MANAGEMENT DEVICE, MANAGEMENT METHOD, AND MANAGEMENT PROGRAM
Claim of Priority
The present application claims priority from Japanese
patent application JP 2019-139086 filed on July 29, 2019, the
content of which is hereby incorporated by reference into this
application.
BACKGROUND
[0001]
The present invention relates to a management device which
manages a target object to be managed, a management method, and
a management program.
[0002]
Conventionally, a learning inference system using a DNN
(Deep Neural Network) requires a lot of datasets for learning.
In the normal learning flow, a dataset acquired by an edge
terminal is transferred to a cloud server. The cloud server
generates a learned model in accordance with an edge environment
in order to improve the accuracy, using the dataset.
[0003]
As an example of the background art, for example, U.S.
Patent Application Publication No. 2019-42955 discloses various
systems and methods for starting and executing contextualized AI
inferencing. In an example of the systems and methods,
operations performed with a gateway computing device to invoke
an inferencing model include receiving and processing a request
for an inferencing operation, selecting an implementation of the
inferencing model on a remote service based on a model
specification and contextual data from the edge device, and
executing the selected implementation of the inferencing model,
3
such that results from the inferencing model are provided back
to the edge device. Operations performed with an edge computing
device to request an inferencing model include collecting
contextual data, generating an inferencing request, transmitting
the inference request to a gateway device, and receiving and
processing the implementation results.
SUMMARY
[0005]
However, in some edge terminals, transferring of datasets
is difficult from the viewpoints of communication costs and
rights for transferring the datasets. According to the
conventional technique of U.S. Patent Application Publication
No. 2019-42955, what is requested is a learned model which is
managed by a data center being a cloud environment, based on the
type of AI inferencing model, the sensor identifier, or
specification of the edge device, in association with vehicles
as environments of the edge device. However, it does not suggest
generation of a learned model corresponding to a new edge
environment.
[0006]
An object of the present invention is to attain learning
with high accuracy, while avoiding transferring of datasets from
the edge terminal to the cloud server.
[0007]
According to an aspect of the present invention disclosed
in this application, there is provided a management device
accessible to a target object to be managed, the device
including: a processor which executes a program; a storage
device which stores the program; and a communication interface
4
which can communicate with the target object to be managed, in
which the processor executes a reception process for receiving
first environmental information representing a first environment
of the target object to be managed, a first generation process
for generating relevant information representing relevancy
between the first environmental information received by the
reception process and second environmental information
representing a second environment of the target object to be
managed, a second generation process for generating a first
learned model to be used for inference by the target object to
be managed in the first environment, based on the relevant
information generated by the first generating process and a
second learned model to be used for inference by the target
object to be managed in the second environment, and a
transmission process for transmitting the first learned model
generated by the second generation process to the target object
to be managed.
[0008]
According to the typical preferred embodiment of the
present invention, it is possible to attain the learning with
high accuracy, while avoiding transferring of datasets from the
edge terminal to the cloud server. Those objects,
configurations, and effects other than those described above
will be apparent from the descriptions of the preferred
embodiments as described below.
BRIED DESCRIPTION OF THE DRAWINGS
FIG. 1 is an explanatory diagram illustrating a system
configuration example of a management system according to a
first embodiment.
5
FIG. 2 is an explanatory diagram illustrating an example of
an environmental information DB illustrated in FIG. 1.
FIG. 3 is an explanatory diagram illustrating an example of
an explanatory diagram illustrating an example of a “matched
number counting table” according to the first embodiment.
FIG. 4A and FIG. 4B are explanatory diagrams illustrating
respectively a model blending condition and a learned model,
according to the first embodiment.
FIG. 5A, FIG. 5B, and FIG. 5C are explanatory diagrams
illustrating examples of a neural network process which is
present in one layer of a CNN.
FIG. 6 is an explanatory diagram illustrating an example of
weight filters of the entire layers (L number of layers) of the
CNN.
FIG. 7 is an explanatory diagram illustrating an example of
a learned model including weight filter values of the entire
layers of the CNN.
FIG. 8 is a block diagram illustrating an example of a
hardware configuration of a computer.
FIG. 9 is a sequence diagram illustrating a sequence
example of a management system according to the first
embodiment.
FIG. 10 is an explanatory diagram illustrating an example
of an environmental information DB after updated.
FIG. 11 is an explanatory diagram illustrating an example
of a matched number counting table after updated.
FIG. 12 is an explanatory diagram illustrating an example
of a matched number counting table according to a second
embodiment.
FIG. 13A and FIG. 13B are explanatory diagrams respectively
6
illustrating model blending conditions and a learned model,
according to the second embodiment.
FIG. 14 is an explanatory diagram illustrating an example
of a system configuration of a management system according to a
third embodiment.
FIG. 15 is a sequence diagram illustrating a sequence
example of the management system according to the third
embodiment.
FIG. 16 is a graph illustrating the transition of
recognition accuracy of the CNN.
FIG. 17 is an explanatory diagram illustrating examples of
learned models stored in a model DB.
FIG. 18 is an explanatory illustrating an example of an
environmental information DB according to a fourth embodiment.
FIG. 19 is an explanatory diagram illustrating a matched
number counting table in association with edge terminals.
FIG. 20A and FIG. 20B are explanatory diagrams illustrating
a model blending condition and a learned model, according to the
fourth embodiment.
DETAILED DESCRIPTION
First Embodiment
[0010]

