JP6276727B2 - Statistical model creation device, statistical model creation method, and statistical model creation program - Google Patents

Description

  The present invention relates to a statistical model creation device, a statistical model creation method, and a statistical model creation program.

  Conventionally, the maintenance of equipment has been “Time-Based Maintenance (TBM)” based on equipment operation time. However, with the development of sensing technology for equipment operating status, “condition-based maintenance (CBM)” based on equipment operating status is becoming widespread. The CBM monitors the state of the device using a statistical model corresponding to the operating state of the device, and issues a warning or the like when it is determined that the device is abnormal to prompt early maintenance of the device. There are various types of devices to be diagnosed, and there are various usage environments of the devices. Therefore, in CBM, it is necessary to create a statistical model customized for each device to be diagnosed. On the other hand, creating a statistical model from scratch for each device (scratch development) leads to an increase in cost.

  The electronic control system of Patent Document 1 reduces the cost of developing such a scratch. When a transient fault such as noise or seasonal variation occurs in a sensor, the same type of transient fault often occurs in other sensors regardless of the type of sensor. Therefore, the electronic control system of Patent Document 1 separates diagnostic processing for detecting transient faults from diagnostic processing for detecting fixed faults that should be detected, and performs diagnostic processing for detecting transient faults for a plurality of diagnostic targets. Standardize on equipment.

JP2012-252412 (paragraph 0019)

  In order to improve development efficiency, it is common practice to divert part of an existing statistical model and newly change only the other part for another device. Assume that a part of the statistical model for device A is changed and a statistical model for device B is created. At this time, there is a risk that information that makes the design knowledge of the device A (for example, an abnormality determination threshold, a sensor detection time zone or a cycle, etc.) imagine in the statistical model for the device B may remain. Usually, the user of the device A distinguishes information that the user of the device B wants to keep secret from information that may be disclosed. Similarly, the user of the device B distinguishes information that should be kept secret from the user of the device A and information that may be disclosed.

The design knowledge for each device is information that the user of the device does not want to disclose. However, in fact, such a distinction cannot be made clear, so it becomes too cautious for publicly disclosing information, and the possibility of partial sharing (reuse) of statistical models is hindered. Sometimes. The electronic control system of Patent Document 1 is not particularly aimed at security management (management of what kind of information a person who knows what kind of information may be known) on the premise of partial sharing of a statistical model.
Therefore, an object of the present invention is to divert an existing statistical model without leaking design knowledge inherent in the existing statistical model, and to reduce the development cost of a new statistical model.

The statistical model creation apparatus according to the present invention includes an operation procedure for handling equipment and a storage unit storing a statistical model including operation conditions corresponding to each of the operation procedures, and handling the first equipment with reference to the storage unit. The first statistical model for the first device is acquired, a second statistical model for handling the second device is newly created based on the acquired first statistical model, and the operation procedure of the first statistical model and When the operating conditions and the operating procedure and operating conditions of the second statistical model are output to the user in a comparison format, at least a part of the operating conditions of the first statistical model or the second statistical model A control unit that outputs at least a part of the operating conditions after changing the content of the information to be concealed according to the position of the user of the output destination.
Other means will be described in the embodiment for carrying out the invention.

  According to the present invention, it is possible to divert an existing statistical model without leaking design knowledge inherent in the existing statistical model, and to reduce the development cost of a new statistical model.

It is a figure explaining the structure of a statistical model creation apparatus. It is a figure explaining an example of a statistical model. (A) is a figure explaining an example of access authority. (B) is a figure explaining an example of a model difference. It is a flowchart of a pre-processing model selection processing procedure. It is a flowchart of an operation condition processing procedure. It is a flowchart of a post-processing model creation processing procedure. It is a flowchart of a model difference confirmation processing procedure. It is a flowchart of a model mask processing procedure. It is a flowchart of an incentive processing procedure. It is a figure which shows an example of a model model selection screen. It is a figure which shows an example of the model difference display screen for model creators. It is a figure which shows an example of the model difference display screen for model providers. It is a figure which shows an example of the model difference display screen for model users. (A) is a figure showing an example of a point display screen. (B) is a figure showing an example of a user detailed information display screen.

  Hereinafter, two embodiments for carrying out the present invention will be described in detail with reference to the drawings. The two embodiments consist of the first embodiment as a basic type and the second embodiment as an application type. The details will be described later, but the difference between them is the presence or absence of an incentive processing procedure (detailed later) performed by the point granting unit. First, a first embodiment having no incentive processing procedure will be described, and then a second embodiment having an incentive processing procedure will be described by paying attention to differences from the first embodiment. The two embodiments are examples of diagnosing device abnormality, but the present invention can be applied not only to device diagnosis but also to overall handling of the device.

(First embodiment)
(Background of the embodiment, etc.)
Assume that company A uses statistical model A to diagnose its own device A. The statistical model A has a certain operational track record. On the other hand, Company B needs a new statistical model B for the device B in order to start using the device B and to diagnose the device B. Company C is contracted by Company B to develop statistical model B. Company C has a track record of undertaking development of statistical model A from Company A. Therefore, company C wants to develop statistical model B based on statistical model A so that the design knowledge of device A is not known to company B at the lowest possible cost. Further, company C wants to feed back statistical model B to company A so that the design knowledge of device B is not known. Although details will be described later, the statistical model is a processing procedure for machine learning of past data and determining the state of the device based on the result of the machine learning.

(Pre-processing model, post-processing model, model creator, model provider, and model user)
The pre-processing model is a statistical model that is a basis for developing a new statistical model, and corresponds to the statistical model A in the above example. The post-processing model is a statistical model newly developed based on the pre-processing model, and corresponds to the statistical model B in the above example. The “first statistical model” and the “second statistical model” correspond to a pre-processing model and a post-processing model, respectively.
The model creator is an entity that develops a post-process model based on the pre-process model, and corresponds to the company C (employee) in the above example. A model provider is an entity that allows a model creator to use (reuse) a pre-processing model for his / her device in order to create a post-processing model for another device. Is equivalent to Company A (employee). The model user is a person who uses the processed model for his / her equipment, and corresponds to company B (employee) in the above example. Note that “a person related to the first device” and “a person related to the second device” correspond to a model provider and a model user, respectively.

