JP7119901B2 - Analysis device, analysis method and analysis program - Google Patents

Description

The present invention relates to analyzers and the like.

There is a technique for estimating context using sensor data, which is the measurement result of a sensor. Various contexts are included in the inferred context. For example, an acceleration sensor is used to estimate a worker's "overturn", and a temperature and humidity sensor is used to estimate a "heat level". In addition, there is also a method that uses a sensor to detect the position of the worker and estimate the working efficiency of the worker.

JP 2018-28906 A JP 2018-124639 A JP 2011-180946 A JP 2017-21711 A

However, the conventional technology described above has a problem that it is impossible to specify a combination of sensor data used for estimating the context.

In recent years, context has been used as information that forms the basis of application applications using IoT (Internet of Things), such as visualization of work, optimization of work, improvement of QoL (Quality of Life) for the elderly, and preventive care. and the content of the context is diversified. For example, as a context, estimating human activity labels such as "unloading" and "displaying goods", estimating "stress amount", and estimating "IADL (Instrumental Activities of Daily Living) index". is required.

In order to estimate the context that is the basis of the above application, it is necessary to recognize work behavior and human conditions, and various factors are related. Difficult to infer context. Moreover, when a context to be estimated is determined, it is difficult to specify a combination of sensor data used for estimating the context.

Here, it is conceivable to generate a model for estimating context by simply performing learning with a deep neural network or the like using all sensor data. However, in such a model, the context is estimated by inputting all sensor data, and the amount of calculation, data communication, etc. at the time of estimation becomes enormous, and it is difficult to perform real-time processing at the time of operation.

In one aspect, an object of the present invention is to provide an analysis device, an analysis method, and an analysis program capable of specifying a combination of sensor data used for estimating context.

In a first alternative, the analysis device comprises a splitter, an evaluator and a selector. The dividing unit obtains time-series information measured by a plurality of sensors, and divides the time-series information into a plurality of time intervals to generate a plurality of measured values. The evaluation unit evaluates a combination of a plurality of measured values corresponding to the first sensor and a plurality of measured values corresponding to the second sensor among the plurality of sensors. The selection unit selects a set of sensors to be combined based on the evaluation result of the evaluation unit.

It is possible to specify necessary sensor data or a combination of sensor data without specifying individual measurement targets required for context estimation. In addition, the amount of data communication can be reduced by specifying necessary acquisition target sensor data.

FIG. 1 is a diagram showing the configuration of the system during learning according to this embodiment. FIG. 2 is a functional block diagram showing the configuration of the analyzer according to this embodiment. FIG. 3 is a diagram illustrating an example of a data structure of a sensor DB; FIG. 4 is a diagram showing an example of the data structure of the correct context DB. FIG. 5 is a diagram showing an example of the data structure of a weight table. FIG. 6 is a diagram showing an example of the data structure of the selected sensor table. FIG. 7 is a diagram illustrating an example of an intermediate estimator generated by the dividing unit; FIG. 8 is a diagram for explaining processing for generating a feature amount vector. FIG. 9 is a diagram for explaining weight vectors. FIG. 10 is a flow chart showing the processing procedure of the analyzer according to the present embodiment. FIG. 11 is a diagram showing the configuration of the system during operation according to this embodiment. FIG. 12 is a functional block diagram showing the configuration of the estimation device according to this embodiment. FIG. 13 is a flow chart showing the processing procedure of the estimation device according to the present embodiment. FIG. 14 is a diagram showing the configuration of another system during learning according to this embodiment. FIG. 15 is a diagram showing the configuration of another system during operation according to this embodiment. FIG. 16 is a diagram showing an example of the hardware configuration of a computer that implements the same functions as the analysis apparatus according to this embodiment. FIG. 17 is a diagram illustrating an example of the hardware configuration of a computer that implements the same functions as the estimation device according to this embodiment.

Hereinafter, embodiments of the analysis apparatus, analysis method, and analysis program disclosed in the present application will be described in detail based on the drawings. In addition, this invention is not limited by this Example.

FIG. 1 is a diagram showing the configuration of the system during learning according to this embodiment. As shown in FIG. 1, this system has sensors 10a, 10b, 10c and an analyzer 100. As shown in FIG. Sensors 10 a , 10 b , 10 c and analysis device 100 are interconnected via network 50 . For the sake of convenience, only the sensors 10a to 10c are shown here, but it is assumed that the analysis device 100 is further connected to other sensors.

The sensors 10a to 10c are any of various sensors such as image sensors, sound sensors, and motion sensors. For example, sensors 10a-10c are installed at the work site. Note that the installation location of the sensor is not limited to the work site. For example, it may be worn by a worker, attached to a machine operated by the worker, or combined with sensors attached to the environment, people, and objects. The sensors 10a to 10c transmit sensor data, which are measurement results, to the analyzer 100. FIG. It is assumed that sensor data includes information in which sensor values and times are associated in chronological order. The sensor data header contains sensor identification information that uniquely identifies the sensor that is the source of the transmission.

The sensor 10a has a sensing module 11a and a communication control section 12a. The sensing module 11a is a sensor that measures image information, sound information, or motion information. In this embodiment, as an example, the sensing module 11a is a module that senses image information. The sensing module 11a outputs sensor data to the communication control unit 12a.

The communication control unit 12a is a processing unit that receives sensor data from the sensing module 11a and controls processing for transmitting the sensor data to the analysis device 100 and an estimation device (not shown). The communication control unit 12a transmits sensor data to the analysis device 100 in the learning phase. In the operation phase, the communication control unit 12a transmits sensor data to the estimation device when receiving a transmission command from the estimation device. In the operation phase, the communication control unit 12a suppresses the transmission of the sensor data when the transmission command is not received. The communication control unit 12a corresponds to a communication device.

The sensor 10b has a sensing module 11b and a communication controller 12b. The sensing module 11b is a sensor that measures image information, sound information or motion information. In this embodiment, as an example, the sensing module 11b is a module that senses sound information. The sensing module 11b outputs sensor data to the communication control unit 12b.

The communication control unit 12b is a processing unit that receives sensor data from the sensing module 11b and controls processing for transmitting the sensor data to the analysis device 100 and an estimation device (not shown). The communication control unit 12b transmits sensor data to the analysis device 100 in the learning phase. In the operation phase, the communication control unit 12b transmits sensor data to the estimation device when receiving a transmission command from the estimation device. In the operation phase, the communication control unit 12b suppresses transmission of sensor data when no transmission command is received. The communication control unit 12b corresponds to a communication device.

