Comprehensive analysis method, comprehensive analysis device, and comprehensive analysis program

Abstract

The comprehensive analysis method according to one aspect of the present invention includes the following steps: each client device performs an operation for solving the correlation between each element of each local sample included in local learning data; the server device obtains the results of the operations based on each client device; the server device calculates a comprehensive result representing the correlation between each element of all local samples included in all local learning data by comprehensively analyzing the results of the operations obtained from each client device; the server device derives one or more principal components from the calculated comprehensive result by performing principal component analysis; and the server device outputs information related to the derived one or more principal components.

Description

Technical Field

[0001] The present invention relates to a comprehensive analysis method, a comprehensive analysis device and a comprehensive analysis program. Background Art

[0002] Principal component analysis can be used in various applications of analyzing data. For example, according to principal component analysis, the features of multidimensional data can be extracted and the amount of information of the data can be compressed. In addition, for example, by using the local space obtained by principal component analysis, it is possible to implement predetermined inferences such as class recognition (local space method) for object data. As an example of predetermined inference, a method for determining the quality of a product shown in an observation image by a local space method is proposed in non-patent document 1.

[0003] Prior art literature

[0004] Non-patent literature

[0005] Non-patent document 1: Kenta Toyota, Kazuhiro Horita, "Automatic determination of defective parts using local space method and robust statistics", SSII2016, IS3-22, June 10, 2016 Summary of the invention

[0006] Problems to be solved by the invention

[0007] The inventors of the present invention have discovered that the existing methods using principal component analysis have the following problems. That is, the learning data that are the subject of principal component analysis are collected one by one. In order to improve the quality of data analysis based on principal component analysis, it is expected that each user will collect enough learning data. However, collecting enough learning data one by one is costly and difficult. Therefore, if learning data are collected one by one, sample deviation is likely to occur, which may lead to poor quality of data analysis based on principal component analysis. For example, in the case of the above-mentioned data compression, the quality of the compression model obtained by principal component analysis is poor, which may lead to the deletion of originally useful information (for example, information that is useful for other users' tasks). In addition, for example, in the case of implementing the above-mentioned predetermined inference, information that is useful for the inference is not taken into account, which may lead to poor accuracy of the inference.

[0008] Here, in order to ensure sufficient learning data, it is considered to aggregate the learning data collected one by one in a system composed of one or more computers. However, aggregating the learning data collected one by one results in huge costs (such as communication costs) for data exchange. In addition, if principal component analysis is performed on a large amount of aggregated learning data, the calculation cost becomes large, which may lead to the occurrence of undesirable conditions such as insufficient memory used for calculation processing and failure of calculation processing to end within the scheduled time.

[0009] One aspect of the present invention has been completed in view of such actual circumstances, and an object thereof is to provide a technique for suppressing an increase in cost and improving the quality of data analysis based on principal component analysis.

[0010] Method for solving problems

[0011] In order to solve the above problems, the present invention adopts the following structure.

[0012] That is, the comprehensive analysis method according to one aspect of the present invention includes the following steps: a plurality of client devices respectively perform an operation for solving the correlation between respective elements of each local sample included in the local learning data on the local learning data; the server device acquires the result of the operation from each of the client devices; the server device calculates a comprehensive result representing the correlation between respective elements of all local samples included in all the local learning data by synthesizing the results of the operation acquired from each of the client devices; the server device derives one or more principal components from the calculated comprehensive result by performing principal component analysis; and the server device outputs information related to the derived one or more principal components.

[0013] In the comprehensive analysis method according to this structure, the results of operations related to the correlation between elements of local learning data, which are not the local learning data itself, are aggregated in the server device. Thereby, it is possible to reduce the cost of exchanging data between each client device and the server device. In addition, a part of a series of calculation processes for deriving one or more principal components from all local learning data is borne by each client device. Thereby, it is possible to reflect the locally learned data collected one by one in the principal component analysis, and it is possible to reduce the calculation cost related to the server device. Therefore, according to this structure, it is possible to suppress an increase in cost and improve the quality of data analysis based on principal component analysis.

[0014] It should be noted that the form of the operation result related to the correlation between elements of local learning data may not be particularly limited as long as it is not the local learning data itself and can derive the principal components of all local learning data. The operation result may be constituted by, for example, the autocorrelation matrix of local learning data. In addition, the form of the comprehensive result may not be particularly limited as long as it is a form obtained in the process of deriving the principal components of all local learning data. The comprehensive result may be constituted by, for example, a variance-covariance matrix, a correlation coefficient matrix, or the like.

[0015] The comprehensive analysis method involved in the above aspect may also be that the operation for solving the correlation consists of the following steps: obtaining the average value of each element of all local samples included in all the local learning data; normalizing (centralizing) each local sample by subtracting the obtained average value from the value of each element of each local sample included in the local learning data; and calculating the autocorrelation matrix of the local learning data based on the normalized local samples. It may also be that obtaining the result of the operation is constituted by obtaining the calculated autocorrelation matrix. It may also be that the integration of the results of the operation is constituted by summing the autocorrelation matrices obtained from the respective client devices. According to this structure, one or more principal components of all the local learning data can be appropriately derived.

[0016] The comprehensive analysis method involved in the above aspect may also be that it further includes: the comprehensive analysis method further includes the step of each client device accepting the designation of the importance of each local sample. It may also be that each local sample is weighted according to the designated importance. It may also be that the average value of each element of all the local samples is a weighted average value weighted according to the importance. It may also be that in the calculation step, the server device calculates the variance-covariance matrix of all the local learning data as the comprehensive result by dividing the sum of the autocorrelation matrices by the sum of the weights corresponding to the importance. According to this structure, by reflecting the importance of each local sample designated in each client device in the principal component analysis, an improvement in the quality of data analysis based on the principal component analysis can be achieved.

[0017] In the comprehensive analysis method involved in the above aspect, it may also be that the average value of each element of all the local samples is calculated by using secret calculation of the number of local samples and the average value of each element obtained from each client device. If the number of local samples and the average value of each element are made public, the confidentiality of the local learning data may be compromised. In this structure, by using secret calculation, the number of local samples and the average value of each element can be kept confidential, and the average value of each element of all the local samples can be obtained. Therefore, according to this structure, the confidentiality of the local learning data can be ensured during a series of calculation processes for deriving one or more principal components from all the local learning data.

[0018] In the comprehensive analysis method involved in the above aspect, it may also be that the integration of the results of the operation is performed by secret calculation. According to this structure, the confidentiality of the local learning data can be ensured during a series of calculation processes for deriving one or more principal components from all the local learning data.

[0019] The comprehensive analysis method involved in the above aspect may also be that the comprehensive analysis method further includes a step in which each of the client devices accepts the designation of two or more elements from among the multiple elements that make up each of the local samples. It may also be that, in the calculating step, the server device calculates the comprehensive result by synthesizing the calculation results obtained from each of the client devices for the designated two or more elements. It may also be that, in the deriving step, the server device derives one or more principal components from the calculated comprehensive result for the designated two or more elements by performing principal component analysis. According to this structure, principal component analysis for the elements designated in each client device can be implemented.

[0020] The comprehensive analysis method involved in the above aspect may also be that the comprehensive analysis method further includes a step in which the server device distributes each of the client devices to at least any one of a plurality of groups based on the degree of agreement of the designated two or more elements. It may also be that, in the calculating step, the server device calculates the comprehensive result by synthesizing the calculation results obtained from each of the client devices for the designated two or more elements within the same group. It may also be that, in the deriving step, the server device derives one or more principal components from the comprehensive result calculated for the designated two or more elements within the same group by performing principal component analysis. According to this structure, each client device can be grouped, and principal component analysis can be performed for each group based on the designated elements.

[0021] The comprehensive analysis method involved in the above aspect may also be that the comprehensive analysis method further includes a step in which the server device distributes each of the client devices to at least any one of a plurality of groups. It may also be that, in the calculating step, the server device calculates the comprehensive result by synthesizing the calculation results obtained from each of the client devices within the same group. It may also be that, in the deriving step, the server device derives one or more principal components from the comprehensive result calculated within the same group by performing principal component analysis. According to this structure, each client device can be grouped, and principal component analysis can be performed for each group.

[0022] In the comprehensive analysis method involved in the above aspect, it may also be that, in the distributing step, the server device sends a list indicating the plurality of groups to each of the client devices, selects one or more groups from among the plurality of groups shown in the list, and distributes each of the client devices to the selected one or more groups. According to this structure, each client device can be grouped by a simple method.

[0023] In the comprehensive analysis method related to the above-mentioned one aspect, it may also be that the server device acquires attribute data related to the local learning data from each of the client devices, clusters the attribute data acquired from each of the client devices, and based on the result of the clustering, assigns each of the client devices to at least any one of the multiple groups. According to this structure, each client device can be grouped according to the attributes of the local learning data.

[0024] In the comprehensive analysis method related to the above-mentioned one aspect, it may also be that outputting information related to the one or more principal components is constituted by the server device sending the information related to the derived one or more principal components to each of the client devices. According to this structure, in each client device, the result of the principal component analysis for all local learning data can be utilized.

[0025] In the comprehensive analysis method related to the above-mentioned one aspect, it may also be that the local learning data is constituted by image data reflecting a product or measurement data obtained by measuring the attributes of a product. According to this structure, regarding the data that can be used for product appearance inspection, an increase in cost can be suppressed, and an improvement in the quality of data analysis based on principal component analysis can be achieved.

[0026] In the comprehensive analysis method related to the above-mentioned one aspect, it may also be that the local learning data is constituted by sensing detection data obtained by a sensor that observes the state of an object person. According to this structure, regarding the data that can be used for estimating the state of an object person, an increase in cost can be suppressed, and an improvement in the quality of data analysis based on principal component analysis can be achieved.

[0027] In addition, as another aspect of the comprehensive analysis method related to the above-mentioned various forms, one aspect of the present invention may also be a computer system constituted by the above-mentioned client devices and server device. Or, one aspect of the present invention may also be any one of one or more devices that implement all or a part of the above-mentioned structures, an information processing method based on the device, a program, and a computer-readable storage medium such as a computer and other devices and machines that store such a program. Here, the computer-readable storage medium is a medium that stores information such as a program through electrical, magnetic, optical, mechanical, or chemical actions.

[0028] For example, the comprehensive analysis device according to one aspect of the present invention includes: an acquisition unit that acquires, from a plurality of client devices respectively, the results of operations that are operations for solving the correlations between respective elements of each local sample included in local learning data collected by the plurality of client devices respectively; a synthesis unit that calculates a synthesis result representing the correlations between respective elements of all local samples included in all local learning data by synthesizing the results of the operations acquired from the respective client devices; an analysis unit that derives one or more principal components from the calculated synthesis result by performing principal component analysis; and an output unit that outputs information related to the derived one or more principal components.

[0029] In addition, for example, the comprehensive analysis program according to one aspect of the present invention is a program for causing a computer to execute the following steps: acquiring, from a plurality of client devices respectively, the results of operations that are operations for solving the correlations between respective elements of each local sample included in local learning data collected by the plurality of client devices respectively; calculating a synthesis result representing the correlations between respective elements of all local samples included in all local learning data by synthesizing the results of the operations acquired from the respective client devices; deriving one or more principal components from the calculated synthesis result by performing principal component analysis; and outputting information related to the derived one or more principal components.

[0030] Effects of the Invention

[0031] According to the present invention, it is possible to suppress an increase in cost and achieve an improvement in the quality of data analysis based on principal component analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 An example of a situation where the present invention is applied is schematically illustrated.

[0033] Figure 2 An example of the hardware structure of the comprehensive analysis device according to the embodiment is schematically illustrated.

[0034] Figure 3 An example of the hardware structure of the client device according to the embodiment is schematically illustrated.

[0035] Figure 4 An example of the software structure of the comprehensive analysis device according to the embodiment is schematically illustrated.

[0036] Figure 5A An example of the software structure of the client device according to the embodiment is schematically illustrated.

[0037] Figure 5B Schematically illustrate an example of the software structure of the client device related to the embodiment.

[0038] Figure 5C Schematically illustrate an example of the software structure of the client device related to the embodiment.

[0039] Figure 6A Illustrate an example of the processing sequence related to the collection of local learning data based on the client device involved in the embodiment.

[0040] Figure 6B Illustrate an example of the processing sequence related to the calculation of the correlation between elements of the local learning data involved in the embodiment.

[0041] Figure 7 Illustrate an example of the processing sequence of the comprehensive analysis device involved in the embodiment.

[0042] Figure 8 Schematically illustrate an example of the packetization process involved in the embodiment.

[0043] Figure 9 Illustrate an example of the processing sequence related to the packetization of the client device based on the comprehensive analysis device involved in the embodiment.

[0044] Figure 10 Illustrate an example of the processing sequence related to the packetization of the client device based on the comprehensive analysis device involved in the embodiment.

[0045] Figure 11 Illustrate an example of the processing sequence related to the data compression of the client device involved in the embodiment.

[0046] Figure 12 Illustrate an example of the processing sequence related to the predetermined inference of the client device involved in the embodiment.

[0047] Figure 13 Schematically illustrate an example of other situations where the present invention is applied.

[0048] Figure 14 Schematically illustrate an example of other situations where the present invention is applied.

[0049] Figure 15 Schematically illustrate an example of other situations where the present invention is applied.

[0050] Figure 16 Schematically illustrate an example of other situations where the present invention is applied.

[0051] Figure 17An example schematically illustrating another case where the present invention is applied.

[0052] Figure 18 An example schematically illustrating the software structure of the client device involved in the modification example.

[0053] Figure 19 An example illustrating the processing sequence related to the collection of local learning data of the client device involved in the modification example.

[0054] Figure 20 An example schematically illustrating a screen for specifying the acceptance importance and the elements to be analyzed.

[0055] Figure 21 An example illustrating the processing sequence related to the grouping of client devices of the comprehensive analysis device involved in the modification example.

[0056] Figure 22 An example schematically illustrating the software structure of the client device involved in the modification example.

[0057] Figure 23 An example schematically illustrating a case where secret calculation is performed in the modification example.

