JP2020065225A - Computer, sensing system, and data communication method - Google Patents

Abstract

To appropriately reduce the amount of sensor data to eliminate communication bottlenecks.SOLUTION: A sensing system that collects sensor data includes a sensor terminal that generates sensor data on the basis of the measurement result by a sensor, and an edge node that receives the sensor data from the sensor terminal. The edge node includes: a dictionary learning unit that learns a dictionary composed of a matrix of base vectors and changes the dictionary in time series; and a sparse restoration processing unit that performs sparse restoration processing that restores the sensor data from data with a reduced data amount, which is received from the sensor terminal on the basis of sparseness using the learned dictionary.SELECTED DRAWING: Figure 4

Description

  The present invention relates to a sensing system that collects sensor data from a sensor terminal.

  As a background art of this technical field, an underdetermined solution search method using sparsity has been proposed (Non-Patent Document 1). For example, compressed sensing is a method of restoring original information by using sparsity from randomly thinned observation values, and mainly uses image processing of MRI and sample rate using non-equidistant sampling of A / D converter. It is used in a system that randomly obtains some observation values, such as reduction of.

  In addition, as a similar method, a method called sparse coding and dictionary learning for reducing the noise by extracting the feature quantity of the observation value by using the relationship between the transformation matrix of the underdetermined system equation and the sparse signal is proposed. (Non-patent document 2).

  On the other hand, in recent years, an application called IoT (Internet of Things), which provides an advanced information service using a plurality of sensors, has been receiving attention.

D. L. Donoho, "Compressed sensing," IEEE Trans. Inf. Theory, vol. 52, no. 4, pp. 1289-1306, Apr. 2006. A. M. Tillmann, "On the Computational Intractability of Exact and Approximate Dictionary Learning", IEEE Signal Processing Letters 22 (1), 2015: 45-49.

  In the conventional sensing system, the sensor data measured by the sensor terminal is collected at the edge node, and if the number of sensor terminals and the amount of sensor data increase, the communication band between the sensor terminal and the edge node becomes a bottleneck. Therefore, there is a problem that desired sensor data cannot be collected. In this case, in the conventional sensing system, it is necessary for the sensor terminal to extract the feature amount and to reduce the communication data amount by the intermittent operation of the sensor terminal. Usually, the sensor terminal has a lower computing capability than the edge node, so that the analysis of sensor data and the extraction of the feature amount reduce the performance of the entire system.

  Therefore, there is a demand for a system that can appropriately reduce the amount of sensor data and eliminate communication bottlenecks.

  A typical example of the invention disclosed in the present application is as follows. That is, a sensing system that collects sensor data, including a sensor terminal that generates sensor data based on a measurement result by a sensor, and an edge node that receives sensor data from the sensor terminal, wherein the edge node is a base. Based on sparseness, using a dictionary learning unit that learns a dictionary composed of a matrix of vectors and changes the dictionary in time series, and the learned dictionary, the data amount received from the sensor terminal is And a sparse restoration processing unit that performs sparse restoration processing that restores sensor data from the reduced data.

  According to one aspect of the present invention, an efficient sensing system can be realized. Problems, configurations, and effects other than those described above will be clarified by the following description of the embodiments.

It is a figure which shows the structure of the sensing system of the Example of this invention. It is a figure which shows the data communication method of a present Example. It is a figure which shows the data communication method using the dictionary which changes in a time series of a present Example. It is a figure which shows the functional block of the sensing system of a present Example. It is a sequence diagram which shows the communication procedure of the sensor terminal and edge node of a present Example. It is a figure which shows the communication procedure of a sensor terminal and an edge node when using the thinning rate as a control parameter in a present Example. It is a figure which shows the state of communication data and an observed value when using the thinning rate as a control parameter in a present Example. It is a figure which shows an example of the frame format at the time of explicitly measuring signal quality of a present Example. It is a figure which shows the frame structure when using the fixed frame format of a present Example. It is a figure which shows the relationship between the data communication amount and the signal quality restored by the edge node of a present Example. It is a figure which shows the method of estimating the restoration quality, restoring the thinned compressed data in a present Example.