The sensor 10c has a sensing module 11c and a communication controller 12c. The sensing module 11c is a sensor that measures image information, sound information or motion information. In this embodiment, as an example, the sensing module 11c is a module that senses exercise information. The sensing module 11c outputs sensor data to the communication control unit 12c. For example, the sensor 10c is installed on a worker and senses the worker's exercise information.

The communication control unit 12c is a processing unit that receives sensor data from the sensing module 11c and controls processing for transmitting the sensor data to the analysis device 100 and an estimation device (not shown). The communication control unit 12c transmits sensor data to the analysis device 100 in the learning phase. In the operation phase, the communication control unit 12c transmits sensor data to the estimation device when receiving a transmission command from the estimation device. In the operation phase, the communication control unit 12c suppresses the transmission of the sensor data when the transmission command is not received. The communication control unit 12c corresponds to a communication device.

Analysis device 100 receives sensor data from each of sensors 10a-10c (and other sensors not shown) and converts each sensor data into an intermediate estimator. Analysis device 100 uses the result of multiplying each intermediate estimator to learn parameters that can estimate a given context, and based on the learned parameters, analyzes useful sensor information for estimating a given context. Choose a combination.

FIG. 2 is a functional block diagram showing the configuration of the analyzer according to this embodiment. As shown in FIG. 2 , this analysis device 100 has a communication section 110 , an input section 120 , a display section 130 , a storage section 140 and a control section 150 .

The communication unit 110 is a processing unit that receives sensor data from the sensors 10a to 10c (and other sensors not shown) via the network 50. FIG. Communication unit 110 outputs the sensor data to control unit 150 . The communication unit 110 corresponds to a communication device. Also, the communication unit 110 transmits and receives data to and from an estimation device (not shown). A control unit 150 , which will be described later, exchanges data with each sensor and estimation device via the communication unit 110 .

The input unit 120 is an input device for inputting various kinds of information to the analysis device 100 . For example, input unit 120 corresponds to a keyboard, mouse, touch panel, and the like.

Display unit 130 is a display device for displaying information output from control unit 150 . For example, display unit 130 corresponds to a liquid crystal display, an organic display, a touch panel, or the like.

The storage unit 140 has a sensor DB 140a, a correct context DB 140b, a weight table 140c, and a selected sensor table 140d. The storage unit 140 corresponds to semiconductor memory devices such as RAM (Random Access Memory), ROM (Read Only Memory), flash memory, and storage devices such as HDD (Hard Disk Drive).

The sensor DB 140a is a DB (Data Base) that holds sensor data measured by the sensors 10a to 10c and other sensors (not shown). FIG. 3 is a diagram illustrating an example of a data structure of a sensor DB; As shown in FIG. 3, the sensor DB 140a associates sensor identification information with sensor data. The sensor identification information is information that uniquely identifies the sensor. The sensor data is sensor data measured by each sensor in chronological order.

The correct context DB 140b is a DB that associates a context to be estimated with a time zone in which the context occurs. FIG. 4 is a diagram showing an example of the data structure of the correct context DB. As shown in FIG. 4, the correct context DB 140b associates context contents with time periods. The context content indicates the content of the context to be estimated. The time zone indicates the time zone in which the relevant context occurs. For example, the record shown in the first line indicates that the context content "unloading" is occurring in the time period "time Ts1 to Te1".

The weight table 140c is a table that holds weight vector information for estimating the context. FIG. 5 is a diagram showing an example of the data structure of a weight table. As shown in FIG. 5, the weight table 140c associates context contents with weight vectors. The context content indicates the content of the context to be estimated. A weight vector is a learned weight vector for estimating the relevant context, and consists of a plurality of elements.

The selected sensor table 140d is a table that holds sensor information used for estimating the context. FIG. 6 is a diagram showing an example of the data structure of the selected sensor table. As shown in FIG. 6, this selected sensor table 140d associates context contents with selected sensors. The context content indicates the content of the context to be estimated. A selected sensor indicates a sensor that is selected when estimating the relevant context.

Returning to the description of FIG. The control unit 150 has a reception unit 150a, a division unit 150b, an evaluation unit 150c, a selection unit 150d, and a notification unit 150e. The control unit 150 can be realized by a CPU (Central Processing Unit), an MPU (Micro Processing Unit), or the like. The control unit 150 can also be realized by hardwired logic such as ASIC (Application Specific Integrated Circuit) and FPGA (Field Programmable Gate Array).

The receiving unit 150a is a processing unit that receives sensor data from the sensors 10a to 10c and other sensors (not shown). The receiving unit 150a stores each sensor data in the sensor DB 140a. The receiving unit 150a refers to the sensor identification information included in the header of each sensor data, associates the sensor data with the corresponding sensor identification information, and stores the sensor data in the sensor DB 140a.

The receiving unit 150a also receives correct context data and stores the received correct context data in the correct context DB 140b. The receiving unit 150 a may receive correct context data from the input unit 120 or from an external device (not shown) via the network 50 .

The dividing unit 150b is a processing unit that divides the sensor data and generates an intermediate estimator based on the divided data. As described below, the dividing unit 150b performs dividing processing and intermediate estimation processing for each sensor data associated with each sensor identification information.

An example of the dividing process by the dividing unit 150b will be described. The dividing unit 150b acquires sensor data to be divided from the sensor DB 140a, and divides the acquired sensor data into certain time window widths (time intervals). Note that each sensor may have a plurality of time window widths. Also, the time window width may be different for each sensor. In the following description, each piece of data obtained by dividing sensor data by a certain time window width is referred to as "divided sensor data".

The dividing unit 150b holds each divided sensor data obtained by dividing the sensor data of the sensor identification information "sensor 10a" in association with the sensor identification information "sensor 10a". The dividing unit 150b holds each divided sensor data obtained by dividing the sensor data of the sensor identification information "sensor 10b" in association with the sensor identification information "sensor 10c". The dividing unit 150b holds each divided sensor data obtained by dividing the sensor data of the sensor identification information "sensor 10c" in association with the sensor identification information "sensor 10c". The dividing unit 150b also stores each divided sensor data obtained by dividing the sensor data corresponding to other sensor identification information in association with the corresponding sensor identification information. It should be noted that the method of dividing sensor data into time window widths in the present invention is not limited to the method of handling data as separate data for each time interval as described above, and is divided according to an algorithm for processing sensor data. Including cases. For example, in a neural network having a recurrent structure, specifically RNN (Recurrent Neural Network), LSTM (Long Short Term Memory), GRU (Gated Recurrent Unit), the degree of delay of the time signal for recurring input is defined as the time window width. Also included is one that arranges time-series signals one-dimensionally, performs convolution calculations with a CNN (Convolutional Neural Network), and gives the filter size of the input part as the time window width.