FIG. 1 is an explanatory diagram illustrating an example of
a system configuration of a management system according to the
first embodiment. The management system 100 has a management
device 101 and an edge environment 102. The management device
101 and the edge environment 102 are connected and communicable
with each other through a network 103 (regardless of
7
wire/wireless), such as the Internet, a LAN (Local Area
Network), a WAN (Wide Area Network), or the like. The management
system 100 has only one edge environment 102, but may include a
plurality of edge environments 102.
[0011]
The management device 101 is a cloud environment. The edge
environment 102 is a target to be managed by the management
device 101, and is composed of an edge server and one or more
edge terminals Ti (“i” is an integer satisfying 1 ≤ i ≤ n). In
FIG. 1, in the edge environment 102, “n-1” edge terminals T1 to
Tn-1 are arranged, and an edge terminal Tn is newly added to the
edge environment 102. The edge server is a gateway of the edge
environment 102, and communicably connects the management device
101 and each of the edge terminals Ti.
[0012]
In FIG. 1, the edge environment 102 is, for example, a
factory, and the edge terminal Ti is a work machine, a robot, or
a vehicle. The edge terminal Ti is not necessarily a mobile
body. The edge environment 102 is not limited to a factory. For
example, the edge environment 102 may be a warehouse in and from
which commodities are warehoused/shipped. Functions of the edge
server may be included in each of the edge terminals Ti.
[0013]
The management device 101 generates a blend recipe from
each environmental information of the edge terminals T1 to Tn-1,
and blends the generated blend recipe with learned models LM1 to
LMn, thereby generating the learned model LMn of the newly-added
edge terminal Tn without executing a learning process for a
dataset.
[0014]
8
Specifically, for example, the management device 101 has a
model DB 110, an environmental information DB 111, a first
communication unit 112, a calculation unit 113, and a blending
unit 114. The model DB 110 stores learned models LMi in
association with each edge terminal Ti. The learned models LMi
are weight parameters of the DNN that can be acquired by
learning datasets for learning. However, because the management
device 101 does not have datasets, it does not learn datasets.
Thus, the learned model LMi is externally-prepared data.
[0015]
The environmental information DB 111 stores environmental
information, in association with the edge terminals Ti. Note
that the environmental information represents the environment of
the edge terminal Ti. The environment of the edge terminal Ti is
the situation of the edge terminal Ti itself or its surrounding
situation. It includes, for example, use of DNN implemented in
the edge terminal Ti, an arrangement position of the edge
terminal Ti, a type of the edge terminal Ti, a user operating
the edge terminal Ti, the temperature inside or outside the edge
terminal Ti, the behavior of the edge terminal Ti, and a work
time zone of the edge terminal Ti. A part of the environmental
information is detected by a sensor 123, such as a camera or the
like, connected to each edge terminal Ti. The environmental
information DB will be described later in detail with reference
to FIG. 2.
[0016]
The first communication unit 112 receives environmental
information En from the newly-added edge terminal Tn, sends the
generated learned model LMn of the edge terminal Tn to the edge
terminal Tn.
9
[0017]
The calculation unit 113 calculates a model blending
condition Cn with regard to the new edge terminal Tn. The model
blending condition Cn is the above-described blend recipe. The
model blending condition Cn is relevant information which is
determined based on the relevancy between existing environmental
information E1 to En (see FIG. 2) of the respective edge
terminals T1 to Tn-1 and the environmental information En of the
edge terminal Tn. In particular, the relevancy may be the
matched number therebetween.
[0018]
The blending unit 114 blends the model blending condition
Cn with the existing learned models LM1 to LMn-1 to generate a
learned model LMn of the edge terminal Tn. As the abovedescribed matched number is large, parameters of the learned
model LMn of the edge terminal Tn would be close to weight
parameters of the existing learned models LM1 to LMn-1.
[0019]
FIG. 1 illustrates an equation 1 as a model blending
example. The left side of the equation represents the learned
model LMn of the edge terminal Tn. The denominator on the right
side of the equation represents the number of edge terminals T1
to Tn-1. A term Ri of the numerator on the right side thereof
represents a weight based on the matched number between
environmental information Ei of the existing edge terminal Ti
and that of the edge terminal Tn, and would be one element of
the model blending condition Cn. As the matched number is high,
the weight becomes large, and its learned model LMi has a high
degree of effect on the learned model LMn.
[0020]
10
The edge server 120 has a second communication unit 121.
The second communication unit 121 transfers environmental
information En from the edge terminal Tn to the management
device 101, and transfers the learned model LMn from the
management device 101 to the edge terminal Tn.
[0021]
The edge terminal Ti has the learned model LMi, an
inference unit 122, and the sensor 123. The inference unit 122
is, for example, a convolutional neural network (CNN) as an
example of the DNN. The inference unit 122 applies the learned
model LMi to a CNN, inputs image data from a camera as the
sensor 123 to the CNN, and outputs an inference result. The edge
terminal Ti controls the behavior of the edge terminal Ti itself
or any connected device, using the inference result.
[0022]
In this manner, the management device 101 generates a model
blending condition Cn as a blend recipe from the environmental
information of the edge terminals T1 to Tn-1, blends the
generated model blending condition Cn with the learned models
LM1 to LMn-1, thereby generating the learned model LMn of the
newly-added edge terminal Tn without executing learning of
datasets.
[0023]
The edge terminal Tn executes inference using the learned
model LMn. Then, the edge terminal Tn does not need to upload a
dataset to the management device 101, thus not incurring excess
communication cost for transferring the dataset. Because there
is no need to upload the dataset to the management device 101,
leakage of the dataset is prevented. The management device 101
does not need to implement a learning function using the
11
dataset, thus attaining a reduction in calculation load.
[0024]