(Equipment configuration)
The configuration of the statistical model creation device will be described with reference to FIG. The statistical model creation system 1 includes a statistical model creation device 2 and a terminal device 3. These can be connected via the network 4. The statistical model creation device 2 is a general computer and includes a central control device 11, a main storage device 12, an auxiliary storage device 13, and a communication device 14. These are connected to each other by a bus. The auxiliary storage device 13 stores a pre-process model DB (Date Base) 31, a post-process model DB 32, and a model difference DB 33. The pre-processing model DB 31 stores a statistical model 41 (which is a pre-processing model). The post-processing model DB 32 stores a statistical model 42 (which is a post-processing model). The model difference DB 33 stores an access authority 43 and a model difference 44 (details will be described later).

The auxiliary storage device 13 for storing the pre-processing model DB 31 and the post-processing model DB 32 is independent from the statistical model creation device 2, and the statistical model creation device 2 is connected to the pre-processing model DB 31 via the network 4. And you may decide to access model DB32 after processing.
The pre-processing model selection unit 21, the operation condition processing unit 22, the post-processing model creation unit 23, the model difference confirmation unit 24, and the point provision unit 25 in the main storage device 12 are programs. Thereafter, when the subject is described as “XX section”, the central control device 11 reads each program from the auxiliary storage device 13 and loads it into the main storage device 12, and then the function of each program (detailed later). Shall be realized.

  The terminal device 3 is also a general computer and includes a communication device 15, an input device 16, an output device 17, a central control device 18, a main storage device 19, and an auxiliary storage device 20. These are connected to each other by a bus. The terminal device 3 is operated by a model creator, a model provider, or a model user. Note that the statistical model creation device 2 according to the first embodiment does not have the point assigning unit 25 as a program. The statistical model creation apparatus 2 according to the second embodiment has all the five programs described above (for the convenience of explanation, the maximum configuration is shown in FIG. 1).

(Statistical model)
The statistical model 41 will be described with reference to FIG. In FIG. 2, the solid line arrows indicate the flow of processing, and the broken line arrows indicate the flow of data. The statistical model 41 itself is read into an arbitrary system (not shown; hereinafter referred to as “diagnostic system”) different from the statistical model creation apparatus 2 to determine whether the device is abnormal.

  In the present embodiment, the device is a general device capable of measuring a time-series physical quantity with a sensor, for example, a copier. The physical quantity includes, for example, values such as temperature, noise, vibration, operating time, rotation speed, current, voltage, liquid flow rate and pressure, information processing amount, and values calculated by calculating these values. . As described above, the statistical model is a processing procedure for machine learning of past data and determining the state of the device based on the result of the machine learning, and includes a diagnosis procedure 101, operating conditions 102, model management information 103, and It is composed of operating condition rewriting information 104.

The diagnosis procedure 101 is a set of a plurality of programs executed by the diagnosis system in a predetermined order, and corresponds to an “operation procedure”. The diagnosis procedure 101 here includes a sensor input process Si, a learning process St, and an identification process Sd, and each process is executed in this order.
The sensor input process Si is a process in which the diagnostic system reads the measured value of the sensor. For example, it is a process of sampling (sampling) the value of the vibration sensor every predetermined period.
The learning process St is a process in which the diagnostic system performs machine learning based on past data and optimizes the statistical model 41.
The identification process Sd is a process in which the diagnostic system determines an apparatus abnormality using the optimized statistical model 41.

  The operation condition 102 includes a sensor input operation condition Pi, a learning operation condition Pt, an identification operation condition Pd, and a code book CB. Among these, the sensor input operation condition Pi, the learning operation condition Pt, and the identification operation condition Pd correspond to the sensor input process Si, the learning process St, and the identification process Sd on a one-to-one basis. It is a parameter that determines the type or algorithm type and operation.

Specifically, as the sensor input operation condition Pi, the type of sensor (vibration sensor or temperature sensor), the number of sensors, the sampling period (sampling rate) for each sensor, and the like are designated.
As the learning operation condition Pt, for example, “k-means method”, “n-layer neuro method” or the like as a machine learning algorithm is designated. Note that n = 2, 3,..., And “neuro” is an abbreviation for “neural network”. In addition, as the learning operation condition Pt, for example, when the machine learning algorithm is “k-means method”, the number of cluster divisions is specified, and when the machine learning algorithm is “n-layer neuro method”, The number of units in the output layer and hidden layer is specified.

  As the identification operation condition Pd, “k-means method”, “n-layer neuro method” or the like as an identification algorithm is designated. In the case of the “k-means method”, the identification process Sd obtains the distance between the machine-learned cluster information (for example, the center of gravity of the cluster indicating abnormality) and the sensor value, and the magnitude relationship between the distance and a predetermined threshold value. Based on the above, determine whether the device is abnormal. In many cases, the algorithm type of the learning process St matches the algorithm type of the identification process Sd. For example, if the type of learning process is “k-means method”, the type of identification process is often “k-means method”. Similarly, if the type of learning process is “n-layer neuro method”, the type of identification process is often “n-layer neuro method”.

In addition, as the identification operation condition Pd, for example, when the identification algorithm is “k-means method”, an abnormality determination threshold value that is applied to the distance between the cluster centroid and the sensor value is designated. . When the identification algorithm is the “n-layer neuro method”, an abnormality determination threshold value between output units to be applied to the evaluation value of the abnormal output unit is specified.
The code book CB is a result of machine learning, and includes, for example, cluster centroid information in the case of the “k-means method”, a coupling load between units in the case of the “n-layer neuro method”, and the like.

The model management information 103 is meta information attached to the entire statistical model 41. The model management information 103 includes a device type 103a, a sensor type 103b, a model creator 103c, a use approval state 103d, and a model application record 103e.
The device type 103a is the type (model) of the device to be diagnosed.
The sensor type 103b is a type of sensor.
The model creator 103c is the name of the person who developed the statistical model.

The usage approval state 103d is “approved” indicating that the model user is approved to use the post-processing model created based on the statistical model, or “unapproved” indicating that the model user has not been approved. Approved.
The model application record 103e is “present (m times)” indicating that there is a record selected as a pre-processing model (detailed later) and the number of times regardless of the above “approved” or “unapproved”. Or, “None” indicating that there is no record. However, m = 1, 2, 3,...

