JP6593519B1 - Sensor data compression system, sensor terminal, model construction device, and program - Google Patents

Abstract

To provide a sensor data compression system, a sensor terminal, a model construction device, and a program that achieve both maintenance of data analysis accuracy and reduction in communication cost. In a sensor data compression system, a sensor terminal compresses sensor data using a sensor that acquires sensor data and a mathematical model that acquires a compression ratio related to compression sensing based on a feature amount of the acquired sensor data. A data compression unit that compresses by ratio, a data communication unit that transmits sensor data compressed by the data compression unit, and a degree of divergence that detects the degree of divergence between the acquired sensor data and the sensor data used to construct the mathematical model A detection unit. When the detected degree of divergence exceeds a threshold, the data communication unit reconstructs the mathematical model, and a model transmission unit that transmits the mathematical model reconstructed by the compression ratio model construction unit to the sensor terminal . [Selection] Figure 4

Description

  The present invention relates to a sensor data compression system, a sensor terminal, a model construction device, and a program.

  2. Description of the Related Art In recent years, techniques for determining various states related to a detection target by acquiring sensor data with a sensor terminal installed on the detection target and analyzing the sensor data with a server have become widespread. In addition, a method has been proposed for reducing communication costs by performing data compression when transmitting sensor data from a sensor terminal to a server.

  For example, Patent Document 1 discloses data restored from compressed data by an information processing apparatus provided separately from the sensor terminal in order to dynamically optimize the compression encoding of sensor data transmitted from the sensor terminal to the server. A technique for generating compressed encoded information based on the statistical information and using the compressed encoded information in a sensor terminal is disclosed.

Japanese Patent Laid-Open No. 2006-63297

D.L.Donoho, “Compressed sensing”, IEEE Transactions on Information Theory, vol.52, no.4, pp.1289--1306, Apr. 2006. Timothy P. Lillicrap, 7 others, “Continuous control with deep reinforcement learning”, September 5, 2015, [Online], [October 15, 2018 search], Internet <https://arxiv.org/pdf /1509.02971.pdf>

  However, in Patent Document 1, the optimization of the compression ratio is not described, and the optimization of the compression ratio using a mathematical model or the optimal compression ratio is dynamically performed on the sensor terminal side having a low calculation capability. There is no discussion about how to set.

  Therefore, the present invention has been made in view of the above, and an object of the present invention is a new and improved sensor capable of both maintaining data analysis accuracy and reducing communication costs. The object is to provide a data compression system, a sensor terminal, a model construction device, and a program. Here, the data analysis accuracy is the degree to which a data analysis result with less error can be obtained without reducing the amount of information necessary for the analysis according to the purpose.

  In order to solve the above-described problem, according to an aspect of the present invention, a sensor that acquires sensor data, and a mathematical model that acquires a compression ratio related to compression sensing based on a feature amount of the acquired sensor data are used. A data compression unit that compresses the sensor data at the compression ratio, a data communication unit that transmits the sensor data compressed by the data compression unit, and the acquired sensor data and the mathematical model. A divergence degree detection unit that detects a divergence degree from the sensor data, and the data communication unit models the sensor data that is not compressed by the compression ratio when the detected divergence degree exceeds a threshold value. Using a sensor terminal that is transmitted to the construction device and the sensor data that is not compressed by the compression ratio received from the sensor terminal A model construction apparatus comprising: a compression ratio model construction unit for reconstructing a mathematical model; and a model transmission unit for transmitting the mathematical model reconstructed by the compression ratio model construction unit to the sensor terminal. A sensor data compression system is provided.

  The mathematical model is a compression ratio learning model constructed by using a machine learning method so as to dynamically change the compression ratio for the sensor data according to an evaluation index including a restoration error and the compression ratio. The compression ratio model construction unit reconstructs the compression ratio learning model by using the sensor data that has not been compressed by the compression ratio received from the sensor terminal, and the model transmission unit performs the compression The compression ratio learning model reconstructed by the ratio model construction unit may be transmitted to the sensor terminal.

  Further, the divergence degree detection unit may detect the divergence degree using a divergence degree detection learning model constructed using the sensor data used for constructing the compression ratio learning model.

  The model construction device further includes a deviation degree detection model construction unit that constructs the deviation degree detection learning model using the sensor data that is not compressed by the compression ratio received from the sensor terminal, The model transmitting unit may transmit the deviation degree detection learning model constructed by the deviation degree detection model construction unit to the sensor terminal.

  In addition, the compression ratio model construction unit may construct the compression ratio learning model using any one of regression learning, unsupervised learning, or reinforcement learning, which is one of supervised learning.

  Further, the compression ratio model construction unit obtains the compression ratio for optimizing the evaluation index, and performs the compression ratio learning by regression learning using the sensor data as explanatory variables and the optimum compression ratio for the sensor data as an objective variable. You may build a model.

  Further, the compression ratio model construction unit may obtain the compression ratio for optimizing the evaluation index by any one of a random search, a grid search, and a hill climbing method.

  Further, the compression ratio model construction unit uses an unsupervised learning to obtain the compression ratio to be optimized by updating the parameters of the neural network by the error back propagation method using a loss function in which the loss becomes smaller as the data compression efficiency is higher. The compression ratio learning model may be constructed.

  Further, the compression ratio model construction unit may construct the compression ratio learning model by reinforcement learning using the sensor data as a state, the compression ratio as an action, and data compression efficiency as a reward.

  The compression ratio model construction unit may perform reinforcement learning using time-domain data or frequency-domain data related to the sensor data as a state.

  The compression ratio model construction unit may divide the range of the compression ratio by a natural number N, and perform reinforcement learning using the compression ratio of each representative value as a discrete value action.

  Further, the compression ratio model construction unit may perform reinforcement learning using the compression ratio as a continuous value action.

  In addition, the compression ratio model construction unit may perform reinforcement learning using an evaluation index including the compression ratio or a restoration error as a reward.

  In order to solve the above problems, according to another aspect of the present invention, a sensor that acquires sensor data, and a mathematical function that acquires a compression ratio related to compression sensing based on a feature amount of the acquired sensor data. Using the model, a data compression unit that compresses the sensor data at the compression ratio, a data communication unit that transmits the sensor data compressed by the data compression unit, the acquired sensor data, and the mathematical model A divergence degree detecting unit that detects a divergence degree from the sensor data used for the construction, and the data communication unit is compressed by the compression ratio when the detected divergence degree exceeds a threshold value. There is provided a sensor terminal that transmits the non-sensor data to the model construction device and receives the mathematical model reconstructed using the sensor data.

  In order to solve the above problem, according to another aspect of the present invention, the sensor data is compressed using a mathematical model that obtains a compression ratio related to compression sensing based on a feature amount of the obtained sensor data. A compression ratio model constructing unit for reconstructing the mathematical model using the sensor data when the sensor data not compressed by the compression ratio is received from a sensor terminal that compresses by the ratio; and the compression ratio model The mathematical model reconstructed by the construction unit is transmitted to a sensor terminal that has transmitted at least the sensor data not compressed by the compression ratio among a plurality of sensor terminals connected via a network. A model construction device comprising a model transmission unit is provided.