[0058] Figure 24 An example schematically illustrating a case where secret calculation is performed in the modification example. Detailed implementation manners

[0059] Hereinafter, an embodiment related to one aspect of the present invention (hereinafter, also referred to as "this embodiment") will be described based on the drawings. However, the embodiment described below is merely an illustration of the present invention in all aspects. It goes without saying that various improvements and modifications can be made without departing from the scope of the present invention. In other words, when implementing the present invention, a specific structure corresponding to the embodiment can be appropriately adopted. It should be noted that natural language is used to describe the data that appears in this embodiment, but more specifically, computer-recognizable simulation languages, instructions, parameters, machine languages, etc. are used to specify.

[0060] §1 Application examples

[0061] Figure 1 An example schematically illustrating a case where the present invention is applied. As Figure 1 shown, the system 100 according to this embodiment includes a comprehensive analysis device 1 and a plurality of client devices 2.

[0062] Each client device 2 is a computer configured to collect local learning data 3. The type of the local learning data 3 is not particularly limited as long as it can be an object of principal component analysis, and can be appropriately selected according to the implementation. The local learning data 3 can be, for example, image data, sound data, numerical data, text data, or other measurement data obtained through various sensors. Hereinafter, the measurement data obtained through sensors is also referred to as "sensing detection data".

[0063] In the present embodiment, each client device 2 can use the sensor S to collect the local learning data 3. The sensor S can be, for example, an image sensor (camera), an infrared sensor, a sound sensor (microphone), an ultrasonic sensor, a light sensor, a pressure sensor, a barometric pressure sensor, a temperature sensor, etc. In addition, the sensor S can be, for example, an environmental sensor, a life sensor, a medical examination device, a vehicle-mounted sensor, a home security sensor, etc. The environmental sensor can be, for example, a barometer, a thermometer, a hygrometer, a sound pressure meter, a sound sensor, an ultraviolet sensor, an illuminometer, a rain gauge, a gas sensor, etc. The life sensor can be, for example, a sphygmomanometer, a pulse meter, a heart rate meter, an electrocardiograph, an electromyograph, a thermometer, a galvanic skin response meter, a microwave sensor, an electroencephalograph, a magnetoencephalograph, an activity meter, a blood glucose meter, an eye potential sensor, an eye movement detector, etc. The medical examination device can be, for example, a CT (Computed Tomography) device, an MRI (Magnetic Resonance Imaging) device, etc. The vehicle-mounted sensor can be, for example, an image sensor, a Lidar (light detection and ranging) sensor, a millimeter wave radar, an ultrasonic sensor, an acceleration sensor, etc. The home security sensor can be, for example, an image sensor, an infrared sensor, an activity (voice) sensor, a gas (CO2, etc.) sensor, a current sensor, a smart meter (a sensor for measuring the power consumption of home appliances, lighting, etc.).

[0064] The local learning data 3 is composed of a plurality of local samples. Each local sample includes a plurality of elements. Each element of the sample can be directly obtained from the data, such as each pixel of the image data, or can be obtained by performing a certain information process (i.e., indirectly from the data), such as the size of the object reflected in the image data.

[0065] Each client device 2 performs an operation for solving the correlation between elements of each local sample included in the local learning data 3 on the local learning data 3. As a result, each client device 2 generates a result 51 of the operation related to the correlation between elements of the local learning data 3. The form of the result 51 of the operation is not particularly limited as long as it is not the local learning data 3 itself and can be used for principal component analysis.

[0066] The comprehensive analysis device 1 is a computer configured to perform principal component analysis. The comprehensive analysis device 1 is an example of the "server device" of the present invention. The comprehensive analysis device 1 obtains the result 51 of the operation from each client device 2. The comprehensive analysis device 1 calculates a comprehensive result 40 representing the correlation between elements of all local samples included in all local learning data by synthesizing the results 51 of the operation obtained from each client device 2. All local learning data refers to all local learning data 3 obtained by each client device 2 and being the object of principal component analysis. All local samples refer to all local samples being the object of principal component analysis. The form of the comprehensive result 40 can be arbitrarily determined in a manner that can be used for principal component analysis.

[0067] The comprehensive analysis device 1 derives one or more principal components 41 from the calculated comprehensive result 40 by performing principal component analysis. The operation content of the principal component analysis, that is, the method for deriving the principal components 41, can be arbitrarily selected. As the method for deriving the principal components 41, known methods such as singular value decomposition, eigenvalue decomposition, and KL expansion can be adopted.

[0068] The comprehensive analysis device 1 outputs the derived one or more principal components 41. The output form and use of the principal components 41 can be arbitrarily selected respectively. In the present embodiment, the derived one or more principal components 41 can be provided to each client device 2. As a result, each client device 2 can utilize the one or more principal components 41 derived from all local learning data for various purposes such as data compression and predetermined inference, for example.

[0069] "Inference" can also be referred to as "estimation". Inference can be, for example, either deriving a discrete value (e.g., a class corresponding to a specific feature) by grouping (classification, recognition) or deriving a continuous value (e.g., the probability of occurrence of a specific feature) by recursion. Inference can also include any kind of determination such as detection and decision-making based on the result of the grouping or recursion. Additionally, inference can also include prediction.

[0070] Note that the name of the system 100 can be appropriately changed according to the information processing performed within the system 100, the utilization form of the main component 41, etc. For example, the system 100 can also be referred to as an analysis system, a compression system, an inference system, etc. In the computer within the system 100, when the derived main component 41 is utilized for inference, the name of the system 100 can be appropriately changed according to the content of the inference, such as a inspection system, a monitoring system, a diagnostic system, a detection system, a prediction system, etc. Similarly, each client device 2 can be referred to as a compression device, an inference device, etc. When the derived main component 41 is utilized for the inference of each client device 2, the name of each client device 2 can be appropriately changed according to the content of the inference, such as an inspection device, a monitoring device, a diagnostic device, a detection device, a prediction device, etc.

[0071] As described above, in the present embodiment, the results 51 of operations related to the correlations between elements of the local learning data 3 rather than the local learning data 3 itself are aggregated in the comprehensive analysis device 1. Thereby, the cost of exchanging data between each client device 2 and the comprehensive analysis device 1 can be reduced. In addition, each client device 2 can be made to bear a part of a series of computational processes for deriving one or more main components 41 from all the local learning data. Thereby, the local learning data 3 collected one by one by each client device 2 can be reflected in the principal component analysis, and the computational cost related to the comprehensive analysis device 1 can be reduced. Therefore, according to the present embodiment, an increase in cost can be suppressed, and an improvement in the quality of data analysis based on principal component analysis can be achieved.

[0072] Note that Figure 1 In the example of, there are three client devices 2a to 2c within the system 100. Hereinafter, for convenience of explanation, when distinguishing each client device, further reference numerals such as a, b, c are attached, and when not, these reference numerals are omitted as in "client device 2". Each of the client devices 2a to 2c collects local learning data 3a to 3c and generates results 51a to 51c of operations related to the correlations between elements of the local learning data 3a to 3c. The generated results 51a to 51c of the operations are aggregated in the comprehensive analysis device 1. The comprehensive analysis device 1 calculates a comprehensive result 40 based on the results 51a to 51c of the operations and derives one or more main components 41 from the calculated comprehensive result 40. Thereby, the comprehensive analysis device 1 can derive one or more main components 41 related to the local learning data 3a to 3c collected by the three client devices 2a to 2c. However, the number of client devices 2 is not limited to three and can be arbitrary.

[0073] In addition, Figure 1In the example, the comprehensive analysis device 1 and each client device 2 are interconnected via a network. The type of the network can be appropriately selected from, for example, the Internet, a wireless communication network, a mobile communication network, a telephone network, a private network, etc. However, the method of exchanging data between the comprehensive analysis device 1 and each client device 2 is not limited to such an example and can be appropriately selected according to the embodiment. For example, data can also be exchanged between the comprehensive analysis device 1 and each client device 2 by using a storage medium, an external storage device, etc.

[0074] In addition, Figure 1 In the example, the comprehensive analysis device 1 and the multiple client devices 2 are each constituted by one computer. However, the structure of the system 100 according to the present embodiment is not limited to such an example and can be appropriately determined according to the embodiment. For example, at least any one of the comprehensive analysis device 1 and the multiple client devices 2 can be constituted by multiple computers.

[0075] §2 Structural Example

[0076] [Hardware Structure]

[0077] <Comprehensive Analysis Device>

[0078] Figure 2 An example of the hardware structure of the comprehensive analysis device 1 according to the present embodiment is schematically illustrated. As Figure 2 shown, the comprehensive analysis device 1 according to the present embodiment is a computer in which a control unit 11, a storage unit 12, a communication interface 13, an input device 14, an output device 15, and a driver 16 are electrically connected. In Figure 2 , the interface is denoted as "I / F".

[0079] The control unit 11 includes a CPU (Central Processing Unit) as a hardware processor, a RAM (Random Access Memory), a ROM (Read Only Memory), etc., and is configured to perform information processing based on programs and various data. The storage unit 12 is an example of a memory and is constituted by, for example, a hard disk drive, a solid state drive, etc. In the present embodiment, the storage unit 12 stores various information such as a comprehensive analysis program 81, multiple pieces of operation result data 221, principal component information 121, a group list 123, allocation information 124, etc.

[0080] The comprehensive analysis program 81 is for causing the comprehensive analysis device 1 to execute the information processing related to the derivation of the principal component 41 described later ( Figure 7 , Figure 9 and Figure 10) program. The comprehensive analysis program 81 includes a series of commands for this information processing. The calculation result data 221 of each piece represents the result 51 of a calculation related to the correlation between elements of the local learning data 3 collected by each client device 2. The principal component information 121 includes information related to one or more derived principal components 41. The principal component information 121 is generated as a result of executing the comprehensive analysis program 81. The group list 123 represents a list of multiple groups that are candidates for allocating each client device 2. The allocation information 124 represents the correspondence between each client device 2 and each group.

[0081] The communication interface 13 is, for example, a wired LAN (Local Area Network) module, a wireless LAN module, etc., and is an interface for performing wired or wireless communication via a network. The comprehensive analysis device 1 can perform data communication via the network with other information processing devices by using the communication interface 13.

[0082] The input device 14 is, for example, a device for input such as a mouse and a keyboard. In addition, the output device 15 is, for example, a device for output such as a display and a speaker. An operator such as a user can operate the comprehensive analysis device 1 by using the input device 14 and the output device 15.

[0083] The drive 16 is, for example, a CD drive, a DVD drive, etc., and is a drive device for reading various information such as programs stored in the storage medium 91. The storage medium 91 is a medium that stores the programs and other information by electrical, magnetic, optical, mechanical, or chemical action in a manner that enables a computer and other devices and machines to read the stored programs and other information. At least any one of the above comprehensive analysis program 81, multiple pieces of calculation result data 221, group list 123, and allocation information 124 may be stored in the storage medium 91. The comprehensive analysis device 1 may also obtain at least any one of the above comprehensive analysis program 81, multiple pieces of calculation result data 221, group list 123, and allocation information 124 from the storage medium 91. It should be noted that in Figure 2 , as an example of the storage medium 91, disc-shaped storage media such as CDs and DVDs are exemplified. However, the type of the storage medium 91 may not be limited to the disc type and may also be in a form other than the disc type. As a storage medium other than the disc type, for example, a semiconductor memory such as a flash memory can be cited. The type of the drive 16 can be arbitrarily selected according to the type of the storage medium 91.

[0084] It should be noted that regarding the specific hardware structure of the comprehensive analysis device 1, the omission, replacement, and addition of components can be appropriately carried out according to the implementation manner. For example, the control unit 11 may also include multiple hardware processors. The hardware processor may be composed of a microprocessor, an FPGA (field-programmable gate array), a DSP (digital signal processor), etc. The storage unit 12 may also be composed of the RAM and ROM included in the control unit 11. At least any one of the communication interface 13, the input device 14, the output device 15, and the driver 16 may be omitted. The comprehensive analysis device 1 may also be composed of multiple computers. In this case, the hardware structures of each computer may be the same or different. In addition, the comprehensive analysis device 1 may be a general computer device such as a PC (Personal Computer) or a general server device in addition to the information processing device designed specifically for the provided services.

[0085] <Client device>

[0086] Figure 3 An example of the hardware structure of each client device 2 according to this embodiment is schematically illustrated. As Figure 3 shown, each client device 2 according to this embodiment is a computer in which a control unit 21, a storage unit 22, a communication interface 23, an input device 24, an output device 25, a driver 26, and an external interface 27 are electrically connected.

[0087] The control unit 21 to the driver 26 and the storage medium 92 of each client device 2 may be configured in the same manner as the control unit 11 to the driver 16 and the storage medium 91 of the above-mentioned comprehensive analysis device 1, respectively. The control unit 21 includes a CPU, a RAM, a ROM, etc. as hardware processors, and is configured to execute various information processes based on programs and data. The storage unit 22 is composed of, for example, a hard disk drive, a solid state drive, etc. The storage unit 22 stores various information such as a collection program 85, a compression program 86, an inference program 87, local learning data 3, operation result data 221, and principal component information 121.

[0088] The collection program 85 is a program for causing each client device 2 to execute the information process described later ( Figure 6A and Figure 6B ), and the information process described later involves the generation of the result 51 of operations related to the collection of local learning data 3 and the correlation. The local learning data 3 and the operation result data 221 are generated as a result of executing the collection program 85. The compression program 86 is a program for causing each client device 2 to execute the information process described later related to data compression using the derived one or more principal components 41 ( Figure 11) The program. The inference program 87 is a program for causing each client device 2 to execute information processing (described later) related to a predetermined inference using one or more extracted principal components 41. Figure 12 ) The program. The name of the inference program 87 can be appropriately changed according to the content of the inference, such as a check program, a monitoring program, a diagnostic program, a detection program, a prediction program, etc. Each of the programs 85 to 87 includes a series of commands for each information processing. At least any one of the collection program 85, the compression program 86, the inference program 87, and the principal component information 121 can be stored in the storage medium 92. In addition, each client device 2 can acquire at least any one of the collection program 85, the compression program 86, the inference program 87, and the principal component information 121 from the storage medium 92.

[0089] The external interface 27 is, for example, a USB (Universal Serial Bus) port, a dedicated port, etc., and is an interface for connecting to an external device. The type and number of the external interfaces 27 can be arbitrarily selected. Each client device 2 can be connected to the sensor S for obtaining a sample via at least one of the communication interface 23 and the external interface 27.