An example of processing for generating an intermediate estimator by the dividing unit 150b will be described. The dividing unit 150b generates one intermediate estimator based on one divided sensor data. For example, the dividing unit 150b divides the divided sensor data into a plurality of partial data, and extracts feature amounts from each partial data. The dividing unit 150b clusters each partial data based on the feature amount extracted from each partial data. The dividing unit 150b generates, as an intermediate estimator, a histogram in which each clustered group is plotted on the horizontal axis and the number (frequency) of partial data belonging to the group is plotted on the vertical axis.

Here, when the divided sensor data (sensor data) is image information, the feature amount extracted from the partial data is edge direction, pixel value, position, gradation, and the like. When the divided sensor data is sound information, the feature amount extracted from the partial data is the waveform of sound and time, frequency characteristics, spectrum, and the like. When the divided sensor data is exercise information, it is a waveform of acceleration (or velocity, angular velocity, angular acceleration) and time, strength in the direction of gravity, and the like.

Note that the clustering process described in the explanation of the dividing unit 150b is a narrowly defined term of “unsupervised learning in which sensor data is assigned to the nearest neighbor basis and the basis is learned from the assigned sensor data”, which is generally used. It is not a thing, but a broad definition that has the function of ``acquiring a basis represented by a representative value from sensor data, and outputting a value calculated from the obtained basis and sensor data when sensor data is input'' is. Specifically, as clustering techniques, the dividing unit 150b uses algorithms such as the k-means method and the k-meadian method, which are widely used as clustering techniques in the narrow sense, and unsupervised deep neural networks such as the AutoEncoder method and the Deep Boltzmann Machine. A learning algorithm that divides the feature amount space into grids and defines a basis in advance may be used. In addition to the unsupervised learning algorithm, for example, an intermediate layer output of a neural network learned by supervised learning may be extracted and used as a clustering algorithm.

In addition, the histogram described so far is not in the narrow sense of "the frequency of occurrence for each basis" as is generally used, but in the broad sense of "a vector value having a value corresponding to each basis". It is. Specifically, in the histogram according to the present invention, the sum of all values included in the histogram does not have to be 1, and a certain value included in the histogram may take a negative value. Also, the values contained in the histogram may not have a one-to-one correspondence with the basis. For example, a plurality of values may correspond to a certain basis, or a certain value may be calculated from a plurality of basis.

The method of generating an intermediate estimator from clustering results is not limited to the examples described above. For example, the dividing unit 150b may generate an intermediate estimator using the Bag of Feature method, the n-gram model, the Fisher Vector Encoding method, the sparse coding method, the Locality Constrained Linear Coding method, and the VLAD method. Moreover, when a neural network is used as the clustering technique, the intermediate layer output of the neural network may be used as the intermediate estimator.

Furthermore, in general, a post-processing called L1 normalization is added so that the sum of all frequencies becomes 1 so that the histogram can be treated as a probability value. good. For example, the dividing unit 150b may perform L2 normalization and Hellinger normalization processing, or may directly use the result of clustering without performing post-processing.

FIG. 7 is a diagram illustrating an example of an intermediate estimator generated by the dividing unit; The intermediate estimators 20a, 21a, 22a are generated from divided sensor data obtained by dividing the sensor data of the sensor 10a. Time points corresponding to the intermediate estimators 20a, 21a, and 22a are assumed to be time points T-2, T-1, and T, respectively. The times corresponding to the intermediate estimators 20a, 21a, and 22a may be set in any way. and the end time.

The intermediate estimators 20b, 21b, and 22b are generated from divided sensor data obtained by dividing the sensor data of the sensor 10b. Time points corresponding to the intermediate estimators 20b, 21b, and 22b are assumed to be time points T-2, T-1, and T, respectively. The times corresponding to the intermediate estimators 20b, 21b, and 22b may be set in any way. It may be an intermediate time between the time and the end time.

The intermediate estimators 20c, 21c, and 22c are generated from the divided sensor data obtained by dividing the sensor data of the sensor 10b. Time points corresponding to the intermediate estimators 20c, 21c and 22c are assumed to be time points T-2, T-1 and T, respectively. The times corresponding to the intermediate estimators 20c, 21c, and 22c may be set in any way. It may be an intermediate time between the time and the end time.

The division unit 150b outputs information in which the sensor identification information and each intermediate estimator are associated with each other to the evaluation unit 150c. For example, referring to FIG. 7, the dividing unit 150b outputs to the evaluating unit 150c information that associates the sensor identification information "sensor 10a" with the intermediate estimators 20a, 21a, and 22a. The dividing unit 150b outputs information in which the sensor identification information “sensor 10b” is associated with the intermediate estimators 20b, 21b, and 22b to the evaluating unit 150c. Information that associates the sensor identification information “sensor 10c” with the intermediate estimators 20c, 21c, and 22c is output to the evaluation unit 150c.

The evaluation unit 150c is a processing unit that evaluates each combination of intermediate estimators corresponding to each sensor identification information. For example, the evaluation unit 150c performs a process of generating a feature amount vector and a process of learning a weight vector based on each intermediate estimation amount corresponding to each sensor identification information.

An example of processing for generating a feature amount vector by the evaluation unit 150c will be described. FIG. 8 is a diagram for explaining processing for generating a feature amount vector. The evaluation unit 150c arranges the intermediate estimators corresponding to the sensor identification information in chronological order. The evaluation unit 150c calculates a feature amount vector by using, as a vector element, a product obtained by multiplying each intermediate estimation amount between different sensor identification information. For example, a feature amount vector obtained by multiplying each intermediate estimation amount included in a time period with time t as a reference is denoted as φ(t).

Which intermediate estimator is multiplied by which intermediate estimator (cartesian product) is determined as which element of the feature vector is determined according to a preset rule. The evaluation unit 150c may use the direct product of three or more intermediate estimators as elements of the feature amount vector. Also, multilinear representation may be used when performing multiplication of each intermediate estimator.

For example, in the feature amount vector φ(t), each element included in the portion 31 is a direct product of an intermediate estimated amount of certain sensor identification information and an intermediate estimated amount of another sensor identification information. For example, the evaluation unit 150 c places the direct product of the intermediate estimator 20 a and the intermediate estimator 20 b in one of the elements of the portion 31 .

Each element included in portion 32 contains the direct product of the intermediate estimator of the first sensor identity, the intermediate estimator of the second sensor identity, and the intermediate estimator of the third sensor identity. are placed respectively. For example, the evaluation unit 150c places the direct product of the intermediate estimator 20a, the intermediate estimator 20b, and the intermediate estimator 20c in one of the elements of the portion 32. FIG.