  In order to solve the above problem, according to another aspect of the present invention, a computer is provided with a sensor for acquiring sensor data, and a compression ratio related to compression sensing based on a feature amount of the acquired sensor data. Using a mathematical model to be acquired, a data compression unit that compresses the sensor data at the compression ratio, a data communication unit that transmits the sensor data compressed by the data compression unit, the acquired sensor data, and the A divergence degree detecting unit that detects a divergence degree from the sensor data used for constructing the mathematical model, and when the detected divergence degree exceeds a threshold, the data communication unit performs compression by the compression ratio. A sensor terminal that transmits the sensor data that has not been performed to the model construction device and receives the mathematical model reconstructed using the sensor data. Program for operating is provided.

  In order to solve the above-described problem, according to another aspect of the present invention, the sensor data is obtained using a mathematical model that obtains a compression ratio related to compression sensing based on a feature amount of the obtained sensor data. A compression ratio model constructing unit that reconstructs the mathematical model using the sensor data when the sensor data that has not been compressed by the compression ratio is received from the sensor terminal that compresses the compression ratio at the compression ratio; The sensor terminal that has transmitted the sensor data that is not compressed by at least the compression ratio among the plurality of sensor terminals connected to the mathematical model reconstructed by the compression ratio model construction unit via a network There is provided a program for functioning as a model construction device, comprising:

  As described above, according to the present invention, it is possible to maintain both data analysis accuracy and reduce communication costs.

It is a figure which shows the assumption environment where the sensor data compression system which concerns on one Embodiment of this invention can be applied. It is a block diagram showing an example of composition of a sensor data compression system concerning the embodiment. It is a block diagram which shows the structural example of the sensor terminal for learning which concerns on the same embodiment. It is a block diagram which shows the structural example of the sensor terminal which concerns on the same embodiment. It is a block diagram which shows the structural example of the data receiver which concerns on the same embodiment. It is a flowchart which shows the flow of a process of the sensor terminal for learning which concerns on the embodiment. It is a figure which shows the outline | summary of the compression ratio learning model construction by the regression learning which concerns on the same embodiment. It is an example of the sensor data to be compressed according to the embodiment. FIG. 9 is an enlarged view of sensor data in a period A in FIG. 8 according to the embodiment. FIG. 9 is an enlarged view of sensor data in a period B of FIG. 8 according to the embodiment. It is a figure which shows the relationship between the compression ratio which concerns on each sensor data in the period A of FIG. 8, and the period B of FIG. 8 which concern on the same embodiment, and the data size after compression. It is a figure which shows the relationship between the compression ratio which concerns on each sensor data in the period A of FIG. 8, and the period B of FIG. 8 which concern on the embodiment. It is a figure which shows the relationship between the compression ratio and data compression efficiency which concern on each sensor data in the period A and the period B of FIG. 8 which concern on the embodiment. FIG. 9 is a diagram showing a comparison between a calculated value of an optimal compression ratio for the sensor data shown in FIG. 8 according to the embodiment and an optimal compression ratio estimated by a compression ratio learning model constructed by regression learning. It is a figure which shows the outline | summary of the compression ratio learning model construction by the unsupervised learning which concerns on the same embodiment. It is a figure which shows the outline | summary of the compression ratio learning model construction by the reinforcement learning which concerns on the same embodiment. It is a sequence diagram which shows the flow of the process in the data compression and decompression | restoration stage which concerns on the embodiment. It is a block diagram which shows the hardware structural example of the model construction apparatus which concerns on one Embodiment of this invention. It is a figure for demonstrating the outline | summary of the data compression using compression sensing.

  Exemplary embodiments of the present invention will be described below in detail with reference to the accompanying drawings. In the present specification and drawings, components having substantially the same functional configuration are denoted by the same reference numerals, and redundant description is omitted.

<1. One Embodiment>
<< 1.1. Overview >>
First, the outline of the present invention will be described. The present invention relates to vibration data, video data, audio data, etc. in which the degree of sparsity (characteristic that there are many 0 components), which is an index indicating the amount of information of sensor data (hereinafter sometimes referred to simply as data), varies. This is a technique for efficiently compressing the broadband data at the edge-side sensor terminal and transmitting the data through a low-band link such as a radio.

  Here, the degree of sparsity refers to the minimum data size required to reproduce the original data with respect to the original data size. In general, even for data having the same data size, the sparseness is higher as the amount of information included in the data is larger, and the sparseness is lower as the information amount included in the data is smaller.

  When sensor data is collected and data mining is performed, it is generally necessary to increase the amount of sensor data to be collected in order to realize highly accurate data mining. However, an edge terminal that transmits sensor data, that is, a sensor terminal generally has large restrictions on memory capacity, storage capacity, calculation capacity, and the like. In particular, when sensor data acquired by a sensor terminal is sent to the analysis server via a wireless link, a large restriction occurs on the network bandwidth. In this way, there is a trade-off relationship between the amount of data information and the waste of resources, so the data transmission amount is reduced while maintaining the data information amount by compressing the data efficiently on the sensor terminal side. It is required to do.

  Here, as a technique for effectively performing data compression, for example, a technique disclosed in Patent Document 1 can be cited. However, as described above, in the technique disclosed in Patent Document 1, it is necessary to provide an information processing device separately from the sensor terminal in order to generate compression-coded information, which increases the cost of system construction. Is assumed.

  Patent Document 1 does not discuss the optimization of the compression ratio, particularly the optimization of the compression ratio using a mathematical model. When unknown sensor data is acquired, efficient data compression is performed. Difficult to do.

  On the other hand, in the sensor data compression system according to an embodiment of the present invention, when data having a relatively high sampling frequency or data having a large data transfer amount is transmitted, an edge having a large resource restriction such as a calculation capacity or a battery capacity is large. One of the features is that an optimal compression ratio is dynamically acquired by the sensor terminal on the side, and sensor data is compressed using the compression ratio.

  At this time, the sensor terminal according to the embodiment of the present invention implements control for dynamically changing the optimal compression ratio according to the information amount of the sensor data by using a mathematical model constructed in advance. According to the above control, it is possible to set an optimal compression ratio with a lightweight process even at an edge side sensor terminal having a low calculation capability.

  Here, the above-mentioned compression ratio is the ratio of the size of the data after compression to the size of the original data, and the smaller the compression ratio, the smaller the data size after compression. Further, the optimum compression ratio may be defined as a compression ratio that can minimize the size of data to be transmitted while minimizing the loss of information amount of data.

  Furthermore, in the sensor data compression system according to the present embodiment, even when unknown sensor data is acquired by the sensor terminal on the edge side, a sensor that is not used at the time of constructing the mathematical model in the sensor terminal The degree of deviation of the acquired sensor data is detected, and when the degree of deviation exceeds the threshold value, the accuracy of the compression ratio setting is maintained by reconstructing the mathematical model using sensor data with a high degree of deviation. Is one of the features.