[0090] It should be noted that, regarding the specific hardware structure of each client device 2, components can be appropriately omitted, replaced, and added according to the embodiment. For example, the control unit 21 may also include multiple hardware processors. The hardware processor may be composed of a microprocessor, an FPGA, a DSP, etc. The storage unit 22 may be composed of the RAM and ROM included in the control unit 21. At least any one of the communication interface 23, the input device 24, the output device 25, the driver 26, and the external interface 27 can be omitted. Each client device 2 may be composed of multiple computers. In this case, the hardware structures of the respective computers may be the same or different. In addition, each client device 2 may be a general-purpose server device, a general-purpose PC, a PLC (programmable logic controller), a tablet terminal, etc., in addition to the information processing device designed specifically for the provided service.

[0091] [Software Structure]

[0092] <Comprehensive Analysis Device>

[0093] Figure 4 An example of the software structure of the comprehensive analysis device 1 according to the present embodiment is schematically illustrated.

[0094] The control unit 11 of the comprehensive analysis device 1 expands the comprehensive analysis program 81 stored in the storage unit 12 in the RAM. Moreover, the control unit 11 interprets and executes the commands included in the comprehensive analysis program 81 expanded in the RAM through the CPU to control each component. Thus, as Figure 4 shown, the comprehensive analysis device 1 according to the present embodiment operates as a computer, and the computer includes an acquisition unit 111, a synthesis unit 112, an analysis unit 113, an output unit 114, and a packetization unit 115 as software modules. That is, in the present embodiment, each software module of the comprehensive analysis device 1 is implemented by the control unit 11 (CPU).

[0095] The acquisition unit 111 acquires the operation result data 221, and the operation result data 221 represents the result 51 of an operation for solving the correlation between the elements of each local sample included in the local learning data 3 performed on the local learning data 3 collected by each client device 2. The synthesis unit 112 calculates a synthesis result 40 representing the correlation between the elements of all local samples included in all local learning data by synthesizing the operation results 51 obtained from each client device 2. The analysis unit 113 derives one or more principal components 41 from the calculated synthesis result 40 by performing principal component analysis. The output unit 114 outputs principal component information 121 related to the derived one or more principal components 41. The packetization unit 115 assigns each client device 2 to at least one of a plurality of groups.

[0096] <Client Device>

[0097] Figures 5A - 5C An example of the software structure of each client device 2 according to the present embodiment is schematically illustrated.

[0098] Similar to the above comprehensive analysis device 1, the control unit 21 of each client device 2 interprets and executes the commands included in the collection program 85 through the CPU. Thus, as Figure 5A shown, each client device 2 according to the present embodiment operates as a computer, and the computer includes a collection unit 201, an operation unit 202, and an output unit 203 as software modules. Similarly, the control unit 21 interprets and executes the commands included in the compression program 86 through the CPU. Thus, as Figure 5B shown, each client device 2 according to the present embodiment operates as a computer, and the computer includes an acquisition unit 211, a compression unit 212, and an output unit 213 as software modules. The control unit 21 interprets and executes the commands included in the inference program 87 through the CPU. Thus, as Figure 5CAs shown, each client device 2 according to this embodiment operates as a computer, and this computer includes an acquisition unit 215, an inference unit 216, and an output unit 217 as software modules. That is, in this embodiment, similar to the above-described comprehensive analysis device 1, each software module for each information process of each client device 2 is implemented by a control unit 21 (CPU).

[0099] As Figure 5A shown, a collection unit 201 collects local learning data 3 composed of a plurality of local samples 30. An arithmetic unit 202 performs an operation for solving the correlation between the respective elements of each local sample 30 included in the local learning data 3 on the local learning data 3. An output unit 203 outputs arithmetic result data 221 representing the result 51 of the operation generated by the arithmetic unit 202.

[0100] As Figure 5B shown, an acquisition unit 211 acquires object data 223 (sample) that is an object to be reduced in information amount (i.e., compressed). A compression unit 212 compresses the object data 223 by using the derived principal component 41 with reference to the principal component information 121. Thereby, the compression unit 212 generates compressed data 224. An output unit 213 outputs the generated compressed data 224.

[0101] As Figure 5C shown, an acquisition unit 215 acquires object data 226 (sample) that is an object to be inferred. An inference unit 216 performs a predetermined inference on the object data 226 by using the derived principal component 41 with reference to the principal component information 121. The inference method can be arbitrarily selected. In this embodiment, as an example of inference, the inference unit 216 performs class recognition of the features included in the object data 226 based on the comparison between the object data 226 and a data group 227. Specifically, the data group 227 is composed of a plurality of samples 228. At least any one of the plurality of samples 228 can use the local samples 30 that constitute the local learning data 3. Each sample 228 includes features corresponding to the respective categories. Each sample 228 is projected onto a local space by using the principal component 41, and thus is converted into a feature amount 2281. Thereby, within the local space, the range belonging to the class of the object can be determined. The inference unit 216 projects the object data 226 onto the local space by using the principal component 41, and thus converts the object data 226 into a feature amount 2261. The inference unit 216 compares the obtained feature amount 2261 with each feature amount 2281 within the local space. The inference unit 216 determines whether the features included in the object data 226 belong to the class of the object based on the result of this comparison. An output unit 217 outputs information related to the result of the inference.

[0102] <Others>

[0103] Regarding each software module of the comprehensive analysis device 1 and each client device 2, it will be described in detail by the operation examples described later. It should be noted that in this embodiment, an example in which each software module of the comprehensive analysis device 1 and each client device 2 is implemented by a general-purpose CPU will be described. However, a part or all of the above software modules may also be implemented by one or more dedicated processors. In addition, regarding the software structure of each of the comprehensive analysis device 1 and each client device 2, the omission, replacement, and addition of software modules may be appropriately performed according to the embodiment.

[0104] §3 Operation Example

[0105] (1) Data Collection

[0106] Figure 6A It is a flowchart showing an example of the processing sequence related to the collection of the local learning data 3 based on each client device 2 according to this embodiment. It should be noted that the processing sequence described below is only an example, and each step can be changed as much as possible. In addition, for the processing sequence described below, the omission, replacement, and addition of steps can be appropriately performed according to the embodiment.

[0107] (Step S101)

[0108] In step S101, the control unit 21 operates as the collection unit 201 and collects the local learning data 3.

[0109] The local learning data 3 consists of a plurality of local samples 30. Each local sample 30 can be appropriately obtained. For example, in the real space or the imaginary space, data is generated under various conditions. The generated data can be obtained as each local sample 30. In this embodiment, it may also be that by observing an object under various conditions using the sensor S, sensing detection data is generated. The object to be observed can be appropriately selected according to the use purpose of the local sample 30. The generated sensing detection data can be obtained as each local sample 30.

[0110] Each local sample 30 can be automatically generated by the operation of a computer or manually generated by at least partially including the operation of an operator. In addition, the generation of each local sample 30 can be performed by each client device 2 or by other computers other than each client device 2. When each local sample 30 is generated by each client device 2, the control unit 21 can obtain each local sample 30 by automatically or manually performing the above generation process using the operation of the operator via the input device 24. On the other hand, when each local sample 30 is generated by other computers, the control unit 21 can obtain the local sample 30 generated by other computers via a network, a storage medium 92, etc., for example. It is also possible that a part of the local samples 30 are generated by each client device 2 and the other local samples 30 are generated by one or more other computers.

[0111] The number of local samples 30 can be arbitrarily selected. If the local learning data 3 is collected, the control unit 21 advances the process to the next step S102.

[0112] (Step S102)

[0113] In step S102, the control unit 21 acts as the arithmetic unit 202 and performs an operation for solving the correlation between the elements of each local sample 30 included in the local learning data 3 on the local learning data 3. As a result, the control unit 21 generates the result 51 of the operation related to the correlation between the elements of the local learning data 3. The calculation content can be appropriately determined according to the form of the result 51.

[0114] Figure 6B It is a flowchart showing the processing order of the sub-process related to the calculation of the correlation between the elements of the local learning data 3 according to the present embodiment. The processing of step S102 according to the present embodiment includes the following processing of step S1021 to step S1023. However, the processing order described below is only an example, and each process can be changed as much as possible. In addition, for the processing order described below, steps can be appropriately omitted, replaced, and added according to the embodiment.

[0115] (Step S1021)

[0116] In step S1021, the control unit 21 obtains the average value of each element of all local samples.

[0117] The calculation method of the average value of each element of all local samples can be appropriately determined. As an example, the local learning data 3 of each client device 2 can be expressed by the following Equation 1. The local samples 30 included in the local learning data 3 can be expressed by the following Equation 2. The average value of each element of all local samples can be calculated based on the number of local samples 30 included in each local learning data 3 and the average value of each element.

[0118] [Equation 1]

[0119]

[0120] [Equation 2]

[0121]

[0122] X (P) represents the local learning data 3 collected by the P-th client device 2. N (P) represents the number of local samples 30. X n (P) represents the n-th local sample 30. d represents the number of elements (dimensions). x n#i (P) represents the i-th element of the n-th local sample 30. The control unit 21 calculates the average value of each element of the local samples 30 included in the local learning data 3 by performing the operation of the following Equation 3.

[0123] [Equation 3]

[0124]

[0125] U (P) represents the average value of each element of the local samples 30 included in the local learning data 3 collected by the P-th client device 2. u i (P) represents the average value of the i-th element. Hereinafter, the average value of each element of the local sample 30 is also referred to as "average of local samples".

[0126] Each client device 2 notifies the average value and the number of its own local samples to other client devices 2. The notification method can be arbitrarily selected. For example, the control unit 21 can use the communication interface 23 to notify the average value and the number of its own local samples to other client devices 2 via the network. Moreover, the control unit 21 performs the operation of the following Equation 4 using the average value and the number of its own local samples and the average value and the number of local samples obtained from other client devices 2.

[0127] [Equation 4]

[0128]

[0129] U represents the average value of each element of all local samples. u i i represents the average value of the i-th element of all local samples. Thus, the control unit 21 of each client device 2 can obtain the average value U of each element of all local samples. If the average value U is obtained, the control unit 21 advances the process to the next step S1022.

[0130] It should be noted that the operation process of the average value U is not limited to such an example. As an example, the operation of Equation 4 can be executed by other computers. The other computer can be the comprehensive analysis device 1. In this case, the control unit 21 of each client device 2 notifies the other computer of the average and quantity of its own local samples. The other computer calculates the average value U of each element of all local samples by performing the operation of the above Equation 4 using the average and quantity of the local samples obtained from each client device 2. The other computer notifies the calculated average value U to each client device 2. Thus, the control unit 21 of each client device 2 can obtain the average value U of each element of all local samples.

[0131] (Step S1022)

[0132] In step S1022, the control unit 21 subtracts the obtained average value from the value of each element of each local sample 30 included in the local learning data 3 as shown in Equation 5 below. Thus, the control unit 21 normalizes (centralizes) each local sample 30.

[0133] [Equation 5]

[0134]

[0135] X C (P) C (P) represents the local learning data 3 collected and normalized by the P-th client device 2. If each local sample 30 is normalized, the control unit 21 advances the process to the next step S1023.

[0136] (Step S1023)

[0137] In step S1023, the control unit 21 performs the operation of Equation 6 below. Thus, the control unit 21 calculates the autocorrelation matrix of the local learning data 3 based on the normalized local samples 30.

[0138] [Equation 6]

[0139]

[0140] Q (P)It represents the autocorrelation matrix calculated in the P-th client device 2. Thus, the control unit 21 can obtain the autocorrelation matrix as the result 51 of the operation related to the correlation between the elements of the local learning data 3. If the autocorrelation matrix is calculated, the control unit 21 ends the sub-process related to the calculation of the correlation involved in the present embodiment and advances the process to the next step S103.

[0141] (Step S103)

[0142] In step S103, the control unit 21 acts as the output unit 203 and outputs the operation result data 221 representing the generated operation result 51.

[0143] The output form can be appropriately determined according to the embodiment. For example, it can be that the control unit 21 outputs the operation result data 221 to the output device 25 as the process of step S103. Additionally, for example, it can be that the control unit 21 stores the operation result data 221 in a predetermined storage area as the process of step S103. The predetermined storage area can be, for example, the RAM in the control unit 21, the storage unit 22, an external storage device, a storage medium, or a combination thereof. The storage medium can be, for example, a CD, a DVD, etc., and the control unit 21 can store the operation result data 221 in the storage medium via the drive 26. The external storage device can be, for example, a data server such as a NAS (Network Attached Storage). In this case, the control unit 21 can store the operation result data 221 in the data server via the network using the communication interface 23. Additionally, the external storage device can be, for example, an external storage device connected to each client device 2 via the external interface 27.

[0144] Thus, if the output of the operation result data 221 is completed, the control unit 21 ends the series of processes related to the collection of the local learning data 3.

[0145] (2) Principal Component Analysis

[0146] Figure 7 It is a flowchart showing an example of the processing sequence of the comprehensive analysis device 1 according to the present embodiment. The following-described processing sequence is an example of the comprehensive analysis method. The comprehensive analysis method can include the above-described processing sequence of data collection. However, the following-described processing sequence is merely an example, and each step can be changed as much as possible. Additionally, for the following-described processing sequence, steps can be appropriately omitted, replaced, and added according to the embodiment.

[0147] (Step S201)

[0148] In step S201, the control unit 11 operates as an acquisition unit 111 and acquires operation result data 221 representing the operation result 51 of each client device 2.

[0149] The operation result data 221 of each client device 2 can be provided to the comprehensive analysis device 1 at any time point. For example, the client device 2 can transmit the operation result data 221 to the comprehensive analysis device 1 as the processing of the above step S103 or separately from the processing of step S103. The control unit 11 can acquire the operation result data 221 of each client device 2 by receiving this transmission. Additionally, for example, the control unit 11 can access each client device 2 or the data server via the network using the communication interface 13 to acquire the operation result data 221 of each item. Additionally, for example, the control unit 11 can acquire the operation result data 221 of each item via the storage medium 91 or an external storage device. Additionally, for example, it can also be that the operator inputs the operation result 51 output to the output device 25 of each client device 2 via the input device 14, whereby the control unit 11 acquires the operation result data 221 of each item. In the present embodiment, the control unit 11 acquires the operation result data 221 representing the autocorrelation matrix of the local learning data 3 as the operation result 51. If the operation result data 221 of each item is acquired, the control unit 11 advances the process to the next step S202.

