WO2020161835A1

WO2020161835A1 - Management system and machine learning device therefor and managing method

Info

Publication number: WO2020161835A1
Application number: PCT/JP2019/004272
Authority: WO
Inventors: 唐橋　聡; ヨンテカン
Original assignee: オーウエル株式会社; コエバーアイ・アンド・ティーカンパニー，リミテッド
Priority date: 2019-02-06
Filing date: 2019-02-06
Publication date: 2020-08-13
Also published as: JP6600120B1; KR20200098471A; JPWO2020161835A1; KR102242476B1

Abstract

The present invention creates a cause-and-effect diagram on the basis of expert knowledge about association between an item to be managed and a plurality of causes. In the cause-and-effect diagram, the plurality of causes are data acquired in the field of manufacture or service in which the item to be managed is set, and the degree of importance for the item to be managed is imparted to each of the plurality of causes. In machine learning, first, a link structure for a probability model is set on the basis of the relationship between the item to be managed and the plurality of causes in the cause-and-effect diagram, and a value for each node of the probability model is set on the basis of the degree of importance imparted to each of the plurality of causes. Then, the link structure or the value for each node is learned using the data acquired in the field of manufacture or service.

Description

Management system, machine learning device therefor, and management method

The present invention relates to a management system suitable for management of items that can be affected by various factors in the field of manufacturing or service, and a machine learning device and management method therefor.

As prior art documents disclosing the prior art related to the present invention, for example, International Publication No. 2016/170574 can be cited. In the system disclosed in International Publication No. 2016/170574, operation data (big data) of a large number of distributed target bases are collected in a central base, and the operation data is utilized to detect anomaly sign. The operation data is sensor data acquired by various sensors. In this system, a model for detecting the sign of abnormality is constructed based on the collected learning sensor data. Then, while utilizing the constructed model and the sensor data for predictive analysis, the predictive analysis of the abnormality is executed, and the cause analysis of the abnormality is executed based on the sensor data for the abnormal analysis.

International Publication No. 2016/170574

Today, various management target items are set at the manufacturing or service sites. For example, an abnormality whose sign is detected in the above-mentioned prior art is also one item to be managed. When a managed item is affected by some factor, if the relationship between the managed item and the factor is known in advance, it is possible to predict the state or numerical value of the managed item from changes in the factor, and vice versa. In addition, it is possible to estimate the factor that is the cause of the change from the state or numerical value of the managed item. In the above-mentioned prior art, the relationship between the management target item and a plurality of factors is modeled, and the sign of abnormality that is the management target item is detected using the learning sensor data and the model corresponding to the factors.

However, it is known that it takes a lot of time to build a model that shows the relationship between managed items and multiple factors. The larger the number of factors that can cause the change in the state or numerical value of the managed item, the longer the time required to build the model. Today, machine learning is used to build such models, but the time required for model building and the accuracy of the model depend largely on the scale of computer resources used for machine learning. In an environment where large-scale computer resources cannot be used, it takes a lot of time to obtain the relationship between the managed item and multiple factors. In addition, a large amount of data is required to build a highly accurate model, and a large amount of computer resources are inevitably required to process the large amount of data.

However, preparing a large-scale computer resource requires a great deal of cost. Depending on the manufacturing or service site, it is often the case that only a small computer resource can be prepared due to cost constraints, and if the same result can be obtained, the cost to invest in computer resources should be kept low. For this reason, even if it is a small-scale computer resource, it is necessary to obtain results in a short time and with a small amount of data in order to perform machine learning of the relationship between the management target item and multiple factors in the manufacturing or service site. Is sought in.

The present invention has been made in view of the above problems, and an object of the present invention is to enable efficient machine learning of the relationship between a management target item and a plurality of factors at a manufacturing or service site.

As a means for achieving the above object, in the present invention, in machine learning for associating a managed item with a plurality of factors, expert knowledge of the relationship between the managed item and a plurality of factors is included in the initial modeling. Take advantage of. The expert knowledge is knowledge of an expert who is familiar with the site of manufacturing or service in which items to be managed are set, and is knowledge formalized by a method described later. The present invention provides a management system, a machine learning device therefor, and a management method. The management system provided by the present invention is a system that enables utilization of expert knowledge for machine learning, and the machine learning device for the management system provided by the present invention utilizes machine knowledge to perform machine learning. The management method provided by the present invention is a device that enables utilization of expert knowledge for machine learning.