The operation condition rewriting information 104 is also meta information attached to the statistical model 41 as a whole. The operating condition rewriting information 104 includes a sensor input 104a, learning 104b, and identification 104c.
The sensor input 104a is either “Yes” indicating that the sensor input operation condition has been rewritten or “No” indicating that the sensor input operation condition has not been rewritten.
The learning 104b is either “Yes” indicating that the learning operation condition has been rewritten, or “No” indicating that the learning operation condition has not been rewritten.
The identification 104c is either “Yes” indicating that the identification operation condition has been rewritten or “No” indicating that the identification operation condition has not been rewritten.
The auxiliary storage device 13 stores a plurality of statistical models 41 in the pre-processing model DB 31.

  The access authority 43 will be described with reference to FIG. The access authority 43 is a matrix in which the vertical axis is the model creator, the model provider, and the model user, and the horizontal axis is the pre-processing model and the post-processing model. An access authority flag is stored in the cell at the intersection of the vertical axis and the horizontal axis. The access authority flag is either “◯” or “×”. “◯” indicates that the user on the vertical axis can completely know the type of operating condition and the operating condition parameter of the model on the horizontal axis. “X” indicates that it cannot be completely understood (details will be described later).

The model difference 44 will be described with reference to FIG. The model difference 44 has a pre-processing model column 152 and a post-processing model column 153 in association with the diagnosis procedure column 151.
The diagnosis procedure column 151 stores the above-described diagnosis procedure.
The pre-processing model column 152 includes an operation condition type column 152a and an operation condition parameter column 152b. Similarly, the post-processing model column 153 also includes an operation condition type column 153a and an operation condition parameter column 153b. In the operation condition type field 152a, the sensor type is stored for the sensor input process, and the algorithm type is stored for the learning process and the identification process. In the parameter column 152b of the operating conditions, parameters such as the number of sensors and sampling period and specific values of the parameters are stored for the sensor input process. For the learning process and the identification process, parameters used in the algorithm and specific values of the parameters are stored. The same applies to the operation condition type field 153a and the operation condition parameter field 153b.

(Model difference display screen for model creator)
For convenience of explanation, the model difference display screen 51 for the model creator will be described along FIG. 11 by skipping FIGS. The model difference column 201 of the model difference display screen 51 for model creator has the same configuration as the model difference 44 of FIG.
The model creator model difference display screen 51 is a screen visually recognized by the model creator. The model difference confirmation unit 24 of the statistical model creation device 2 displays the pre-machining model and the post-machining model in a comparison format on the model creator model difference display screen 51. The model creator knows, for example, the following by visually recognizing the model difference display screen 51 for model creator.

(1: About diagnostic procedure)
Both the pre-processing model and the post-processing model include “sensor input processing”, “learning processing”, and “identification processing” in the order of execution as diagnostic procedures.
(2: Sensor input processing)
-The sensor type of the “sensor input process” in the pre-processing model is “vibration sensor”, and the sensor type of the “sensor input process” in the post-processing model is also “vibration sensor”. That is, there is no difference in the sensor type of “sensor input processing” between the two.
-The number of sensors in the model before machining is "3", whereas the number of sensors in the model after machining is "2".
The sampling period (frequency) of the pre-processing model is “1 kHz”, whereas the sampling period of the post-processing model is “2 kHz”.

(3: About learning process)
-The pre-processing model uses the “k-means method” while the post-processing model uses the “three-layer neuro method”.
-The number of clusters in the “k-means method” in the pre-processing model is “5”, whereas the number of intermediate layer nodes in the “three-layer neuro method” in the post-processing model is “10”.

(4: About identification processing)
-The pre-processing model uses the “k-means method” while the post-processing model uses the “three-layer neuro method”.
-The number of clusters in the “k-means method” in the pre-processing model is “5”, whereas the number of intermediate layer nodes in the “three-layer neuro method” in the post-processing model is “10”.
As apparent from the above, the “type of operation condition” and “parameter of operation condition” that can be visually recognized by the model creator are not particularly limited. This is because the model creator has a confidentiality obligation for each of the model provider and the model user. This corresponds to the fact that in the access authority 43 (FIG. 3A), the access authority flags in the “model creator” line are both “◯”.

  The model provider model difference display screen 52 will be described with reference to FIG. The model difference column 221 on the model provider model difference display screen 52 has the same configuration as the model difference 44 in FIG. However, the displayed information is different from that shown in FIG. The model provider model difference display screen 52 is a screen visually recognized by the model provider. The model difference confirmation unit 24 of the statistical model creation device 2 displays the pre-processing model and the post-processing model in a comparison format on the model provider model difference display screen 52. The model provider knows, for example, the following by visually recognizing the model provider model difference display screen 52.

(1: Model before processing)
The information displayed as the pre-processing model 223 is information that the self provided to the model creator in the past.
-The model creator is trying to develop a post-processing model for a subject (which is a model user) based on the pre-processing model.

(2: Model after processing)
In the post-processing model 224, the sensor type of “sensor input processing” is “mechanical sensor”, which is a superordinate concept of the “vibration sensor” of the pre-processing model.
-It can be seen that the post-processing model uses some sensors to sample some "machine quantity" at a certain period, but the specific number of sensors and the specific sampling period are unknown.
The algorithm type of the “learning process” of the model after processing is “other than the k-means method” and not the “k-means method” that is the algorithm type of the model before processing.
・ Don't know at all about the “parameters of the operating conditions” of the “learning process” of the processed model.

-The algorithm type of the “discriminating process” of the model after processing is “other than the k-means method” and not the “k-means method” that is the algorithm type of the model before processing.
・ Never know about the “parameters of the operating conditions” in the “identification process” of the model after machining.
・ After all, you cannot know the model user's secret (design knowledge).
As is apparent from the above, the model provider cannot visually recognize the “type of operation condition” and “parameter of operation condition” belonging to the secret (design knowledge) of the model user. This means that in the access authority 43 (FIG. 3A), the access authority flag in the row of “model provider” is “◯” for the pre-process model, whereas “×” for the post-process model. It corresponds to that.

  The model user model difference display screen 53 will be described with reference to FIG. The model difference column 241 of the model user model difference display screen 53 has the same configuration as the model difference 44 of FIG. However, the displayed information is different from that shown in FIG. The model user model difference display screen 53 is a screen visually recognized by the model user. The model difference confirmation unit 24 of the statistical model creation apparatus 2 displays the pre-processing model and the post-processing model in a comparison format on the model user model difference display screen 53. The model user knows, for example, the following by visually recognizing the model user model difference display screen 53.