  Embodiments of the present invention will be described below with reference to the drawings.

  FIG. 1 is a diagram showing the configuration of the sensing system of the present embodiment.

  The sensing system of this embodiment includes a cloud 11, an edge node 13 connected to the cloud 11, an edge node 14 connected to local resources such as a computer (PC) 15 and a storage 12, and a sensor terminal 16. The sensor terminal 16 is connected to a network (not shown) through the edge nodes 13 and 14, and the sensor terminal 16 and the edge nodes 13 and 14, the edge nodes 13 and 14 each other, and the link between the cloud 11 and the edge node 13 are connected. Each communication band can become a bottleneck. Therefore, it is desired to reduce the communication data amount of each link. In the embodiment of the present invention, the focus is mainly on the reduction of the data communication amount between the sensor terminal 16 and the edge nodes 13 and 14 which have low computing ability. However, the link between the edge nodes 13 and 14 and the edge node The method of the present invention can also be applied to the connection between the cloud 13 and the cloud 11.

  The edge nodes 13 and 14 are configured by a computer having a processor (CPU), a memory, an auxiliary storage device, and a communication interface.

  The processor is a computing device that executes a program stored in the memory. Various functions of the edge nodes 13 and 14 are realized by the processor executing various programs. A part of the processing performed by the processor by executing the program may be executed by another arithmetic device (for example, hardware such as FPGA (Field Programmable Gate Array) and ASIC (Application Specific Integrated Circuit)).

  The memory includes a ROM which is a non-volatile storage element and a RAM which is a volatile storage element. The ROM stores an immutable program (for example, BIOS) and the like. The RAM is a high-speed and volatile storage element such as a DRAM (Dynamic Random Access Memory), and temporarily stores a program executed by the processor and data used when the program is executed.

  The auxiliary storage device is a large-capacity and non-volatile storage device such as a magnetic storage device (HDD) or a flash memory (SSD). Further, the auxiliary storage device stores data used by the processor when executing the program, and programs executed by the processor (dictionary learning program, sparse restoration processing program, feature amount extraction processing program, control program, etc.). That is, the program is read from the auxiliary storage device, loaded into the memory, and executed by the processor to realize the respective functions of the edge nodes 13 and 14.

  The communication interface is a network interface device that controls communication with other devices according to a predetermined protocol.

  The program executed by the processor is provided to the edge nodes 13 and 14 via a network or removable medium (CD-ROM, flash memory, etc.), and is stored in a non-volatile auxiliary storage device which is a non-transitory storage medium.

  The edge nodes 13 and 14 are computer systems physically configured on one computer or on a plurality of logically or physically configured computers, and are constructed on a plurality of physical computer resources. It may operate on a virtual computer.

  FIG. 2 is a diagram showing a data communication method of this embodiment.

  In the present embodiment, the sensor terminal 16 randomly performs data thinning 24, reduces the data communication amount, and transmits the data to the edge nodes 13 and 14. The edge nodes 13 and 14 perform sparse restoration processing 209, and the observed value 200 To restore. The sparse restoration process is a minimum value search problem expressed by Equation (1), where y is a thinned observation value, x is a sparse representation of the original observation value, and A is a conversion matrix.

  A general-purpose dictionary (for example, a matrix composed of basis vectors) such as DCT or wavelet transform is used as a transformation matrix of general sparse restoration processing, but in the sparse restoration processing of the present embodiment, it is specialized to sensor data in the field. By learning this dictionary in the field, we obtain more sparse expressions and improve the decimation rate of recoverable data. Dictionary learning from sensor data is represented by Formula (2), where D is a dictionary to be learned, Y is a thinned observation value, and x is a sparse representation of the observation value.