The management system provided by the present invention includes a data acquisition means, a characteristic factor diagram creation means, a probability model initial setting means, and a probability model learning means. The data acquisition means acquires data at the manufacturing or service site where the management target item is set. The characteristic factor diagram creating means generates a characteristic factor diagram showing the relationship between the management target item and a plurality of factors that are presumed to affect it, based on the expert knowledge of the relationship between the management target item and the plurality of factors. create. However, the plurality of factors in the characteristic factor diagram are at least a part of the data acquired by the data acquisition unit, and each of the plurality of factors is given a degree of importance with respect to the management target item. The probabilistic model initial setting means sets the link structure of the probabilistic model based on the relationship between the management target item and the plurality of factors in the characteristic factor diagram, and sets the probabilistic model based on the importance attached to each of the multiple factors Set the value of each node. The probabilistic model learning means learns the link structure of the probabilistic model or the value of each node using the data acquired by the data acquisition means.

The management system may further include characteristic factor diagram updating means for updating the characteristic factor diagram based on the learned link structure of the probabilistic model or the value of each node. In addition, the management system further includes a monitoring unit that, when an abnormality is detected in the management target item, identifies a factor that is presumed to be the cause of the abnormality in the management target item based on the probabilistic model, and notifies the operator of the identified factor. You can also

In another embodiment, a management system includes a data acquisition device that acquires data at a manufacturing or service site in which items to be managed are set, and a terminal for creating the characteristic factor diagram based on expert knowledge. And a machine learning device that learns a probabilistic model related to the relationship between a managed item and a plurality of factors using data acquired by the data acquisition device. However, the machine learning device initializes the link structure of the probabilistic model based on the relationship between the managed item in the characteristic factor diagram and the plurality of factors, and the probabilistic model based on the importance assigned to each of the plurality of factors. Is configured to initialize the value of each node in the.

A machine learning device for a management system provided by the present invention is provided with a data input unit to which data acquired at a manufacturing or service site in which a management target item is set and the characteristic factor diagram described above are set. A characteristic factor diagram setting unit, a probabilistic model initial setting unit that initializes a probabilistic model, and a probabilistic model learning unit that learns a probabilistic model are provided. The probabilistic model initial setting unit sets the link structure of the probabilistic model based on the relationship between the managed item and the plurality of factors in the characteristic factor diagram, and based on the importance attached to each of the plurality of factors, It is configured to set the value of each node. The probabilistic model learning unit is configured to learn the link structure or the value of each node using the data input to the data input unit.

The management method provided by the present invention includes the following first to fourth steps. The first step is a step of acquiring data at the manufacturing or service site where the management target item is set. The second step is a step of creating the above-mentioned characteristic factor diagram based on expert knowledge of the relationship between the management target item and a plurality of factors. The third step is to set the link structure of the probabilistic model based on the relationship between the managed items in the characteristic factor diagram and the plurality of factors, and to set each of the probabilistic models based on the importance assigned to each of the plurality of factors. This is the step of setting the value of the node. The fourth step is a step of learning the link structure of the probabilistic model or the value of each node using the acquired data.

According to the present invention, the relationship between the management target item and a plurality of factors in the field of manufacturing or service is represented by a probabilistic model, and the characteristic of formalizing expert knowledge about the relationship between the management target item and a plurality of factors is provided. By utilizing the factor diagram for the initial modeling of the probabilistic model, it is possible to efficiently machine-learn the relationship between the management target item and a plurality of factors. Details of the effects will be apparent from the drawings briefly described below and the embodiments described in detail in connection with the drawings.