[0150] (Step S202)

[0151] In step S202, the control unit 11 operates as a synthesizing unit 112 and calculates a comprehensive result 40 representing the correlation between elements of all local samples included in all local learning data by synthesizing the operation results 51 represented by the operation result data 221 acquired from each client device 2.

[0152] The content of the synthesizing operation can be appropriately determined according to the form of the operation result 51. In the present embodiment, the operation result 51 is represented by the above autocorrelation matrix. Here, the control unit 11 can sum the autocorrelation matrices obtained from each client device 2. Additionally, the control unit 11 can appropriately acquire information representing the number of local samples 30 of each client device 2. Moreover, the control unit 11 can divide the sum of the autocorrelation matrices by the sum of the numbers of local samples 30. Thereby, the control unit 11 can calculate the variance - covariance matrix C as shown in the following Equation 7. The control unit 11 can acquire the calculated variance - covariance matrix C as the comprehensive result 40. If the comprehensive result 40 is calculated, the control unit 11 advances the process to the next step S203.

[0153] [Equation 7]

[0154]

[0155] It should be noted that the form of the comprehensive result 40 is not limited to such an example. As an example of other forms, the comprehensive result 40 can be represented by a correlation coefficient matrix. In this case, the control unit 11 can calculate the correlation coefficient matrix of all local samples based on the autocorrelation matrix of the local learning data 3 of each client device 2. The control unit 11 can obtain the calculated correlation coefficient matrix as the comprehensive result 40.

[0156] (Step S203)

[0157] In step S203, the control unit 11 acts as the analysis unit 113, and derives one or more principal components 41 from the calculated comprehensive result 40 by performing principal component analysis.

[0158] The operation content of the principal component analysis can be appropriately determined according to the embodiment. In the present embodiment, the control unit 11 can obtain the variance-covariance matrix C of all local samples as the comprehensive result 40. Here, the control unit 11 can derive one or more principal components 41 by performing KL expansion on the variance-covariance matrix C. Alternatively, the control unit 11 performs eigenvalue decomposition of the variance-covariance matrix as shown in Equation 8 below. The control unit 11 can obtain the eigenvalues λ obtained by eigenvalue decomposition as the principal components 41.

[0159] [Equation 8]

[0160] C = VΛV T …(Equation 8)

[0161] [Equation 9]

[0162] Λ = diag(λ1,..., λ r )…(Equation 9)

[0163] [Equation 10]

[0164] V = (v1,..., v r )…(Equation 10)

[0165] V represents the eigenvector (hereinafter, also referred to as "principal component vector"). Λ represents the eigenvalue matrix. The eigenvector V can be represented by the above Equation 9. In addition, the eigenvalue matrix Λ can be represented by the above Equation 10. diag in Equation 9 indicates a diagonal matrix. r represents the number of eigenvalues (principal components 41). Each component v of the eigenvector V can be calculated during the eigenvalue decomposition process.

[0166] The number of the derived principal components 41 can be arbitrarily selected. For example, it can also be that the control unit 11 calculates the cumulative contribution rate and derives the principal components 41 until the calculated cumulative contribution rate exceeds a threshold value. The threshold value can be appropriately determined. Thus, the control unit 11 can derive one or more principal components 41. If one or more principal components 41 are derived, the control unit 11 advances the process to the next step S204.

[0167] It should be noted that the method for deriving the principal components 41 is not limited to the above method. As an example of another method, when the correlation coefficient matrix of all local samples is obtained as the comprehensive result 40, the control unit 11 performs eigenvalue decomposition of the correlation coefficient matrix. The control unit 11 can obtain the eigenvalues obtained by eigenvalue decomposition as the principal components 41. Alternatively, the control unit 11 calculates the variance-covariance matrix or the deviation matrix of the correlation coefficient matrix, and performs singular value decomposition of the calculated deviation matrix. The control unit 11 can calculate the principal components 41 based on the singular values obtained by singular value decomposition. The derivation of the principal components 41 can appropriately adopt a known method.

[0168] (Step S204)

[0169] In step S204, the control unit 11 acts as the output unit 114 and outputs principal component information 121 related to the one or more derived principal components 41.

[0170] As long as the computer can utilize the derived principal components 41 or principal component vectors by referring to the principal component information 121, the form of the principal component information 121 is not particularly limited and can be appropriately determined according to the embodiment. For example, the principal component information 121 can be constituted by at least any one of the derived principal components 41 themselves and the above principal component vectors.

[0171] In addition, the output form of the principal component information 121 can also be appropriately determined according to the embodiment. For example, the control unit 11 can output the principal component information 121 to the output device 15 as the process of step S204. Additionally, for example, the control unit 11 can save the principal component information 121 in a predetermined storage area as the process of step S204. The predetermined storage area can be, for example, the RAM in the control unit 11, the storage unit 12, an external storage device, a storage medium, or a combination thereof.

[0172] In addition, for example, the control unit 11 may send (transmit) the principal component information 121 to each client device 2 as the process of step S204. The sending method can be arbitrarily selected. As an example, the control unit 11 may directly send the principal component information 121 to each client device 2 via a network. Alternatively, the control unit 11 may indirectly send the principal component information 121 to each client device 2 via another computer such as a data server. Each client device 2 can obtain the principal component information 121 by receiving this transmission.

[0173] Among them, the providing method and the time point of providing are not limited to such examples. As other examples, the principal component information 121 may be provided to each client device 2 via a storage medium 92 or an external storage device. Alternatively, it may be that the principal component information 121 is input and output to the output device 15 of the comprehensive analysis device 1 by an operator via the input device 24, thereby providing the principal component information 121 to each client device 2. The control unit 11 may provide the principal component information 121 to each client device 2 separately from the process of step S204. In addition, the control unit 11 may provide the principal component information 121 to other computers that utilize the derived principal component 41 in addition to each client device 2.

[0174] Thereby, when the output of the principal component information 121 ends, the control unit 11 ends a series of processes related to principal component analysis.

[0175] (3) Grouping

[0176] Figure 8 An example is schematically illustrated of the situation of grouping each client device 2 according to the present embodiment. For example, when the types of data are completely different between one local learning data 3 and other local learning data 3, it is difficult to comprehensively analyze the operation results 51 obtained from them respectively in the comprehensive analysis device 1. Here, the control unit 11 may act as a grouping unit 116 and allocate each client device 2 to at least one of a plurality of groups.

[0177] Each group can be appropriately set according to the type of local learning data 3 (local sample 30), the purpose of utilization, etc. Figure 8 In the example, each client device 2 is allocated to either one of two groups, the first group and the second group. However, the number of groups is not limited to two and can be arbitrarily determined. The control unit 11 stores the allocation result of the group for each client device 2 as allocation information 124. The allocation information 124 can be saved in a predetermined storage area, for example. The predetermined storage area can be a RAM within the control unit 11, the storage unit 12, an external storage device, a storage medium, or a combination thereof.

[0178] Accordingly, the control unit 11 performs the processes of step S201 to step S204 for each group. In step S202 above, the control unit 11 synthesizes the calculation results 51 obtained from each client device 2 within the same group, thereby calculating the synthesis result 40. In step S203 above, the control unit 11 can derive one or more principal components 41 from the synthesis result 40 calculated within the same group by performing principal component analysis. Thus, principal component analysis can be performed for each group.

[0179] The method of grouping can be not particularly limited and can be appropriately determined according to the embodiment. In the present embodiment, the control unit 11 can allocate each client device 2 to at least any one of a plurality of groups by any one of the following two methods. It should be noted that grouping each client device 2 can be regarded as synonymous with grouping the users of the client device 2, grouping the local learning data 3, and the like.

[0180] (3-1) First grouping method

[0181] Figure 9 FIG. is a flowchart showing an example of the processing order of the first grouping method. In the first grouping method, the control unit 11 allocates each client device 2 to at least any one of a plurality of groups by selecting a desired group from the list of groups. It should be noted that when the group allocation method adopts the first grouping method, allocating each client device 2 to at least any one of a plurality of groups is constituted by the following steps S211 to step S213. However, the processing order described below is merely an example, and each process can be changed as much as possible. In addition, for the processing order described below, steps can be appropriately omitted, replaced, and added according to the embodiment.

[0182] (Step S211)

[0183] In step S211, the control unit 11 sends the group list 123 showing an overview of multiple groups to each client device 2. The sending method can be arbitrarily selected. As an example, the control unit 11 can directly send the group list 123 to each client device 2 via the network. Alternatively, the control unit 11 can indirectly send the group list 123 to each client device 2 via another computer such as a data server. Thus, the control unit 11 selects one or more groups from the multiple groups shown in the group list 123 with respect to each client device 2. Each group can be appropriately set according to the local learning data 3, the client device 2, the user of the client device 2, etc. For example, when the local sample 30 can be used for appearance inspection, each group can be set according to attributes such as the production line number, the factory name, and the company name. In addition, new groups can be set in the group list 123 according to the requirements from each client device 2. The operator of each client device 2 can refer to the group list 123 output to the output device 25, operate the input device 24, and select one or more desired groups from the group list 123. Each client device 2 can select two or more groups. When the selection is completed, the control unit 21 of each client device 2 returns an answer of the group selection to the comprehensive analysis device 1.

[0184] (Steps S212 and S213)

[0185] In step S212, the control unit 11 obtains the answer of the group selection from each client device 2. Moreover, in step S213, the control unit 11 assigns each client device 2 to the one or more selected groups based on the obtained answer. When the assignment of the one or more groups is completed, the control unit 11 ends a series of processes related to the assignment of the groups based on the first grouping method. According to this first grouping method, each client device 2 can be grouped by a simple method.

[0186] (3-2) Second grouping method

[0187] Figure 10 is a flowchart showing an example of the processing sequence of the second grouping method. In the second grouping method, the control unit 11 assigns each client device 2 to an appropriate group according to the attributes of the local learning data 3. It should be noted that when the group assignment method adopts the second grouping method, the assignment of each client device 2 to at least any one of the multiple groups is constituted by the following steps S221 to S223. However, the processing sequence described below is only an example, and each process can be changed as much as possible. In addition, for the processing sequence described below, steps can be appropriately omitted, replaced, and added according to the embodiment.

[0188] (Step S221)

[0189] In step S221, the control unit 11 acquires attribute data related to the local learning data 3 from each client device 2. The method for acquiring the attribute data can be appropriately determined according to the embodiment.

[0190] The attribute data may include all information related to the local learning data 3. For example, the attribute data may include information indicating the data type of the local sample 30, information indicating the features represented by the local sample 30, information indicating the purpose of use of the local sample 30, and the like. The attribute data can be generated when the local learning data 3 is collected through the above step S101. If the attribute data is acquired, the control unit 11 advances the process to the next step S222.

[0191] (Steps S222 and S223)

[0192] In step S222, the control unit 11 performs clustering on the attribute data acquired from each client device 2. The method of clustering is not particularly limited and can be appropriately selected according to the embodiment. Clustering can adopt a well-known method such as the k-means method (k-means clustering).

[0193] In step S223, based on the result of the clustering, the control unit 11 assigns each client device 2 to at least one of a plurality of groups. As an example, the control unit 11 assigns the client devices 2 whose acquired attribute data is assigned to the same class to the same group. In this case, each group can be set according to the class of the attribute data. In addition, the control unit 11 can assign the client device 2 to two or more groups based on the result of the clustering.

[0194] If the assignment of the groups based on the result of the clustering is completed, the control unit 11 ends a series of processes related to the assignment of the groups based on the second grouping method. According to this second grouping method, the control unit 11 can assign each client device 2 to an appropriate group according to the attributes of the local learning data 3.

[0195] In the present embodiment, the control unit 11 can assign each client device 2 to at least one of a plurality of groups by adopting at least one of the above two methods. However, the method of grouping is not limited to these examples and can be appropriately determined according to the embodiment.

[0196] (4) Utilization of Principal Components: Data Compression

[0197] Figure 11It is a flowchart showing an example of the processing sequence related to data compression of each client device 2 according to the present embodiment. Data compression is an example of the use of the derived principal component 41. It should be noted that the processing sequence described below is merely an example, and each step can be changed as much as possible. In addition, for the processing sequence described below, steps can be appropriately omitted, replaced, and added according to the embodiment.

[0198] (Step S301)

[0199] In step S301, the control unit 21 acts as the acquisition unit 211 and acquires the object data 223 (sample) to be compressed. The object data 223 is data of the same type as the local sample 30. The object data 223 can be acquired by any method. In the present embodiment, the control unit 21 can generate sensed detection data by observing the object using the sensor S. The object to be observed can be appropriately selected according to the embodiment. The control unit 21 can acquire the generated sensed detection data as the object data 223. If the object data 223 is acquired, the control unit 21 advances the process to the next step S302.

[0200] (Step S302)

[0201] In step S302, the control unit 21 acts as the compression unit 212, and by referring to the principal component information 121, acquires the principal component vector (eigenvector V) obtained from one or more principal components 41 derived by the above-mentioned comprehensive analysis device 1. The control unit 21 projects the object data 223 onto the local space by the acquired principal component vector. That is, the control unit 21 calculates the product of the principal component vector and the object data 223. The control unit 21 can generate the compressed data 224 by compressing the object data 223 using this calculation. The compressed data 224 corresponds to the object data 223 converted in a way that reduces the amount of information. If the compression of the object data 223 is completed, the control unit 21 advances the process to the next step S303.

[0202] (Step S303)

[0203] In step S303, the control unit 21 acts as the output unit 213 and outputs the generated compressed data 224.

[0204] The output form of the compressed data 224 can be appropriately determined according to the embodiment. For example, the control unit 21 can output the compressed data 224 to the output device 25. In addition, for example, the control unit 21 can save the compressed data 224 in a predetermined storage area. The predetermined storage area can be, for example, the RAM in the control unit 21, the storage unit 22, an external storage device, a storage medium, or a combination thereof. The generated compressed data 224 can be provided to other computers.

[0205] Accordingly, if the output of the compressed data 224 ends, the control unit 21 ends a series of processes related to data compression.

[0206] (5) Utilization of principal components: Inference

[0207] Figure 12 FIG. is a flowchart showing an example of the processing sequence related to the predetermined inference based on each client device 2 according to the present embodiment. Inference is an example of the use of the derived principal components 41. It should be noted that the processing sequence described below is merely an example, and each step can be changed as much as possible. In addition, for the processing sequence described below, steps can be appropriately omitted, replaced, and added according to the embodiment.