The evaluation unit 150c calculates a feature amount vector φ(t) for each different time t. For example, the evaluation unit 150c calculates feature amount vector φ(t ₁ ) at time t ₁ , feature amount vector φ(t ₂ ) at time t ₂ , . . . , feature amount vector φ(t _n ) at time t _n as calculate.

It should be noted that when the evaluation unit 150c performs the above-described multiplication, intermediate estimators generated at different times t may be multiplied. For example, as shown in FIG. 8, the evaluation unit 150c assigns an intermediate estimator 20a and an intermediate estimator 21b generated at different times t−2 and t−1 to one of the elements of the portion 31. Place a Cartesian product.

Also, the multiplication may be performed on all combinations of intermediate estimators, or may be performed on a limited number of combinations. Examples of the limiting method include a method of limiting the time difference in which the intermediate estimator is generated, and a method of limiting the distance of the position where the sensor used to generate the intermediate estimator is installed within a certain value. Alternatively, a method of limiting combinations to predetermined rules may be used. In addition, by using a frequent pattern mining algorithm or an enumeration algorithm, combinations may be limited to combinations in which the number of times the norm of the intermediate estimator takes a value equal to or greater than a certain value is equal to or greater than a certain value.

Furthermore, when the multiplied feature amount φ(t) is used for learning/estimation, φ(t) may be used directly, or post-processing may be added. For example, φ(t) obtained by applying a dimensional reduction algorithm such as principal component analysis may be used for learning/estimation.

An example of processing for learning weight vectors by the evaluation unit 150c will be described. The evaluation unit 150c learns weight vectors using a structural regularization algorithm. Structural regularization algorithms are trained to include more 0s in the weight vector elements. For example, the evaluation unit 150c learns the weight vector w so that the value of the evaluation function L shown in Equation (1) becomes a smaller value.

“l(φ(t), w, y)” included in Equation (1) is the loss function defined by Equation (2). "λ" included in the formula (1) is an adjustment parameter whose value is set in advance. “Ω(w)” is a regularization term that indicates the magnitude of the weight vector w and is defined by Equation (3).

In equation (2), “y” is set to “1” when the target context occurs in the time zone corresponding to time t of the feature amount vector φ(t). If not, "-1" is set. The evaluation unit 150c compares the time t and the correct context DB 140b to determine whether or not the target context has occurred.

As an example, the case where the evaluation unit 150c learns the weight vector w for estimating the context “unloading” will be described. As shown in FIG. 4, the context "unloading" occurs in the time period "time Ts1 to time Te1", so if time t is included in the time period "time Ts1 to time Te1", "y=1", and "y=-1" if not included.

"w" included in formulas (1) and (2) indicates a weight vector. FIG. 9 is a diagram for explaining weight vectors. As shown in FIG. 9, the weight vector w has multiple elements. The number of elements of the weight vector w and the number of elements of the feature amount vector φ(t) match and are associated with each other. For example, each weight of the portion 41 of the weight vector w is a weight corresponding to each element of the portion 31 of the feature quantity vector φ(t). Each weight of the portion 42 of the weight vector w is a weight corresponding to each element of the portion 32 of the feature amount vector φ(t).

The evaluation unit 150c, for each time t (t ₌ t ₁ , t ₂ , . =1 ory =-1) are input into equation (1) to learn the weight vector w. The evaluation unit 150c registers the learned weight vector w in the weight table 140c.

Note that the loss function is not limited to the form shown in Equation (2). As other examples, logistic loss, cross-entropy loss, hinge loss, smoothed hinge loss, and Huber loss, for example, may be used as the loss function. In the above description, context estimation deals with the problem of determining whether or not specific context content is performed, that is, the classification problem. May be used when handling For example, learning is performed using L2 loss and log Poisson loss as a loss function for the problem of estimating a "stress amount" that takes a continuous value as context estimation, as an embodiment of the present invention. Furthermore, the regularization term is not limited to the form shown in equation (3). As other examples, for example, an L0 regularization term, an Lp regularization term, a group L0 regularization term, and a group L1 regularization term may be used. In addition, as a method of defining groups of these regularization terms, multiple overlaps may be allowed. Furthermore, as a regularization term, as in Elastic Net, using the sum of the regularization term described above and multiple different regularization terms including the L2 regularization term and the L∞ regularization term Also good.

When there are a plurality of types of context content, the evaluation unit 150c repeatedly executes the above processing for each context content, calculates a unique weight vector for estimating the context, and registers it in the weight table 140c. For example, in the example shown in FIG. 5, the weight vector for estimating the context content "unloading" is "w1", and the weight vector for estimating the context content "product on display" is "w2". , the weight vector for estimating the context content "packing" is "w3".

In addition, in the explanation so far, an example in which clustering processing is first performed and then weight learning is described, but the learning method shown in the present invention is not limited to this example. For example, a method in which clustering processing and weight learning are alternately and repeatedly learned, or an End- Learning may be performed in a to-end format.

The selection unit 150d is a processing unit that selects a set of sensors to be used when estimating each context based on the weight table 140c. The selection unit 150d registers the selected result in the selected sensor table 140d.

As an example, a process of selecting a set of sensors to be used when the selection unit 150d estimates the context content “unloading” will be described. The selection unit 150d acquires the learned weight vector w1 corresponding to the context content “unloading” from the weight table 140c, and detects the elements that are 0 among the elements of the acquired weight vector.

When the selection unit 150d detects an element that becomes 0, the selection unit 150d adds 1 as a "penalty value" to the sensor corresponding to the detected element. For example, when the weight (element) of the direct product of the intermediate estimator 20a and the intermediate estimator 20b described with reference to FIG. 1 is added as a penalty value to each corresponding sensor 10b.

When the weight (element) of the direct product of the intermediate estimator 20a, the intermediate estimator 20b, and the intermediate estimator 20c is 0, the sensor 10a corresponding to the intermediate estimator 20a and the sensor 10a corresponding to the intermediate estimator 20b 1 is added as a penalty value to each of the sensor 10b and the sensor 10c corresponding to the intermediate estimated amount 20c.

The selection unit 150d calculates the penalty value of each sensor by repeatedly executing the above processing for the elements that are 0. The selecting unit 150d selects a sensor whose penalty value is less than a preset penalty value from among the sensors to be used when estimating the context content “unloading”. The selection unit 150d sorts the sensors in ascending order according to the penalty value, and selects the top M sensors with the lowest penalty values as the sensors to be used when estimating the context content “unloading”. may M is a preset natural number.

The selection unit 150d also selects a set of sensors to be used when estimating other context content in the same manner as in the case of the context content "unloading", and stores the selected information in the selected sensor table 140d. sign up.