FIG. 2 is an explanatory diagram illustrating an example of
the environmental information DB 111 illustrated in FIG. 1. The
environmental information DB 111 has a p-number (p is an integer
of 1 or higher) of edge information items e1 to ep in
association with edge numbers 200. For example, the edge
information items e includes items of, for example, a use of DNN
e1, an edge environment e2, 3M information e3, Man information
e4, …, and time information ep, as environmental items. A
combination of edge information items in one row is the
environmental information Ei of an edge terminal Ti specified by
its corresponding edge number 200.
[0025]
The “3M information” e3 includes at least one of three
items of “machine”, “method”, and “material”. The “Man
information” e4 represents the user of the edge terminal Ti (for
example, the number, physical appearance, or sex of users). In
this embodiment, the environmental information E1 to En-1 are
stored in the environmental information DB 111. The edge
terminal Tn is added to the edge environment 102, and the
management device 101 receives environmental information En from
the edge terminal Tn. This is a state in which the environmental
information En is additionally registered.
[0026]

FIG. 3 is an explanatory diagram illustrating an example of
a matched number counting table according to the first
embodiment. The matched number counting table 300 counts a
12
matched number 301 of edge information between the existing
environmental information Ei and the environmental information
En of the new edge terminal Ti. When edge information items are
matched respectively between the existing environmental
information Ei and the environmental information En of the new
edge terminal Ti, a value of “1” is defined. When the
information items are not matched therebetween, a value of “0”
is defined. The total value in the row direction represents a
matched number hi with respect to the environmental information
Ei.
[0027]

FIG. 4A and FIG. 4B are explanatory diagrams illustrating
respectively a model blending condition and a learned model,
according to the first embodiment. FIG. 4A illustrates the model
blending condition Cn of the edge terminal Tn, while FIG. 4B
illustrates an example of an equation for the learned model LMn.
The model blending condition Cn is a combination of weights R1
to Rn-1. The denominator Σ(h) of the weight Ri is the total sum
of matched numbers h1 to hn-1. In the equation of 4B, the model
blending condition Cn is substituted in “Ri” of Equation 1 of
FIG. 1.
[0028]