[0208] (Step S311)

[0209] In step S311, the control unit 21 operates as an acquisition unit 215 and acquires object data 226 (sample) to be the object of inference. The object data 226 is data of the same type as the local sample 30. Step S311 can be the same as the above step S301. In the present embodiment, the control unit 21 can acquire the object data 226 through the sensor S. If the object data 226 is acquired, the control unit 21 advances the process to the next step S312.

[0210] (Steps S312 to S314)

[0211] In steps S312 to S314, the control unit 21 operates as an inference unit 216 and uses one or more derived principal components 41 to infer the features included in the object data 226. In the present embodiment, as an example of inference, based on the comparison between the object data 226 in the local space and each sample 228 included in the data group 227, the class of the features included in the object data 226 is identified.

[0212] Specifically, in step S312, the control unit 21 acquires a principal component vector (feature vector V) obtained from one or more principal components 41 derived by the above comprehensive analysis device 1 by referring to the principal component information 121. The control unit 21 projects the object data 226 onto the local space by the acquired principal component vector. Accordingly, the control unit 21 acquires the feature quantity 2261.

[0213] Similarly, each sample 228 is also converted into a feature quantity 2281 by the principal component vector. The conversion of each sample 228 can be performed in advance. Each sample 228 contains features corresponding to the respective categories. For example, in the case of appearance inspection by class recognition, each sample 228 can use image data of a product showing a defect of a type corresponding to the category of the object. Alternatively, each sample 228 can use image data of a product showing no defect according to the category of "qualified". The number of set categories can also be arbitrary.

[0214] In step S313, the control unit 21 compares the feature quantity 2261 obtained in the local space with each feature quantity 2281. In step S314, the control unit 21 identifies whether the feature included in the object data 226 belongs to the category of the object based on the result of the comparison. The comparison method can be appropriately determined. In the present embodiment, based on each feature quantity 2281 obtained from each sample 228, it is possible to determine a range belonging to the category of the object in the local space. Here, it is also possible to set a boundary for defining a range belonging to the category of the object based on each feature quantity 2281 obtained from each sample 228. The boundary can be appropriately expressed by a function or the like. The control unit 21 can determine whether the obtained feature quantity 2261 is included in the range of the category of the object based on the set boundary. The control unit 21 can also identify whether the feature included in the object data 226 belongs to the category of the object based on the result of this determination. Alternatively, the control unit 21 calculates the distance between the obtained feature quantity 2261 and each feature quantity 2281 obtained from each sample 228 belonging to the category of the object. The control unit 21 can identify whether the feature included in the object data 226 belongs to the category of the object based on the calculated distance. When the inference for the object data 226 ends, the control unit 21 advances the process to the next step S315.

[0215] (Step S315)

[0216] In step S315, the control unit 21 operates as the output unit 217 and outputs information related to the inference result.

[0217] The output destination and the content of the output information can be appropriately determined according to the embodiment. For example, the control unit 21 may directly output the recognition result of step S314 to the output device 25. Additionally, for example, the control unit 21 may perform a certain information process based on the recognition result of step S314. Moreover, the control unit 21 may output the result of having performed this information process as information related to the inference result. The output of the result of having performed this information process may include outputting a specific message according to the inference result, controlling the operation of the controlled device according to the inference result, and the like. The output destination may be, for example, the output device 25, the output device of another computer, the controlled device, and the like.

[0218] When the output of the information related to the inference result ends, the control unit 21 ends a series of processes related to the predetermined inference. It should be noted that it may also be the case that, within a predetermined period, the control unit 21 continuously and repeatedly executes a series of information processes of steps S311 to S315. The repetition timing can be arbitrary. Thereby, each client device 2 can continuously perform the predetermined inference.

[0219] [Features]

[0220] As described above, in the present embodiment, the results 51 of operations related to the correlations between elements of the local learning data 3 rather than the local learning data 3 itself are aggregated in the comprehensive analysis device 1. Thereby, in the above step S201, it is possible to reduce the cost of exchanging data between each client device 2 and the comprehensive analysis device 1. Additionally, it is possible to make each client device 2 bear a part of a series of computational processes for deriving one or more principal components 41 from all the local learning data. In the present embodiment, through the above step S102, it is possible to make each client device 2 bear the process of calculating the autocorrelation matrix of the local learning data 3. Thereby, it is possible to reflect the local learning data 3 collected one by one by each client device 2 in the principal component analysis, and it is possible to reduce the computational cost related to the comprehensive analysis device 1 by an amount corresponding to the amount of calculation until the autocorrelation matrix is calculated. Therefore, according to the present embodiment, it is possible to suppress an increase in cost and to improve the quality of data analysis based on principal component analysis. By using the derived one or more principal components 41, in the case of the above data compression, in step S302, it is difficult to delete useful information. Additionally, in the case of the above inference, it is possible to improve the accuracy of the inference based on steps S312 to S314.

[0221] §4 Variation

[0222] As described above, the embodiments of the present invention have been described in detail, but the description so far is merely illustrative of the present invention in all aspects. It goes without saying that various improvements or modifications can be made without departing from the scope of the present invention. For example, the following changes can be made. Note that hereinafter, for the same structural elements as those in the above embodiments, the same reference numerals are used, and for the same points as those in the above embodiments, the description is appropriately omitted. The following modification examples can be combined as appropriate.

[0223] <4.1>

[0224] The system 100 according to the above embodiment can be applied to all cases where one or more principal components are derived from local learning data collected for various purposes. The purpose of collecting the local learning data 3 can be, for example, tasks such as appearance inspection, monitoring of cultivation status, monitoring of the state of an object person, and monitoring of the state of a machine. Hereinafter, modification examples that limit the application cases are illustrated.

[0225] (A) Case of appearance inspection

[0226] Figure 13 An example of the application case of the inspection system 100A according to the first modification example is schematically illustrated. This modification example is an example in which the above embodiment is applied when performing principal component analysis on data obtained by observing the state of the product RA. The inspection system 100A according to this modification example includes a comprehensive analysis device 1 and a plurality of inspection devices 2A. Similar to the above embodiment, the comprehensive analysis device 1 and each inspection device 2A can be connected via a network.

[0227] In this modification example, the local learning data 3A is composed of image data showing the product RA or measurement data obtained by measuring the attributes of the product RA. The image data can be obtained by photographing the product RA using the camera SA. In this case, each pixel of the image data corresponds to each element of the local sample. In addition, the measurement data can be composed of, for example, measurement values of each attribute calculated by observing each attribute of the product RA using a sensor such as the camera SA and based on the obtained sensing detection data. The attributes of the product RA to be measured can be appropriately selected. The attributes of the product RA can be, for example, width, thickness, shape, color, inclination, unevenness, texture, etc. The texture of the product RA can be defined by touch (e.g., rough / smooth, degree of surface roughness), material (e.g., metal / plastic), etc. In this case, the measurement value of each attribute corresponds to each element of the local sample. It can also be that, except for these limitations, the inspection system 100A according to this modification example is configured in the same manner as the system 100 according to the above embodiment.

[0228] It should be noted that the product RA can be, for example, a product transported by a production line such as an electronic device, an electronic component, an automotive component, a drug, or food. The electronic component can be, for example, a base, a surface-mounted capacitor, a liquid crystal, a winding of a relay, etc. The automotive component is, for example, a connecting rod, a shaft body, an engine block, an electric window switch, a control panel, etc. The drug can be, for example, a packaged tablet, an unpackaged tablet, etc. The product RA can be a final product generated after the end of the manufacturing process, an intermediate product generated in the middle of the manufacturing process, or an initial product prepared before the manufacturing process. In addition, defects that are the object of appearance inspection can be, for example, scratches, dirt, cracks, impact marks, burrs, uneven color, foreign matter inclusion, etc. The inference related to the defect can be constituted by, for example, determining whether the product RA contains a defect, determining the probability that the product RA contains a defect, identifying the type of defect contained in the product RA, determining the range of the defect contained in the product RA, or a combination thereof.

[0229] (Inspection device)

[0230] Each inspection device 2A according to this modification example corresponds to each client device 2 according to the above-described embodiment. The hardware structure and software structure of each inspection device 2A according to this modification example can be the same as those of each client device 2 according to the above-described embodiment. Thus, the information processing of each inspection device 2A can be executed in the same order as that of each of the above-described client devices 2.

[0231] In the above step S101, each inspection device 2A collects local learning data 3A. The local learning data 3A is composed of a plurality of local samples of image data reflecting the product RA or measurement data obtained by measuring the attributes of the product RA. The acquisition of each local sample can use the camera SA. In the above step S102, each inspection device 2A calculates the result 51A of an operation related to the correlation between elements of the local learning data 3A. Each inspection device 2A can calculate the autocorrelation matrix of the local learning data 3A as the result 51A of the operation by executing the processing of the above step S1021 to step S1023. In step S103, each inspection device 2A outputs the calculated result 51A of the operation.

[0232] (Comprehensive analysis device)

[0233] In this modified example, the comprehensive analysis device 1 derives one or more principal components 41A by comprehensively analyzing the image data showing the product RA or the measurement data obtained by measuring the attributes of the product RA. Specifically, in the above step S201, the comprehensive analysis device 1 obtains the result 51A of the operation related to the correlation from each inspection device 2A. In the above step S202, the comprehensive analysis device 1 calculates the comprehensive result 40A representing the correlation between the elements of all local samples included in all local learning data by synthesizing the operation results 51A obtained from each inspection device 2A. Through the above operation process, the comprehensive analysis device 1 can obtain the variance-covariance matrix of all local learning data as the comprehensive result 40A. In the above step S203, the comprehensive analysis device 1 derives one or more principal components 41A from the calculated comprehensive result 40A by performing principal component analysis. In step S204, the comprehensive analysis device 1 outputs information related to the one or more derived principal components 41A.

[0234] (Utilization of Principal Components)

[0235] The one or more derived principal components 41A can be utilized in any application. In addition, the information related to the one or more derived principal components 41A can be provided to each inspection device 2A at any time point. Each inspection device 2A can utilize the one or more derived principal components 41A to compress the object data through the processing of the above steps S301 to S303. In addition, each inspection device 2A can utilize the one or more calculated principal components 41A to identify the state of the product RA of the object data through the processing of the above steps S311 to S315.

[0236] When the data of the normal product RA without defects is utilized for each sample constituting the data group, in step S314, when it is identified that the object data belongs to the object category, each inspection device 2A can determine that the product RA of the object data has no defects (i.e., the product RA is normal). On the other hand, when it is identified that the object data does not belong to the object category, each inspection device 2A can determine that the product RA of the object data has defects.

[0237] In addition, when the data of the product RA containing specific defects is utilized for each sample constituting the data group, in step S314, when it is identified that the object data belongs to the object category, each inspection device 2A can determine that the product RA of the object data has defects of the type corresponding to the object category. On the other hand, when it is identified that the object data does not belong to the object category, each inspection device 2A can determine that the product RA of the object data has no defects.

[0238] In the above step S315, each inspection device 2A outputs information related to the result inferred for the defect of the product RA. For example, each inspection device 2A can directly output the result inferred for the defect of the product RA to the output device. Additionally, for example, when it is determined that the product RA has a defect, each inspection device 2A can output a warning for notifying this situation to the output device. Additionally, for example, when each inspection device 2A is connected to the conveyor device that transports the product RA, each inspection device 2A controls the conveyor device based on the result inferred for the defect, so as to transport the defective products and non-defective products on different production lines.

[0239] (Feature)

[0240] According to this modified example, with respect to the data that can be used for appearance inspection, an increase in cost can be suppressed, and an improvement in the quality of data analysis based on principal component analysis can be achieved. Thus, in the case of the above data compression, by using one or more derived principal components 41A, it is difficult to delete information beneficial for appearance inspection. Additionally, in the case of the above inference, an improvement in the accuracy of appearance inspection based on steps S312 to S314 can be achieved.

[0241] (B) Case of monitoring the cultivation status

[0242] Figure 14 An example of the application scenario of the monitoring system 100B according to the second modified example is schematically illustrated. This modified example is an example in which the above-described embodiment is applied when performing principal component analysis on the observation data related to the plant RB. The monitoring system 100B according to this modified example includes a comprehensive analysis device 1 and a plurality of monitoring devices 2B. Similar to the above-described embodiment, the comprehensive analysis device 1 and each monitoring device 2B can be connected via a network.

[0243] In this modified example, the local learning data 3B is observation data related to the plant RB. The observation data can be composed of, for example, sensing detection data obtained by observing the state of the plant RB using the environmental sensor SB, observation data of the plant RB obtained by the input of an operator, or a combination thereof. The type of the environmental sensor SB is not particularly limited as long as it can observe the cultivation status of the plant RB, and can be appropriately selected according to the implementation manner. The environmental sensor SB can be, for example, a barometer, a thermometer, a hygrometer, a sound pressure meter, a sound sensor, an ultraviolet sensor, an illuminometer, a rain gauge, a gas sensor, etc. The type of the plant RB can be arbitrarily selected. The monitored cultivation status can be related to any element of cultivating the plant RB. The cultivation status can be determined, for example, by the growth environment and growth state up to the cultivation time. It can also be that the growth environment is related to the condition for growing the plant RB, and is specified, for example, by the time of irradiating light on the plant RB, the temperature around the plant RB, the amount of water given to the plant RB, etc. The growth state can be specified, for example, by the growth degree of the plant RB. The observation data can be composed of, for example, operation record data, environmental record data, or a combination thereof. The operation record data can be composed of information indicating the presence or absence, execution time, amount, etc. of operations such as flower picking, leaf picking, and bud picking. In addition, the environmental record data can be composed of information indicating the result of the operator observing the environment (such as climate, temperature, humidity, etc.) around the plant RB. It can also be that, except for these limitations, the monitoring system 100B according to this modified example is configured in the same manner as the system 100 according to the above-described embodiment.

[0244] (Monitoring device)

[0245] Each monitoring device 2B according to this modified example corresponds to each client device 2 according to the above-described embodiment. The hardware structure and software structure of each monitoring device 2B according to this modified example can be the same as those of each client device 2 according to the above-described embodiment. Thus, the information processing of each monitoring device 2B can be executed in the same order as that of each of the above-described client devices 2.