  In order to realize the above feature, in the present embodiment, first, a mathematical model for sequentially and dynamically setting an optimal compression ratio in accordance with the amount of data information is constructed in advance by a terminal having a high calculation capability. The mathematical model is a compression ratio learning model constructed using a machine learning method so as to dynamically change the compression ratio with respect to sensor data according to an evaluation index including a restoration error rate and a compression ratio. There may be.

  In the present embodiment, for example, the compression ratio learning model constructed as described above is transplanted to the edge-side sensor terminal and operated, so that the compression ratio optimization is sequentially performed on the sensor terminal side. It is possible to reduce the calculation cost at the sensor terminal and the overhead of the gateway and control data.

  In addition, the sensor terminal at the edge end uses the above compression ratio learning model to dynamically change the optimal compression ratio according to the data, thereby maintaining both data analysis accuracy and communication cost reduction. As a result, data compression with high compression efficiency becomes possible.

  Furthermore, when machine learning is used for model construction, the estimation accuracy for unknown data generally deteriorates, but in this embodiment, the degree of divergence between learned data and data to be compressed is detected. When the degree is high, it is regarded as unknown data, and re-learning is performed with a terminal having high calculation capability to update the learning model of the sensor terminal, thereby preventing deterioration in accuracy of estimation of the optimum compression ratio.

  Note that a compression sensing technique as described in Non-Patent Document 1 may be used for compression of sensor data by the sensor terminal according to the present embodiment. Compressed sensing is a technique for restoring a high-dimensional signal having sparsity from few observations. In compression sensing, the calculation cost of the compression process on the transmission side is generally low, but the calculation cost of the restoration process on the reception side is high. Therefore, it can be said to be a technique suitable for data compression by the sensor terminal on the edge side having a low calculation capability.

  Here, an outline of compressed sensing will be described. FIG. 19 is a diagram for explaining an outline of data compression using compression sensing. For example, as shown in FIG. 19, when the original data is defined as X, the observation matrix for compressing the data is defined as Φ, and the observation data after compression is defined as Y, the compressed observation data is expressed by the following formula (1). expressed.

  Here, if the size of the original data X is N and the size of the observation data Y after compression is M, the size of the observation matrix Φ is M × N. In the compressed sensing according to the present embodiment, it is possible to adjust the compression ratio M / N related to the observation data Y after compression by changing the M, that is, the number of rows of the observation matrix Φ. Note that Non-Patent Document 1 does not discuss dynamic change of the compression ratio, nor does it suggest adjustment of the compression ratio by changing M. For this reason, the compression sensing according to the present embodiment has an excellent effect compared to the technique described in Non-Patent Document 1 in that a dynamic compression ratio can be easily set.

  Next, the restoration in the compressed sensing technology will be described. In the above formula (1), since the unknown number> the number of formulas, a unique solution cannot usually be obtained, but the original data X contains many zeros in the element as X = Ψx. If it can be represented by the product of the sparse matrix x and the transformation matrix Ψ, there is a possibility that the observation data Y after compression can be completely restored. Specifically, it can be obtained by reducing to the problem of minimizing the L1 norm of x under the condition satisfying the following formula (2).

  At this time, the original data X can be restored by calculating X = Ψx using the obtained x. However, if a sufficient compression ratio is not set for the data to be compressed, it is also assumed that an error occurs between the restored data and the original data. For this reason, in this embodiment, said concern is solved by utilizing the evaluation parameter | index containing a decompression | restoration error and a compression ratio for learning.

  The outline according to the embodiment of the present invention has been described above. Hereinafter, the configuration for realizing the above and the effects achieved by the configuration will be described in more detail.

<< 1.2. Configuration >>
As described above, the sensor data compression system according to the present embodiment compresses sensor data with an optimal compression ratio set at a low calculation cost on the sensor terminal side in an environment where a large number of sensor terminals are installed, Allows data to be restored. Here, the compression ratio is an index calculated by (data size before compression) / (data size after compression). The higher the compression ratio, the smaller the data size after compression.

  Although it is theoretically possible to set various compression ratios for the obtained data and find the optimal compression ratio from the relationship between the restoration error and compression ratio of each, the data restoration processing is generally computational cost. However, a sensor terminal with a high resource limitation and a large processing delay may occur, which is not suitable. Further, performing optimization processing on the server side via the network requires transmission of raw data that is not compressed to the server via the network, resulting in wasted network bandwidth and processing delay.

  For this reason, in the sensor data compression system according to the present embodiment, the optimization control is not performed sequentially at the sensor terminal on the edge side, but is handled by using a mathematical model that is built in advance and that obtains an optimal compression ratio. .

  In the present embodiment, the compression ratio may be dynamically changed in the sensor terminal in accordance with the restoration error due to the information amount of the sensor data. For example, when the information amount of data is low, the compression ratio is decreased, and conversely, when the information amount is large, the compression ratio is increased. Here, as an index of the amount of information, for example, data sparsity can be cited. Even with the same data size, the greater the value of 0, the higher the sparsity. Conversely, the greater the value other than 0, the lower the sparsity. For this reason, it is possible to achieve both a high compression ratio and high restoration accuracy of data by dynamically changing the compression ratio in consideration of the information amount and the restoration on the restoration side.

  However, in a sensor terminal that compresses sensor data, it is difficult to analytically obtain sparsity to obtain an optimal compression ratio and to obtain a restoration error because the calculation amount is enormous. Therefore, in this embodiment, the optimal compression ratio is set using, for example, a mathematical model constructed using a machine learning method, thereby realizing high-speed calculation as compared with the case of obtaining the optimal solution analytically. However, it is possible to optimize the compression ratio even in the sensor terminal at the edge end having a low calculation capability.

  Further, in the sensor data compression system according to the present embodiment, when unknown data that is not used in the construction of the mathematical model is acquired by the sensor terminal on the edge side, the unknown data is transmitted to a device having a high calculation capability, By reconstructing the mathematical model, it is possible to maintain the accuracy of the optimal compression ratio estimation.

  In general, in machine learning, accuracy with respect to unknown data that has not been learned is deteriorated, but in this embodiment, the sensor terminal detects the degree of deviation of the data to be compressed from data that has been learned, and the deviation is detected. Re-learning of unknown data is performed with a device having high calculation ability by regarding high-frequency data as unknown data. In addition, it is possible to prevent deterioration in accuracy by transferring the updated learning model to the sensor terminal and estimating the optimum compression ratio at the sensor terminal.

  FIG. 1 is a diagram illustrating an assumed environment to which the sensor data compression system according to the present embodiment can be applied. Here, it is assumed that the sensor data acquired by the plurality of sensor terminals 70a to 70e is transmitted to the gateway 80 via a wireless link, and the data analysis server 90 restores the data via the gateway 80 to perform data analysis. As a data compression method, for example, the above-described compressed sensing is assumed. In general, the data analysis server 90 that performs data mining or the like needs to collect a large amount of data from a plurality of sensor terminals 70a to 70e in order to increase data analysis accuracy. However, in a normal case, the sensor terminals 70a to 70e waste a battery, and the network side may waste a wireless band and a storage for storing data.