FIGs. 5 are explanatory diagrams respectively illustrating
examples of a neural network process is present in one layer of
the CNN. FIG. 5A illustrates an input characteristic map 501,
FIG. 5B illustrates a weight filter 502, and FIG. 5C illustrates
an output characteristic map 503. The output characteristic map
503 is a multiplication result of the input characteristic map
13
501 and the weight filter 502, and is the input characteristic
map 501 of the next layer.
[0029]
The input characteristic map 501 is composed of a matrix of
an N-number of H*W rows and columns, where N is the number of
input channels. For example, when the input characteristic map
501 is composed of rows and columns derived from image data, the
number N of input channels is 3, that is, R (red), G (green),
and B (blue). “H” represents the height of the input
characteristic map 501 (the number of elements in column
direction), while “W” represents the width of the input
characteristic map 501 (the number of elements in row
direction).
[0030]
The weight filter 502 is a matrix of k*k. The weight
filter 502 is composed of a matrix of N-number which is the
number of input channels. The input characteristic map 501 is
raster scanned by the weight filter 502 in association with the
input channels, thereby obtaining the output characteristic map
503 in association with the input channels. “H’” represents the
height of the output characteristic map 503 (the number of
elements in the column direction), while “W’” represents the
width of the output characteristic map 503 (the number of
elements in the row direction).
[0031]
FIG. 6 is an explanatory diagram illustrating an example of
the weight filter 502 of the entire layers (the number of Layers
L). The weight filter 502 of the first layer is composed of a
matrix of k1*k1 corresponding to the number N1 of input channels.
“M1” represents the output channel number. The weight filter
14
502-2 of the second layer is composed of a matrix of k2*k2
corresponding to the number N2 of the input channels. “M2”
represents the output channel number. The weight filter 502-L of
the L-th layer is composed of a matrix of kL*kL corresponding to
the number NL of the input channels. “ML” represents the output
channel number.
[0032]
FIG. 7 is an explanatory diagram illustrating an example of
the learned model LMi with values of the weight filter 502 of
the entire layers (L-number of layers) of the CNN. A value 706
includes a one-dimensional vector of a layer number 701, an
input channel number 702, an output channel number 703, and a
weight (vertical) 705. The matrix (weight filter 502) having the
entire one-dimensional vectors arranged in the column direction
is the learned model LMi.
[0033]

Descriptions will now be made to an example of the hardware
configuration of the computer. The computer may be any of the
management device 101, the edge server 120, and the edge
terminal Ti, which are illustrated in FIG. 1.
[0034]
FIG. 8 is a block diagram illustrating an example of a
hardware configuration of the computer. The computer 800 has a
processor 801, a storage device 802, an input device 803, an
output device 804, and a communication interface (communication
IF) 805. The processor 801, the storage device 802, the input
device 803, the output device 804, and the communication IF 805
are connected with each other through a bus 806.
[0035]
15
The processor 801 controls the computer 800. The processor
801 includes a CPU (Central Processing Unit) and a GPU (Graphics
Processing Unit). The storage device 802 is a work area of the
processor 801. The storage device 802 is a non-temporary or
temporary storage medium which stores various programs or data.
The storage device 802 may be any of a ROM (Read Only Memory),
for example, a RAM (Random Access Memory), a HDD (Hard Disk
Drive), and a flash memory.
[0036]
The input device 803 inputs data. The input device 803 may
be any of a keyboard, for example, a mouse, a touch panel, a
ten-keyboard, and a scanner. The output device 804 outputs data.
For example, the output device 804 may be any of a display and a
printer. The communication IF 805 is connected to the network
103 to transmit and receive data.
[0037]
In the management device 101, the first communication unit
112, the calculation unit 113, and the blending unit 114 are
realized by controlling the processor 801 to execute the
programs stored, for example, in the storage device 802. The
model DB 110 and the environmental information DB 111 are
realized, for example, by the storage device 802 illustrated in
FIG. 8. In the edge server 120, the second communication unit
121 is realized by controlling the processor 801 to execute the
programs stored, for example, in the storage device 802. In the
edge terminal Tn, the inference unit 122 is realized by
controlling the processor 801 to execute the programs stored,
for example, in the storage device 802.
[0038]