It is a figure which shows the whole structure of the management system which concerns on embodiment of this invention. It is a figure for demonstrating the structure of a fish bone diagram. It is a figure for explaining a creation method of a fishbone diagram. It is a figure for explaining a creation method of a fishbone diagram. It is a figure for explaining a creation method of a fishbone diagram. It is a figure for explaining the outline of machine learning. It is a figure for demonstrating the search method of a latent factor. It is a figure for demonstrating the learning method of a Bayesian network. It is a figure for demonstrating the creation method of the data for machine learning. It is the figure which represented the whole flow of the machine learning performed in the embodiment of this invention with the image. It is a figure for demonstrating the update method of a fish bone diagram. It is a figure for demonstrating the update example of a fishbone diagram. It is a figure for demonstrating the update example of a fishbone diagram.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings. However, in the following embodiments, when reference is made to the number of each element, the number, the amount, the range, etc., the reference is made unless otherwise specified or in principle clearly specified by the number. The present invention is not limited to the number. Further, the structures and processes described in the following embodiments are not necessarily essential to the present invention, unless otherwise specified or clearly specified in principle.

1. Configuration of Management System The present invention is applicable to various manufacturing and service sites. Manufacturing sites include, for example, food and machine manufacturing lines, painting lines, chemical plants, and the like. The service site includes, for example, a distribution center, a cleaning plant, a restaurant, and the like. That is, a suitable site using the present invention is a site where at least one management target item exists and a plurality of factors that may affect the management target item may exist. In this embodiment, an example in which the present invention is applied to a coating line among these sites will be described.

FIG. 1 is a diagram showing the overall configuration of the management system according to the present embodiment. The management system 1 is applied to the coating line 10. The coating line 10 includes a pretreatment device 12 for pretreating the product 11, a coating device 14 for coating the product 11, a drying device 16 for drying the coated product 11, and a pretreatment device 12 to the coating device 14. A transport device 13 that transports the product 11, a transport device 15 that transports the product 11 from the coating device 14 to the drying device 16, a utility facility 17, and a mixing chamber 18 that mixes the paint.

Each of the equipments 11-18 constituting the coating line 10 is provided with one or more sensors (not shown). The sensor includes various sensors such as a temperature sensor, a pressure sensor, a flow rate sensor, a weight sensor, an atmospheric pressure sensor, an outside air temperature sensor, and a wind speed sensor. In each facility 11-18, sensor data is acquired by these sensors. However, the data acquisition means is not limited to the sensor. At least some of the facilities 11-18 acquire device data such as the rotation speed of the motor. Further, the operator who manages the site may obtain the manually input data by inputting the data on a tablet terminal or the like. [Xi] in the figure represents various data (including sensor data, device data, and manual input data) acquired at various points of the coating line 10.

The management system 1 includes an FBD creation terminal 2, a machine learning device 3, and a monitoring device 4, in addition to the above-described data acquisition device that acquires various data. The FBD creation terminal 2 is a terminal for creating a fishbone diagram (FBD) described later based on expert knowledge. The machine learning device 3 is a device that performs machine learning for finding a relationship between the data acquired by the data acquisition device and the customer task in the coating line 10. The data acquired on the painting line 10 is input to the machine learning device 3. At that time, data obtained at the same time and in the same space are grouped, and the data is managed in group units. Details of machine learning will be described later.

The monitoring device 4 is a device that uses the relationship between the data obtained by machine learning and customer issues to identify the factor that caused the abnormality. The monitoring device 4 includes a monitor, and dynamically displays the location and contents of the abnormality and the factor that caused the abnormality. However, the monitoring device 4 is not essential, and a system including the FBD creation terminal 2 and the machine learning device 3 can also be called the management system 1.

2. Structure and Creation Method of Fishbone Diagram In the present embodiment, a fishbone diagram (hereinafter, referred to as FBD), that is, a characteristic factor diagram is used as a tool for formalizing knowledge possessed by an expert. The FBD is a diagram systematically organizing and associating the relationship between a task (characteristic) and a plurality of factors that are presumed to affect it. As shown in FIG. 2, in the FBD, a large bone is drawn obliquely with respect to the spine representing the task. Bone means a large classification of factors. In this embodiment, 5 major 4M1E, that is, 5 major bones of MEN (person), MAC (mechanical equipment), MAT (material), MET (working method), and ENV (environment) are drawn. There is. MEN, MAC, MAT, and MET are work conditions on site, and ENV is environmental conditions on site.

Data acquired from the coating line 10 is classified into any of the above 4M1E. For example, in the example shown in FIG. 2, three middle bones are drawn with respect to the large bones of MEN. These middle bones mean the factors that influence the MEN, and the corresponding data are given for each middle bone. An expert determines which data (factor) is classified into 4M1E based on his/her own knowledge.