  In general, when data is transmitted from a transmission source to a destination via a network, it is preferable to perform data compression at a terminal close to the transmission source because the total amount of data flowing through the network is reduced. In particular, since the sensor terminals 70a to 70e may be connected by a low-band wireless link, data compression is most effective for reducing the communication cost in the sensor terminals 70a to 70e as the transmission sources. However, when dynamic control of the compression ratio is performed by the sensor terminals 70a to 70e, calculating the optimum compression ratio in real time requires a huge amount of calculation for analytically obtaining sparsity or obtaining a restoration error. Therefore, it may be difficult due to resource constraints.

  Therefore, in the present embodiment, an optimal compression ratio setting using a mathematical model constructed using a machine learning method or the like is performed on the sensor terminal side. In general, machine learning requires a large computing resource when creating a learning model, but can be performed with light processing when performing inference using the constructed learning model. For this reason, the terminal that learns the optimal compression ratio may be a terminal that is separate from the sensor terminal that is installed at the time of inference.

  FIG. 2 is a block diagram illustrating a configuration example of the sensor data compression system according to the present embodiment. For example, as shown in FIG. 2, the sensor data compression system according to the present embodiment includes a learning sensor terminal 20, a plurality of sensor terminals 30 a, and a data reception device 40. Here, the sensor terminal for learning 20 and the sensor terminal 30 are attached to the detection target 10.

(Detection target 10)
The detection target 10 according to the present embodiment is a target for acquiring various sensor data such as vibrations, images, and sounds. The detection target 10 according to the present embodiment may be, for example, various equipment installed in a factory, various equipment installed in an office, various constructions / buildings, and the like.

(Learning sensor terminal 20)
The learning sensor terminal 20 according to the present embodiment constructs a compression ratio learning model (an example of a mathematical model) for obtaining an optimal compression ratio, and a deviation degree detection learning model for obtaining a deviation degree between acquired sensor data and learning data. Terminal. The learning sensor terminal 20 is an example of a model construction device according to the present invention.

(Sensor terminal 30)
The sensor terminal 30 according to the present embodiment uses the compression ratio learning model constructed by the learning sensor terminal 20 to set a compression ratio for the acquired sensor data, and receives the sensor data compressed using the compression ratio. It is a terminal that transmits to the device 40. As shown in FIG. 2, a plurality of sensor terminals 30 according to the present embodiment may be installed for a single detection target 10.

(Data receiving device 40)
The data receiving device 40 according to the present embodiment is a device that receives the sensor data compressed by the sensor terminal 30 and restores it as necessary. Note that the data receiving device 40 according to the present embodiment may further include a learning model restructuring function and a data analysis function described later.

  Next, the configuration of the learning sensor terminal 20 will be described in more detail in the present embodiment. FIG. 3 is a block diagram illustrating a configuration example of the learning sensor terminal 20 according to the present embodiment. Referring to FIG. 3, the learning sensor terminal 20 according to the present embodiment includes a sensor 210, an AD conversion unit 220, a data preprocessing unit 230, a compression ratio model construction unit 240, a deviation degree detection model construction unit 250, and model transmission. Part 260 is provided.

((Sensor 210))
The sensor 210 according to the present embodiment can be various sensors such as a vibration sensor, a strain sensor, an acoustic sensor, and an image sensor, for example.

((AD converter 220))
The AD conversion unit 220 according to the present embodiment performs analog / digital conversion when the sensor data acquired by the sensor 210 is an analog signal.

((Data preprocessing unit 230))
The data preprocessing unit 230 according to the present embodiment performs noise conversion filtering, measurement value conversion, and frequency conversion processing such as FFT (Fast Fourier Transform) on the sensor data converted into the digital signal. .

((Compression ratio model construction unit 240))
The compression ratio model construction unit 240 according to the present embodiment constructs a mathematical model that acquires a compression ratio related to compression sensing based on the feature amount of the input sensor data. The mathematical model may be a learning model constructed by a machine learning method, for example. For example, the compression ratio model construction unit 240 according to the present embodiment uses, as one of supervised learning, a compression ratio learning model that dynamically changes the compression ratio for sensor data according to an evaluation index including a restoration error and a compression ratio. It can be constructed by some regression learning, unsupervised learning, reinforcement learning, or the like. On the other hand, the mathematical model according to the present embodiment is not limited to the above example, and may be, for example, a graph or data indicating the correspondence between the feature amount of the sensor data and the optimum compression ratio. Details of the functions of the compression ratio model construction unit 240 according to this embodiment will be described later.

((Deviation degree detection model construction unit 250))
The divergence degree detection model construction unit 250 according to the present embodiment detects the degree of divergence between the input sensor data and the sensor data (learning data) used to construct the compression ratio learning model (mathematical model). Build a detection learning model. At this time, the divergence degree detection model construction unit 250 according to the present embodiment may construct One-Class SVM using the learning data, an auto encoder, or the like as the divergence degree detection learning model.

((Model transmission unit 260))
The model transmission unit 260 according to the present embodiment uses a compression ratio learning model constructed by the compression ratio model construction unit 240 or a divergence degree detection learning model constructed by the divergence degree detection model construction unit 250 as a single or a plurality of sensor terminals. 30.

  Next, the configuration of the sensor terminal 30 in this embodiment will be described in more detail. FIG. 4 is a block diagram illustrating a configuration example of the sensor terminal 30 according to the present embodiment. Referring to FIG. 4, the sensor terminal 30 according to the present embodiment includes a sensor 310, an AD conversion unit 320, a data preprocessing unit 330, a divergence degree detection unit 340, a data compression unit 350, and a data communication unit 360. The sensor 310, the AD conversion unit 320, and the data preprocessing unit 330 may have the same functions as the sensor 210, the AD conversion unit 220, and the data preprocessing unit 230 of the learning sensor terminal 20 described above. Detailed description is omitted.

((Deviation degree detector 340))
The divergence degree detection unit 340 according to the present embodiment uses the divergence degree detection learning model constructed by the divergence degree detection model construction unit 250 of the learning sensor terminal 20 to construct input sensor data and a compression ratio learning model. The degree of divergence from the learning data used in is detected. Details of the functions of the divergence degree detection unit 340 according to the present embodiment will be described in detail separately.

((Data compression unit 350))
The data compression unit 350 according to the present embodiment uses the mathematical model constructed by the compression ratio model construction unit 240 of the learning sensor terminal 20 to obtain an optimal compression ratio for the input sensor data, and based on the compression ratio. Perform data compression.

((Data communication unit 360))
The data communication unit 360 according to the present embodiment performs data communication with the learning sensor terminal 20 and the data receiving device 40 via the network. For example, the data communication unit 360 according to the present embodiment transmits the sensor data compressed by the data compression unit 350 to the data reception device 40. In addition, the data communication unit 360 according to the present embodiment is, for example, a state in which the data compression unit 350 does not compress sensor data that the divergence degree detection unit 340 determines that the divergence degree from the learning data exceeds a threshold value. To the model construction device such as the learning sensor terminal 20. Details of the functions of the data communication unit 360 according to the present embodiment will be described later.