16
FIG. 9 is a sequence diagram illustrating a sequence
example of the management system 100 according to the first
embodiment. The edge terminal Tn detects the environmental
information En from the sensor 123, such as a temperature
sensor, an environment sensor 123A, a camera 123B, or the like.
The terminal Tn transmits the environmental information En to
the first communication unit 112 of the management device 101,
through the second communication unit 121 of the edge server 120
(Step S901). Note that the user of the management device 101
inputs the environmental information En from a management
terminal 900 (Step S900), thereby transmitting it to the first
communication unit 112 of the management device 101 (Step S901).
The first communication unit 112 outputs the received
environmental information En to the calculation unit 113.
[0039]
The calculation unit 113 calculates the model blending
condition Cn, and outputs it to the blending unit 114 (Step
S902). The blending unit 114 mixes the model blending condition
Cn with the existing learned models LM1 to LMn-1, to generate
the learned model LMn of the edge terminal Tn (Step S903). The
blending unit 114 registers the generated learned model LMn in
the model DB 110 (Step S904). When a new learned model LMn+1 is
generated, the learned model LMn is handled as an existing
learned model LMi.
[0040]
The blending unit 114 outputs a deploy instruction
including the learned model LMn to the first communication unit
112 (Step S905). The first communication unit 112 transmits the
deploy instruction to the edge terminal Tn through the second
communication unit 121 of the edge server 120 (S906). Upon
17
reception of the deploy instruction, the edge terminal Tn
applies the learned model LMn to the CNN of the inference unit
122. When image data is input from the camera 123B to the CNN,
the inference unit 122 outputs the inference result (Step S907).
Then, the terminal Tn controls, as a terminal Ti, the behavior
of the edge terminal Ti itself or other connected devices using
the inference result.
[0041]
In this manner, the management device 101 generates a model
blending condition Cn as a blend recipe from each environmental
information of the edge terminals T1 to Tn-1, and blends the
generated model blending condition Cn with the learned model LM1
to LMn-1. By so doing, it is possible to generate the learned
model LMn of the newly-added edge terminal Tn, without executing
learning of datasets.
[0042]
The edge terminal Tn executes inference using the learned
model LMn. Hence, the edge terminal Tn does not need to upload
the dataset to the management device 101. Thus, it requires no
communication cost in transferring the dataset. Besides, there
is no need to upload the dataset to the management device 101,
thus preventing leakage of the dataset. The management device
101 does not need to implement the learning function using the
dataset, thus attaining a reduction in calculation load.
[0043]
In the first embodiment, the descriptions have been made to
the case where the edge terminal Tn is newly added to the edge
environment 102. The blending may be performed in the same
manner as that for newly adding the edge terminal Tn, even when
updating the environmental information Ei of the existing edge
18
terminal Ti.
[0044]
FIG. 10 is an explanatory diagram illustrating an example
of the environmental information DB 111 after updated. FIG. 10
illustrates a state in which the edge information e2, the 3M
information e3, and the time information ep of the environmental
information En have been updated.
[0045]
FIG. 11 is an explanatory diagram illustrating an example
of the matched number counting table 300 after updated. FIG. 11
illustrates a state in which the edge environment e2, the 3M
information e3, and the time information ep in relation with the
edge number (#1) have been updated to “1”, and in which the time
information ep in relation with the edge number (#n-1) has been
updated to “1”, based on the environmental information En after
updated as illustrated in FIG. 10.
[0046]
In this manner, for example, even when the environmental
information En of the edge terminal Tn has been updated, the
management device 101 generates a new learned model LMn in the
same manner as that for newly adding the environmental
information En, to update the model DB 110. The same applies to
a case where any other environmental information E1 to En-1 has
been updated.
[0047]
Assumed is a case where the edge terminal T3 is removed
from the edge environment 102. In this case, the management
device 101 deletes the learned model LM3 from the model DB 110.
The management device 101 may use the learned models LM1, LM2,
LM4 to LMn as they are. The management device 101 may update the
19
learned models LM4 to LMn whose edge number # (i) is greater
than 3, in order to eliminate the effect of the edge terminal
T3.
[0048]
For example, the management device 101 may control the
calculation unit 113 to calculate the matching degree of the
environmental information E4 with other environmental
information E1, E2, E5 to En, and may control the blending unit
114 to generate a new learned model LM4. The same applies to the
environmental information E5 to En.
[0049]
As a result, like the above-described new adding steps, the
management device 101 can update the learned model LMi without
executing learning of the datasets. The edge terminal Ti
executes inference using the learned model LMi after updated.
Hence, the edge terminal Ti does not need to upload the dataset
to the management device 101. Thus, it requires no communication
cost in transferring the dataset. Besides, there is no need to
upload the dataset to the management device 101, thus preventing
leakage of dataset. The management device 101 does not need to
implement the learning function using the dataset, thus
attaining a reduction in calculation load.
Second Embodiment
[0050]
Descriptions will now be made to a second embodiment. The
second embodiment introduces an example in which the matching
degree is weighted in accordance with the edge information.
Hereinafter, because the second embodiment will mainly be
described, those parts that are common to those of the first
embodiment are identified by the same reference numerals, and
20
thus will not be described over and over.
[0051]