[0246] In the above step S101, each monitoring device 2B collects the local learning data 3B. The local learning data 3B is composed of a plurality of local samples of observation data related to the plant RB. Each local sample can be obtained by at least one of the environmental sensor SB and the input of the operator. In the above step S102, each monitoring device 2B calculates the result 51B of an operation related to the correlation between elements of the local learning data 3B. Each monitoring device 2B can calculate the autocorrelation matrix of the local learning data 3B as the result 51B of the operation by executing the processing of the above step S1021 to step S1023. In step S103, each monitoring device 2B outputs the calculated result 51B of the operation.

[0247] (Comprehensive analysis device)

[0248] In this variant, the comprehensive analysis device 1 outputs one or more principal components 41B with respect to the observation data related to the plant RB. Specifically, in the above step S201, the comprehensive analysis device 1 obtains the result 51B of the operation related to the correlation from each monitoring device 2B. In the above step S202, the comprehensive analysis device 1 calculates the comprehensive result 40B representing the correlation between the elements of all local samples contained in all local learning data by integrating the result 51B of the operation obtained from each monitoring device 2B. Through the above calculation processing, the comprehensive analysis device 1 is able to obtain the variance-covariance matrix of all local learning data as the comprehensive result 40B. In the above step S203, the comprehensive analysis device 1 derives one or more principal components 41B from the calculated comprehensive result 40B by performing principal component analysis. In step S204, the comprehensive analysis device 1 outputs information related to the derived one or more principal components 41B.

[0249] (Use of main component)

[0250] The derived one or more principal components 41B can be used for any purpose. In addition, information related to the derived one or more principal components 41B can be provided to each monitoring device 2B at any time point. Through the processing of the above-mentioned steps S301 to S303, each monitoring device 2B can use the derived one or more principal components 41B to compress the object data. In addition, through the processing of the above-mentioned steps S311 to S315, each monitoring device 2B can use the calculated one or more principal components 41B to infer the cultivation conditions of the plant RB of the object data. It should be noted that the inferred cultivation conditions may include at least one of the growth environment and the operation content in which the harvest amount is estimated to be the maximum, the inferred optimal operation content in the current growth environment observed, etc.

[0251] In the above step S315, each monitoring device 2B outputs information related to the result inferred for the cultivation status of the plant RB. For example, each monitoring device 2B can directly output the result inferred for the cultivation status of the plant RB to the output device. In this case, each monitoring device 2B can urge the operator to improve the cultivation status of the plant RB by outputting the inferred result such as at least one of the growth environment and the operation content that maximizes the yield, and the best operation content in the currently observed growth environment to the output device. Additionally, for example, each monitoring device 2B can be connected to the cultivation device CB. The cultivation device CB is configured to control the growth environment of the plant RB. In this case, each monitoring device 2B can determine the control command to be given to the cultivation device CB based on the result of inferring the cultivation status. The correspondence relationship between the cultivation status and the control command can be given by reference information in the form of a table or the like. This reference information can be stored in a RAM, ROM, storage unit, storage medium, external storage device, etc., and each monitoring device 2B can determine the control command based on the inferred cultivation status by referring to this reference information. Moreover, it can also be that each monitoring device 2B controls the operation of the cultivation device CB by giving the determined control command to the cultivation device CB. Additionally, for example, it can also be that each monitoring device 2B outputs the information indicating the determined control command to the output device and urges the manager of the plant RB to control the operation of the cultivation device CB.

[0252] It should be noted that the type of the cultivation device CB is not particularly limited as long as it can control the growth environment of the plant RB, and it can be appropriately selected according to the embodiment. The cultivation device CB can be, for example, a curtain device, a lighting device, an air conditioning device, a sprinkler device, etc. The curtain device is configured to open and close the curtain installed on the window of the building. The lighting device is, for example, LED (light emitting diode) lighting, fluorescent lamp, etc. The air conditioning device is, for example, an air conditioner, etc. The sprinkler device is, for example, a sprinkler, etc. The curtain device and the lighting device are used to control the time of light irradiation on the plant RB. The air conditioning device is used to control the temperature around the plant RB. The sprinkler device is used to control the amount of water given to the plant RB.

[0253] (Feature)

[0254] According to this modification example, regarding the observation data that can be used for monitoring the cultivation status of the plant RB, it is possible to suppress the increase in cost and improve the quality of data analysis based on principal component analysis. Thus, in the case of the above data compression, by using the one or more derived principal components 41B, it is difficult to delete the information beneficial to monitoring the cultivation status of the plant RB. Additionally, in the case of the above inference, it is possible to improve the accuracy of inferring the cultivation status of the plant RB in steps S312 to S314.

[0255] (C) Case of diagnosing health status

[0256] Figure 15 An example of the application case of the diagnostic system 100C according to the third modification example is schematically illustrated. This modification example is an example in which the above-described embodiment is applied in the case of performing principal component analysis on the sensing detection data acquired by the sensor SC for observing the state of the subject RC. The diagnostic system 100C according to this modification example includes a comprehensive analysis device 1 and a plurality of diagnostic devices 2C. Similar to the above-described embodiment, the comprehensive analysis device 1 and each diagnostic device 2C may also be connected via a network.

[0257] In this modification example, the local learning data 3C is the sensing detection data obtained by the sensor SC. The type of the sensor SC is not particularly limited as long as it can observe the state of the subject RC, and can be appropriately selected according to the embodiment. The sensor SC may be, for example, a vital sensor, a medical examination device, etc. The vital sensor may be, for example, a sphygmomanometer, a pulse meter, a heart rate meter, an electrocardiograph, an electromyograph, a thermometer, a galvanic skin response meter, a microwave sensor, an electroencephalograph, a magnetoencephalograph, an activity meter, a blood glucose meter, an eye potential sensor, an eye movement meter, etc. The medical examination device may be, for example, a CT device, an MRI device, etc. Inferring the health status may be constituted by, for example, determining whether one is healthy, determining whether there are signs of illness, identifying the type of health status, determining the probability of suffering from a target disease, or a combination thereof. Alternatively, except for these limitations, the diagnostic system 100C according to this modification example is configured in the same manner as the system 100 according to the above-described embodiment.

[0258] (Diagnostic device)

[0259] Each diagnostic device 2C according to this modification example corresponds to each client device 2 according to the above-described embodiment. The hardware structure and software structure of each diagnostic device 2C according to this modification example may be the same as those of each client device 2 according to the above-described embodiment. Thus, the information processing of each diagnostic device 2C can be executed in the same order as that of each of the above-described client devices 2.

[0260] In the above step S101, each diagnostic device 2C collects the local learning data 3C. The local learning data 3C is constituted by a plurality of local samples of the sensing detection data obtained by the sensor SC. In the above step S102, each diagnostic device 2C calculates the result 51C of an operation related to the correlation between elements of the local learning data 3C. Each diagnostic device 2C can calculate the autocorrelation matrix of the local learning data 3C as the result 51C of the operation by executing the processing of the above step S1021 to step S1023. In step S103, each diagnostic device 2C outputs the calculated result 51C of the operation.

[0261] (Comprehensive analysis device)

[0262] In this modification example, the comprehensive analysis device 1 derives one or more principal components 41C with respect to the sensing detection data obtained by the sensor SC. Specifically, in the above step S201, the comprehensive analysis device 1 acquires the result 51C of the operation related to the correlation from each diagnostic device 2C. In the above step S202, the comprehensive analysis device 1 calculates the comprehensive result 40C representing the correlation between the elements of all local samples included in all local learning data by synthesizing the operation results 51C acquired from each diagnostic device 2C. Through the above operation processing, the comprehensive analysis device 1 can acquire the variance-covariance matrix of all local learning data as the comprehensive result 40C. In the above step S203, the comprehensive analysis device 1 derives one or more principal components 41C from the calculated comprehensive result 40C by performing principal component analysis. In step S204, the comprehensive analysis device 1 outputs information related to the one or more derived principal components 41C.

[0263] (Utilization of principal components)

[0264] The one or more derived principal components 41C can be utilized in any application. In addition, the information related to the one or more derived principal components 41C can be provided to each diagnostic device 2C at any time point. Each diagnostic device 2C can utilize the one or more derived principal components 41C to compress the target data through the processing of the above steps S301 to S303. In addition, each diagnostic device 2C can utilize the one or more calculated principal components 41C to identify the health status of the target person RC of the target data through the processing of the above steps S311 to S315.

[0265] In the above step S315, each diagnostic device 2C outputs information related to the result of inferring the health status of the target person RC. For example, each diagnostic device 2C can directly output the result of inferring the health status of the target person RC to the output device. In addition, for example, it can also be that, in the case where the inferred health status of the target person RC indicates a sign of a predetermined disease, each diagnostic device 2C outputs a message urging a hospital examination to the output device. In addition, for example, each diagnostic device 2C can send the result of inferring the health status of the target person RC to the terminal of the registered hospital. It should be noted that the information of the terminal to be the destination can be stored in a predetermined storage area such as a RAM, a ROM, a storage unit, a storage medium, and an external storage device.

[0266] (Features)

[0267] According to this modification example, for the sensing detection data that can be used for monitoring the health status of the subject RC, it is possible to suppress the increase in cost and achieve an improvement in the quality of data analysis based on principal component analysis. Thus, in the case of the above data compression, by using one or more derived principal components 41C, it is difficult to delete information beneficial to monitoring the health status of the subject RC. In addition, in the case of the above inference, it is possible to improve the accuracy of inferring the health status of the subject RC in steps S312 to S314. It should be noted that the situation of diagnosing the health status of the subject RC involved in this modification example is an example of the situation of inferring the status of the subject. However, the situation of inferring the status of the subject may not be limited to the situation of diagnosing the health status. In addition, between the situation of obtaining local learning data and the situation of using principal components, the subjects may not be the same.

[0268] (D) Situation of monitoring the driver's status

[0269] Figure 16 An example of the application situation of the monitoring system 100D according to the fourth modification example is schematically illustrated. This modification example is a case where the above-described embodiment is applied in the case of performing principal component analysis on the sensing detection data obtained by the sensor SD for observing the status of the driver RD. The situation of monitoring the status of the driver RD involved in this modification example is another example of the situation of inferring the status of the above-mentioned subject. The monitoring system 100D involved in this modification example includes a comprehensive analysis device 1 and a plurality of monitoring devices 2D. Similar to the above-described embodiment, the comprehensive analysis device 1 and each monitoring device 2D may be connected via a network.

[0270] In this modification example, the local learning data 3D is the sensing detection data obtained by the sensor SD. The type of the sensor SD is not particularly limited as long as it can observe the status of the driver RD, and it can be appropriately selected according to the embodiment. The sensor SD can be, for example, a camera, an infrared sensor, a microphone, a life sensor, etc. The status of the driver RD can include, for example, posture, behavior, drowsiness level, fatigue level, relaxation level, etc. The drowsiness level represents the degree of drowsiness of the driver RD. The fatigue level represents the degree of fatigue of the driver RD. The relaxation level represents the degree of relaxation with respect to the driving of the driver RD. It can also be that, except for these limitations, the monitoring system 100D involved in this modification example is configured in the same manner as the system 100 involved in the above-described embodiment.

[0271] (Monitoring device)

[0272] Each monitoring device 2D according to this modification example corresponds to each client device 2 according to the above-described embodiment. The hardware structure and software structure of each monitoring device 2D according to this modification example may be the same as those of each client device 2 according to the above-described embodiment. Thus, the information processing of each monitoring device 2D can be executed in the same order as that of each of the above-described client devices 2.

[0273] In the above step S101, each monitoring device 2D collects local learning data 3D. The local learning data 3D is constituted by a plurality of local samples of the sensing detection data obtained by the sensor SD. In the above step S102, each monitoring device 2D calculates the result 51D of an operation related to the correlation between elements of the local learning data 3D. Each monitoring device 2D can calculate the autocorrelation matrix of the local learning data 3D as the result 51D of the operation by executing the processing of the above step S1021 to step S1023. In step S103, each monitoring device 2D outputs the calculated result 51D of the operation.

[0274] (Comprehensive analysis device)

[0275] In this modification example, the comprehensive analysis device 1 derives one or more principal components 41D with respect to the sensing detection data obtained by the sensor SD. Specifically, in the above step S201, the comprehensive analysis device 1 acquires the result 51D of an operation related to the correlation from each monitoring device 2D. In the above step S202, the comprehensive analysis device 1 calculates the comprehensive result 40D representing the correlation between elements of all local samples included in all local learning data by synthesizing the results 51D of the operations acquired from each monitoring device 2D. Through the above operation processing, the comprehensive analysis device 1 can acquire the variance-covariance matrix of all local learning data as the comprehensive result 40D. In the above step S203, the comprehensive analysis device 1 derives one or more principal components 41D from the calculated comprehensive result 40D by performing principal component analysis. In step S204, the comprehensive analysis device 1 outputs information related to the derived one or more principal components 41D.

[0276] (Utilization of principal components)

[0277] The derived one or more principal components 41D can be utilized for any purpose. In addition, the information related to the derived one or more principal components 41D can be provided to each monitoring device 2D at any time point. Each monitoring device 2D can utilize the derived one or more principal components 41D to compress the target data through the processing of the above step S301 to step S303. In addition, each monitoring device 2D can utilize the calculated one or more principal components 41D to identify the state of the driver RD in the target data through the processing of the above step S311 to step S315.

[0278] In the above step S315, each monitoring device 2D outputs information related to the result of inferring the state of the driver RD. For example, each monitoring device 2D can directly output the result of inferring the state of the driver RD to the output device. Alternatively, for example, when the result of inferring the state of the driver RD based on at least one of the drowsiness level and the fatigue level exceeding a threshold is determined to be that driving should preferably not continue, each monitoring device 2D outputs a warning to the output device to urge the driver RD to stop the vehicle and take a rest. Additionally, for example, when each monitoring device 2D is connected to a control device (not shown) that controls the operation of the vehicle, each monitoring device 2D determines an instruction for instructing the vehicle to perform a desired operation based on the result of inferring the state of the driver RD. Each monitoring device 2D can control the operation of the vehicle by giving the determined instruction to the control device. It should be noted that each monitoring device 2D and the control device can be constituted by an integrated computer.