  Next, the configuration of the data receiving device 40 according to the present embodiment will be described in more detail. FIG. 5 is a block diagram illustrating a configuration example of the data receiving device 40 according to the present embodiment. Referring to FIG. 5, the data receiving device 40 according to the present embodiment includes a data receiving unit 410 and a data restoring unit 420.

((Data receiving unit 410))
The data receiving unit 410 according to the present embodiment receives the compressed sensor data from the sensor terminal 30.

((Data restoration unit 420))
The data restoration unit 420 according to the present embodiment restores the compressed sensor data received by the data reception unit 410.

  The configuration of the sensor data compression system according to this embodiment has been described in detail above. In addition, said structure demonstrated using FIGS. 2-5 is an example to the last, and the structure of the sensor data compression system which concerns on this embodiment is not limited to the example which concerns. For example, as described above, the data receiving device 40 according to the present embodiment may further include a learning model reconstruction function and a data analysis function. On the other hand, the model construction device for reconstructing the learning model may be provided separately from the learning sensor terminal 20 and the data receiving device 40, or an analysis server having an analysis function may be provided separately. . The configuration of the sensor data compression system according to the present embodiment can be flexibly modified according to specifications and operations.

<< 1.3. Process flow >>
Next, the flow of processing by the sensor data compression system according to this embodiment will be described in detail. The processing by the sensor data compression system according to the present embodiment mainly includes a learning stage for constructing a mathematical model and a deviation degree detection learning model for setting an optimal compression ratio, and detecting unknown data and using the unknown data This includes a data compression / decompression stage in which a mathematical model and a deviation detection learning model are reconstructed, and data compression is performed using a mathematical model constructed in the learning stage.

  First, the flow of the learning stage according to the present embodiment will be described. FIG. 6 is a flowchart showing a process flow of the learning sensor terminal 20 in the learning stage according to the present embodiment. In the following, a case where the learning sensor terminal 20 constructs a compression ratio learning model by a machine learning method will be described as an example of a mathematical model.

  Referring to FIG. 6, first, the sensor 210 of the learning sensor terminal 20 acquires sensor data, and the AD conversion unit 220 performs analog / digital conversion on the sensor data (S1101).

  Next, the data preprocessing unit 230 performs preprocessing such as filtering processing for noise removal on the digital signal obtained in step S1101 and FFT processing for converting time-series data into frequency data every predetermined time. Processing is performed (S1102). Note that the data preprocessing unit 230 may perform wavelet transform processing instead of the FFT processing.

  Next, the compression ratio model construction unit 240 constructs a compression ratio learning model using the sensor data preprocessed in step S1102. At this time, the compression ratio model construction unit 240 according to the present embodiment may construct a compression ratio learning model using any one of regression learning, unsupervised learning, or reinforcement learning.

  When regression learning is employed (S1103: regression learning), the compression ratio model construction unit 240 according to the present embodiment searches for the compression ratio that optimizes the evaluation index for the sensor data, and uses a neural network or the like. The relationship between the sensor data and the optimum compression ratio for the sensor data is learned by the regression analysis method.

  At this time, the compression ratio model construction unit 240 according to the present embodiment obtains a compression ratio for optimizing the evaluation index by, for example, a random search, a grid search, a hill climbing method, or the like (S1104).

  Next, the compression ratio model construction unit 240 according to the present embodiment constructs a compression ratio learning model by regression learning using sensor data as explanatory variables and the optimal compression ratio for the sensor data obtained in step S1104 as an objective variable. (S1105).

  According to the regression learning as described above, it is possible to learn after strictly obtaining the compression ratio at which the compression efficiency is optimal.

  On the other hand, when unsupervised learning is employed (S1103: unsupervised learning), the compression ratio model construction unit 240 according to the present embodiment performs learning using only sensor data, unlike the above-described regression learning. At this time, the compression ratio model construction unit 240 according to the present embodiment sets a loss function in which the loss becomes smaller as the compression efficiency is higher, and the loss is minimized when the sensor data is input and the compression ratio is output. Thus, unsupervised learning is performed (S1106). That is, the compression ratio model construction unit 240 according to the present embodiment constructs a compression ratio learning model by obtaining a compression ratio to be optimized by updating the parameters of the neural network by the error back propagation method.

  According to the unsupervised learning as described above, the compression ratio can be estimated directly from the sensor data without explicitly obtaining the optimum compression ratio as in regression learning, and the calculation amount can be reduced. is there.

  On the other hand, when the reinforcement learning is employed (S1103: reinforcement learning), the compression ratio model construction unit 240 according to the present embodiment does not strictly determine the compression ratio at which the evaluation index is optimal, and performs sensor data under trial and error. Learn the optimal compression ratio for.

  At this time, the compression ratio model construction unit 240 according to the present embodiment constructs a compression ratio learning model by reinforcement learning using the sensor data as a reward, such as data compression efficiency defined from the state, the compression ratio, and the restoration error (S1107). ).

  According to reinforcement learning as described above, it is expected to gradually converge to an optimal compression ratio by learning by changing the compression ratio by trial and error without obtaining a strict compression ratio as in regression learning. The On the other hand, when reinforcement learning is adopted, there is a possibility that learning may take time compared to regression learning.

  Next, the divergence degree detection model construction unit 250 according to the present embodiment uses the sensor data (learning data) used for the construction of the compression ratio learning model by any of the methods described above, and uses the divergence degree detection learning model. Is constructed (S1108). At this time, the divergence degree detection model construction unit 250 according to the present embodiment may construct One-Class SVM using the learning data, an auto encoder, or the like as the divergence degree detection learning model.

  Next, the model transmission unit 260 according to the present embodiment transmits the constructed compression ratio learning model and the deviation degree detection learning model to the sensor terminal (S1109), and ends the process.

  The flow of the learning stage according to the present embodiment has been described above using FIG. Subsequently, the learning method according to the present embodiment will be described in more detail.

  First, the construction of a compression ratio learning model by regression learning according to this embodiment will be described in more detail. FIG. 7 is a diagram showing an outline of the compression ratio learning model construction by regression learning according to the present embodiment. In the regression learning according to the present embodiment, first, sensor data is collected, and a plurality of samples of sensor data to be compressed are extracted by performing AD conversion and preprocessing. Next, the data compression efficiency including the compression ratio and the restoration error as elements is set as an evaluation index.

The data compression efficiency according to the present embodiment may be set, for example, as in the following formula (3) or formula (4).

Compression efficiency = compression ratio × (−α) + restoration error × (−β) (where α> 0, β> 0)
... (3)
Compression efficiency = 1 / (compression ratio to the power of α) / (restoration error to the power of β) (where α> 0, β> 0)
... (4)

  Here, the restoration error indicates, for example, an error between data before compression and data after restoration, and may be, for example, a mean square error. In addition, α and β respectively indicate the compression ratio and the weight of the restoration error, and the larger the value, the greater the weight and is regarded as important. That is, the data compression efficiency according to the present embodiment is an index that increases as the compression ratio is lower and as the restoration error is smaller.