  Expression (2) can obtain a converged solution by alternately repeating the process of fixing the dictionary D to be learned to obtain the sparse expression x and the process of fixing the sparse expression x to obtain the dictionary D. . When a dictionary specialized for sensor observation values is created by such a method, if the state of the observation values does not change when the dictionary is created, the sparsity will be higher than when a general-purpose dictionary is used. It is possible to obtain a sparse representation that is strong, that is, can be restored even if the thinning rate is large. On the other hand, when learning the dictionary and when the state of the observed value changes, correct restoration becomes difficult, and the characteristics deteriorate as compared with a general-purpose dictionary. In FIG. 2, dictionary learning is performed on state 1 of the observed value and used for sparse restoration. In this case, the restored state 1 waveform 210 is restored in a form close to the original state 1 waveform 201 even for transmission data with a high decimation rate, but the restored state 2 waveform 211 is the original state. The waveform is likely to be different from the waveform 211 in the state 2 of.

  Therefore, in the present embodiment, a dictionary that changes in time series is learned on the spot, and by following changes in the state of the waveform, a high thinning rate is realized while maintaining the quality of the restored waveform.

  FIG. 3 is a diagram showing a data communication method using a dictionary that changes in time series according to this embodiment.

  In FIG. 3, the dynamic dictionary generation unit detects a change in the state of the observed value and relearns and switches the dictionary. By dynamically generating the dictionary, the restored state 2 waveform 34 can be expected to be a waveform closer to the original state 2 waveform 202.

  FIG. 4 is a functional block diagram of the sensing system of this embodiment.

  The sensor terminal 16 includes a sensor 412, an analog front end 411, a thinning-out processing unit 410, and a control unit 413, and based on control signals from the edge nodes 13 and 14, the thinning-out ratio and the operating speed and band of the analog front end 411. At least one is changed and the observed value is transmitted to the edge nodes 13 and 14. The sensor 412 outputs a measured value of a physical quantity (temperature, vibration, etc.) around the sensor terminal 16. The analog front end 411 includes an AD converter that converts the measurement value output by the sensor 412 into a digital signal (sensor data). The thinning-out processing unit 410 reduces the data amount by thinning out the sensor data at random. The control unit 413 transmits and receives control parameters to and from the edge nodes 13 and 14 and controls the operation of the sensor terminal 16. It should be noted that the thinning process in 410 may be practically non-equidistant, and may be an arbitrary fixed pattern that is not equidistant or a pseudo random pattern.

  The edge nodes 13 and 14 include a dictionary learning unit 407, a sparse restoration processing unit 408, a feature amount extraction processing unit 409, and a control unit 414. When performing dictionary learning, the dictionary learning unit 407 receives sensor data without thinning and learns a dictionary. Further, the dictionary learning unit 407 sends learning data to the cloud 11 as needed and receives a known dictionary from the cloud 11. When performing the data communication after learning, the sparse restoration processing unit 408 restores the thinned-out data received from the sensor terminal using the learned dictionary. Further, the sparse restoration processing unit 408 controls the thinning rate according to the quality of the restored signal, and notifies the sensor terminal 16 via the control unit 414. The feature amount extraction processing unit 409 analyzes the data using the restored observation value, extracts the feature amount, and notifies the cloud 11 of the feature amount. The control unit 414 transmits and receives control parameters to and from the sensor terminal 16 and controls the operations of the edge nodes 13 and 14. When the cloud 11 performs data analysis and feature amount extraction, the edge nodes 13 and 14 send the restored waveform itself to the cloud 11.

  The cloud 11 has a dictionary storage unit 404, a data analysis unit 405, and a control unit 406. The dictionary storage unit 404 also stores the dictionary learned by the dictionary learning unit 407, and exchanges the dictionary with the edge nodes 13 and 14 as necessary. The data analysis unit 405 analyzes the feature amount and data transmitted from the edge according to the application. The control unit 406 controls the operation of the cloud 11.

  FIG. 5 is a sequence diagram showing a communication procedure between the sensor terminal 16 and the edge nodes 13 and 14 of this embodiment.

