posterior distribution 36a, j4 can be set to a desired value.

[0093] Returning to Fig. 6, as shown in Fig. 6, the knowledge model storage unit 38 stores the conditional probability table 36b generated by the generation unit 36 ​​as relationship information 24 between the parent factor A and the child factor C, and stores the knowledge model 4a including the relationship information 24.

[0094] The model edit processing unit 39 is configured to receive input from an operator and to be able to edit, in accordance with the operator's input, the conditional probability table 36b constituting the knowledge model 4a stored in the knowledge model storage unit 38. The model edit processing unit 39 makes it possible to adjust each element value of the conditional probability table 36b shown in the upper part of Fig. 12 in accordance with the operator's input. Note that the model edit processing unit 39 can also create the conditional probability table 36b from scratch in accordance with the operator's input.

[0095] Here, in addition to the above processing, when the observation data acquisition unit 32 acquires new observation data 32a, the generation unit 36 ​​can generate a new conditional probability table 36b and update the knowledge model 4a stored in the knowledge model storage unit 38.

[0096] In this case, the prior distribution storage unit 35 stores the element values ​​Nb obtained in the process of generating the posterior distribution 36a by the generation unit 36. jk (expressed in equation (12)) is used as the parameter α of the following prior distribution 35a jk Then, the generator 36 stores the parameter α of the prior distribution 35a newly stored in the prior distribution storage unit 35. jk and the observation data 32a newly acquired by the observation data acquisition unit 32. The generation unit 36 ​​calculates the expected value θb of the posterior distribution 36a based on the newly calculated posterior distribution 36a. jk A conditional probability table 36b can be generated in which each element value is

[0097] 13 to 15, the procedure for updating the conditional probability table constituting the knowledge model 4a will be described in detail. As shown in FIG. 13, when there is no observed data 32a, the parameter α jk In other words, the element values ​​of the conditional probability table 35b are calculated based on the expected value θa of the prior distribution 35a. jk At this point, the knowledge model 4a stored in the knowledge model storage unit 38 is the parameter α jkThe conditional probability table 35b is defined by the conditional probability table 35b, which is expressed only by

[0098] Next, assume that observed data 32a is acquired. As shown in FIG. 14, the generator 36 generates a parameter α jk Based on the observation data 32a, the parameter γ of the posterior distribution 36a is calculated using Bayesian estimation. jk Calculate the expected value θb of the posterior distribution 36a jk Then, the generating unit 36 ​​generates the expected value θb of the posterior distribution 36a. jk Then, the generation unit 36 ​​updates the knowledge model 4a stored in the knowledge model storage unit 38.

[0099] Next, the generation unit 36 ​​generates the element values ​​Nb jk The parameter α of the new prior distribution 35a is stored in the prior distribution storage unit 35. jk That is, as shown in FIG. 15, the parameter α jk However, each element value Nb obtained in the previous generation process of the posterior distribution 36a jk Then, assume that new observation data 32a is acquired. The generator 36 updates the parameter α jk and the new observation data 32a, a new posterior distribution 36a is calculated, and the expected value θb' of the new posterior distribution 36a is calculated. jk Then, the generating unit 36 ​​generates the expected value θb' of the new posterior distribution 36a. jk As a result, the generation unit 36 ​​updates the knowledge model 4a stored in the knowledge model storage unit 38 again.

[0100] 5. Effects of the knowledge model creation support device 30 According to the knowledge model creation support device 30, the conditional probability tables 35b and 36b are used as relationship information 24 between factors constituting the knowledge model 4a. The conditional probability tables 35b and 36b are defined by the probability of the factor value of child factor C occurring when the factor values ​​of parent factors A and B are set as conditions. In other words, the conditional probability tables 35b and 36b do not directly associate the factor values ​​of parent factors A and B with the factor value of child factor C, but instead express them using probabilities. Therefore, the relationship between factors can be flexibly defined.

[0101] Furthermore, the generation unit 36 ​​generates a conditional probability table 36b as relationship information 24 between factors based on the observation data 32a. If new observation data 32a is acquired after the generation of the conditional probability table 36b, the generation unit 36 ​​generates a new conditional probability table 36b based on the new observation data 32a. Then, the generation unit 36 ​​updates the knowledge model 4a stored in the knowledge model storage unit 38 to a knowledge model 4a that uses the newly generated conditional probability table 36b as relationship information 24. Therefore, since the knowledge model 4a is updated, even if the accuracy of the knowledge model 4a is low initially, it gradually improves. In this way, the knowledge model 4a can be created efficiently.