Furthermore, the expert decides the importance of each data. The degree of importance is the magnitude of the influence of each data on the task, and the larger the influence, the larger the number assigned. In this embodiment, a number from 0 to 5 and X are set as the rank of importance. Ranks 1 to 5 mean that the higher the number, the higher the importance. Rank 0 means that the degree of importance is unknown, and rank X means that the degree of importance is judged to be zero by an expert. However, the data of rank X is not displayed on the FBD. In the example shown in FIG. 2, the data (2nd_3rd_4th_5th_001) is assigned the importance of rank 5, and the data (2nd_3rd_4th_5th_002) is assigned the importance of rank 3. The FBD creation terminal 2 creates an FBD in which each factor is ranked, that is, a ranked FBD (R-FBD).

In the ranked FBD, the major issues are coating quality, safety, facility maintenance, paints and chemicals, productivity, etc., and specific issues are set under these major issues. For example, as shown in FIG. 3, specific problems such as film thickness, coating NV, and dust spots are set under the major problem of coating quality. These specific problems are items to be managed by the management system 1. The ranked FBD is created for each task. The example shown in FIG. 3 indicates that a ranked FBD is created for each of the film thickness, the coating NV, and the dust spots. Further, when the problem is a problem that occurs independently for each facility, the ranked FBD is created for each facility and for each problem. In the example shown in FIG. 2, it is shown that a ranked FBD having a preventive maintenance problem is created for each of the three facilities PT11, PT31, and PT01.

4 and 5 show specific examples of the core of the ranked FBD. In the example shown in FIG. 4, two middle bones are drawn with respect to the large bones of MET. These backbones are the factors that experts have determined to affect MET. Here, the chemical surface of the solution of the preliminary degreasing device of the facility PT11 and the free alkalinity of the solution of the preliminary degreasing device of the facility PT11 are listed as factors for MET. On the other hand, in the example shown in FIG. 5, two middle bones are drawn with respect to the large bones of MEN. Then, a small bone is drawn diagonally to one middle bone. The data corresponding to the middle bone (PT01-1410) is the clogging, and the data corresponding to the small bone (PT01-1210) is the pressure. As shown in the table, such a ranked FBD is created when an expert determines that clogging and pressure are related to results and factors.

As described above, in the present embodiment, creation of a ranked FBD is performed by the FBD creation terminal 2. In correspondence with the claims, the FBD creation terminal 2 corresponds to the characteristic factor diagram creation means described in the claims.

3. Contents of Machine Learning In the machine learning of the present embodiment, a method of searching for a causal relationship based on space-time is used. FIG. 6 is a diagram for explaining the outline of machine learning. Here, the problem corresponding to the product j is represented by Y _j , the set of data acquired from the coating line 10 is represented by X _i, and the number of data is 100, the problem Y _j is the set X _i ={X ₁ , X ₂ , X ₃ ... X ₃₅ , X ₃₆ ... X ₉₉ , X ₁₀₀ }. The task Y _j is quantified. The set X _i corresponding to the task Y _j is a group of data acquired at the same time as the task Y _j .

In machine learning, a parameter ΔY indicating a change in the task Y and a parameter ΔX indicating a change in each data X are calculated. As a method of calculating the parameters ΔY and ΔX, there are a method of calculating the differential value shown in Expression 1 and a method of calculating the change rate shown in Expression 2. Either calculation method may be used.

As shown in the upper part of FIG. 6, ΔY _j and ΔX _i are calculated and grouped for each product. Since products flow on the coating line at regular intervals, calculation for each product is equivalent to calculation for every fixed time. In the figure, the parameter surrounded by O means that the value exceeds the limit value. For example, for product No. 19, ΔX ₂ exceeds the limit value. On the other hand, for the 20th product, ΔY _20, ΔX _2, ΔX _{3, and} ΔX ₁₀₀ exceed the limit values. Since ΔY ₂₀ exceeds the limit value when ΔX ₃ and ΔX ₁₀₀ exceed the limit values, it can be estimated that X ₃ and X ₁₀₀ affect the task Y.