  In the present embodiment, first, the control unit 414 notifies the sensor terminal 16 of the training request 51 for learning from the edge nodes 13 and 14. Upon receiving the training request 51, the sensor terminal 16 sends the training signal 52 from the analog front end 411 to the edge nodes 13 and 14. The dictionary learning unit 407 of the edge nodes 13 and 14 performs the dictionary learning 56 using the training signal 52. After the learning is completed, the control unit 414 notifies the sensor terminal 16 of the data request 53. The data request 53 notifies the sensor terminal 16 of the thinning rate. When the sensor terminal 16 receives the data request 53, the thinning processing unit 410 reduces the data amount according to the thinning rate notified by the data request 53, and transmits the data 54.

  The sparse restoration processing unit 408 of the edge nodes 13 and 14 successively evaluates the restoration quality of the data, and when the deterioration of the signal quality is detected, it is determined that the state of the observed value has changed (57), and the control unit again. 414 transmits the training request 57 for learning, and the dictionary learning unit 407 learns and updates the dictionary using the training signal 52. The training request 57 may be notified as an explicit control signal, but since the training signal 52 for learning is the observed value itself of the sensor terminal 16, both the training request 57 and the data request 53 are notified by the notification of the thinning rate. The format of can be shared.

  FIG. 6 is a sequence diagram showing a communication procedure between the sensor terminal 16 and the edge nodes 13 and 14 when the thinning rate is used as a control parameter in this embodiment.

  The control unit 414 of the edge nodes 13 and 14 transmits a training request 61 for learning, which is a control parameter for requesting data transmission without thinning, to the sensor terminal 16. Upon receiving the notification, the sensor terminal 16 transmits the data 62 without thinning out from the analog front end 411 to the edge nodes 13 and 14. The dictionary learning unit 407 of the edge nodes 13 and 14 performs the dictionary learning 66 using the data without thinning. After the learning is completed, the control unit 414 of the edge nodes 13 and 14 transmits the control parameter 63 specifying the thinning rate to the sensor terminal 16. In the sensor terminal 16, the thinning processing unit 410 reduces the amount of data according to the thinning rate notified by the control parameter 63, and transmits the thinned data 64.

  The sparse restoration processing unit 408 of the edge nodes 13 and 14 sequentially evaluates the restoration quality of the data, and when the deterioration of the signal quality is detected, it is determined that the state of the observed value has changed (67), and the control unit again. 414 transmits the training request 68 for learning, and the dictionary learning unit 407 learns and updates the dictionary using the data 62 without thinning.

  When this procedure is used, the sensor terminal 16 does not need to grasp the states of the edge nodes 13 and 14, and the data amount can be reduced by simple control for changing the thinning rate.

  FIG. 7 is a diagram showing the states of communication data and observed values when the thinning rate is used as a control parameter in this embodiment.

  In FIG. 7, the training signal 71 is data in which the ratio of the original data to the transmission data is 1: 1 without thinning. At the time of data transmission after learning, the thinned data 72, 73, 74 are transmitted to the edge nodes 13, 14. The edge nodes 13 and 14 monitor the state change of the observed value based on the restoration quality of the thinned data, and when the state change is detected, a data request without thinning is sent from the edge nodes 13 and 14 to the sensor terminal 16. Notice.

  In FIG. 7, the timing of the non-decimation data request 77 is synchronized with the change from the state 75 to the state 76, but in reality, the edge nodes 13 and 14 estimate the change of state based on the restoration quality. Therefore, even if the state of the observed value changes, if high restoration quality is obtained, no thinning-out data is not required.

  For restoration quality monitoring in the edge nodes 13 and 14, either an explicit method or an implicit method may be used depending on the situation. In the explicit method, the sensor terminal 16 inserts undecimated data having a length sufficiently shorter than that of the training signal as a pilot signal at a constant rate when transmitting the decimated data. The edge nodes 13 and 14 perform thinning processing and sparse restoration processing on the pilot signal and directly monitor the quality of the restored signal. In the implicit method, the edge nodes 13 and 14 estimate the restoration quality using the data whose data amount has been reduced by thinning. Specifically, the restoration quality is estimated based on the sparseness at the time of sparse restoration and the size of the objective function, and the change in state is estimated depending on whether the current observed value matches the dictionary.

  Here, the sparsity is the first-order norm in the above-mentioned mathematical expression (1), the absolute value of the variable x, and the magnitude of the objective function is the mathematical expression (1) itself.