  At this time, in the regression learning according to the present embodiment, the compression ratio is changed from a small value, and the compression ratio that maximizes the data compression efficiency is defined as the optimum compression ratio.

  Here, a description will be given with a specific example of sensor data. FIG. 8 is an example of sensor data to be compressed according to the present embodiment. 9 and 10 are enlarged views of sensor data in period A and period B in FIG. 8, respectively. Here, period A indicates a period from time ta to ta ′, and period B indicates a period from time tb to tb ′.

  FIG. 11 is a diagram showing the relationship between the compression ratios related to the respective sensor data in the period A and the period B and the data size after compression. FIG. 12 is a diagram showing the relationship between the compression ratio and the restoration error rate related to the respective sensor data in the period A and the period B. FIG. 13 is a diagram illustrating the relationship between the compression ratio and the data compression efficiency relating to the respective sensor data in the period A and the period B. The restoration error rate shown in FIG. 12 is a value obtained by dividing the mean square error of the restoration error by the standard deviation of the amplitude, and α and β are α = 1 and β = 5, respectively.

  Here, referring to FIG. 11, it can be seen that the respective sensor data in the period A and the period B have a data size after compression that is proportional to the compression ratio regardless of the contents of the data. Referring to FIG. 12, when the compression ratio increases, the restoration error rate of each sensor data in the period A and the period B increases. However, even in the same compression ratio, the restoration error in the period A and the period B. It can be seen that the rates are different.

  As a result, as shown in FIG. 13, the compression efficiency of each sensor data in the period A and the period B is improved when the compression ratio increases to a certain value, but when the certain ratio is exceeded, the data compression efficiency is increased. Will decrease. Further, it can be seen that the compression ratio at which the data compression efficiency becomes highest in each of the sensor data in the period A and the period B is different, and it is necessary to set an optimum compression ratio according to the sensor data.

  Therefore, in the regression learning according to the present embodiment, an optimal compression ratio for a certain data is obtained by random search, grid search, hill climbing method, etc., and sensor data to be compressed is an explanatory variable, and the optimal compression for the sensor data. Learning with the compression ratio as the objective variable. At this time, the compression ratio model construction unit 240 according to the present embodiment may perform regression learning using deep learning such as LSTM (Long Short-Term Memory) and CNN (Convolutional Neural Network).

  FIG. 14 is a diagram showing a comparison between the calculated value of the optimum compression ratio for the sensor data shown in FIG. 8 and the optimum compression ratio estimated by the compression ratio learning model constructed by regression learning. In the upper part of FIG. 14, the optimum compression ratio obtained by grid search for the sensor data to be compressed shown in FIG. 8 is shown in time series. In the lower part of FIG. 14, the optimal compression ratio estimated by the compression learning model when the sensor data to be compressed shown in FIG. 8 is input is shown in time series.

  Here, when both are compared, it turns out that the compression ratio is substantially in agreement. Thus, according to the compression ratio learning model constructed by the regression learning according to the present embodiment, it is possible to accurately estimate the optimal compression ratio for the input sensor data.

  Next, the construction of a compression ratio learning model by unsupervised learning according to this embodiment will be described in detail. FIG. 15 is a diagram showing an outline of the compression ratio learning model construction by unsupervised learning according to the present embodiment.

  The construction of the compression ratio learning model by unsupervised learning according to the present embodiment does not obtain the optimal compression ratio for the sensor data in advance as in the above-described regression learning, and the loss function so that the loss is minimized at the optimal compression ratio. This is a method for obtaining an optimal compression ratio directly from sensor data.

  At this time, the loss function is set so that the loss becomes smaller as the data compression efficiency is higher, for example, by taking the reciprocal of the data compression efficiency shown in the above formulas (3) and (4).

  The compression ratio model construction unit 240 according to the present embodiment uses the loss function set as described above to learn parameter update by the error back propagation method so that the loss function is minimized with respect to the output of the neural network. I do. According to such learning, the optimum compression ratio output from the neural network converges on the input sensor data, and a compression ratio learning model can be constructed without obtaining the optimum compression ratio in advance. Is possible.

  Next, the construction of a compression ratio learning model by reinforcement learning according to this embodiment will be described in detail. FIG. 16 is a diagram showing an outline of the compression ratio learning model construction by reinforcement learning according to the present embodiment. In reinforcement learning, learning is performed to control a system that maximizes a numerical value indicating a long-term goal.

  At this time, as shown in FIG. 16, the compression ratio model construction unit 240 according to the present embodiment sets the sensor data as the state, the compression ratio as the action, and is set as in the above formulas (3) and (4). Reinforcement learning with compression efficiency as a reward may be performed.

  More specifically, the compression ratio model construction unit 240 according to the present embodiment may use time domain data or frequency domain data related to sensor data as a state.

  Further, the compression ratio model construction unit 240 according to the present embodiment may treat the compression ratio corresponding to the action as a discrete value or a continuous value. At this time, when the compression ratio model construction unit 240 regards the compression ratio as a discrete value, for example, when the deep reinforcement learning method such as Deep-Q-Network (DQN) is regarded as a continuous value, For example, a deep reinforcement learning method such as Deep Deterministic Policy Gradient (DDPG) as described in Non-Patent Document 2 may be adopted. By using the deep reinforcement learning method as described above, it is possible to obtain an optimum action (compression ratio) by approximation even when the number of states is close to an infinite number.

  More specifically, when the compression ratio is handled as a discrete value, the compression ratio model construction unit 240 divides the possible range of the compression ratio by a natural number N and strengthens the compression ratio of each representative value as a discrete value action. You may learn. For example, when the range that the compression ratio can take is 0.1 to 0.9, when the five division is performed, the compression ratio of each representative value is 0.1, 0.3, 0.5, 0.7 0.9.

  On the other hand, when the compression ratio is regarded as a continuous value, the compression ratio model construction unit 240 performs reinforcement learning using an arbitrary value extracted from a possible range of the compression ratio as an action.

  Further, the reward according to the present embodiment is not limited to the data compression efficiency as shown in FIG. 16, and can be changed flexibly. The compression ratio model construction unit 240 according to the present embodiment can also perform reinforcement learning using an evaluation index including a compression ratio (data transmission amount) or a restoration error as a reward, for example.

  The learning stage flow and learning method according to the present embodiment have been described in detail above. Next, the flow of the data compression / decompression stage according to the present embodiment will be described in detail. In the data compression / decompression stage according to the present embodiment, the sensor data is compressed by the sensor terminal 30 using the compression ratio learning model constructed in the learning stage.

  However, if sensor data of a type that was not used for learning the compression ratio learning model at the learning stage is acquired as sensor data, there is a possibility that the compression ratio cannot be set correctly for the sensor data. Therefore, in the sensor data compression system according to the present embodiment, when unknown data that is not used as learning data is detected at the time of constructing the compression ratio learning model in the data compression / decompression stage, the learning sensor terminal 20 or a separate sensor data compression system is separately provided. One of the features is that the compression ratio learning model using the unknown data is re-learned by a model construction apparatus realized by a server or the like, and the re-learned compression ratio learning model is applied to the sensor terminal 30.