The lower part of FIG. 6 shows the physical structure of the coating line 10. The data acquired from the painting line 10 is also classified by the acquired space. For example, X ₁ , X ₂ , and X ₃ are data obtained by the coating device, X ₇ , X ₈ , and X ₉ are data obtained in the mixing chamber, and X ₁₁ is a transfer device between the coating device and the drying device. The data obtained in, are classified as. In addition, the arrow lines connecting the facilities in FIG. 6 indicate the overlap of spaces on the data. In the example shown in FIG. 6, the coating device 14 and the mixing chamber 18 are regarded as the same space in terms of data. As described above, in the machine learning of the present embodiment, a set of data obtained at the same time and the same space is created, and the machine learning is performed in units of the set.

In the middle of FIG. 6, the logical structure of the painting line 10 is drawn. In the logical structure, {Pro○, ProΔ△} is a set of products included in one batch, and, for example, {Pro19, Pro20} represents a set of 19th and 20th products. The frame drawn with a broken line represents the classification of the data X that changed when the task Y changed. In the case of the example shown in FIG. 6, in {Pro19, Pro20}, changes occur in the set {X ₁ , X ₂ , X ₃ }, the set {X ₇ , X ₈ , X ₉ }, and the set {X ₁₁ }, and { In Pro44, Pro45}, the set {X ₁ , X ₂ , X ₃ } and the set {X ₇ , X ₈ , X ₉ } and the set {X ₁₁ } change, and in {Pro 102, Pro 103}, the set {X ₇ , X ₈ , X ₉ } and the set {X ₁₁ } have changed. From the example shown in FIG. 6, each of the set {X ₁ , X ₂ , X ₃ }, the set {X ₇ , X ₈ , X ₉ }, and the set {X ₁₁ } is a cause of repeated occurrence, and , Set {X ₁ , X ₂ , X ₃ } and set {X ₇ , X ₈ , X ₉ } are the causes of overlapping in the same space.

By performing the machine learning as described above on the data obtained from the painting line 10, the relationship between the problem and a plurality of factors can be sorted out, and the factor causing the problem can be searched. Becomes Here, a learning method of a probabilistic model and a searching method of latent factors in machine learning will be described in more detail with reference to FIG. 7. FIG. 7 shows time-series data of ΔY _j and the corresponding set {ΔX ₁ , ΔX ₂ ... ΔX ₃₀ , ΔX ₃₁ , ΔX ₃₂ ... ΔX ₉₉ , ΔX ₁₀₀ }. In machine learning, a change in ΔY _j is determined for each time. If ΔY _j is Boolean data, the change from true to false is determined as the change in ΔY _j . When ΔY _j is a numerical value, it is determined that the numerical value (measured value) exceeds the limit as a change in ΔY _j .

When a change occurs in ΔY _j , the set {ΔX ₁ , ΔX ₂ ... ΔX ₃₀ , ΔX ₃₁ , ΔX ₃₂ ... ΔX ₉₉ , ΔX ₁₀₀ } at that time is searched for an X _i that has a difference equal to or larger than the limit value. Counting is performed and the counting result Count(X _i |Y _j ) is obtained. Then, Count(X _i |Y _j ) is converted into discrete data, and the probabilistic model is learned based on it. In this embodiment, a Bayesian network model is used as the probabilistic model. The problem of finding the link structure and the value of each node in the Bayesian network model can be solved by integer programming. In machine learning, latent factors are searched using a Bayesian network model. In this case, a new factor (variable) is searched by the belief propagation method based on the time lag effect based on space.

Next, a method for learning a Bayesian network by the machine learning device 3 will be described with reference to FIG. As described in the section of Bayesian pattern recognition algorithm, a Bayesian network is a physical/chemical phenomenon that has a number of infinite cases, and a computer-understandable structure with a finite number of cases. It is a model represented by. Specifically, although the factors omega _i of the actual phenomena in coating line 10 can not be observed and observable factor X _i is present, the observable factor X _i is a Bayesian network, a causal relationship between the results and the factor , And the causal relationship between factors is represented by a link structure.

Figure 8 shows a graphical statistical model of a recurrent neural network. The update structured model in this graphical statistical model is a model for updating the structure of the Bayesian network, that is, the link structure and the value of each node. The set of data obtained at the current time on the painting line 10 and the output of the updated structured model at the previous time are input to the updated structured model. The output of the updated structured model is the previous value of the unobservable factor mentioned above. The above-described belief propagation method is used for learning in the update structured model.