[0102] In particular, the generation unit 36 ​​generates the conditional probability table 36b using Bayesian estimation. In other words, by effectively utilizing the prior distribution 35a in Bayesian estimation, the knowledge model 4a can be made highly accurate even at a stage when there is little observation data 32a. However, depending on the accuracy of the prior distribution 35a, there is a possibility that the knowledge model 4a cannot be made highly accurate. Therefore, the generation unit 36 ​​uses the weight W j Using the expected value θb of the posterior distribution 36a jk By calculating the above, it is possible to generate a knowledge model 4a according to the degree of certainty of the prior distribution 35a.

[0103] Then, when the generation unit 36 ​​updates the conditional probability table 36b using Bayesian estimation, and a large amount of observation data 32a is obtained, the expected value θb of the posterior distribution 36a jkTherefore, when a large amount of observation data 32a can be secured, the influence rate of the highly reliable observation data 32a is increased, thereby making it possible to generate a highly accurate knowledge model 4a.

[0104] And the weight W j The weight W can be set to a different value for each factor value of the parent factors A and B, i.e., for each discrete interval j of the parent factors A and B. j represents the confidence level of the prior distribution 35a. The confidence level may differ depending on the discrete interval j of the parent factors A and B. For example, in a discrete interval j that the operator has had many experiences with, the prior distribution 35a can be determined with high confidence. However, in other cases, the prior distribution 35a may not be determined with high confidence. In such cases, the weight W j By setting for each discrete interval j, it is possible to generate an appropriate conditional probability table 36b, and therefore a knowledge model 4a.

[0105] 6. How to create knowledge model 4a A method for creating the knowledge model 4a will be described with reference to Figs. 16 to 18. Figs. 16 to 18 show the drawing GUI window 40 on the display screen of the knowledge model creation support device 30. First, as shown in Fig. 16, the operator specifies the GUI element of "factor" in the drawing GUI window 40 and places it at an arbitrary position in the drawing area. Then, a node graphic 11 (shown in Fig. 1) representing the factor is displayed in the drawing area.

[0106] 17, the operator connects the factors by specifying the GUI element "connect" in the drawing area of ​​the drawing GUI window 40. Then, a link graphic 12 connecting the factors is displayed in the drawing area.

[0107] Next, as shown in FIG. 18, the operator specifies a link graphic 12 in the drawing area of ​​the drawing GUI window 40. Then, a new input GUI window 41 is displayed, separate from the drawing GUI window 40. The new input GUI window 41 that is displayed displays selectable connection types. In addition to the above-mentioned "conditional probability table," there are also well-known connection types such as "rank value-rank value," "rank value-range value," and "range value-range value." By selecting a desired connection type, the operator can input information corresponding to the selected connection type. When the operator selects "conditional probability table" in the input GUI window 41, a conditional probability table is automatically generated, as described above.

[0108] 7. Automatic update method for knowledge model 4a Initially, the method of creating the knowledge model 4a is performed using a drawing GUI window 40 as shown in Figures 16 to 18. Once the knowledge model 4a is created, it can be updated by acquiring new observation data 32a.

[0109] 1, it is assumed that the knowledge model execution device 4 includes a knowledge model creation support device 30, and that the knowledge model execution device 4 is configured to be able to communicate with a plurality of industrial machines 2 and 3. In this case, a generation unit 36 ​​constituting the knowledge model creation support device 30 can acquire observation data 32a from the industrial machines 2 and 3 at a set timing. Then, the generation unit 36 ​​can generate a new conditional probability table 36b at a set timing according to factors, and update the knowledge model 4a stored in the knowledge model storage unit 38.

[0110] In this case, the knowledge model execution device 4 can acquire observation data 32a of multiple industrial machines 2 and 3, so the generation unit 36 ​​that constitutes the knowledge model creation support device 30 can generate and update a knowledge model 4a common to the multiple industrial machines 2 and 3, and store the knowledge model 4a in the knowledge model storage unit 38.

[0111] Alternatively, different from the above, the knowledge model execution device 4 may include a knowledge model creation support device 30, which may be provided for each of the industrial machines 2 and 3. In this case, the observation data acquisition unit 32 constituting the knowledge model creation support device 30 acquires observation data 32a for each of the industrial machines 2 and 3. The generation unit 36 ​​generates and updates a conditional probability table 36b corresponding to each of the industrial machines 2 and 3, and stores the knowledge model 4a in the knowledge model storage unit 38. In this way, the knowledge model 4a corresponds to each of the industrial machines 2 and 3 individually. Therefore, it is possible to generate a knowledge model 4a that appropriately corresponds to each of the industrial machines 2 and 3. [Explanation of symbols]

[0112] 4a Knowledge Model 21,22,23 factors 24 Relationship Information 30 Knowledge model creation support device 31 Target Factor Acquisition Section 31a Target Factor 32 Observation data acquisition section 32a Observation data 36 Generation part 36b Conditional Probability Tables 38 Knowledge Model Storage A,B parent factor C child factor