[0279] (Feature)

[0280] According to this modification example, with respect to the sensing detection data that can be used for monitoring the state of the driver RD, an increase in cost can be suppressed, and an improvement in the quality of data analysis based on principal component analysis can be achieved. Thus, in the case of the above data compression, by using one or more derived principal components 41D, it is difficult to delete information beneficial to monitoring the state of the driver RD. Additionally, in the case of the above inference, an improvement in the accuracy of inferring the state of the driver RD in steps S312 to S314 can be achieved.

[0281] (E) Case of detecting an abnormality of a machine

[0282] Figure 17 An example of an application scenario of the detection system 100E according to the fifth modification example is schematically illustrated. This modification example is a case where the above-described embodiment is applied in the case of performing principal component analysis on sensing detection data obtained by a sensor SE for observing the state of a machine RE. The detection system 100E according to this modification example includes a comprehensive analysis device 1 and a plurality of detection devices 2E. Similar to the above-described embodiment, the comprehensive analysis device 1 and each detection device 2E may be connected via a network.

[0283] In this modified example, the local learning data 3E is sensing detection data obtained by the sensor SE. The type of the sensor SE is not particularly limited as long as it can observe the state of the machine RE, and can be appropriately selected according to the implementation manner. The sensor SE can be, for example, a microphone, an acceleration sensor, a vibration sensor, etc. Inferring the state of the machine RE can be constituted, for example, by determining whether the machine RE has an abnormality, determining the probability of the machine RE having an abnormality, identifying the type of the abnormality generated by or with signs of generation by the machine RE, determining the part where the abnormality has occurred, or a combination thereof. The type of the machine RE and the abnormality is not particularly limited and can be appropriately selected according to the implementation manner. The machine RE can be, for example, a device constituting a production line such as a conveyor device, an industrial robot, etc. The machine RE can be the entire device or a part of the device such as a motor. The abnormality can be, for example, a failure, the mixing of foreign matter, the adhesion of dirt, the wear of structural components. It can also be that, except for these limitations, the detection system 100E according to this modified example is configured in the same way as the system 100 according to the above-mentioned implementation manner.

[0284] (Detection device)

[0285] Each detection device 2E according to this modified example corresponds to each client device 2 according to the above-mentioned implementation manner. The hardware structure and software structure of each detection device 2E according to this modified example can be the same as those of each client device 2 according to the above-mentioned implementation manner. Thus, the information processing of each detection device 2E can be executed in the same order as that of each of the above-mentioned client devices 2.

[0286] In the above step S101, each detection device 2E collects the local learning data 3E. The local learning data 3E is constituted by a plurality of local samples of the sensing detection data obtained by the sensor SE. In the above step S102, each detection device 2E calculates the result 51E of the operation related to the correlation between the elements of the local learning data 3E. Each detection device 2E can calculate the autocorrelation matrix of the local learning data 3E as the result 51E of the operation by executing the processing of the above step S1021 to step S1023. In step S103, each detection device 2E outputs the calculated result 51E of the operation.

[0287] (Comprehensive analysis device)

[0288] In this modification example, the comprehensive analysis device 1 derives one or more principal components 41E with respect to the sensing detection data obtained by the sensor SE. Specifically, in the above step S201, the comprehensive analysis device 1 obtains the results 51E of operations related to correlation from each detection device 2E. In the above step S202, the comprehensive analysis device 1 calculates the comprehensive result 40E representing the correlation between the elements of all local samples included in all local learning data by comprehensively processing the results 51E of operations obtained from each detection device 2E. Through the above arithmetic processing, the comprehensive analysis device 1 can obtain the variance-covariance matrix of all local learning data as the comprehensive result 40E. In the above step S203, the comprehensive analysis device 1 can derive one or more principal components 41E from the calculated comprehensive result 40E by performing principal component analysis. In step S204, the comprehensive analysis device 1 outputs information related to the one or more derived principal components 41E.

[0289] (Utilization of Principal Components)

[0290] The one or more derived principal components 41E can be utilized in any application. In addition, the information related to the one or more derived principal components 41E can be supplied to each detection device 2E at any time point. Each detection device 2E can utilize the one or more derived principal components 41E to compress the target data through the processing of the above steps S301 to S303. In addition, each detection device 2E can utilize the one or more calculated principal components 41E to identify the state of the machine RE in the target data through the processing of the above steps S311 to S315.

[0291] In the above step S315, each detection device 2E outputs information related to the result of inferring the state of the machine RE. For example, each detection device 2E can directly output the result of inferring the state of the machine RE to the output device. In addition, for example, it can also be that, in the case where an abnormality of the machine RE is detected based on the result of inferring the state of the machine RE, each detection device 2E outputs a warning for notifying the occurrence of the abnormality to the output device. Moreover, in the case where each detection device 2E is configured to be able to control the operation of the machine RE, each detection device 2E can stop the operation of the machine RE corresponding to the detection of the occurrence of an abnormality. In addition, each detection device 2E can also output the type of the abnormality generated by the machine RE and information indicating the maintenance method for dealing with the abnormality to the output device. In this case, the information indicating the maintenance method for dealing with the abnormality can be stored in a predetermined storage area such as a storage unit, a storage medium, an external storage device, a storage medium, etc. Each detection device 2E can appropriately obtain the information indicating the maintenance method for dealing with the abnormality from the predetermined storage area.

[0292] (Features)

[0293] According to this modification example, for the sensing detection data that can be used for detecting anomalies in the machine RE, an increase in cost can be suppressed, and an improvement in the quality of data analysis based on principal component analysis can be achieved. Thus, in the case of the above data compression, by using one or more derived principal components 41E, it is difficult to delete information beneficial to the detection of anomalies in the machine RE. In addition, in the case of the above inference, an improvement in the accuracy of detecting anomalies in the machine RE in steps S312 to S314 can be achieved.

[0294] <4.2>

[0295] In at least any one of the multiple client devices 2 according to the above embodiment, each local sample 30 can be weighted according to its importance. In addition, in at least any one of the multiple client devices 2 according to the above embodiment, two or more elements that are the objects of principal component analysis can be specified from among the multiple elements constituting each local sample 30.

[0296] Figure 18 An example of the software structure of the client device 2J according to this modification example is schematically illustrated. The hardware structure of the client device 2J according to this modification example can be the same as that of the client device 2 according to the above embodiment. The control device of the client device 2J interprets and executes the commands included in the collection program through the CPU. Thus, the client device 2J operates as a computer, and this computer further includes a first reception unit 205 and a second reception unit 206 as software modules. The first reception unit 205 receives the specification of the importance of each local sample 30. The second reception unit 206 receives the specification of two or more elements from among the multiple elements constituting each local sample 30. It may be that, except for this point, the client device 2J according to this modification example is configured in the same manner as the client device 2 according to the above embodiment.

[0297] Figure 19 It is a flowchart showing an example of the processing sequence related to the collection of the local learning data 3 based on the client device 2J according to this modification example. It should be noted that the processing sequence described below is merely an example, and each step can be changed as much as possible. In addition, for the processing sequence described below, steps can be appropriately omitted, replaced, and added according to the embodiment.

[0298] In step S101, similar to the above-described embodiment, the control unit of the client device 2J acts as the collection unit 201 and collects local learning data 3. In step S111, the control unit acts as the first reception unit 205 and receives the designation of the importance of each local sample 30. In step S112, the control unit acts as the second reception unit 206 and receives the designation of two or more elements from among the multiple elements that make up each local sample 30. The order of processing in step S111 and step S112 is not limited to such an example and can be appropriately determined according to the embodiment.

[0299] Figure 20 Schematically illustrate an example of a screen 250 for receiving the designation of importance and elements to be analyzed. In the case where the client device 2J includes a display as an output device, Figure 20 the illustrated screen 250 can be displayed on the display. The screen 250 includes an input field 251 and a checkbox 252. Each input field 251 is set according to the local sample 30. In addition, each checkbox 252 is set according to the elements of the local sample 30.

[0300] Figure 20 In the example, the input field 251 is configured to be able to designate the importance of each local sample 30 in five levels. Sample A is designated as importance "5", and sample B is designated as importance "1". However, the method of designating importance and the number of levels are not limited to such an example and can be appropriately determined according to the embodiment. Importance can be designated by discrete values or continuous values. In addition, Figure 20 In the example, by checking the checkbox 252, the elements to be analyzed can be designated. However, the method of designating elements is not limited to such an example and can be appropriately determined according to the embodiment.

[0301] An operator can designate the importance of each local sample 30 by operating each input field 251 via an input device. If the importance of each local sample 30 is designated, each local sample 30 is weighted according to the designated importance. Similarly, an operator can designate two or more elements to be analyzed by operating each checkbox 252 via an input device. It should be noted that the elements to be analyzed can be designated according to the task. As an example, in the case of the above-described appearance inspection, different elements between the first task of detecting the first defect and the second task of detecting the second defect can also be designated as the elements to be analyzed. Thereby, elements suitable for the execution of the task can be designated as the elements to be analyzed.

[0302] If the designation of importance and the elements to be analyzed are completed, the control unit advances the process to the next step S102. In the above calculation, X n (P)Replace it with the following formula 11, and the above N (P) Replace it with the following formula 12. In addition, the elements that are not specified (selected) are removed from the operation. Except for these aspects, the control unit performs the process of step S102 in the same manner as in the above-described embodiment.

[0303] [Formula 11]

[0304]

[0305] [Formula 12]

[0306]

[0307] w n (P) represents the importance (weight) specified for the nth local sample 30 of the local learning data 3 collected by the Pth client device 2. Through the process of step S102, the control unit can generate the result 51 of the operation related to the correlation reflecting the importance for the specified elements. In this modified example, the average value of each element of all the local samples obtained in the above step S1021 is a weighted average weighted according to the importance. By executing the processes of the above step S1021 to step S1023, the control unit can calculate the autocorrelation matrix reflecting the importance as the result 51 of the operation for the specified elements. It should be noted that the removal of the elements that are not specified may be performed not by the client device 2J but by the comprehensive analysis device 1. In step S103, the control unit outputs the calculated result 51 of the operation.

[0308] (Comprehensive Analysis Device)

[0309] Corresponding to the specified importance, in the above step S202, the control unit 11 can calculate the variance-covariance matrix of all the local learning data as the comprehensive result 40 by dividing the sum of the autocorrelation matrices by the sum of the weights corresponding to the importance. In addition, by specifying the elements to be analyzed, in the above step S202, the control unit 11 synthesizes the obtained results 51 of the operation for two or more specified elements to calculate the comprehensive result 40. In addition, in the above step S203, the control unit 11 derives one or more principal components 41 from the comprehensive result 40 calculated for two or more specified elements by performing principal component analysis. It may be that, except for these aspects, the information processing of the comprehensive analysis device 1 according to this modified example is the same as that in the above-described embodiment.

[0310] (Feature)

[0311] According to this modification example, it is possible to assign superiority or inferiority to each local sample 30 based on importance. The importance of important local samples 30 can be increased, and the importance of unimportant local samples 30 can be decreased. By reflecting the importance of each specified local sample 30 in this way in the principal component analysis, it is possible to improve the quality of data analysis based on the principal component analysis. In addition, according to this modification example, it is possible to select two or more elements from among a plurality of elements of the local sample 30 to be the object of the principal component analysis. Thus, by removing elements with low relevance to the purpose such as the execution of a task from the analysis object, it is possible to reduce the computational cost involved in the principal component analysis and derive principal components 41 suitable for the purpose. It should be noted that in the software structure of the client device 2J according to this modification example, at least one of the first reception unit 205 and the second reception unit 206 can be omitted. Accordingly, at least one of the specification of importance and the specification of target elements can be omitted.

[0312] (Grouping)

[0313] The more similar the elements specified as the analysis object are, the higher the likelihood that the utilization purposes of the collected local learning data 3 are similar. Therefore, it is assumed that if the operation results 51 calculated from the local learning data 3 collected between client devices 2J with high similarity of the specified elements are integrated, one or more principal components 41 can be appropriately derived from the obtained integrated result 40. Here, in this modification example, it is also possible that the control unit 11 of the integrated analysis device 1 acts as a grouping unit 115 and groups each client device 2J based on the element specification result.

[0314] Figure 21 is a flowchart showing an example of the processing sequence of the grouping method using the element specification result. When the group allocation method adopts the Figure 21 grouping method shown, the allocation of each client device 2J to at least one of a plurality of groups is constituted by the following step S251 and step S252. However, the processing sequence described below is merely an example, and each process can be changed as much as possible. In addition, for the processing sequence described below, steps can be appropriately omitted, replaced, and added according to the embodiment.

[0315] In step S251, the control unit 11 obtains the results specifying two or more elements from each client device 2J. The method for obtaining the specified results can be appropriately determined according to the embodiment. In step S252, the control unit 11 assigns each client device 2J to at least one of a plurality of groups based on the degree of agreement of the two or more elements specified in each client device 2J. As an example, the control unit 11 can assign the client devices 2J with exactly the same specified elements to the same group. Alternatively, the control unit 11 can assign the client devices 2J with the degree of agreement of the specified elements exceeding a threshold to the same group. The threshold can be set appropriately. Thus, it is possible to group each client device 2J by using the specified results of the elements to be analyzed.

[0316] According to this grouping, in the above step S202, the control unit 11 synthesizes the calculation results 51 obtained from each client device 2J for the two or more specified elements within the same group, thereby calculating the synthesis result 40. Moreover, in the above step S203, the control unit 11 derives one or more principal components 41 from the synthesis result 40 calculated for the two or more specified elements within the same group by performing principal component analysis.

[0317] It should be noted that, when there are client devices 2J with different specified elements within the same group, in the above step S202, the control unit 11 can synthesize the calculation results 51 obtained from each client device 2J for the elements including the elements specified by each client device 2J. Moreover, in the above step S203, the control unit 11 can derive one or more principal components 41 from the synthesis result 40 calculated for the included elements. Alternatively, in the above step S203, the control unit 11 can derive one or more principal components 41 from the synthesis result 40 calculated for the two or more elements specified for each client device 2J.

[0318] <4.3>

[0319] In the above embodiment, each client device 2 performs three information processes of collecting local learning data 3, data compression, and predetermined inference. However, the structure of each client device 2 is not limited to such an example. At least any one of the plurality of client devices 2 can be constituted by multiple computers. In this case, each information process can be executed by different computers.