The machine learning device 3 performs machine learning by the above method. However, although the number of data is 100 here, the actual number of data is large, and there are many problems to be managed, so that a small computer resource requires a long time for calculation. What is used there is the above-mentioned ranked FBD created by an expert. As shown in FIG. 9, the machine learning device 3 converts the ranked FBD created by the FBD creation terminal 2 into an orthogonal table. Alternatively, the ranked FBD may be converted into an orthogonal table in advance and the orthogonal table may be input to the machine learning device 3. In the orthogonal table, in addition to the relationship between the task (result) and the factor (cause), the importance of each factor and the causal relationship between the factors are input.

The machine learning device 3 initializes the Bayesian network based on the orthogonal table. Since the orthogonal table is converted from the ranked FBD, the link structure of the Bayesian network is initialized based on the relationship between the issues in the ranked FBD and multiple factors, and the important factors assigned to each of the multiple factors The value of each node of the Bayesian network will be initialized based on the degree. There are many factors that can influence the task, but the factors that the expert judges to have no influence are not displayed in the ranked FBD and are not reflected in the Bayesian network. As a result, the factor is excluded from the target of machine learning, so that machine learning can be efficiently performed and small-scale computer resources can be processed.

The learned Bayesian network is used in the monitoring device 4. When an abnormality is detected in the management target item (issue), the monitoring device 4 identifies a factor that is presumed to be the cause of the abnormality in the management target item based on the learned Bayesian network, and notifies the identified factor to the operator. .. In correspondence with the claims, the monitoring device 4 corresponds to the monitoring means described in the claims.

4. Updating of Fishbone Diagram As described above, in the present embodiment, the ranked FBD in which expert knowledge is formalized is applied to machine learning using Bayesian network. Machine learning will clarify the relationship between the issues that are the items to be managed and various factors, and by feeding back the results of machine learning to the ranked FBD, it is possible to further improve the accuracy of the ranked FBD. .. FIG. 10 is a diagram showing an image of the overall flow of machine learning executed in this embodiment.

The expert stores knowledge of the relationship between the observable data (factors) obtained on the painting line 10 and the task (step S1). Then, a ranked FBD is created based on the knowledge. Observable data obtained from the coating line 10 is associated with the middle and small bones of the ranked FBD (step S2).

In the machine learning device 3, the link structure of the Bayesian network and the value of each node are initialized based on the ranked FBD created by the expert (step S3). Further, learning is performed based on the observable data obtained on the coating line 10 (step S4). Learning is performed by a recursive neural network, and unobservable data is used as a hidden layer. The learning result is reflected in the Bayesian network, and the value of each node changes or the link structure changes according to the learning result (step S5). Furthermore, the learning result is reflected in the ranked FBD via the Bayesian network, and the importance of the bone structure and each factor is updated based on the learned link structure or the value of each node (step S6).

Here, a method of updating the ranked FBD will be specifically described with reference to FIG. There are a total of 100 data changes in the predetermined time when the task Y has changed, of which 4 data X ₁ , 3 data X ₂ , 10 data X _{3, and 3} data X ₄ are included. It was supposed to be. Unless the relation with the task Y is taken into consideration, the probability of the data X ₃ is higher than the probability of the data X ₂ , so that the importance of the data X ₃ is generally higher than that of the data X _2. Is also considered high. Moreover, since the probability of the data X ₄ is not high, it is generally judged that the importance of the data X ₄ is low. Here, it is assumed that the expert has created the ranked FBD according to such a determination.

However, in machine learning using the Bayesian network, the relation between the task Y and each data is statistically calculated. According to FIG. 11, the probability P(Y|X ₂ ) that the task Y changed when the data X ₂ changed was 1/3, while the task Y changed when the data X ₃ changed. The probability P(Y|X ₃ ) is 2/10. Therefore, the machine learning device 3 determines that the importance of the data X ₂ is higher than the importance of the data X ₃ . According to this judgment, the values of the nodes corresponding to the data X ₂ and X ₃ are changed in the Bayesian network. Further, the change in the Bayesian network is also reflected in the ranked FBD, and the importance of the factors corresponding to the data X ₂ and X ₃ is changed.