In the description so far, the selection unit selects sensors by giving a penalty to the number of sensors to be used, but the embodiment of the present invention is not limited to this method. Specifically, the selection unit estimates the context content of a sensor in which any of the weight elements related to a specific sensor is not 0, or the minimum absolute value of the weight elements is a certain value or more. It may be selected as a sensor. Alternatively, the same sensor may be selected for all context contents without changing the set of sensors used at the time of recognition for each context contents. As a specific method thereof, for example, for sensors previously selected for each context content, a union of the sensors may be selected as the sensor for estimating the context content. Furthermore, the selected sensor may be fixed, re-learning may be performed for a different adjustment parameter λ or a different regularization term, and the weight w may be updated.

The notification unit 150e is a processing unit that notifies the weight table 140c and the selected sensor table 140d to an estimation device, which will be described later.

Next, an example of the processing procedure of the analyzer 100 according to this embodiment will be described. FIG. 10 is a flow chart showing the processing procedure of the analyzer according to the present embodiment. As shown in FIG. 10, the receiving unit 150a of the analysis device 100 receives time-series sensor data from each of the sensors 10a to 10c (other sensors) and registers them in the sensor DB 140a (step S101). The receiving unit 150a receives the correct context data and registers it in the correct context DB 140b (step S102).

The dividing unit 150b of the analysis device 100 divides each sensor data into predetermined time intervals to generate a plurality of divided sensor data (Step S103). The dividing unit 150b generates a plurality of intermediate estimators based on the plurality of divided sensor data (step S104).

The evaluation unit 150c of the analysis device 100 arranges each intermediate estimator in chronological order for each sensor identification information (Step S105). The evaluation unit 150c calculates a feature amount vector by multiplying intermediate estimation amounts between different sensor identification information (step S106). The evaluation unit 150c learns each element of the weight vector so that the value of the evaluation function L becomes smaller, and registers the learned weight vector in the weight table 140c (step S107).

The selection unit 150d of the analysis device 100 selects sensors to be used when estimating the context based on the weight table 140c (step S108). The selection unit 150d updates the selected sensor table 140d based on the selection result (step S109). The notification unit 150e of the analysis device 100 notifies the estimation device of the weight table 140c and the selected sensor table 140d (step S110).

Next, an estimation device that estimates content based on the weight table 140c and the selected sensor table 140d notified from the analysis device 100 will be described.

FIG. 11 is a diagram showing the configuration of the system during operation according to this embodiment. As shown in FIG. 11, this system has sensors 10a, 10b, 10c and an estimating device 200. As shown in FIG. Sensors 10 a , 10 b , 10 c and estimation device 200 are interconnected via network 50 . For convenience, only the sensors 10a to 10c are shown here, but the estimating device 200 is assumed to be further connected to other sensors. It is assumed that each sensor connected to the analysis device 100 shown in FIG. 1 and each sensor connected to the estimation device 200 are the same sensor.

The sensors 10a to 10c are any of various sensors such as image sensors, sound sensors, and motion sensors. The description of the sensors 10a-10c is the same as the description of the sensors 10a-10c described with reference to FIG.

Upon receiving the specification of the context to be estimated, the estimation device 200 identifies a set of sensors to be used based on the selected sensor table 140d, and transmits a transmission command to the identified sensors. The estimating apparatus 200 receives sensor data from the sensor that has transmitted the transmission command, and uses the weights in the weight table 140c to determine whether or not the context to be estimated is occurring.

FIG. 12 is a functional block diagram showing the configuration of the estimation device according to this embodiment. As shown in FIG. 12 , this estimation device 200 has a communication section 210 , an input section 220 , a display section 230 , a storage section 240 and a control section 250 .

The communication unit 210 is a processing unit that receives sensor data from the selected sensor via the network 50 . Communication unit 110 outputs the sensor data to control unit 150 . The communication unit 210 corresponds to a communication device. Also, the communication unit 210 transmits and receives data to and from the analysis device 100 . A control unit 250 , which will be described later, exchanges data with each sensor and the analysis device 100 via the communication unit 210 .

The input unit 220 is an input device for inputting various kinds of information to the estimation device 200 . For example, input unit 220 corresponds to a keyboard, mouse, touch panel, and the like.

Display unit 230 is a display device for displaying information output from control unit 250 . For example, display unit 230 corresponds to a liquid crystal display, an organic display, a touch panel, or the like.

The storage unit 240 has a sensor DB 240a, a weight table 140c, and a selected sensor table 140d. The storage unit 240 corresponds to semiconductor memory elements such as RAM, ROM, and flash memory, and storage devices such as HDD.

The sensor DB 240a is a DB that holds sensor data measured by the sensors 10a to 10c and other sensors (not shown). The data structure of the sensor DB 240a is the same as the data structure of the sensor DB 140a described in FIG. Sensor data of sensors selected by the communication control unit 250b, which will be described later, is registered in the sensor DB 240a.

The weight table 140c is a table that holds weight vector information for estimating the context. The weight table 140c holds information on weight vectors learned by the analysis device 100 . The data structure of the weight table 140c is the same as the data structure of the weight table 140c shown in FIG.

The selected sensor table 140d is a table that holds sensor information used for estimating the context. The data structure of the selected sensor table 140d is the same as the data structure of the selected sensor table 140d shown in FIG.

The control unit 250 has a receiving unit 250a, a communication control unit 250b, a dividing unit 250c, and a score calculating unit 250d. The control unit 250 can be implemented by a CPU, MPU, or the like. Also, the control unit 250 can be realized by hardwired logic such as ASIC and FPGA.

The receiving unit 250a is a processing unit that receives sensor data from each sensor selected by the communication control unit 250b. The receiving unit 250a stores each sensor data in the sensor DB 240a. The receiving unit 250a refers to the sensor identification information included in the header of each sensor data, associates the sensor data with the corresponding sensor identification information, and stores the sensor data in the sensor DB 240a.

The communication control unit 250b is a processing unit that selects the sensor used when estimating the context based on the selected sensor table 140d. The communication control unit 250b transmits a transmission request to the selected sensor. In addition, the communication control unit 250b may transmit a transmission stop request to sensors other than the selected sensor to suppress transmission of sensor data.

For example, the communication control unit 250b receives specification of context content to be estimated from the input unit 220 or the like. The communication control unit 250b compares the designated context content with the selected sensor table, and identifies the sensor used when estimating the context.

The dividing unit 250c is a processing unit that divides sensor data and generates an intermediate estimator based on the divided data. The process of generating the intermediate estimator by the dividing unit 250c is the same as the process of generating the intermediate estimator by the dividing unit 150b of the analysis apparatus 100. FIG. The division unit 250c outputs information in which the sensor identification information and each intermediate estimator are associated with each other to the score calculation unit 250d.