[0320] Figure 22An example of the structure of the client device 2K according to this modification is schematically illustrated. In this modification, the client device 2K includes a collection device 2001, a first utilization device 2002, and a second utilization device 2003. The hardware structures of the collection device 2001, the first utilization device 2002, and the second utilization device 2003 can be the same as those of the respective client devices 2 according to the above-described embodiment. The collection device 2001 operates as a computer having a collection unit 201, an arithmetic unit 202, and an output unit 203 as software modules by executing a collection program 85. The first utilization device 2002 operates as a computer having an acquisition unit 211, a compression unit 212, and an output unit 213 as software modules by executing a compression program 86. The second utilization device 2003 operates as a computer having an acquisition unit 215, an inference unit 216, and an output unit 217 as software modules by executing an inference program 87.

[0321] It should be noted that the computer that utilizes one or more derived principal components 41 is not limited to each client device 2. One or more derived principal components 41 can be utilized by a computer other than each client device 2. The other computer can include the above-described comprehensive analysis device 1. In addition, the use of one or more derived principal components 41 is not limited to the above-described data compression and predetermined inference. One or more derived principal components 41 can be utilized for any purpose.

[0322] <4.4>

[0323] In the above-described embodiment, in the process of calculating the operation result 51 in step S102, the number of local samples 30 of each client device 2 and the average value of each element are used to calculate the average value U of each element of all local samples. In addition, in the above step S202, the operation results 51 obtained from each client device 2 are integrated. These data are related to the local learning data 3. Therefore, if these data are disclosed, the confidentiality of the local learning data 3 of each client device 2 may be damaged. Here, in order to improve the confidentiality of the local learning data 3, each operation can use secure computing. The average value U of each element of all local samples can be calculated by secure computing using the number of local samples 30 obtained from each client device 2 and the average value of each element. In addition, the integration of the operation results 51 can be performed by secure computing. The method of secure computing is not particularly limited and can be appropriately selected according to the embodiment. In this modification, the control unit 11 can perform secure computing by any one of the following two methods.

[0324] (A) Using the method of secret sharing

[0325] Figure 23An example is schematically illustrated of a case where secret computation is performed using secret sharing. In the method using secret sharing, the first server 61 and the second server 62 are respectively set up as reliable third-party devices in the network. The first server 61 and the second server 62 are each a computer having a hardware processor and a memory, similar to the comprehensive analysis device 1 and the like.

[0326] In this method, first, when the control unit 21 of each client device 2 sends its own operation result to another computer, it generates a random number. The cases of sending its own operation result to another computer are the case of sending the number of local samples 30 and the average value of each element (hereinafter, also referred to as "the first case") and the case of sending the operation result 51 (hereinafter, also referred to as "the second case") in the above-described embodiment. In the first case, the other computer is a computer that calculates the average value U of each element of all local samples (for example, the comprehensive analysis device 1, other client devices 2). In addition, in the second case, the other computer is the comprehensive analysis device 1. The method of generating a random number can be appropriately selected according to the embodiment.

[0327] Next, the control unit 21 calculates the difference between the value of the sent operation result and the generated random number. In the first case, the sent operation results are the product of the number of local samples 30 and the average value of each element and the number of local samples 30. The control unit 21 calculates the difference between each of them and the random number. The random numbers having a difference between them can be the same or different. In the second case, the sent operation results are the operation result 51 related to the correlation and the number of local samples 30. Similar to the first case, the control unit 21 calculates the difference between each of them and the random number. The random numbers having a difference between them can be the same or different. Then, the control unit 21 sends the calculated difference to the first server 61, and on the other hand, sends the generated random number to the second server 62.

[0328] Accordingly, the first server 61 calculates the sum of the differences received from each client device 2 as shown in Equation 13 below. On the other hand, the second server 62 calculates the sum of the random numbers received from each client device 2 as shown in Equation 14 below.

[0329] [Equation 13]

[0330] ∑P(Y(P)-s(P)) …(Equation 13)

[0331] [Equation 14]

[0332] ∑Ps(P)…(Equation 14)

[0333] It should be noted that Y (P)Indicates the value based on the calculation result of the P-th client device 2. In the first case, Y (P) is the product of the number of local samples 30 and the average value of each element (N (P) U (P) ) and the number of local samples 30, N (P) These two are calculated separately. On the other hand, in the second case, Y (P) is the result 51 of the operation related to the correlation (Q (P) ) and the number of local samples 30, N (P) These two are calculated separately. s (P) Indicates the random number generated by the P-th client device 2.

[0334] The first server 61 and the second server 62 respectively send the calculation results of the sum to other computers. The other computers add the calculation results of the sum received from the first server 61 and the second server 62 respectively. Thus, it is possible to prevent the calculation results of each client device 2 from being determined in other computers, and the other computers calculate the sum of each calculation result. In the first case, each client device 2 can obtain the calculation result of Equation 4. In the second case, the comprehensive analysis device 1 can obtain the calculation result of Equation 7.

[0335] It should be noted that the method of secret sharing may not be particularly limited and can be appropriately selected according to the implementation manner. For example, the method of the international standard (ISO / IEC 19592-2:2017) can be used for the method of secret sharing. As long as the comprehensive analysis device 1 is a reliable server, the comprehensive analysis device 1 can act as either the first server 61 or the second server 62. In addition, the first server 61 and the second server 62 can be composed of the same computer.

[0336] (B) Method using homomorphic encryption

[0337] Figure 24 Schematically illustrates an example of a case where secret calculation is performed using homomorphic encryption. In the method using homomorphic encryption, the server 65 is set as a reliable third-party device in the network. The server 65 is a computer that has a hardware processor and a memory in the same way as the comprehensive analysis device 1 and the like.

[0338] In this method, first, the server 65 issues a public key and a private key. The public key is generated to have homomorphic properties. In other words, when two ciphertexts encrypted by the public key are given, the public key is generated in such a way that the two ciphertexts can be added while maintaining the encrypted state. The server 65 sends the public key among the issued public key and private key to each client device 2.

[0339] The control b21 of each client device 2 encrypts its own calculation result using the received public key. Moreover, the control unit 21 sends the encrypted calculation result to other computers. Other computers calculate the sum of the values of the calculation results received from each client device 2 while keeping them encrypted as shown in Equation 15 below.

[0340] [Equation 15]

[0341] H(∑PY(P))…(Equation 15)

[0342] Note that H indicates encryption using the public key.

[0343] Other computers send the encrypted sum to the server 65. The server 65 decrypts the encrypted sum received from other computers using the private key. Moreover, the server 65 returns the sum of the decrypted calculation results to other computers. Thus, it is possible to prevent the calculation results of each client device 2 from being determined in other computers, and other computers can calculate the sum of each calculation result. In the first case, each client device 2 can obtain the calculation result of Equation 4. In the second case, the comprehensive analysis device 1 can obtain the calculation result of Equation 7.

[0344] Note that the method of homomorphic encryption is not particularly limited and can be appropriately selected according to the embodiment. For example, modified-ElGamal encryption, Paillier encryption, etc. can be used as the method of homomorphic encryption. In addition, if the comprehensive analysis device 1 is a reliable server, the comprehensive analysis device 1 can act as the server 65.

[0345] As described above, according to this modification example, by any of the above two methods, it is possible to calculate the sums of Equation 4 and Equation 7 respectively using secure multi-party computation. Thus, it is possible to improve the confidentiality of the local learning data 3 of each client device 2.

[0346] <4.5>

[0347] In the above embodiment, the process of grouping each client device 2 can also be omitted. Accordingly, the grouping unit 115 can also be omitted from the software structure of the comprehensive analysis device 1. In addition, the processes of the above steps can also be executed by different computers. For example, the process of the above step S101 and the process of the above step S102 can be executed by different computers.

[0348] Description of Reference Numerals

[0349] 1…Comprehensive analysis device; 11…Control unit; 12…Storage unit; 13…Communication interface; 14…Input device; 15…Output device; 16…Driver; 111…Acquisition unit; 112…Integration unit; 113…Analysis unit; 114…Output unit; 115…Packetization unit; 121…Principal component information; 123…Group list; 124…Allocation information; 81…Comprehensive analysis program; 91…Storage medium; 2…Client device; 21…Control unit; 22…Storage unit; 23…Communication interface; 24…Input device; 25…Output device; 26…Driver; 27…External interface; 201…Collection unit; 202…Arithmetic unit; 203…Output unit; 211…Acquisition unit; 212…Compression unit; 213…Output unit; 215…Acquisition unit; 216…Inference unit; 217…Output unit; 221…Arithmetic result data; 223…Object data; 224…Compressed data; 226…Object data; 227…Data group; 228…Sample; 85…Collection program; 86…Compression program; 87…Inference program; 92…Storage medium; 3…Local learning data; 30…Local sample; 40…Comprehensive result; 41…Principal component; 51…(Arithmetic) result.

Claims

1. A comprehensive analysis method, comprising the following steps: A plurality of client devices respectively perform operations on local learning data, and the operations are used to solve the correlation between various elements of each local sample included in the local learning data; The server device obtains the results of the operations from each of the client devices; The server device calculates a comprehensive result representing the correlation between various elements of all local samples included in all local learning data by synthesizing the results of the operations obtained from each of the client devices; The server device derives one or more principal components from the calculated comprehensive result by performing principal component analysis; And The server device outputs information related to the derived one or more principal components, The server device includes a grouping unit, The client device includes a first receiving unit for receiving the designation of the importance of each local sample and a second receiving unit for receiving the designation of two or more elements from among the multiple elements constituting each local sample, The comprehensive analysis method further includes a step in which the grouping unit assigns each of the client devices to at least any one of a plurality of groups based on the designation result of the elements, In the calculating step, the server device calculates the comprehensive result by synthesizing the results of the operations obtained from each of the client devices within the same group, In the deriving step, the server device derives one or more principal components from the comprehensive result calculated within the same group by performing principal component analysis, In the assigning step, a group list showing a list of the plurality of groups that are candidates for assigning each of the client devices is sent to each of the client devices, and an operator of each of the client devices refers to the group list output by the output device, operates the input device, selects one or more desired groups from the group list, and assigns each of the client devices to the selected one or more desired groups.

2. The comprehensive analysis method according to claim 1, wherein The operation for solving the correlation is composed of the following steps: Obtain the average value of each element of all local samples included in all local learning data; Normalize each local sample by subtracting the obtained average value from the value of each element of each local sample included in the local learning data; And According to the normalized local samples, calculate the autocorrelation matrix of the local learning data, Obtaining the result of the operation is constituted by obtaining the calculated autocorrelation matrix, Synthesizing the results of the operations is constituted by summing the autocorrelation matrices obtained from each of the client devices.

3. The comprehensive analysis method according to claim 2, wherein Each local sample is weighted according to the designated importance, The average value of each element of all local samples is a weighted average value weighted according to the importance, In the step of the calculation, the server device calculates the variance-covariance matrix of all the local learning data as the comprehensive result by dividing the sum of the autocorrelation matrices by the weighted sum corresponding to the importance.

4. The comprehensive analysis method according to claim 2 or 3, wherein, The average value of each element of all the local samples is calculated by using the number of the local samples obtained from each of the client devices and the secret calculation of the average value of each element.

5. The comprehensive analysis method according to claim 1, wherein, The integration of the operation results is performed by secret calculation.

6. The comprehensive analysis method according to claim 1, wherein, In the step of the calculation, the server device calculates the comprehensive result by integrating the operation results obtained from each of the client devices for the specified two or more elements. In the step of the derivation, the server device derives one or more principal components from the calculated comprehensive result for the specified two or more elements by performing principal component analysis.

7. The comprehensive analysis method according to claim 1, wherein, Outputting information related to the one or more principal components is constituted by the server device sending information related to the derived one or more principal components to each of the client devices.

8. The comprehensive analysis method according to claim 1, wherein, The local learning data is composed of image data reflecting a product or measurement data obtained by measuring the attributes of a product.

9. The comprehensive analysis method according to claim 1, wherein, The local learning data is composed of sensing detection data obtained by a sensor observing the state of an object person.

10. A comprehensive analysis device, comprising: An acquisition unit that respectively acquires the results of operations from a plurality of client devices: the operations are operations for solving the correlation between each element of each local sample included in the local learning data performed on the local learning data respectively collected by the plurality of client devices; A synthesis unit that calculates a comprehensive result representing the correlation between each element of all the local samples included in all the local learning data by synthesizing the operation results obtained from each of the client devices; An analysis unit that derives one or more principal components from the calculated comprehensive result by performing principal component analysis; An output unit that outputs information related to the derived one or more principal components; And A grouping unit, The client device includes a first reception unit that receives the designation of the importance of each local sample and a second reception unit that receives two or more elements from among the plurality of elements constituting each local sample, The grouping unit distributes each of the client devices to at least one of a plurality of groups based on the element designation result, Calculates the comprehensive result by synthesizing the operation results obtained from each of the client devices within the same group, Derives one or more principal components from the comprehensive result calculated within the same group by performing principal component analysis, Send a group list showing a list of the multiple groups that are candidates for allocation to each of the client devices to each of the client devices. An operator of each client device refers to the group list output by the output device, operates the input device, selects one or more desired groups from the group list, and allocates each client device to the selected one or more desired groups.

11. A storage device stores a comprehensive analysis program that causes a computer to perform the following steps: Obtain the results of operations respectively from multiple client devices, where the operations are for solving the correlations between respective elements of each local sample included in local learning data collected by the multiple client devices. Calculate a comprehensive result representing the correlations between all elements of all local samples included in all local learning data by synthesizing the results of the operations obtained from each client device. Derive one or more principal components from the calculated comprehensive result by performing principal component analysis. Output information related to the derived one or more principal components. The client device includes a first reception unit that receives the designation of the importance of each local sample and a second reception unit that receives the designation of two or more elements from among the multiple elements constituting each local sample. The comprehensive analysis program further causes the computer to perform the following step: Allocate each client device to at least any one of multiple groups based on the designation result of the elements. In the calculating step, the server device calculates the comprehensive result by synthesizing the results of the operations obtained from each client device within the same group. In the deriving step, the server device derives one or more principal components from the comprehensive result calculated within the same group by performing principal component analysis. In the allocating step, the server device sends a group list showing a list of the multiple groups that are candidates for allocation to each client device to each client device. An operator of each client device refers to the group list output by the output device, operates the input device, selects one or more desired groups from the group list, and allocates each client device to the selected one or more desired groups.

Images