  FIG. 8 is a diagram showing an example of a frame format when the signal quality is explicitly measured according to the present embodiment.

  The frame transmitted from the sensor terminal includes a training signal 81, thinned-out data 82, and a pilot signal 83 that is sufficiently short for training signal processing. As with the training signal 81, it is desirable to use data without thinning out as the pilot signal 83, but data with a thinning rate sufficiently larger than the thinning data for data communication may be used.

  In this method, as shown in FIGS. 5 and 6, it is necessary to exchange control signals for reducing the data communication amount between the sensor terminal 16 and the edge nodes 13 and 14. When adopting a configuration that emphasizes controllability rather than efficiency, a fixed frame format is set in which the transmission periods of the training signal 81, the thinned-out data 82, and the pilot signal 83 are set at fixed predetermined intervals. Can be used. In this case, there is a delay in detecting the state change due to the transmission interval of the training signal, but the sensor terminal 16 does not need to interpret the notification from the edge nodes 13 and 14 and control the transmission signal. Can be simplified. On the other hand, there is a trade-off between the length of time that a high-quality restored waveform cannot be obtained because a state change cannot be detected, and the effective amount of data communication reduction. Inferior to frame format.

  FIG. 9 is a diagram showing a frame structure when the fixed frame format of this embodiment is used.

  In the fixed frame format, the distance between the training signal 91 and the data / pilot signal 92 is fixed. The sensor terminal 16 repeatedly transmits the training signal 91 and the data 92 at a predetermined cycle. When the edge nodes 13 and 14 detect the deterioration of the signal quality, the edge nodes 13 and 14 perform dictionary learning using the training signal 91 transmitted next.

  FIG. 10 is a diagram showing the relationship between the data communication amount and the signal quality restored by the edge node according to the present embodiment.

  In FIG. 10, the solid line indicates the restoration quality when the sparse restoration process is performed using the dynamic dictionary learning specialized for the sensor data of the present embodiment, and the broken line indicates the sparse restoration process using a general-purpose discrete cosine transform dictionary. Indicates the quality of restoration performed. The horizontal axis of the figure represents the number of data points used for the restoration with respect to 1024 original data points, and the signal quality of the vertical axis is the correlation coefficient between the original signal x and the restored signal y. The correlation coefficient ρ is represented by the mathematical expression (3). In Expression (3), the code with an overline is the expected value of the value indicated by each code.

  Comparing with a correlation coefficient value of 0.7, the method of the present embodiment can restore from about 60 data points, whereas the conventional method requires about 170 data points. As described above, the amount of data communication can be reduced by using the method of this embodiment.

  FIG. 11 is a diagram showing a method of estimating the restoration quality while restoring the original signal from the compressed signal thinned out in the present embodiment.

  D is thinned-out data used for signal restoration, and T is a received signal for restoration quality estimation. In the restoration quality estimation method shown in FIG. 11, the original signal is restored from the thinned data. The restoration result R is used as a restoration signal. At this time, in order to estimate the restoration quality, a part of the thinned-out data D is used as the restoration quality estimation data T to generate the restoration signal R excluded from the signal restoration, and the restoration signal R and the restoration signal estimation data T are generated. And are compared to estimate the quality of the reconstructed signal. By the method shown in FIG. 11, the quality of the restored signal can be estimated while restoring the compressed thinned-out data.

  As described above, according to the embodiment of the present invention, the receiver (edge nodes 13 and 14) learns a dictionary composed of a matrix of basis vectors and changes the dictionary in time series. A learning unit 407 and a sparse restoration processing unit 408 that performs sparse restoration processing that restores sensor data from data with a reduced data amount received from the sensor terminal based on sparseness using the learned dictionary. Since it is provided, the communication quality can be reduced while maintaining the restoration quality of the sensing data transmitted by the sensor terminal 16, and an efficient sensing system can be realized.