Also, the probability P that problems Y when the data X ₄ is changed is changed (Y | X ₄₎ is the statistical probability problems Y when the data X ₂ is changed is changed P (Y | X ₂ ) And the probability P(X ₂ |X ₄ ) that the data X ₂ changes when the data X ₄ changes. Therefore, the probability P(Y|X ₄ ) is 1/9 instead of zero. According to this judgment, the link structure is changed in the Bayesian network so as to add the node corresponding to the data X ₄ . Further, the change in the Bayesian network is reflected in the ranked FBD, and a small bone corresponding to the data X ₄ is newly added.

FIG. 12 is a diagram for more specifically explaining an example of updating a ranked FBD, specifically, an example in which the importance of a factor changes. In learning the Bayesian network, based on the history of each change of the data X _i obtained from the painting line 10 and the history of the evaluation result of the task Y _j that is a management target item, each of the data X _i and the task Y _j. The degree of association with is examined, and the value of each node is reviewed according to the degree of association.

In the example shown in FIG. 12, the position of the node of the data X ₂ and the position of the node of the data X ₃ are exchanged before and after the learning. This means that the magnitude relation of the nodes between the data X ₂ and the data X ₃ has changed before and after the learning. When the value of a node is reviewed in the Bayesian network, the ranked FBD corrects the importance assigned to the factor corresponding to the node. In the example shown in FIG. 12, the importance of the data (1st_2nd_3rd_4th_5th_003) corresponding to the middle bone is changed from rank 5 to rank 3 before and after the learning, and as a result, the importance is lower than the data of rank 4 (1st_2nd_3rd_4th_5th_002). There is.

FIG. 13 is a diagram for more specifically explaining an example of updating the ranked FBD, specifically, an example in which the result-factor relationship changes. In learning the Bayesian network, based on the history of each change of the data X _i obtained from the painting line 10 and the history of the evaluation result of the task Y _j that is a management target item, each of the data X _i and the task Y _j. The degree of association with is investigated. Then, when the node corresponding to the data having a high degree of association does not exist in the link structure of the Bayesian network, the node is newly added to the link structure.

In the example shown in FIG. 13, after learning, the node of the data X ₄ is newly added below the node of the data X ₂ . This means that the data X _{4, which} has a relationship between the result and the factor with respect to the data X ₂ , is newly found by machine learning. When a new node is added to the link structure, a factor corresponding to the node is newly added to the ranked FBD. In the example shown in FIG. 13, after learning, a small bone corresponding to the data (1st_2nd_3rd_4th_5th_004) is newly added to the middle bone corresponding to the data (1st_2nd_3rd_4th_5th_002).

The processing described above is performed by the machine learning device 3. In correspondence with the claims, the function of the machine learning device 3 to acquire data from the painting line 10 corresponds to the function as a data input unit described in the claims. The function of the machine learning device 3 to accept the ranked FBD created by the FBD creation terminal 2 corresponds to the function as the characteristic factor diagram setting unit described in the claims. The link structure of the Bayesian network and the function of initializing each node value based on the ranked FBD correspond to the probabilistic model initializing unit described in the claims, and also correspond to the probabilistic model initializing unit described in the claims. To do. The function of learning the link structure of the Bayesian network and each node value based on the data acquired from the painting line 10 corresponds to the probabilistic model learning means described in the claims, and the probabilistic model learning unit described in the claims. Equivalent to. The function of updating the ranked FBD based on the Bayesian network after learning corresponds to the characteristic factor diagram updating means described in the claims.

5. Other Embodiments Some of the functions of the machine learning device 3 in the above embodiments may be transferred to another computer. Further, some or all of the functions of the machine learning device 3 can be transferred to a server on the cloud.