  FIG. 17 is a sequence diagram showing a flow of processing in the data compression / decompression stage according to the present embodiment. FIG. 17 shows an example in which the learning sensor terminal 20 according to the present embodiment performs relearning of a learning model as a model construction device.

  Referring to FIG. 17, first, acquisition of sensor data by the sensor 310 of the sensor terminal 30 and analog / digital conversion by the AD conversion unit 320 are executed (S2101).

  Next, the data preprocessing unit 330 performs preprocessing such as a change to frequency data (S2202).

  Next, the divergence degree detection unit 340 according to the present embodiment uses the divergence degree detection learning model constructed in step S1108 shown in FIG. 6 to acquire the acquired sensor data and the currently applied compression ratio learning. The degree of deviation from the learning data used to construct the model is detected (S2103).

  Here, when the divergence degree detected in step S2103 exceeds the threshold value, the data communication unit 360 according to the present embodiment supplies sensor data that is not compressed using the compression ratio learning model to the learning sensor terminal 20. Transmit (S2104).

  Subsequently, the compression ratio model construction unit 240 and the deviation degree detection model construction unit 250 of the learning sensor terminal 20 use the sensor data received in step S2104 to relearn the compression ratio model and the deviation degree detection learning model ( Reconstruction) is performed (S2105).

  Subsequently, the model transmission unit 260 of the learning sensor terminal transmits the compression ratio model and the deviation degree detection learning model re-learned in step S2105 to the sensor terminal 30 (S2106).

  According to the above processing, even in an environment where unknown data can be acquired, the accuracy of the compression ratio setting can be maintained by reconstructing the learning model each time. At this time, the model transmission unit 260 according to the present embodiment learns the reconstructed sensor terminal 30 that transmitted the sensor data at least in step S2104 among the plurality of sensor terminals 30 connected via the network. You may send a model.

  In addition, the model transmission unit 260 is connected to the same detection target 10 as the sensor terminal 30 that transmitted the sensor data in step S2104, or another detection target having similar characteristics. 10 may be transmitted to the plurality of sensor terminals 30 when there is another sensor terminal 30 installed in association with the sensor terminal 30. According to the processing as described above, it is possible to efficiently maintain the accuracy of the compression ratio setting by the plurality of sensor terminals 30.

  On the other hand, when the divergence detected in step S2103 is equal to or smaller than the threshold, the data compression unit 350 uses the compression ratio learning model transmitted from the learning sensor terminal 20 in the learning stage to calculate the optimum compression ratio for the sensor data. Acquired (S2107), and compresses the sensor data by the compression ratio (S2108).

  Next, the data communication unit 360 transmits the sensor data compressed in step S2108 to the data receiving device 40 (S2109).

  Subsequently, the data restoration unit 420 of the data reception device 40 restores the sensor data received by the data reception unit 410 in step S2109 (S2110). At this time, since the data restoration unit 420 needs to grasp the original data size and the compressed data size, the data communication unit 360, as described above, together with the compressed sensor data in step S2109. Information related to the appropriate data size may be transmitted to the data receiving device 40.

<2. Hardware configuration example>
Next, a hardware configuration example when the model construction device according to the embodiment of the present invention is realized as a server separate from the learning sensor terminal 20 will be described. FIG. 18 is a block diagram illustrating a hardware configuration example of a model construction device according to an embodiment of the present invention. Referring to FIG. 18, the model construction apparatus includes, for example, a processor 871, a ROM 872, a RAM 873, a host bus 874, a bridge 875, an external bus 876, an interface 877, an input unit 878, and an output unit 879. A storage unit 880, a drive 881, a connection port 882, and a communication unit 883. Note that the hardware configuration shown here is an example, and some of the components may be omitted. Moreover, you may further include components other than the component shown here.

(Processor 871)
The processor 871 functions as, for example, an arithmetic processing unit or a control unit, and controls all or part of the operation of each component based on various programs recorded in the ROM 872, the RAM 873, the storage unit 880, or the removable recording medium 901. To do.

(ROM 872, RAM 873)
The ROM 872 is a means for storing a program read by the processor 871, data used for calculation, and the like. In the RAM 873, for example, a program to be read by the processor 871, various parameters that change as appropriate when the program is executed, and the like are temporarily or permanently stored.

(Host bus 874, bridge 875, external bus 876, interface 877)
The processor 871, the ROM 872, and the RAM 873 are connected to each other via, for example, a host bus 874 capable of high-speed data transmission. On the other hand, the host bus 874 is connected to an external bus 876 having a relatively low data transmission speed via a bridge 875, for example. The external bus 876 is connected to various components via an interface 877.

(Input unit 878)
For the input unit 878, for example, a mouse, a keyboard, a touch panel, a button, a switch, a microphone, a lever, and the like are used. Furthermore, as the input unit 878, a remote controller (hereinafter referred to as a remote controller) that can transmit a control signal using infrared rays or other radio waves may be used.

(Output unit 879)
The output unit 879 includes acquired information such as a display device (display device) such as a CRT (Cathode Ray Tube), LCD, or organic EL, an audio output device such as a speaker or headphones, a printer, a mobile phone, or a facsimile. Is a device capable of visually or audibly notifying a user.

(Storage unit 880)
The storage unit 880 is a device for storing various data. As the storage unit 880, for example, a magnetic storage device such as a hard disk drive (HDD), a semiconductor storage device, an optical storage device, a magneto-optical storage device, or the like is used.

(Drive 881)
The drive 881 is a device that reads information recorded on a removable recording medium 901 such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, or writes information to the removable recording medium 901.

(Removable recording medium 901)
The removable recording medium 901 is, for example, a DVD medium, a Blu-ray (registered trademark) medium, an HD DVD medium, or various semiconductor storage media. Of course, the removable recording medium 901 may be, for example, an IC card on which a non-contact IC chip is mounted, an electronic device, or the like.

(Connection port 882)
The connection port 882 is a port for connecting an external connection device 902 such as a USB (Universal Serial Bus) port, an IEEE 1394 port, a SCSI (Small Computer System Interface), an RS-232C port, or an optical audio terminal. is there.

(External connection device 902)
The external connection device 902 is, for example, a printer, a portable music player, a digital camera, a digital video camera, or an IC recorder.

(Communication unit 883)
The communication unit 883 is a communication device for connecting to the network 903. For example, a communication card for wired or wireless LAN, Bluetooth (registered trademark), or WUSB (Wireless USB), a router for optical communication, ADSL (Asymmetric) A digital subscriber line) or a modem for various types of communication. Moreover, you may connect to various telephone networks, such as a public telephone network, a mobile telephone provider network, and an extension telephone network.

<3. Summary>
As described above, the sensor data compression system according to an embodiment of the present invention includes the sensor terminal 30 and the model construction device realized by the learning sensor terminal 20 or a separate server. According to the sensor data compression system according to an embodiment of the present invention, it is possible to achieve both maintenance of data analysis accuracy and reduction of communication costs.

  The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to such examples. It is obvious that a person having ordinary knowledge in the technical field to which the present invention pertains can come up with various changes or modifications within the scope of the technical idea described in the claims. Of course, it is understood that these also belong to the technical scope of the present invention.

