Signal compression device, signal restoration device and signal processing system

Abstract

A measurement signal processing unit (20) comprises: a reception part (21) for receiving a compression signal y acquired, by using an observation matrix Φ, from a time-series detection signal of a measurement target; and a restoration part (22) for restoring the received compression signal y. The restoration part (22) comprises a dictionary matrix Ψ including the observation matrix Φ and a past signal of the measurement target, obtains an estimation vector ^s by inputting the compression signal y and a sensing matrix θ to a restoration algorithm execution module (221), and derives a restoration signal ^x corresponding to a detection signal x by inputting, to a calculation module (222) for obtaining a product, the obtained estimation vector ^s and the dictionary matrix Ψ. Accordingly, restoration with high accuracy and at a high speed is performed by applying, to the dictionary matrix, a past signal of the signal of the same type as the target signal.

Description

[Technical field]

[0001] The present invention relates to a technique for restoring a compressed signal acquired by compressed sensing with high accuracy and high speed. [Background technology]

[0002] In recent years, with the spread of IoT, there has been a growing demand for sensing devices that sense the state of a living body and transmit the sensed biosignal wirelessly to a computer (edge ​​computer). In the case of devices that sense biosignals, wearable devices that are attached to the body are common, so the devices are required to be small and to operate for a long time. Therefore, even with a small battery, it is necessary to continue operating for a long time, so there is a high demand for the development of low power consumption technology for sensing devices. In this type of sensing device, the power consumption of the circuit for wirelessly transmitting the sensed signal accounts for a high proportion of the total power consumption. Therefore, today, compressed sensing technology, which can significantly reduce the power required for wireless transmission by compressing the signal while sensing, has attracted attention, and its incorporation into many sensing devices is being considered.

[0003] The principle of compressive sensing is shown in Fig. 8. In Fig. 8, the shading assigned to each cell determined by rows and columns represents the signal level for convenience, and in particular, white cells represent the 0 level. As shown in Fig. 8(A), in compressive sensing, an observation matrix Φ of m×n (m < n) is multiplied by an observation signal vector x having n elements to obtain a compressed signal vector y having m elements. Generally, under the condition of m < n, it is impossible to accurately obtain x from y and Φ. However, when x can be expressed as the product of a certain dictionary matrix Ψ and a sparse vector s (i.e., a vector containing many elements that can be regarded as the 0 level or the 0 level) (see Fig. 8(B)), if the sensing matrix θ = Φ×Ψ, there are various algorithms for restoring (estimating) the sparse vector s from the compressed signal vector y and θ (see Fig. 8(C)), and the restored signal ^x can be obtained so as to obtain the original signal vector x by the product of the dictionary matrix Ψ and the restored estimated vector ^s. As restoration algorithms, for example, BSBL (Block Sparse Bayesian Learning) shown in Non-Patent Document 1 and OMP (Orthogonal Matching Pursuit) shown in Non-Patent Documents 2 and 3 are well-known. Whether restoration is possible in compressive sensing depends on whether the sensed signal can be converted into a sparse vector. That is, the existence of a dictionary matrix that can be converted into a vector with high sparsity is important.

[0004] Compressive sensing is one of the important technologies for reducing power consumption. However, since it is a type of irreversible compression, maintaining and improving the restoration accuracy are required. However, since there is a trade-off relationship between the restoration accuracy and the restoration time, shortening the required time for restoration is an issue in order to achieve high-precision restoration. For example, the OMP restoration algorithm has a very short restoration time but cannot be expected to have high restoration accuracy. On the other hand, the BSBL restoration algorithm is a high-precision restoration method that utilizes Bayesian estimation, but there is an issue that it takes a long calculation time.