FIG. 12 is an explanatory diagram illustrating an example of the
matched number counting table according to the second
embodiment. In a matched number counting table 1200 of the
second embodiment, what differs from the matched number counting
table 300 of the first embodiment is that an intrinsic weight
coefficient wj is given to each edge information ej (1 ≤ j ≤ p),
and the matched number 301 is changed to a weighted matched
number 1001. The weight coefficient wj has been set in advance.
The weight coefficient wj is an integer of, for example, 0 or
higher. When weight coefficients w1 to wp are all 1, the matched
number counting table 1200 is the same as the matched number
counting table 300 of the first embodiment.
[0052]
It is assumed that a value “xij” (1 or 0) represents
matching or non-matching of edge information ej in the edge
terminal Ti in relation with the edge number i. The weighted
matched number Si is expressed as Si=Σ(wj*xij). That is, only
when “xij” is 1, the weight coefficient wj is reflected on the
weighted matched number Si.
[0053]

FIGs. 13 are explanatory diagrams illustrating the model
blending condition and the learned model, according to the
second embodiment. FIG. 13A illustrates the model blending
condition Cn, while FIG. 13B illustrates an example of an
equation of the learned model LMn. The model blending condition
Cn is a combination of weights R1 to Rn-1. The denominator Σ(S)
21
of the weight Ri is the total sum of matched numbers S1 to Sn-1.
In the equation of 13B, the model blending condition Cn is
substituted in “Ri” of Equation 1 of FIG. 1.
[0054]
According to the second embodiment, the edge information ej
is weighted with a weight coefficient wj, thereby generating a
learned model LMn of the edge terminal Tn which has been
customized for various edge environments 102.
Third Embodiment
[0055]
Descriptions will now be made to a third embodiment. The
edge server 120 of the first and the second embodiments has been
assumed as a communication device which functions as a gateway
of the edge environment 102. However, in the third embodiment,
learning is executed using a dataset from the edge terminal Ti
to generate a learned model LMi as a learning result.
Hereinafter, because the third embodiment will mainly be
described, those parts that are common to those of the first and
the second embodiments are identified by the same reference
numerals, and thus will not be described over and over.
[0056]

FIG. 14 is an explanatory diagram illustrating a system
configuration example of the management system 100 according to
the third embodiment. The edge server 120 has a learning unit
1401. Specifically, the learning unit 1401 is realized by
controlling, for example, the processor 801 to execute the
programs stored in the storage device 802. The learning unit
1401 is composed of a CNN having the same configuration as that
of the inference unit 122.
22
[0057]

FIG. 15 is a sequence diagram illustrating a sequence
example of the management system 100 according to the third
embodiment. Upon reception of a learned model LMn from the
management device 101 (Step S906), one edge terminal (referred
to as an edge terminal Tn) applies it to the inference unit 122,
and inputs image data from the sensor 123, thereby outputting an
inference result (Step S907). When the inference result is
incorrect, the user of the edge terminal Tn allocates correct
data to the inference result to set datasets. The learning unit
1401 provides the CNN with the dataset which has been set to the
user of the edge terminal Tn, to generate a learned model LMn’
(Step S1508).
[0058]
The learning unit 1401 outputs the learned model LMn’ to
the inference unit 122, and also transmits it from the second
communication unit 121 to the first communication unit 112 of
the management device 101 (Step S1509). By outputting the
learned model to the inference unit 122, the edge terminal Tn
can execute inference without waiting for the learned model LMn’
from the management device 101.
[0059]
The management device 101 outputs the learned model LMn’ to
the blending unit 114. The blending unit 114 updates the learned
model LMn in the model DB 110 to the learned model LMn’ (Step
S1510). Then, the management device 101 can apply the learned
model LMn’ into the mixing by the blending unit 114.
[0060]
FIG. 16 is a graph illustrating the transition of
23
recognition accuracy of the CNN. As compared with U.S. Patent
Application Publication No. 2019-42955, a higher recognition
accuracy is attained in the CNN into which the new learned model
LMn has been applied, according to the first and the second
embodiments. Applying the third embodiment can attain a further
higher recognition accuracy in the CNN into which the new
learned model LMn’ has been applied in the third embodiment,
than those of the first and second embodiments, after passage of
a time for preparing a dataset.
[0061]
According to the third embodiment, it is possible to attain
a high accuracy of the learned model LMn’. By executing learning
with the edger server 120 instead of the management device 101,
it is possible to prevent uploading of the dataset to the
management device 101. This requires no communication cost in
transferring the dataset. Besides, there is no need to upload
the dataset to the management device 101, thus preventing
leakage of the dataset. The management device 101 does not need
to implement the learning function using the dataset, thus
attaining a reduction in calculation load.
Fourth Embodiment
[0062]
Descriptions will now be made to a fourth embodiment. In
the first to the third embodiments, the descriptions have been
made to the example in which the management device 101 generates
the learned model LMi without using a dataset, by
increasing/decreasing the number of edge terminals Ti. In this
fourth embodiment, contrarily, descriptions will now be made to
an example of generating a latest learned model LMi, when the
environmental information Ei of the edge terminal Ti is present
24
in the time direction. Hereinafter, because the fourth
embodiment will mainly be described, those parts that are common
to those of the first to the third embodiments are identified by
the same reference numerals, and thus will not be described over
and over.
[0063]