  Further, the sparse restoration processing unit 408 measures the restoration quality from the result of the sparse restoration processing, and the dictionary learning unit 407 changes the dictionary according to the measured restoration quality, so that according to the change in the state of the observed value. By changing the dictionary, efficient communication can be performed at an appropriate data reduction rate (thinning rate).

  In addition, the control unit 414 of the edge nodes 13 and 14 requests the sensor terminal 16 to transmit a training signal according to the measured restoration quality (51, 61), and the sensor terminal 16 responds to the training signal request. The training signals 52 and 62 are transmitted, and the control unit 414 of the edge nodes 13 and 14 requests the sensor terminal 16 to transmit data whose data amount has been reduced (decimated) after dictionary learning (53, 63). The sensor terminal 16 transmits the data 54, 64 with the reduced data amount in response to the data request from the edge nodes 13, 14, so that an unnecessary training signal is not transmitted and the reduction of the effective rate can be suppressed. .

  Further, the training signals 52 and 62 are sensor data that have not been thinned out, and the data 54 and 64 that are transmitted from the sensor terminal 16 after dictionary learning are data whose data amount has been reduced by the thinning processing. The processing load of 16 can be reduced. Further, it is not necessary to prepare a signal for training.

  Further, the sparse restoration processing unit 408 performs thinning processing and sparse restoration processing on undecimated data (pilot signal) and measures restoration quality, so that the quality of the restored signal can be accurately measured, and the restoration quality of the signal can be determined. Can be improved.

  Further, since the sparse restoration processing unit 408 measures the restoration quality by using the thinned data (for example, by matching the sensor data restored from the thinned data and the dictionary), data that has not been thinned (pilot signal ) Need not be sent, and the amount of data sent can be reduced.

  Further, since the sparse restoration processing unit 408 measures the restoration quality from the sparseness calculated in the sparse restoration processing and the size of the objective function of the sparse restoration processing, a pilot signal is not necessary, and thus a reduction in transmission efficiency can be suppressed. . Moreover, the quality of the restored signal can be estimated with a small amount of calculation. Therefore, it is effective when the waveform of the sensor data changes frequently.

  Further, since the sensor terminal 16 reduces the data amount by thinning out the sensor data at random, the sensor terminal 16 does not need to have a receiving function, and the system configuration and the configuration of the sensor terminal 16 can be simplified.

  Further, since the sensor terminal 16 repeatedly transmits the training signals 81 and 91 and the data 82 and 92 with the reduced data amount in a predetermined cycle, the data to be transmitted while suppressing an increase in the processing load of the sensor terminal 16. The amount can be reduced.

  The present invention is not limited to the above-described embodiments, but includes various modifications and equivalent configurations within the spirit of the appended claims. For example, the above-described embodiments have been described in detail in order to explain the present invention in an easy-to-understand manner, and the present invention is not necessarily limited to those having all the configurations described. Further, part of the configuration of one embodiment may be replaced with the configuration of another embodiment. Further, the configuration of another embodiment may be added to the configuration of one embodiment. Further, a part of the configuration of each embodiment may be added / deleted / replaced with another configuration.

  Further, each of the above-described configurations, functions, processing units, processing means, etc. may be realized in hardware by designing a part or all of them in an integrated circuit, for example, and a processor realizes each function. It may be realized by software by interpreting and executing the program.

  Information such as a program, a table, and a file that realizes each function can be stored in a memory, a hard disk, a storage device such as SSD (Solid State Drive), or a recording medium such as an IC card, an SD card, and a DVD.

  Further, the control lines and information lines are shown to be necessary for explanation, and not all the control lines and information lines necessary for mounting are shown. In reality, it can be considered that almost all configurations are connected to each other.