(1: Model after processing)
The information displayed as the post-processing model 244 is information provided to the model creator when the model creator is requested to develop the post-processing model.
-The model maker is trying to develop a post-processing model for himself / herself based on a pre-processing model for a certain subject (model provider).

(2: Model before processing)
In the pre-machining model 243, the sensor type of “sensor input processing” is “mechanical sensor”, which is a superordinate concept of the “vibration sensor” of the post-machining model.
-It can be seen that the pre-processing model uses some sensors to sample some "machine quantity" at a certain period, but the specific number of sensors and the specific period are unknown.
-The algorithm type of the “learning process” of the pre-processing model is “other than the three-layer neuro method” and not the “three-layer neuro method” that is the algorithm type of the post-processing model.
・ Never know about “parameters for operating conditions” in “learning” of the pre-processing model.

-The algorithm type of the “discriminating process” of the pre-processing model is “other than the three-layer neuro method” and not the “three-layer neuro method” that is the algorithm type of the post-processing model.
・ Don't know at all about the “parameters of the operating conditions” of the “identification” of the pre-machining model.
・ After all, you cannot know the secret (machine knowledge) of the model provider.
As is clear from the above, the model user cannot visually recognize the “type of operation condition” and “parameter of operation condition” belonging to the secret (design knowledge) of the model provider. This means that, in the access authority 43 (FIG. 3A), the access authority flag in the row of “model user” is “◯” for the model after processing, but “×” for the model before processing. It corresponds to that.

(Processing procedure)
Hereinafter, the processing procedure of this embodiment will be described. As processing procedures, (1) pre-processing model selection processing procedure, (2) operating condition processing processing procedure, (3) post-processing model creation processing procedure, (4) model difference confirmation processing procedure, and (5) model mask processing Procedure exists. (5) is a subroutine of (4). Each processing procedure is executed in the order of (1) → (2) → (3) → (4).

(Pre-processing model selection processing procedure)
A pre-processing model selection processing procedure will be described with reference to FIG.
In step S301, the pre-processing model selection unit 21 of the statistical model creation device 2 displays a model model selection screen 50 (FIG. 10). Specifically, the pre-processing model selection unit 21 displays the model model selection screen 50 on the output device 17 of the terminal device 3 used by the model creator.

  In step S302, the pre-processing model selection unit 21 receives a device type and a sensor type. Specifically, the pre-processing model selection unit 21 accepts that the model creator inputs a device type and a sensor type to the device type column 191 and the sensor type column 192 of the model model selection screen 50, respectively. Then, it accepts that the model creator presses the search button 193. Of course, the model creator may input only one of the device type and the sensor type. Furthermore, other search conditions such as a model creator may be input.

  In step S303, the pre-processing model selection unit 21 searches for a pre-processing model. Specifically, the pre-processing model selection unit 21 searches the pre-processing model DB 31 and includes a statistical model 41 having the device type and the sensor type received in step S302 as the device type 103a (FIG. 2) and the sensor type 103b, respectively. Get all. At this time, the pre-processing model selection unit 21 does not need to completely match the character string of the search key (device type and / or sensor type), and may partially match. Further, a synonym or synonym may be searched.

  In step S304, the pre-processing model selection unit 21 displays the search result. Specifically, the pre-processing model selection unit 21 displays the device type 103a, sensor type 103b, model creator 103c, and model application record 103e of the statistical model 41 acquired in step S303 in the search result column of the model model selection screen 50. 194. Looking at the example in FIG. 10, it can be seen that the pre-processing model selection unit 21 acquires at least four statistical models 41 and displays at least four records (rows) for each statistical model 41.

  In step S305, the pre-processing model selection unit 21 accepts selection by the model creator. Specifically, in the pre-processing model selection unit 21, the model creator selects one of a plurality of records (candidates for a pre-processing model) in the search result column 194 of the model model selection screen 50 ( Click on the radio button) and accept to press the OK button 195.

In step S306, the pre-processing model selection unit 21 stores a model application record. Specifically, the pre-processing model selection unit 21 determines that “Yes (once)” when the model application record 103e (FIG. 2) of the statistical model 41 corresponding to the record selected in Step S305 is “None”. Rewrite to "". When the model application record is “Yes (once)”, it is rewritten as “Yes (once)”. Then, each time a certain statistical model 41 is selected, the model application record of the statistical model 41 is “none” → “yes (once)” → “yes (twice)” → “yes (three times)” It changes like "→ ...".
Thereafter, the pre-processing model selection processing procedure is terminated.

(Operating condition processing procedure)
The operation condition processing procedure will be described with reference to FIG.
In step S311, the operation condition processing unit 22 of the statistical model creation device 2 rewrites the sensor input operation condition. Specifically, first, the operating condition processing unit 22 acquires the statistical model 41 corresponding to the record selected in step S305. Hereinafter, the statistical model 41 is referred to as a pre-processing model.
Secondly, the operation condition processing unit 22 rewrites the part that specifies the device and sensor in the device type 103a and the sensor type 103b of the target pre-processing model into meaningless symbols or character strings. For example, “absorption type cooling / heating machine” is rewritten as “♭ type dredging machine”, and “ABC vibration sensor” is rewritten as “♭♭ 制 ♭ sensor”. Third, the operating condition processing unit 22 receives the sensor input. The operating condition Pi is initialized with a meaningless number. For example, “sampling period = 2 kHz” is rewritten to “sampling period = 999 kHz”.

  In step S312, the operation condition processing unit 22 rewrites the learning operation condition. Specifically, the operation condition processing unit 22 initializes the parameter value of the learning operation condition Pt of the target pre-processing model with a meaningless number. For example, “cluster number = 5” is rewritten to “cluster number = 999”.

  In step S313, the operation condition processing unit 22 rewrites the identification operation condition. Specifically, the operation condition processing unit 22 initializes the parameter value of the identification operation condition Pd of the target pre-processing model with a meaningless number. For example, “abnormality determination threshold = 10” is rewritten to “abnormality determination threshold = 999”.