FIG. 17 is an explanatory diagram illustrating an example
of the learned models LMi stored in the model DB 110. The
learned models LMi include learned models LMi(t1), LMi(t2), …,
LMi(tj), …, and LMi(t(m-1)). Those symbols t1 to t(m-1)
represent timesteps indicating the time. The smaller the “j”,
the older the time. The learned model LMi(tm) is a learned model
which is newly added based on the environmental information Em
of the time tm.
[0064]

FIG. 18 is an explanatory diagram illustrating an example
of the environmental information DB 111 according to the fourth
embodiment. The environmental information DB 111 stores an
environmental information table 111 (Ti) in association with
edge terminals Ti. The environmental information table 111 (Ti)
stores the environmental information E1 (Ti) in association with
time 1800.
[0065]

FIG. 19 is an explanatory diagram illustrating an example
of a matched number counting table with regard to the edge
terminal Ti. FIG. 19 illustrates, as an example, a matched
number counting table 300 (T1) regarding the edge terminal T1
25
whose edge number (#) is 1 (i=1). In FIG. 3, values of the edge
information items e1 to ep are given in association with the
edge numbers 200. On the other hand, in FIG. 19, values of the
edge information items e1 to ep are given in association with
the time 1800. Like the second embodiment, the weight
coefficients w1 to wp may be applied thereinto.
[0066]
A value “1” represents that the edge information items are
matched between the existing environmental information E1(tj)
and the environmental information E1(tm) of the new edge
terminal T1, while a value “0” represents that they are not
matched. The total value in the row direction represents the
matched number gj with respect to the environmental information
E1(tj).
[0067]

FIGs. 20 are explanatory diagrams illustrating a model
blending condition and a learned model, according to the fourth
embodiment. FIG. 20A illustrates the model blending condition of
an edge terminal T1 whose edge number (#) is 1 (i=1), while FIG.
20B illustrates an example of an equation of the learned model
LM1. The model blending condition C1 is a combination of weights
R1 to Rn-1. The denominator Σ(g) of the weight Ri is the total
sum of matched numbers g1 to gm-1. In the equation of 20B, the
model blending condition C1 is substituted in “Ri” of Equation 1
of FIG. 1.
[0068]
In this manner, the management device 101 generates a model
blending condition Ci which is a blend recipe from the
environmental information Ei(T1) to Ei(t(m-1)) in the time
26
direction of the edge terminal Ti, and blends the generated
model blending condition Ci with the learned models LMi(T1) to
LMi(t(m-1)). By so doing, it generates the learned model LMi(tm)
of the edge terminal Ti into which environmental information
Ei(tm) is newly added, without executing learning of the
dataset.
[0069]
The edge terminal Ti executes inference using the learned
model LMi(tm). As a result, the edge terminal Ti does not need
to upload the dataset to the management device 101. Thus, it
requires not communication cost in transferring the dataset.
Because there is no need to upload the dataset to the management
device 101, it is possible to prevent leakage of the dataset.
Because the management device 101 does not need to implement the
learning function using the dataset, it is possible to attain a
reduction in calculation load.
[0070]
The present invention is not limited to the above-described
embodiments, but rather includes various modifications and
equivalent configurations within the meaning of the scope of the
attached claims. For example, the above-described embodiments
have been described in detail for easy understanding of the
present invention. It is not necessarily needed that the present
invention includes the above-described entire configurations.
The configuration of one embodiment may partially be replaced by
any of the rest of embodiments, or the configuration of one
embodiment may be added to any of the rest of embodiments. The
configuration of each of the embodiments may partially be added
to, deleted from, or replaced by other configurations.
[0071]
27
The above-described configurations, functions, processing
units, and processing means may partially or entirely be
realized with the hardware, by designing it using, for example,
an integrated circuit. Alternatively, they may be realized with
the software, controlling the processor to analyze and execute
the programs realizing the functions.
[0072]
Information of the programs, tables, and files for
realizing the functions may be stored in a storage device, such
as a memory, a hard disk, and an SSD (Solid State Drive), or may
be stored on a recording medium, such as an IC (Integrated
Circuit) card, an SD card, and a DVD (Digital Versatile Disc).
[0073]
Only those control lines and the information lines that are
considered necessary for the descriptions have been illustrated.
Thus, all control lines and information lines necessary for the
implementation are not illustrated. In fact, it can be assumed
that nearly all configurations are connected with each other.