1 Management system 2 FBD creation terminal 3 Machine learning device 4 Monitoring device 10 Painting line

Claims

Data acquisition means for acquiring data at the manufacturing or service site where the management target items are set,
It is a characteristic factor diagram showing the relationship between the management target item and a plurality of factors that are presumed to affect it, wherein the plurality of factors is at least a part of the data acquired by the data acquiring unit, In addition, in order to create a characteristic factor diagram in which each of the plurality of factors is given a degree of importance with respect to the management target item, based on expert knowledge of the relationship between the management target item and the plurality of factors. Characteristic factor diagram creation means of
A link structure of a probabilistic model is set based on the relationship between the management target item and the plurality of factors in the characteristic factor diagram, and each of the probabilistic models is based on the importance assigned to each of the plurality of factors. Probabilistic model initial setting means for setting the value of the node,
A probabilistic model learning unit that learns the value of the link structure or each node using the data acquired by the data acquisition unit.
The management system according to claim 1, further comprising characteristic factor diagram updating means for updating the characteristic factor diagram based on the learned link structure or the value of each node.
The probabilistic model learning means, based on the history of each change of the data acquired by the data acquisition means and the history of the evaluation result of the management target item, the degree of association between each of the data and the management target item. And configured to review the value of each node according to the degree of association,
When the value of each node is reviewed, the characteristic factor diagram updating means is configured to correct the importance assigned to each of the plurality of factors based on the reviewed value of each node. The management system according to claim 2, wherein:
The probabilistic model learning means, based on the history of each change of the data acquired by the data acquisition means and the history of the evaluation result of the management target item, the degree of association between each of the data and the management target item. And if a node corresponding to the highly related data does not exist in the link structure, the node is newly added to the link structure,
The characteristic factor diagram updating means is configured to, when a new node is added to the link structure, newly add a factor corresponding to the node to the characteristic factor diagram. The management system according to 2 or 3.
When an abnormality is detected in the management target item, a factor that is presumed to be the cause of the abnormality in the management target item is specified based on the probabilistic model, and a monitoring unit that notifies the operator of the specified factor is further provided. The management system according to any one of claims 1 to 4.
The data acquired by the data acquisition unit includes data on evaluation of the management target item, data on work conditions on the site, and data on environmental conditions on the site. The management system according to any one of 1 to 5.
The management system according to claim 1, wherein the probabilistic model is a Bayesian network.
A data acquisition device that acquires data at the manufacturing or service site where the management target items are set,
It is a characteristic factor diagram showing the relationship between the management target item and a plurality of factors that are presumed to affect it, wherein the plurality of factors is at least a part of the data acquired by the data acquisition device, In addition, in order to create a characteristic factor diagram in which each of the plurality of factors is given a degree of importance with respect to the management target item, based on expert knowledge of the relationship between the management target item and the plurality of factors. Terminal of
A machine learning device that learns a probabilistic model regarding the relationship between the management target item and the plurality of factors using the data acquired by the data acquisition device,
The machine learning device initializes the link structure of the probabilistic model based on the relationship between the management target item and the plurality of factors in the characteristic factor diagram, and the importance assigned to each of the plurality of factors. The management system is configured to initialize the value of each node of the probabilistic model based on the above.
A data input section for inputting data acquired at the manufacturing or service site where the management target items are set,
It is a characteristic factor diagram created based on expert knowledge about the relationship between the management target item and a plurality of factors that are presumed to affect it, and the plurality of factors are input to the data input unit. A characteristic factor diagram setting unit for setting a characteristic factor diagram, which is at least a part of the data, and each of the plurality of factors is assigned a degree of importance to the management target item
A link structure of a probabilistic model is set based on the relationship between the management target item and the plurality of factors in the characteristic factor diagram, and each of the probabilistic models is based on the importance assigned to each of the plurality of factors. Probabilistic model initial setting part that sets the value of the node,
A probabilistic model learning unit that learns the link structure or the value of each node using the data input to the data input unit, the machine learning device for a management system.
Acquiring data at the manufacturing or service site where the management target items are set,
It is a characteristic factor diagram showing a relation between the management target item and a plurality of factors that are presumed to affect it, wherein the plurality of factors are at least a part of the acquired data, and A step of creating a characteristic factor diagram in which each of the factors is given a degree of importance with respect to the management target item based on expert knowledge about the relationship between the management target item and the plurality of factors;
A link structure of a probabilistic model is set based on the relationship between the management target item and the plurality of factors in the characteristic factor diagram, and each of the probabilistic models is based on the importance assigned to each of the plurality of factors. Setting the value of the node,
Learning the link structure or the value of each node using the acquired data.