[0005] Furthermore, the applicant has proposed a low power consumption technology in which, as shown in Patent Document 1, compression processing is performed by sub-Nyquist rate sampling on the sensing device side, and mixed signals (artifacts) containing noise and the like are separated and removed from the sub-Nyquist rate sampled signals on the wireless receiving side by Independent Component Analysis (ICA), and then restoration processing with a certain degree of accuracy is performed by applying a restoration algorithm such as OMP or BSBL (for example, Non-Patent Documents 4, 5, 6). [Prior art documents] [Patent documents]

[0006] [Patent Document 1] JP 2021-115429 A [Non-patent literature]

[0007] [Non-Patent Document 1] Zhilin Zhang et al, “Extension of SBL Algorithms for the Recovery of Block Sparse Signals with Intra-Block Correlation,” IEEE Transactions on Signal Processing, vol.61, no.8, pp.2009-2015, 2013 [Non-Patent Document 2] P. Dao, XJ Li, A. Griffin, and H. Do, “K SVD dictionary learning applied in clinical EEG compressed sensing,” ATC 2018, Oct. 2018. [Non-Patent Document 3] Kotaro Nagai, Daisuke Kanemoto, and Makoto Ohki,“Applying K-SVD Dictionary Learning for EEG Compressed Sensing Framework with Outlier Detection and Independent Component Analysis,”IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences, Vol. Sep. 10th. 2021.

Outdoor Tools 4

Direct Environment 5

Outdoor Configuration6

[0008] The method of utilizing a dictionary matrix using the OMP restoration algorithm and the singular value decomposition (K-SVD) technique shown in Non-Patent Documents 2 and 3 shows some improvements over the application of a general DCT (Discrete Cosine Transform) matrix to the dictionary matrix, but does not currently reach the accuracy achieved by BSBL.

[0009] However, if high-speed restoration can be achieved in, for example, the BSBL restoration algorithm, which allows high-precision restoration, it can become a useful compressed sensing technology. In the BSBL restoration algorithm, the restoration accuracy tends to improve as the number of columns in the dictionary matrix Ψ increases. However, increasing the number of columns tends to increase the sensing matrix θ used in the restoration algorithm processing, which tends to extend the restoration time, and is affected by the trade-off between restoration accuracy and restoration time as described above.

[0010] Furthermore, Patent Document 1 describes a technique for performing compression processing by sub-Nyquist rate sampling and a technique for removing mixed artifacts (noise) by independent component analysis, but does not describe a technique for performing restoration processing with high accuracy and high speed by focusing on the dictionary matrix Ψ.

[0011] The present invention has been made in consideration of the above, and aims to provide a signal compression device, a signal restoration device, and a signal processing system that enable high-precision, high-speed restoration by applying past signals, which are the same type of signals as the target signal, to a dictionary matrix that converts the target signal into a sparse signal (having sparsity). [Means for solving the problem]

[0012] A signal restoration device according to the present invention includes a receiving unit that receives a compressed signal obtained by compressing a target signal using an observation matrix, and a restoration unit that restores the received compressed signal. The restoration unit includes the observation matrix and a dictionary matrix in which past signals, which are the same type of signal as the target signal and are previously acquired multiple times, are arranged in each column, and is characterized in that the restoration unit inputs the received compressed signal, the observation matrix, and the dictionary matrix to a restoration algorithm execution module to obtain a guess vector, and inputs the obtained guess vector and the dictionary matrix to a calculation module to obtain a product, thereby deriving a restored signal corresponding to the target signal.

[0013] Moreover, the signal compression device according to the present invention includes a compression unit that converts the target signal into a compressed signal compressed using an observation matrix, and a transmission unit that transmits the compressed signal to a signal restoration device.

[0014] The signal processing system according to the present invention includes a compression unit that converts a target signal into a compressed signal compressed using an observation matrix, and a restoration unit that restores the compressed signal compressed by the compression unit. The restoration unit includes the observation matrix and a dictionary matrix in which past signals, which are the same type of signal as the target signal and are previously acquired multiple times, are arranged in each column, and inputs the received compressed signal, the observation matrix, and the dictionary matrix to a restoration algorithm execution module to obtain a guess vector, and inputs the obtained guess vector to a calculation module that obtains a product of the dictionary matrix and the dictionary matrix to derive a restored signal corresponding to the target signal.