  In step S314, the operation condition processing unit 22 stores the rewritten statistical model. Specifically, the operation condition processing unit 22 temporarily stores, in the main storage device 12, the target pre-processing model in which the sensor input operation condition Pi, the learning operation condition Pt, and the identification operation condition Pd are rewritten in steps S311 to S313. To do. Note that the operation condition processing unit 22 preferably deletes the code book CB of the target pre-processing model. Thereafter, the operating condition processing procedure is terminated.

(Processing model creation processing procedure)
The post-processing model creation processing procedure will be described with reference to FIG.
In step S321, the post-processing model creation unit 23 of the statistical model creation device 2 creates a post-processing model. Specifically, the post-processing model creation unit 23 creates the statistical model 41 in which the diagnosis procedure 101, the operation condition 102, the model management information 103, and the operation condition rewrite information 104 are not set. Hereinafter, this statistical model 41 is referred to as “target model after processing”.

  In step S322, the post-process model creation unit 23 creates a post-process model diagnosis procedure. Specifically, the post-machining model creation unit 23 copies the diagnosis procedure of the target pre-machining model to the location of the target post-machining model diagnosis procedure 101.

  In step S323, the post-processing model creation unit 23 creates operating conditions for the post-processing model. Specifically, the post-machining model creation unit 23 copies the operating condition of the target pre-machining model to the location of the operating condition 102 of the target post-machining model. At this stage, the diagnosis procedure for the target post-processing model is the same as the diagnosis conditions for the pre-processing model, and the operation conditions of the target post-processing model are rewritten to initial values (insignificant numbers or symbols).

In step S324, the post-processing model creation unit 23 edits the post-processing model. Specifically, the post-processing model creation unit 23 accepts that the model creator inputs the following information.
-As sensor input operation conditions, the sensor type of the model user's diagnosis target device and the number of sensors-Algorithm type used for learning processing and identification processing-As learning operation conditions and identification conditions, parameters and their values

  In step S325, the processed model creation unit 23 completes the processed model. Specifically, the post-processing model creation unit 23 stores “unapproved” as the use approval state 103 d of the target post-processing model, and stores the target post-processing model in the post-processing model DB 42. Thereafter, the post-processing model creation processing procedure is terminated.

(Model difference confirmation processing procedure)
A model difference confirmation processing procedure will be described with reference to FIG.
In step S331, the model difference confirmation unit 24 of the statistical model creation device 2 determines whether or not the user is a model provider. Specifically, first, the model difference confirmation unit 24 accepts that the user of the statistical model creation device 2 inputs information indicating his / her position via the input device 16 of his / her terminal device 3. The information indicating his / her position is either “model provider”, “model creator”, or “model user”.
Second, if the received information is “model provider” (step S331 “Yes”), the model difference confirmation unit 24 proceeds to step S332, and otherwise (step S331 “No”), step S335. Proceed to

  In step S332, the model difference confirmation unit 24 executes a model mask processing procedure with the processed model as an argument. Details of step S332 will be described later, but as a result, the model difference confirmation unit 24 creates information for displaying the model difference 221 on the model provider model difference display screen 52 (FIG. 12).

  In step S333, the model difference confirmation unit 24 displays the model provider model difference display screen 52 (FIG. 12). Specifically, the model difference confirmation unit 24 displays the model difference display screen 52 for model provider on the output device 17 of the terminal device 3 used by the model provider.

  In step S334, the model difference confirmation unit 24 accepts approval. The model provider pays attention to the processed model column 224 of the model provider model difference display screen 52. Then, in the same way that the user does not know the secret (design knowledge) of the model user, the user confirms that the secret of the model user is also hidden from the model user as in the model column 224 after processing. In addition, the model provider confirms that a post-processing model is newly created based on the pre-processing model. Then, as a proof of the confirmation, the model user selects “Approve” in the approval status column 225a of the model use approval column 225 (clicks a radio button), and inputs his / her name in the final approver column 225c. After that, the “OK” button 226 is pressed. The model difference confirmation unit 24 receives such input information. Then, the model difference confirmation unit 24 displays the current year, month, date and time in the final approval date / time column 225b, and then ends the model difference confirmation processing procedure.

  In step S335, the model difference confirmation unit 24 determines whether or not the user is a model creator. Specifically, the model difference confirmation unit 24 proceeds to step S336 if the information received in “first” in step S331 is “model creator” (step S335 “Yes”), and otherwise ( Step S335 “No”), the process proceeds to Step S337.

In step S336, the model difference confirmation unit 24 displays the model creator model difference display screen 51 (FIG. 11). Specifically, the model difference confirmation unit 24 displays the model difference display screen 52 for model creator on the output device 17 of the terminal device 3 used by the model creator. Thereafter, the model difference confirmation processing procedure ends.
The model difference confirmation unit 24 displays a copy of the model use approval field 225 of the model provider display model difference display screen 52 (FIG. 12) in the model use approval state confirmation field 205. Incidentally, in the example of FIG. 11, the model provider is not approved.

  In step S337, the model difference confirmation unit 24 executes a model mask processing procedure with the pre-process model as an argument. Although details of step S337 will be described later, as a result, the model difference confirmation unit 24 creates information for displaying the model difference 241 on the model user model difference display screen 53 (FIG. 13).

  In step S338, the model difference confirmation unit 24 determines whether or not it has been approved. Specifically, if the model provider selects “Approve” in step S334 (step S338 “Yes”), the model difference confirmation unit 24 proceeds to step S339, and otherwise (step S338 “No”). ), The model difference confirmation processing procedure is terminated.

In step S339, the model difference confirmation unit 24 displays the model user model difference display screen 53 (FIG. 13). Specifically, the model difference confirmation unit 24 displays a model user model difference display screen 53 on the output device 17 of the terminal device 3 used by the model user. Thereafter, the model difference confirmation processing procedure ends.
In addition, although the example of the screen display was described in step S333, S336, and S339, these are only examples. For example, the model difference confirmation unit 24 may output data for creating each screen to a recording medium or the like, or may transmit the data to another device. That is, the model difference confirmation unit 24 only needs to be able to output the model difference in a comparison format by any method.

(Model mask processing procedure)
The model mask processing procedure will be described with reference to FIG. The model mask processing procedure is details of steps S332 and S337 of the model difference confirmation processing procedure. The processing contents of the model mask processing procedure are slightly different depending on whether the argument is a “model after processing” (step S332) or the argument is a “model before processing” (step S337). First, the case where the argument is “model after processing” will be described.
As a large flow, the model difference confirmation unit 24 of the statistical model creation apparatus 2 repeats the iterative process of steps S351 to S354 for the sensor input process Si, the learning process St, and the identification process Sd.