The score calculation unit 250d is a processing unit that calculates a score that indicates the likelihood that a context is occurring. For example, the score calculation unit 250d performs a process of generating a feature amount vector and a process of calculating a score.

The process of generating the feature amount vector by the score calculation unit 250d is the same as the process of generating the feature amount vector by the evaluation unit 150c of the analysis device 100 . For example, the score calculation unit 250d calculates a feature amount vector by using, as a vector element, a product obtained by multiplying each intermediate estimation amount between different sensor identification information.

Processing for calculating the score by the score calculation unit 250d will be described. For example, score calculator 250d calculates score S(t) based on equation (4). In equation (4), "w" is the learned weight vector. The score calculation unit 250d compares the context content specified by the input unit 220 or the like with the weight table 140c, and uses the weight vector corresponding to the specified context content. For example, when the specified context content is "unloading", the score calculation unit 250d substitutes the weight vector w1 into Equation (4). φ(t) in Equation (4) is a feature vector.

The score calculation unit 250d may output and display the score calculation result on the display unit 230, or may notify an external device connected via the network 50 of information on the score calculation result. . When the probability that the context to be estimated occurs is high, the score becomes a large value, and when the probability is low, the score becomes a small value. Furthermore, the score calculation unit 250d may generate a graph showing the score for each time and display it on the display unit 230. FIG.

Next, an example of the processing procedure of the estimation device 200 according to this embodiment will be described. FIG. 13 is a flow chart showing the processing procedure of the estimation device according to the present embodiment. As shown in FIG. 13, the communication control unit 250b of the estimation device 200 receives specification of context content (step S201). The communication control unit 250b selects a sensor based on the selected sensor table 140d and transmits a transmission request to the selected sensor (step S202).

The receiving unit 250a of the estimating device 200 receives the sensor data from the selected sensor and registers it in the sensor DB 240a (step S203). The dividing unit 250c of the estimating device 200 divides each sensor data into predetermined time intervals to generate a plurality of divided sensor data (step S204).

The dividing unit 250c generates a plurality of intermediate estimators based on the plurality of divided sensor data (step S205). The score calculation unit 250d of the estimation device 200 arranges the intermediate estimators in chronological order for each sensor identification information (step S206).

The score calculation unit 250d calculates a feature amount vector by multiplying intermediate estimation amounts between different sensor identification information (Step S207). The score calculator 250d acquires the weight vector corresponding to the estimated context content from the weight table 140c (step S208).

The score calculation unit 250d calculates the score of the probability that the context is generated based on the weight vector and the feature amount vector (step S209). The score calculation unit 250d outputs the score calculation result to the display unit 230 (step S210).

Next, the effect of the analyzer 100 according to this embodiment will be described. The analysis device 100 receives sensor data from each sensor and divides each sensor data into a plurality of time intervals to generate a plurality of divided sensor data. Analysis device 100 evaluates a combination of each divided sensor data of a certain sensor and each divided sensor data of another sensor among a plurality of sensors, and based on the evaluation result, selects sensors to be combined. select a pair. This allows the selection of a useful set of sensors when estimating context.

Analysis device 100 generates a histogram obtained by performing clustering based on the feature amount included in each part of the divided sensor data as an intermediate estimator of the divided sensor data, and combines the generated intermediate estimators. to be evaluated. By converting into an intermediate estimator in this way, evaluation can be performed by combining the measurement results of different types of sensors (image sensor, sound sensor, motion sensor, etc.).

The analysis device 100 generates a feature amount vector whose elements are values obtained by multiplying each intermediate estimator of a certain sensor by each intermediate estimator of another sensor. The analysis device 100 learns the weight vector so that the value obtained by multiplying and adding the feature amount vector and the weight vector approaches the correct value of the context. Analysis device 100 identifies a set of sensors to be combined based on each element of the learned weight vector. This allows identifying a useful set of sensors for estimating context.

Also, in the operation phase, the estimation device 200 according to this embodiment transmits a transmission command to the sensors used when estimating the context, based on the selected sensor table 140d generated by the analysis device 100 . By performing such processing, it is possible to receive sensor data from the sensor that has transmitted the transmission command among the plurality of sensors, and to estimate the context, thereby reducing the amount of data communication.

By the way, in this embodiment, as an example, the case where the analysis device 100 and the estimation device 200 are implemented in separate devices has been described, but the present invention is not limited to this. For example, analysis device 100 may have the configuration of estimation device 200 . That is, the control unit 150 of the analysis device 100 shown in FIG. 2 may have the reception unit 250a, the communication control unit 250b, the division unit 250c, and the score calculation unit 250d of the estimation device 200 shown in FIG.

Further, the system configuration during learning shown in FIG. 1 is an example, and the system configuration shown in FIG. 14 may be used. FIG. 14 is a diagram showing the configuration of another system during learning according to this embodiment. As shown in FIG. 14, this system has sensors 10a-10d, a gateway 60, and an analysis device 100. As shown in FIG.

Sensors 10 a - 10 d are connected to gateway 60 . Although not shown in FIG. 14, it is assumed that the gateway 60 is connected to other sensors. Gateway 60 is connected to analysis device 100 via network 50 .

The sensors 10a-10d are any of various sensors such as image sensors, sound sensors, motion sensors, and the like. Sensors 10 a - 10 d transmit sensor data to analysis device 100 via gateway 60 .

The gateway 60 is a device that relays data communication between each sensor and the analysis device 100 . Sensors 10a to 10d (and other sensors) are connected to each port of the gateway 60, and connection and disconnection of communication can be controlled for each port. It is assumed that a management table indicating which sensor is connected to which port is preset in the gateway 60 . In the learning phase, the gateway 60 connects all ports and transmits sensor data measured by each sensor to the analysis device 100 .

The analysis device 100 is a device that receives sensor data of each sensor transmitted from the gateway 60 and selects a useful combination of sensors when estimating a predetermined context using each sensor data. The rest of the description is the same as the description regarding the analysis device 100 described above.

Next, the system configuration during operation shown in FIG. 11 is an example, and the system configuration shown in FIG. 15 may be used. FIG. 15 is a diagram showing the configuration of another system during operation according to this embodiment. As shown in FIG. 15, this system has sensors 10a-10d, a gateway 60, and an estimating device 200. As shown in FIG.

Sensors 10 a - 10 d are connected to gateway 60 . Although not shown in FIG. 15, it is assumed that the gateway 60 is connected to other sensors. Gateway 60 is connected to estimating device 200 via network 50 .