Claims

1. a knowledge model creation support device configured by a plurality of factors defined by industrial technology terms and information on relationships between the factors; an output value determination unit that determines a factor value of an output factor based on a factor value of an input factor and the knowledge model, The creation support device includes: a target factor acquisition unit that acquires parent factors and child factors as target factors for defining the relationship information; an observation data acquisition unit that acquires observation data regarding the factor values ​​of the parent factors and the factor values ​​of the child factors; a generation unit that generates a conditional probability table that defines the probability of a factor value of the child factor being assigned a factor value of the parent factor based on the observation data; a knowledge model storage unit that stores the knowledge model including the relationship information, using the generated conditional probability table as the relationship information between the parent factor and the child factor, the generation unit generates new conditional probability tables when the observation data acquisition unit acquires new observation data, and updates the knowledge model stored in the knowledge model storage unit; In the conditional probability table, the parent factor and the child factor have their respective factor values ​​divided into a plurality of intervals, The output value determination unit The input factor is the parent factor, and the output factor is the child factor; In the conditional probability table, a probability in each interval of the child factor corresponding to the input factor is obtained; a knowledge model execution device that determines a factor value of the output factor based on a value corresponding to each interval of the child factor and a probability in each interval of the child factor;

2. A knowledge model execution device as described in Claim 1, wherein the output value determination unit determines the value of the child factor corresponding to the section of the child factor in which the probability in each section of the child factor is maximum as the factor value of the output factor.

3. The output value determination unit: calculating a value for each interval of the child factor based on the probability for each interval of the child factor and the value corresponding to each interval of the child factor; Calculating the sum of the values ​​of the calculated child factors in each interval; 2. The knowledge model execution device according to claim 1, wherein the factor value of the output factor is determined based on the calculated sum.

4. The creation support device further includes a prior distribution storage unit for storing parameters of a predetermined prior distribution, The generation unit calculating a posterior distribution using Bayesian estimation based on the parameters of the prior distribution and the observation data; 4. The knowledge model execution device according to claim 1, wherein the conditional probability table is generated with each element value being an expected value of the posterior distribution.

5. the prior distribution storage unit stores elements obtained in a process of generating the posterior distribution by the generation unit as parameters of a next prior distribution; The generation unit calculating a new posterior distribution using Bayesian estimation based on the parameters of the prior distribution newly stored in the prior distribution storage unit and the observation data newly acquired by the observation data acquisition unit; 5. The knowledge model execution device according to claim 4, wherein the conditional probability table is generated with each element value being an expected value of the new posterior distribution.

6. The creation support device further includes a weight storage unit that stores weights representing an influence ratio of a parameter of the prior distribution to the observation data in the Bayesian estimation, 6. The knowledge model execution device according to claim 4, wherein the generation unit calculates the posterior distribution using the Bayesian estimation based on the observation data and parameters of the prior distribution taking into account the weights, and generates the conditional probability table in which each element value is an expected value of the posterior distribution.

7. 7. The knowledge model execution device according to claim 6, wherein the weight can be set to a different value for each factor value of the parent factor.

8. The knowledge model execution device according to any one of claims 4 to 7, wherein the generation unit is configured to be able to select between generating the conditional probability table using the Bayesian estimation and generating the conditional probability table using maximum likelihood estimation based on the observation data.

9. The creation support device is a knowledge model execution device described in any one of claims 1 to 8, further comprising a model editing processing unit configured to accept input from an operator and edit the conditional probability table that constitutes the knowledge model stored in the knowledge model storage unit in accordance with the operator's input.

10. The creation support device further includes an interval determination unit that determines intervals in the conditional probability table into which the factor values ​​of the target factor are unevenly discretized, 10. The knowledge model execution device according to claim 1, wherein said generation unit generates said conditional probability table based on said intervals determined by said interval determination unit.

11. 11. The knowledge model execution device according to claim 10, wherein the interval determination unit determines the interval for one factor of the target factor based on mutual information representing a measure of interdependence of random variables for the one factor constituting the target factor and each of the remaining factors constituting the target factor.

12. the observation data acquisition unit acquires the observation data of the industrial machine, The knowledge model execution device according to any one of claims 1 to 11, wherein the generation unit generates a new conditional probability table at a timing set according to the factor, and updates the knowledge model stored in the knowledge model storage unit.

13. the observation data acquisition unit is configured to be able to communicate with a plurality of industrial machines, and acquires the observation data from the plurality of industrial machines; the generation unit generates one conditional probability table based on the observation data of a plurality of the industrial machines; 13. The knowledge model execution device according to claim 1, wherein said knowledge model storage unit stores said knowledge model common to a plurality of said industrial machines.

14. the observation data acquisition unit acquires the observation data for each industrial machine, 13. The knowledge model execution device according to claim 1, wherein said generation unit generates said conditional probability table individually corresponding to said industrial machine.

Images