In step S351, the model difference confirmation unit 24 determines whether or not the types of operation conditions are the same. Specifically, first, the model difference confirmation unit 24 determines whether or not the types of operation conditions of the target pre-processing model and the target post-processing model are the same. The type of operation condition is a sensor type in the sensor input process Si, and an algorithm type in the learning process St and the identification process Sd.
Secondly, the model difference confirmation unit 24 proceeds to step S352 if the types of operation conditions are the same (step S351 “Yes”), and proceeds to step S353 otherwise (step S351 “No”).

In step S352, the model difference confirmation unit 24 converts the type of operation condition into a higher concept. Specifically, the model difference confirmation unit 24 executes the following processing.
(Processing for sensor input processing Si)
For example, when the sensor type of the target pre-processing model is “vibration sensor” and the sensor type of the target post-processing model is also “vibration sensor”, the model difference confirmation unit 24 sets the type of operation condition of the target post-processing model. Rewrite as “mechanical sensor”.

(Processing for learning process St and identification process Sd)
For example, when the algorithm type of the target pre-processing model is “k-means method” and the algorithm type of the target post-processing model is also “k-means method”, the model difference confirmation unit 24 operates the target post-processing model. Rewrite the condition type to "clustering". “Clustering” is a superordinate concept of the “k-means method”.

In step S353, the model difference confirmation unit 24 masks the type of operation condition. Specifically, the model difference confirmation unit 24 executes the following processing.
(Processing for sensor input processing Si)
For example, when the sensor type of the target pre-processing model is “vibration sensor” and the sensor type of the target post-processing model is “temperature sensor”, the model difference confirmation unit 24 sets the type of operation condition of the target post-processing model. Rewrite as “Other than vibration sensor”.
(Processing for learning process St and identification process Sd)
For example, when the algorithm type of the target pre-processing model is “k-means method” and the algorithm type of the target post-processing model is “n-layer neuro method”, the model difference confirmation unit 24 operates the target post-processing model. Rewrite the condition type to “other than k-means method”.

In step S354, the model difference confirmation unit 24 masks the operating condition parameters. Specifically, the model difference confirmation unit 24 executes the following processing.
(Processing for sensor input processing Si)
For example, the operating condition parameter of the target pre-processing model is “number of sensors = 999” (initialized to a meaningless value), and the operating condition parameter of the target post-processing model is “sensor number = 20” ( Is being edited). In this case, the model difference confirmation unit 24 rewrites the parameter of the operation condition of the target post-processing model as “number of sensors = #”.
As another example, when the operating condition parameter of the target pre-processing model is “sampling period = 999 kHz” and the operating condition parameter of the target post-processing model is “sampling period = 10 kHz”, the model difference confirmation unit 24 rewrites the parameter of the operating condition of the target post-processing model to “sampling period = # kHz”.

(Processing for learning process St and identification process Sd)
The model difference confirmation unit 24 rewrites the parameter of the operation condition of the target post-processing model to “−”.
After exiting the iterative process, the model difference confirmation unit 24 ends the model mask processing procedure and proceeds to step S333.

Next, the model mask processing procedure when the argument is “pre-processing model” will be described. The model mask processing procedure in this case is different only in the following (1) to (3) as compared with the case where the argument is “model after processing”.
(1) In step S352, the type of operation condition to be rewritten is “type of operation condition of target pre-machining model”.
(2) In step S353, the type of operation condition to be rewritten is “type of operation condition of target pre-processing model”.
(3) In step S354, the parameter of the operating condition to be rewritten is “the parameter of the operating condition of the target pre-processing model”.

(Initialization, superordinate conceptualization and mask range)
In the above, as an example in which the operation condition processing unit 22 initializes the operation condition, an example in which a part of a character string of the sensor type is rewritten with a meaningless symbol or character string, and an example in which a parameter value is rewritten with a meaningless number Mentioned. However, the unit to be initialized is not limited to this. For example, a part of the parameters (for example, “cluster” in the “number of clusters”) may be targeted for initialization, or only a part of the parameter values (for example, specific digits) may be targeted for initialization. May be. That is, it is sufficient that at least a part of the operating conditions is initialized. The same applies to the superordinate conceptualization and the mask.

(Modification 1 of the first embodiment)
In the above example, the statistical model creation device 2 rewrites the operation condition type of the target pre-processing model or the target post-processing model according to the difference in algorithm type between the target pre-processing model and the target post-processing model ( Mask or superordinate conceptualization). However, it is more desirable that the model creator can select the target pre-processing model from among the candidates for the statistical model 41 having the same algorithm type.

  For example, the parameters of the “n-layer neuro method” are different from the parameters of the “k-means method”. For example, assume that the target post-processing model uses the “n-layer neuro method”. Then, comparing the case where the target pre-processing model uses the “k-means method” and the case where the target pre-processing model uses the “n-layer neuro method”, the latter case is the model. The development burden is lighter for the creator.

Therefore, in the modification of the first embodiment, the above description is changed as follows.
(1) The model model selection screen 50 (FIG. 10) has an “algorithm type” input field as a field in which an essential search condition is input.
(2) In step S302, the pre-processing model selection unit 21 of the statistical model creation device 2 accepts that the model creator inputs “algorithm type”.
(3) In step S303, the pre-processing model selection unit 21 adds the received “algorithm type” to the “search key”. The pre-processing model selection unit 21 can know the algorithm type from the contents of the learning operation condition Pt and the identification operation condition Pd.

(Modification 2 of the first embodiment)
In step S334 of the model difference confirmation processing procedure (FIG. 7), the model difference confirmation unit 24 accepts approval from the model provider. However, the model difference confirmation unit 24 may accept approval from the model creator. In this case, it is assumed that the model provider has given the authorization for approval to the model creator in advance. That is, the model difference confirmation unit 24 may execute the process in step S334 immediately after the process in step S336 (before the paths from “No” in step S338 join).

(Second Embodiment)
After the model creator X creates a new statistical model 41b based on an existing statistical model 41a, the new statistical model 41b starts to be used. Then, another model creator Y creates a new statistical model 41c based on the statistical model 41b. As the number of model creators creating the underlying statistical model increases, or as the number of statistical models created by a model creator increases, the choice of the underlying statistical model increases for subsequent model builders However, statistical model development is facilitated. Therefore, the second embodiment has an “incentive processing procedure” that gives points to each model creator and motivates the creation of a statistical model.