Gateway 60 corresponds to gateway 60 shown in FIG. In the operation phase, the gateway 60 receives control information including sensor identification information that permits communication and sensor identification information that does not permit communication from the estimation device 200, and based on the received control information, connects ports. , to control disconnection. For example, when the gateway 60 receives control information indicating that the communication of the sensors 10a to 10c is permitted and the communication of the sensor 10d is not permitted, the gateway 60 connects the ports connected to the sensors 10a to 10c, and connects the ports connected to the sensors 10a to 10c. Controls disconnection of connected ports. As a result, the sensor data of the sensors 10 a to 10 c are transmitted to the estimating device 200 and the sensor data of the sensor 10 d are not transmitted to the estimating device 200 .

Upon receiving the specification of the context to be estimated, the estimating device 200 identifies a set of sensors to be used based on the selected sensor table 140d, and transmits control information to the gateway 60 for the identified sensors. The estimator 200 receives sensor data from the selected sensors (via gateway 60) and uses the weights in weight table 140c to determine whether the context to be estimated is occurring. The rest of the description is the same as the description regarding estimation device 200 described above.

Next, an example of the hardware configuration of a computer that implements the same functions as the analysis device 100 and the estimation device 200 shown in the embodiment will be described. FIG. 16 is a diagram showing an example of the hardware configuration of a computer that implements the same functions as the analysis apparatus according to this embodiment.

As shown in FIG. 16, a computer 500 has a CPU 501 that executes various arithmetic processes, an input device 502 that receives data input from a user, and a display 503 . The computer 500 also has a reading device 504 that reads programs and the like from a storage medium, and an interface device 505 that exchanges data with an external device or the like via a wired or wireless network. The computer 500 has a RAM 506 that temporarily stores various information and a hard disk device 507 . Each device 501 - 507 is then connected to a bus 508 .

The hard disk device 507 has a reception program 507a, a division program 507b, an evaluation program 507c, a selection program 507d, and a communication program 507e. The CPU 501 reads the reception program 507a, the division program 507b, the evaluation program 507c, the selection program 507d, and the communication program 507e and develops them in the RAM 506. FIG.

The receiving program 507a functions as a receiving process 506a. Splitting program 507b functions as splitting process 506b. Evaluation program 507c functions as evaluation process 506c. Selection program 507d functions as selection process 506d. Communication program 507e functions as communication process 506e.

The processing of the reception process 506a corresponds to the processing of the reception unit 150a. The processing of the splitting process 506b corresponds to the processing of the splitting unit 150b. Processing of the evaluation process 506c corresponds to processing of the evaluation unit 150c. The processing of the selection process 506d corresponds to the processing of the selection unit 150d. The processing of the communication process 506e corresponds to the processing of the notification unit 150e.

Note that the programs 507a to 507e do not necessarily have to be stored in the hard disk device 507 from the beginning. For example, each program is stored in a “portable physical medium” such as a flexible disk (FD), CD-ROM, DVD disk, magneto-optical disk, IC card, etc. inserted into the computer 500 . Then, the computer 500 may read out and execute each of the programs 507a to 507e.

FIG. 17 is a diagram illustrating an example of the hardware configuration of a computer that implements the same functions as the estimation device according to this embodiment.

As shown in FIG. 17, a computer 600 has a CPU 601 that executes various arithmetic processes, an input device 602 that receives data input from a user, and a display 603 . The computer 600 also has a reading device 604 that reads programs and the like from a storage medium, and an interface device 605 that exchanges data with an external device or the like via a wired or wireless network. The computer 600 has a RAM 606 that temporarily stores various information and a hard disk device 607 . Each device 601 - 607 is then connected to a bus 608 .

The hard disk device 607 has a reception program 607a, a communication control program 607b, a division program 607c, and a score calculation program 607d. The CPU 601 reads the reception program 607a, the communication control program 607b, the division program 607c, and the score calculation program 607d, and develops them in the RAM 606. FIG.

The receiving program 607a functions as a receiving process 606a. The communication control program 607b functions as a communication control process 606b. The splitting program 607c functions as a splitting process 606c. The score calculation program 607d functions as a score calculation process 606d.

The processing of the reception process 606a corresponds to the processing of the reception unit 250a. Processing of the communication control process 606b corresponds to processing of the communication control unit 250b. The processing of the splitting process 606c corresponds to the processing of the splitting unit 250c. Processing of the score calculation process 606d corresponds to processing of the score calculation unit 250d.

Note that the programs 607a to 607d do not necessarily have to be stored in the hard disk device 607 from the beginning. For example, each program is stored in a “portable physical medium” such as a flexible disk (FD), CD-ROM, DVD disk, magneto-optical disk, IC card, etc. inserted into the computer 600 . Then, the computer 600 may read and execute each program 607a to 607e.

The following additional remarks are disclosed regarding the embodiments including the above examples.

(Appendix 1) A division unit that acquires time-series information measured by a plurality of sensors and divides the time-series information into a plurality of time intervals to generate a plurality of measured values;
an evaluation unit that evaluates a combination of a plurality of measured values corresponding to a first sensor and a plurality of measured values corresponding to a second sensor among the plurality of sensors;
and a selection unit that selects a set of sensors to be combined based on the evaluation result of the evaluation unit.

(Appendix 2) The dividing unit divides the measured value and generates a histogram obtained by performing clustering based on the feature amount included in each divided portion as an intermediate estimate of the measured value, Supplementary note 1, wherein the evaluation unit evaluates a combination of a plurality of first intermediate estimators corresponding to the first sensor and a plurality of second intermediate estimators corresponding to the second sensor. analyzer.

(Supplementary Note 3) The evaluation unit generates a feature vector whose elements are values obtained by multiplying the plurality of first intermediate estimators by the plurality of second intermediate estimators, and the feature vector and Each element of the weight vector is learned so that a value obtained by multiplying and adding the weight vector approaches a correct value of the context corresponding to the time-series information, and the selection unit performs the learning of the weight vector. The analysis device according to appendix 2, wherein a set of sensors to be combined is selected based on each element of .

(Appendix 4) The evaluation unit arranges the plurality of first intermediate estimators and the plurality of second intermediate estimators in time series, and arranges the plurality of first intermediate estimators included in a predetermined time range, and the plurality of The analysis device according to appendix 3, wherein the evaluation is performed for a combination with the second intermediate estimator.

(Appendix 5) Receive time-series information measured by the plurality of sensors from the communication device, notify the communication device of the selection result of the selection unit, and time-series measured by sensors that are not subject to combination 1. The analysis apparatus according to appendix 1, further comprising a communication control unit that suppresses receiving information of the above.

(Appendix 6) The computer
Acquiring time-series information measured by a plurality of sensors and dividing the time-series information into a plurality of time intervals to generate a plurality of measured values;
Evaluating a combination of a plurality of measured values corresponding to the first sensor and a plurality of measured values corresponding to the second sensor among the plurality of sensors,
An analysis method characterized by executing a process of selecting a set of sensors to be combined based on an evaluation result.