11: cloud, 12: storage, 13: edge node, 15: control terminal, 16: sensor terminal, 200: observed value, 201: waveform of observed value in state 1, 202: waveform of observed value in state 2, 203: Sensor terminal, 205: transmission data after thinning, 206: edge node, 207: dictionary generator of state 1, 208, dictionary of state 1, 212: restored observation value, 210: restored state 1 observation value , 211: waveform of the observed value of the restored state 2, 31: dynamic dictionary generation unit, 32: dynamically generated time series dictionary, 33: state 1 restored using the time series dictionary Restoration waveform of observation value, 34: Restoration waveform of observation value of state 1 restored using time series dictionary, 404: dictionary storage unit, 405: data analysis, 406: control unit, 407: dictionary learning unit, 408: Sparse restoration Part, 409: feature extraction part, 410: thinning processing part, 411: analog front end, 412: sensor, 51: training request, 52: training signal, 53: data request, 54: data transmission, 55: dictionary learning, 56 : State change detection, 61: Training signal thinning rate notification, 62: No thinning data, 63: Data thinning rate notification, 64: Thinning data, 71: Data ratio 1/1 data, 72: Data ratio 1 / 10 data, 73: data ratio 1/30 data, 74: data ratio 1/100 data, 81: training signal, 82: thinning data, 83: pilot signal, 91: training signal, 92: data and pilot. signal

Claims (13)

A computer that receives data from a sensor terminal,
The calculator is
A dictionary learning unit that learns a dictionary composed of a matrix of base vectors and changes the dictionary in time series,
Using the learned dictionary, based on sparseness, a sparse restoration processing unit that performs sparse restoration processing that restores sensor data from data with a reduced data amount received from the sensor terminal,
A computer comprising: the sensor terminal and a control unit that transmits and receives control parameters. The computer according to claim 1,
The sparse restoration processing unit measures restoration quality from the result of the sparse restoration processing,
The computer, wherein the dictionary learning unit changes the dictionary according to the measured restoration quality. The computer according to claim 2, wherein
The control unit is
Requesting the sensor terminal to transmit a training signal according to the measured restoration quality,
A computer characterized in that after the dictionary is learned, the sensor terminal is requested to transmit the data with the reduced data amount. The computer according to claim 3,
The training signal is undecimated sensor data,
The computer, wherein the data received after learning the dictionary is data whose data amount has been reduced by thinning processing. The computer according to claim 2, wherein
The sparse restoration processing unit performs thinning processing and sparse restoration processing on unthinned data, and measures the restoration quality. The computer according to claim 2, wherein
The computer, wherein the sparse restoration processing unit measures restoration quality using the data whose data amount is reduced. The computer according to claim 6,
The computer, wherein the sparse restoration processing unit measures restoration quality from the sparseness calculated in the sparse restoration processing and the size of the objective function of the sparse restoration processing. A sensing system for collecting sensor data,
A sensor terminal that generates sensor data based on the measurement result by the sensor,
An edge node for receiving sensor data from the sensor terminal,
The edge node is
A dictionary learning unit that learns a dictionary composed of a matrix of base vectors and changes the dictionary in time series,
A sparse restoration processing unit that performs sparse restoration processing that restores sensor data from data with a reduced data amount received from the sensor terminal based on sparseness using the learned dictionary. Sensing system. The sensing system according to claim 8, wherein
The sensor terminal,
Transmitting a training signal in response to a training signal request from the edge node,
A sensing system, which transmits data in which the data amount is reduced in response to a data request from the edge node. The sensing system according to claim 8, wherein
The sensor system reduces the amount of data by thinning out sensor data at random. The sensing system according to claim 8, wherein
The sensing system, wherein the sensor terminal repeatedly transmits a training signal and the data with the reduced data amount at a predetermined cycle. A data communication method in a sensing system for collecting sensor data, comprising:
The sensing system has a sensor terminal that generates sensor data based on a measurement result by a sensor, and an edge node that receives sensor data from the sensor terminal,
The data communication method,
A dictionary learning procedure in which the edge node learns a dictionary composed of a matrix of basis vectors, and the dictionary is changed in time series;
The edge node, using a learned dictionary, based on sparseness, based on sparseness, a sparse restoration procedure for restoring sensor data from data with a reduced amount of data, and a sparse restoration procedure. Communication method. The data communication method according to claim 12, wherein
A procedure in which the sensor terminal transmits a training signal in response to a training signal request from the edge node,
The sensor terminal includes a procedure of transmitting the data with the reduced amount of data in response to a data request from the edge node.

Images