We claim:
1. A management device accessible to a target object to
be managed, the device comprising:
a processor which executes a program;
a storage device which stores the program; and
a communication interface which can communicate with the
target object to be managed, wherein the processor executes
a reception process for receiving first environmental
information representing a first environment of the target
object to be managed,
a first generation process for generating relevant
information representing relevancy between the first
environmental information received by the reception process and
second environmental information representing a second
environment of the target object to be managed,
a second generation process for generating a first learned
model to be used for inference by the target object to be
managed in the first environment, based on the relevant
information generated by the first generating process and a
second learned model to be used for inference by the target
object to be managed in the second environment, and
a transmission process for transmitting the first learned
model generated by the second generation process to the target
object to be managed.
2. The management device according to claim 1, wherein:
in the reception process, the processor receives the first
environmental information representing the first environment of
a first data processing device as one target object to be
managed;
in the first generation process, the processor generates
29
the relevant information representing relevancy between the
first environmental information and the second environmental
information representing the second environment of the second
data processing device as one target object to be managed;
in the second generation process, the processor generates
the first learned model to be used for inference by the first
data processing device in the first environment, based on the
relevant information and the second learned model to be used for
inference by the second data processing device in the second
environment; and
in the transmission process, the processor transmits the
first learned model to the first data processing device.
3. The management device according to claim 2, wherein:
the first environmental information and the second
environmental information have values of a plurality of
environmental items each of which defines an environment common
to the first environmental information and the second
environmental information; and
in the first generation process, the processor generates
the relevant information based on a matched number of values of
same environmental items between the first environmental
information and the second environmental information.
4. The management device according to claim 3, wherein:
each of the plurality of environmental items has a level of
importance which is set in association therewith; and
in the first generation process, the processor generates
the relevant information, based on the matched number of values
of the same environmental items between the first environmental
information and the second environmental information and the
level of importance of the environmental items.
30
5. The management device according to claim 2, wherein
in the reception process, the processor receives the first
environmental information, when the first data processing device
is newly added as the target object to be managed.
6. The management device according to claim 2, wherein
in the reception process, when the first environmental
information has newly been added in the first data processing
device, the processor receives the newly added first
environmental information.
7. The management device according to claim 5, wherein
in the second generation process, the processor stores the
first environmental information as the second environmental
information.
8. The management device according to claim 1, wherein:
the communication interface is communicable with a
communication device which performs learning with a dataset
which is a combination of an inference result of the target
object to be managed and correct data corresponding to the
inference result; and
in the second generation process, the processor updates the
first learned model to a learned result from the communication
device.
9. The management device according to claim 1, wherein
in the first generation process, the processor generates
relevant information representing relevancy between the first
environmental information and the second environmental
information representing the second environment of the target
object to be managed at a time point earlier than a time point
of the first environmental information.
10. The management device according to claim 9, wherein
31
in the first generation process, the processor generates
relevant information representing relevancy between first
environmental information which has newly been received by the
reception process and the second environmental information.
11. The management device according to claim 10, wherein
in the second generation process, the processor stores the
first environmental information as the second environmental
information.
12. A management method to be executed by a management
device accessible to a target object to be managed, the
management device comprising a processor executing a program, a
storage device storing the program, and a communication
interface communicable with the target object to be managed, the
method comprising the following processes to be executed by the
processor, as follows:
a reception process for receiving first environmental
information representing a first environment of the target
object to be managed;
a first generation process for generating relevant
information representing relevancy between the first
environmental information received by the reception process and
second environmental information representing a second
environment of the target object to be managed;
a second generation process for generating a first learned
model to be used for inference by the target object to be
managed in the first environment, based on the relevant
information generated by the first generation process and a
second learned model to be used for inference by the target
object to be managed in the first environment; and
a transmission process for transmitting the first learned
32
model generated by the second generation process to the target
object to be managed.
13. A management program for controlling a processor to
execute processes for managing a target object to be managed,
the processor executing:
a reception process for receiving first environmental
information representing a first environment of the target
object to be managed;
a first generation process for generating relevant
information representing relevancy between the first
environmental information received by the reception process and
second environmental information representing a second
environment of the target object to be managed;
a second generation process for generating a first learned
model to be used for inference by the target object to be
managed in the first environment, based on the relevant
information generated by the first generation process and a
second learned model to be used for inference by the target
object to be managed in the second environment; and
a transmission process for transmitting the first learned
model generated by the second generation process to the target
object to the managed.