(Incentive processing procedure)
The incentive processing procedure will be described with reference to FIG.
The trigger for starting the incentive process is step S305 of the pre-processing model selection process procedure. As described above, in step S305, the pre-processing model selection unit 21 of the statistical model creation device 2 accepts a selection by the model creator. The point granting unit 25 of the statistical model creation device 2 constantly monitors the process. Then, the point assigning unit 25 recognizes the model creator selected at this time (clicked on the radio button) as a “new model creator”, and selects the model creator of the statistical model 41 selected at this time as “create existing model”. As a “person”

In step S361, the point giving unit 25 calculates model points. Specifically, first, the point giving unit 25 sets the “number of selections” to “1”.
Second, the point assigning unit 25 sets a value obtained by multiplying the “number of selections” by a predetermined coefficient as a “model point”.

  In step S362, the point giving unit 25 calculates an operation condition point. Specifically, first, the point giving unit 25 calculates the data amount ΔPi rewritten from the sensor input operation condition Pi of the target pre-processing model to the sensor input operation condition Pi of the target post-processing model. Similarly, a data amount ΔPt rewritten from the learning operation condition Pt of the target pre-processing model to the learning operation condition Pt of the target post-processing model is calculated, and the target post-processing model is identified from the identification operation condition Pd of the target pre-processing model. A data amount ΔPd rewritten to the operating condition Pd is calculated.

Second, the point assigning unit 25 sets “ΔPi + ΔPt + ΔPd” multiplied by a predetermined coefficient as an “operation condition point”. The point giving unit 25 may set a value obtained by multiplying “w _i × ΔPi + w _t × ΔPt + w _d × ΔPd” by a predetermined coefficient as an “operation condition point”. Here, w _i + w _t + w _d = 1, 0 ≦ w _i ≦ 1, 0 ≦ w _t ≦ 1 and 0 ≦ w _d ≦ 1 hold. That is, w _i , w _t, and w _d are weights for ΔPi, ΔPt, and ΔPd, respectively.

  In step S363, the point granting unit 25 calculates incentive points. The incentive point is a value obtained by adding the operating condition point to the template point.

In step S364, the point giving unit 25 adds points. Specifically, first, the point giving unit 25 determines the value of “r ₁ × incentive point”, the value of “r ₁ × model point” as the number thereof, and “r ₁ × operation condition point”. Is assigned to the existing model creator.
Second, the point assigning unit 25 sets the value of “r ₂ × incentive point”, the value of “r ₂ × model point” and the value of “r ₂ × operation condition point” as the new model. Grant to creator.
Here, 0 ≦ r ₂ ≦ r ₁ ≦ 1. That is, a larger point is given to the existing model creator.

The point granting unit 25 repeats the processing of steps S361 to S364 every time it goes through step S305. The point assigning unit 25 adds the value of “r ₁ × incentive point”, the value of “r ₁ × model point”, and the value of “r ₁ × operation condition point” for each model creator. The addition result is stored in the auxiliary storage device 13. Similarly, for each model creator, the value of “r ₂ × incentive point”, the value of “r ₂ × model point”, and the value of “r ₂ × operating condition point” are added to assist the addition result. Store in the storage device 13. Here, the incentive processing procedure is temporarily interrupted.

In step S365, the point giving unit 25 displays the point display screen 54 (FIG. 14A). Specifically, first, the point granting unit 25 accepts that the model creator inputs a “point inquiry request” via the input device 16 of his / her terminal device 3 at an arbitrary time point. The incentive processing procedure is restarted when triggered.
Secondly, the point granting unit 25 displays the point display screen 54 on the output device 17 of the model creator's terminal device 3 that has input the “point inquiry request”.
Thirdly, the point assigning unit 25 reads the addition result from the auxiliary storage device 13 and displays the addition result in the user list field 251 of the point display screen 54.

  In step S366, the point granting unit 25 displays the user detailed information display screen 55 (FIG. 14B). Specifically, first, the point granting unit 25 accepts that the model creator selects an arbitrary record in the user list field 251 (clicks on a radio box) and presses an “OK” button 252. Now, it is assumed that the model creator has selected a record of “Suzuki ○ male”.

Second, the point granting unit 25 searches the pre-processing model DB 31 and acquires all the statistical models 41 whose model creator 103c is “Suzuki ○ male”. Then, all the device types 103a and sensor types 103b of the acquired statistical model 41 are acquired.
Third, the point granting unit 25 displays the user detailed information display screen 55 on the output device 17 of the model creator's terminal device 3 that has input the “point inquiry request”.

Fourth, the point granting unit 25 displays “Suzuki ○ male” in the user name field 261a of the user detailed information display screen 55. Then, the current value (addition result) of the incentive point, model point, and operation condition point of “Suzuki ○ male” is displayed in the incentive point field 261b, the model point field 261c, and the operation condition point field 261d, respectively.
Fifth, the point providing unit 25 displays the device type and the sensor type acquired in “second” in the device column 261e that can be diagnosed and the sensor column 261f that can be handled, respectively.
Thereafter, the incentive processing procedure is terminated.

(Effect of embodiment)
The statistical model creation device 2 of the present embodiment has the following effects.
(1) Development efficiency of a new statistical model is improved, and design knowledge included in an existing statistical model can be visually recognized in a manner according to the user's position.
(2) It is possible to keep design knowledge about operating conditions secret between model providers and model users.
(3) The outline of design knowledge that is relatively less necessary to be kept secret by making a part of the operating conditions higher-level concept is disclosed, and the need to keep it secret by masking a part of the operating conditions The relatively high design knowledge can be kept completely secret.
(4) By initializing some of the operating conditions in advance, the disclosure of design conditions can be limited to a minimum even if the conception or masking fails.
(5) It is possible to keep secret the algorithm type and parameter value that are important for handling the device.
(6) An abnormal state of the device can be detected.