  Further, the effects described in the present specification are merely illustrative or exemplary and are not limited. That is, the technology according to the present disclosure can exhibit other effects that are apparent to those skilled in the art from the description of the present specification in addition to or instead of the above effects.

  In addition, it is possible to create a program for causing hardware such as a CPU, ROM, and RAM incorporated in a computer to exhibit functions equivalent to the configuration of the learning sensor terminal 20 (model construction device) and the sensor terminal 30. Also, a non-transitory recording medium that can be read by a computer and that records the program can be provided.

  In addition, each step related to the processing of the sensor data compression system of the present specification does not necessarily have to be processed in time series in the order described in the flowcharts and sequence diagrams. For example, the steps related to the processing of the sensor data compression system may be processed in an order different from the order described in the flowcharts and sequence diagrams, or may be processed in parallel.

DESCRIPTION OF SYMBOLS 10 Detection target 20 Learning sensor terminal 240 Compression ratio model construction part 250 Deviation degree detection model construction part 260 Model transmission part 30 Sensor terminal 340 Deviation degree detection part 350 Data compression part 360 Data communication part 40 Data receiving apparatus

Claims (17)

A sensor for obtaining sensor data,
A data compression unit that compresses the sensor data at the compression ratio using a mathematical model that acquires a compression ratio related to compression sensing based on the acquired feature amount of the sensor data.
A data communication unit that transmits the sensor data compressed by the data compression unit; and a divergence degree detection unit that detects a divergence degree between the acquired sensor data and the sensor data used to construct the mathematical model;
With
The data communication unit, when the detected divergence exceeds a threshold, transmits the sensor data that is not compressed by the compression ratio to a model construction device,
A sensor terminal;
A compression ratio model construction unit that reconstructs the mathematical model using the sensor data that has not been compressed by the compression ratio received from the sensor terminal; and the mathematical model reconstructed by the compression ratio model construction unit A model transmission unit for transmitting to the sensor terminal,
Comprising
The model building device;
Sensor data compression system consisting of The mathematical model is a compression ratio learning model constructed using a machine learning method so as to dynamically change the compression ratio for the sensor data according to an evaluation index including a restoration error and the compression ratio.
The compression ratio model construction unit reconstructs the compression ratio learning model using the sensor data that has not been compressed by the compression ratio received from the sensor terminal,
The model transmission unit transmits the compression ratio learning model reconstructed by the compression ratio model construction unit to the sensor terminal;
The sensor data compression system according to claim 1. The divergence degree detection unit detects the divergence degree using a divergence degree detection learning model constructed using the sensor data used for constructing the compression ratio learning model.
The sensor data compression system according to claim 2. The model construction device further includes a deviation degree detection model construction unit that constructs the deviation degree detection learning model using the sensor data that has not been compressed by the compression ratio received from the sensor terminal,
The model transmission unit transmits the deviation degree detection learning model constructed by the deviation degree detection model construction unit to the sensor terminal;
The sensor data compression system according to claim 3. The compression ratio model construction unit constructs the compression ratio learning model using any one of regression learning, unsupervised learning, or reinforcement learning.
The sensor data compression system according to any one of claims 2 to 4. The compression ratio model construction unit obtains the compression ratio for optimizing the evaluation index, and calculates the compression ratio learning model by regression learning using the sensor data as an explanatory variable and the optimal compression ratio for the sensor data as an objective variable. To construct,
The sensor data compression system according to claim 5. The compression ratio model construction unit obtains the compression ratio for optimizing the evaluation index by any of a random search, a grid search, or a hill climbing method.
The sensor data compression system according to claim 6. The compression ratio model structuring unit uses the loss function in which the loss becomes smaller as the data compression efficiency is higher, and the compression ratio is optimized by unsupervised learning to obtain the compression ratio to be optimized by updating the parameters of the neural network by the error back propagation method. Build a ratio learning model,
The sensor data compression system according to claim 5. The compression ratio model construction unit constructs the compression ratio learning model by reinforcement learning using the sensor data as a state, the compression ratio as an action, and data compression efficiency as a reward.
The sensor data compression system according to claim 5. The compression ratio model construction unit performs reinforcement learning using time domain data or frequency domain data related to the sensor data as a state.
The sensor data compression system according to claim 5. The compression ratio model construction unit divides the range of the compression ratio by a natural number N, and performs reinforcement learning using the compression ratio of each representative value as a discrete value action.
The sensor data compression system according to claim 5. The compression ratio model construction unit performs reinforcement learning using the compression ratio as a continuous value action.
The sensor data compression system according to claim 5. The compression ratio model construction unit performs reinforcement learning with an evaluation index including the compression ratio or a restoration error as a reward.
The sensor data compression system according to claim 5. A sensor for acquiring sensor data;
A data compression unit that compresses the sensor data at the compression ratio using a mathematical model that acquires a compression ratio related to compression sensing based on the acquired feature amount of the sensor data;
A data communication unit for transmitting the sensor data compressed by the data compression unit;
A divergence degree detection unit for detecting a divergence degree between the acquired sensor data and the sensor data used to construct the mathematical model;
With
When the detected degree of divergence exceeds a threshold, the data communication unit transmits the sensor data that is not compressed by the compression ratio to a model construction device, and is reconstructed using the sensor data. Receive mathematical models,
Sensor terminal. The sensor that is not compressed by the compression ratio from a sensor terminal that compresses the sensor data at the compression ratio using a mathematical model that acquires a compression ratio related to compression sensing based on the acquired feature amount of the sensor data. When data is received, a compression ratio model construction unit that reconstructs the mathematical model using the sensor data;
The sensor that has transmitted the sensor data that is not compressed by at least the compression ratio among a plurality of sensor terminals connected to the mathematical model reconstructed by the compression ratio model construction unit via a network A model transmitter to transmit to the terminal;
Comprising
Model building device. Computer
A sensor for acquiring sensor data;
A data compression unit that compresses the sensor data at the compression ratio using a mathematical model that acquires a compression ratio related to compression sensing based on the acquired feature amount of the sensor data;
A data communication unit for transmitting the sensor data compressed by the data compression unit;
A divergence degree detection unit for detecting a divergence degree between the acquired sensor data and the sensor data used to construct the mathematical model;
With
When the detected degree of divergence exceeds a threshold, the data communication unit transmits the sensor data that is not compressed by the compression ratio to a model construction device, and is reconstructed using the sensor data. Receive mathematical models,
Sensor terminal,
Program to function as. Computer
The sensor that is not compressed by the compression ratio from a sensor terminal that compresses the sensor data at the compression ratio using a mathematical model that acquires a compression ratio related to compression sensing based on the acquired feature amount of the sensor data. When data is received, a compression ratio model construction unit that reconstructs the mathematical model using the sensor data;
The sensor that has transmitted the sensor data that is not compressed by at least the compression ratio among a plurality of sensor terminals connected to the mathematical model reconstructed by the compression ratio model construction unit via a network A model transmitter to transmit to the terminal;
Comprising
Model building device,
Program to function as.

Images