(Appendix 7) The generating process divides the measured value and generates a histogram obtained by performing clustering based on the feature amount included in each divided portion as an intermediate estimate of the measured value. Supplementary note 6, wherein the evaluating process evaluates a combination of a plurality of first intermediate estimators corresponding to the first sensors and a plurality of second intermediate estimators corresponding to the second sensors. Analysis method described in.

(Supplementary Note 8) The process of evaluating generates a feature amount vector whose elements are values obtained by multiplying the plurality of first intermediate estimators by the plurality of second intermediate estimators, and The process of learning and selecting each element of the weight vector so that the value obtained by multiplying and adding the weight vector is close to the correct value of the context corresponding to the time-series information. The analysis method according to appendix 7, wherein a set of sensors to be combined is selected based on each element of the weight vector.

(Supplementary note 9) The process of evaluating includes arranging the plurality of first intermediate estimators and the plurality of second intermediate estimators in time series, the plurality of first intermediate estimators included in a predetermined time range, and the plurality of The analysis method according to appendix 8, wherein the combination with the second intermediate estimator of is evaluated.

(Appendix 10) Receive time-series information measured by the plurality of sensors from the communication device, notify the communication device of the selection result of the selection unit, and time-series measured by sensors that are not subject to combination 7. The analysis method according to appendix 6, further comprising executing a process of suppressing reception of the information of .

(Appendix 11) Acquiring time-series information measured by a plurality of sensors, and dividing the time-series information into a plurality of time intervals to generate a plurality of measured values,
Evaluating a combination of a plurality of measured values corresponding to the first sensor and a plurality of measured values corresponding to the second sensor among the plurality of sensors,
An analysis program characterized by causing a computer to execute a process of selecting a set of sensors to be combined based on an evaluation result.

(Appendix 12) In the generating process, a histogram obtained by dividing the measured value and performing clustering based on the feature amount included in each divided portion is generated as an intermediate estimate of the measured value. Supplementary note 11, wherein the evaluating process evaluates a combination of a plurality of first intermediate estimators corresponding to the first sensors and a plurality of second intermediate estimators corresponding to the second sensors. analysis program described in .

(Supplementary Note 13) In the evaluating process, a feature amount vector whose elements are values obtained by multiplying the plurality of first intermediate estimators by the plurality of second intermediate estimators is generated, and the feature amount vector The process of learning and selecting each element of the weight vector so that the value obtained by multiplying and adding the weight vector is close to the correct value of the context corresponding to the time-series information. 13. The analysis program according to appendix 12, wherein a set of sensors to be combined is selected based on each element of the weight vector.

(Supplementary Note 14) The process of evaluating includes arranging the plurality of first intermediate estimators and the plurality of second intermediate estimators in time series, the plurality of first intermediate estimators included in a predetermined time range, and the plurality of 14. The analysis program according to appendix 13, characterized in that the combination with the second intermediate estimator of is evaluated.

(Appendix 15) Receive time-series information measured by the plurality of sensors from the communication device, notify the communication device of the selection result of the selection unit, and time-series measured by sensors not to be combined 12. The analysis program according to appendix 11, further executing a process of suppressing reception of the information of .

100 Analyzer 110, 210 Communication Unit 120, 220 Input Unit 130, 230 Display Unit 140, 240 Storage Unit 140a, 240a Sensor DB
140b Correct Context DB
140c weight table 140d selection sensor table 150, 250 control unit 150a, 250a reception unit 150b, 250c division unit 150c evaluation unit 150d selection unit 150e notification unit 200 estimation device 250b communication control unit 250d score calculation unit

Claims (7)

Acquiring time-series information measured by a plurality of sensors, dividing the time-series information into a plurality of time intervals to generate a plurality of measured values, dividing the measured values, and including in each divided part A dividing unit that generates statistical information obtained by performing clustering based on the feature amount obtained as an intermediate estimator of the measured value ;
an evaluation unit that evaluates a combination of a plurality of first intermediate estimators corresponding to the first sensor and a plurality of second intermediate estimators corresponding to the second sensor among the plurality of sensors;
and a selection unit that selects a set of sensors to be combined based on the evaluation result of the evaluation unit. The dividing unit divides the measured value and generates a histogram obtained by performing clustering based on the feature amount included in each divided portion as an intermediate estimator of the measured value, and the evaluation unit , a combination of a plurality of first intermediate estimators corresponding to the first sensors and a plurality of second intermediate estimators corresponding to the second sensors. . The evaluation unit generates a feature vector whose elements are values obtained by multiplying the plurality of first intermediate estimators by the plurality of second intermediate estimators, and combines the feature vector and the weight vector. Each element of the weight vector is learned such that the multiplied and added value approaches the correct value of the context corresponding to the time-series information, and the selection unit selects each element of the learned weight vector. 3. The analysis apparatus according to claim 2, wherein a set of sensors to be combined is selected based on the above. The evaluation unit arranges the plurality of first intermediate estimators and the plurality of second intermediate estimators in time series, and arranges the plurality of first intermediate estimators and the plurality of second intermediate estimators included in a predetermined time range. 4. The analysis device according to claim 3, wherein evaluation is performed for a combination with quantity. Receive time-series information measured by the plurality of sensors from a communication device, notify the communication device of a selection result of the selection unit, and receive time-series information measured by sensors that are not to be combined. 2. The analysis apparatus according to claim 1, further comprising a communication control unit that prevents the analysis. the computer
Acquiring time-series information measured by a plurality of sensors and dividing the time-series information into a plurality of time intervals to generate a plurality of measured values;
dividing the measurements;
Generate statistical information obtained by performing clustering based on the feature amount contained in each divided portion as an intermediate estimator of the measured value,
Of the plurality of sensors, evaluating a combination of a plurality of first intermediate estimators corresponding to the first sensor and a plurality of second intermediate estimators corresponding to the second sensor,
An analysis method characterized by executing a process of selecting a set of sensors to be combined based on an evaluation result. Acquiring time-series information measured by a plurality of sensors and dividing the time-series information into a plurality of time intervals to generate a plurality of measured values;
dividing the measurements;
Generate statistical information obtained by performing clustering based on the feature amount contained in each divided portion as an intermediate estimator of the measured value,
Of the plurality of sensors, evaluating a combination of a plurality of first intermediate estimators corresponding to the first sensor and a plurality of second intermediate estimators corresponding to the second sensor,
An analysis program characterized by causing a computer to execute a process of selecting a set of sensors to be combined based on an evaluation result.

Images