(7) A new design model can be created without ignoring the intention of the model provider.
(8) A new design model can be efficiently created based on a design model having the same type of algorithm.
(9) The design model creator can select a base design model.
(10) A motivation for diverting the design model can be given to both the creator of the existing design model and the new design model.
(11) The creator of an existing design model can be motivated to create a statistical model that is more easily diverted.
(12) By providing more points to the creator of the existing design model, it is possible to avoid a shortage of supply of the existing statistical model.
(13) Since the storage unit for storing the statistical model is independent, cooperation with an external customer is facilitated.

  In addition, this invention is not limited to an above-described Example, Various modifications are included. For example, the above-described embodiments have been described in detail for easy understanding of the present invention, and are not necessarily limited to those having all the configurations described. Further, a part of the configuration of one embodiment can be replaced with the configuration of another embodiment, and the configuration of another embodiment can be added to the configuration of one embodiment. Further, it is possible to add, delete, and replace other configurations for a part of the configuration of each embodiment.

Each of the above-described configurations, functions, processing units, processing means, and the like may be realized by hardware by designing a part or all of them with, for example, an integrated circuit. Each of the above-described configurations, functions, and the like may be realized by software by interpreting and executing a program that realizes each function by the processor. Information such as programs, tables, and files that realize each function can be stored in a recording device such as a memory, a hard disk, or an SSD (Solid State Drive), or a recording medium such as an IC card, an SD card, or a DVD.
In addition, the control lines and information lines are those that are considered necessary for the explanation, and not all the control lines and information lines on the product are necessarily shown. In practice, it may be considered that almost all the components are connected to each other.

DESCRIPTION OF SYMBOLS 1 Statistical model creation system 2 Statistical model creation apparatus 3 Terminal device 11, 18 Central controller (control part)
12, 19 Main storage device (storage unit)
13, 20 Auxiliary storage device (storage unit)
14, 15 Communication device 16 Input device 17 Output device 21 Pre-processing model selection unit 22 Operating condition processing unit 23 Post-processing model creation unit 24 Model difference confirmation unit 25 Point grant unit 41, 42 Statistical model 43 Access authority 44 Model difference

Claims (15)

A storage unit that stores an operation procedure for handling the device and a statistical model including an operation condition corresponding to each of the operation procedures;
A first statistical model for handling the first device is acquired with reference to the storage unit, and a second statistical model for handling the second device is obtained based on the acquired first statistical model. Create a new one
When the operation procedure and the operation condition of the first statistical model and the operation procedure and the operation condition of the second statistical model are output to the user in a comparison format, the first statistical model After changing at least a part of the operation condition of the second statistical model or at least a part of the operation condition of the second statistical model so as to conceal the contents of the information according to the user's position of the output destination A control unit for outputting;
A statistical model creation device comprising: The controller is
When the user of the output destination is a person related to the first device,
Changing at least part of the operating conditions of the second statistical model without changing the operating conditions of the first statistical model so as to conceal the content of the information ;
When the user of the output destination is a person related to the second device,
Changing at least part of the operating conditions of the first statistical model without changing the operating conditions of the second statistical model so as to conceal the content of the information ;
The statistical model creation device according to claim 1. The controller is
When the operating condition of the first statistical model and the operating condition of the second statistical model are the same,
Outputting at least a part of the operating condition of the first statistical model or at least a part of the operating condition of the second statistical model as a superordinate concept;
If the operating condition of the first statistical model and the operating condition of the second statistical model are different,
Masking and outputting at least part of the operating condition of the first statistical model or at least part of the operating condition of a second statistical model;
The statistical model creation apparatus according to claim 2. The controller is
Prior to the change, initializing at least part of the operating conditions of the first statistical model with meaningless data;
The statistical model creation device according to claim 3. Some of the operating conditions to be changed are:
Including the type of algorithm for handling the device, and parameter values used by the algorithm;
The statistical model creation apparatus according to claim 4. The handling is
Including diagnosing whether the device is in an abnormal state,
The statistical model creation apparatus according to claim 5. The controller is
Based on the first statistical model, the creation of the second statistical model is authorized by a person related to the first device or authorized by a person related to the first device. Accepting the approval of the
The statistical model creation apparatus according to claim 6. The controller is
In obtaining the first statistical model for handling the first device,
Obtaining the first statistical model using an algorithm of the same type as the algorithm type of the second statistical model from the storage unit;
The statistical model creation device according to claim 7. The controller is
In obtaining the first statistical model,
Obtaining a plurality of candidates for the first statistical model;
Receiving the user selecting one of the acquired candidates;
The statistical model creation apparatus according to claim 8. The controller is
Giving points to the creator of the selected first statistical model and the person who made the selection;
The statistical model creation apparatus according to claim 9. The controller is
Calculating the point according to the selected number of times and the amount of data rewritten from the first statistical model to the second statistical model;
The statistical model creation device according to claim 10. The controller is
Reducing the points to be given to those who have made the selection compared to the points to be given to the creator of the selected first statistical model;
The statistical model creation apparatus according to claim 11. The storage unit and the control unit are
Each is an independent configuration,
The controller is
Accessing the storage unit via a network;
The statistical model creation apparatus according to claim 12. The storage unit of the statistical model creation device is
Stores a statistical model that includes operating procedures for handling equipment and operating conditions corresponding to each of the operating procedures,
The control unit of the statistical model creation device includes:
A first statistical model for handling the first device is acquired with reference to the storage unit, and a second statistical model for handling the second device is obtained based on the acquired first statistical model. Create a new one
When the operation procedure and the operation condition of the first statistical model and the operation procedure and the operation condition of the second statistical model are output to the user in a comparison format, the first statistical model After changing at least a part of the operation condition of the second statistical model or at least a part of the operation condition of the second statistical model so as to conceal the contents of the information according to the user's position of the output destination Output,
The statistical model creation method of the statistical model creation apparatus characterized by this. For the storage part of the statistical model creation device,
A statistical model including an operation procedure for handling the device and an operation condition corresponding to each of the operation procedures;
For the control unit of the statistical model creation device,
A first statistical model for handling the first device is acquired with reference to the storage unit, and a second statistical model for handling the second device is obtained based on the acquired first statistical model. Create a new
When the operation procedure and the operation condition of the first statistical model and the operation procedure and the operation condition of the second statistical model are output to the user in a comparison format, the first statistical model After changing at least a part of the operation condition of the second statistical model or at least a part of the operation condition of the second statistical model so as to conceal the contents of the information according to the user's position of the output destination To execute the process to output,
Statistical model creation program for functioning a statistical model creation device characterized by

Images