[0015] According to these inventions, by creating a dictionary matrix in which past signals, which are the same type of signal as the target signal and are previously acquired multiple times, are arranged in each column, a dictionary containing data that is correlated or highly correlated with the target signal can be created, and by using this dictionary, it is possible to convert the acquired target signal into a sparse vector. Moreover, since the dictionary matrix has high correlation, even a dictionary with a small number of columns can maintain the required restoration accuracy in addition to shortening the restoration time. Note that the target signal can include detection signals that can be detected by a sensor, physical quantity data containing various information existing in the natural world, and even artificially created signals. Effect of the Invention

[0016] According to the present invention, past signals of the same kind of signal as the target signal are applied to a dictionary matrix, thereby enabling high-precision and high-speed restoration. [Brief description of the drawings]

[0017] [Figure 1] FIG. 1 is a block diagram showing one embodiment of a signal measurement system according to the present invention, in which (A) is a configuration diagram showing the entire system, and (B) is an external view showing an example of a sensor and a sensing unit, showing an overview of the usage state. [Diagram 2] FIG. 2 is a block diagram functionally illustrating a portion of the configuration of the signal measurement system in FIG. 1, where (A) is a diagram of the sensing unit side, and (B) is a diagram of the measurement signal processing unit side. [Diagram 3] 10 is a flowchart showing an example of a procedure of restored signal processing executed by a control unit of the measurement signal processing unit. [Figure 4] These figures show examples of dictionary matrix configurations: (A) is a diagram of a dictionary with 20 columns after sorting, and (B) is a diagram of a smaller dictionary in which the number of columns has been reduced from 20 in (A) to 11 columns based on frequency of occurrence. [Diagram 5] 13 is a graph of test results showing the average NMSE, an index of restoration error for each dictionary, versus the compression ratio. [Figure 6] 13 is a graph of test results showing the average restoration time (for 150 frames) for each dictionary versus compression ratio. [Figure 7] FIG. 4 is a block diagram showing a second embodiment of a signal measurement system according to the present invention. [Figure 8] These figures explain the basic processing and restoration processing of compressed sensing. (A) explains how a compressed signal vector y is obtained from a signal vector x using an observation matrix Φ. (B) shows the signal vector x as the product of a dictionary matrix Ψ and a sparse vector s. (C) shows the relationship between the sensing matrix θ (=Φ×Ψ) and the estimated vector ^s corresponding to the sparse vector s from the compressed signal vector y in the restoration process. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0018] 1 is a block diagram showing an embodiment of a signal measurement system 1 according to the present invention. The signal measurement system 1 includes a sensing unit 10 that measures the state of a measurement target, and a measurement signal processing unit 20 implemented on a PC or the like. The sensing unit 10 includes a sensor 11, an analog circuit 12, a compression unit 13, and a wireless transmission unit 14. Although not shown, the sensing unit 10 can be equipped with a power source such as a battery such as a secondary battery or a power generation element. The incorporation of a battery, power generation element, etc. in the sensing unit 10 enables remote sensing with respect to the measurement signal processing unit 20.

[0019] The measurement signal processing unit 20 includes a wireless receiving section 21, a restoration section 22, and an analysis section 23, as well as a control section 24 including a processor (such as a CPU) and a storage section 25 for storing necessary data. The storage section 25 includes a restoration algorithm (realized by a computer program or a hardware circuit) required for restoration, as well as a memory area for storing each matrix data required for restoration, and a work area for storing signals received from the sensing unit 10 as necessary and temporarily storing data during processing. The control section 24 reads out the restoration algorithm from the storage section 25 and executes restoration processing as described below.

[0020] In this embodiment, the sensor 11 has electrodes for measuring biosignals, for example, electroencephalogram signals (electrical signals). When the sensor 11 is, for example, a pair of adhesive electrodes, or, for example, a configuration in which a plurality of electrodes distributed in advance on a headgear are placed so as to abut on the scalp as shown in Fig. 1(B), it is possible to measure electroencephalogram signals emitted from a desired position on the head by putting the headgear on the head. The analog circuit 12 amplifies the weak electroencephalogram signals detected by the sensor 11 with an amplifier, and is also provided with a filter circuit as necessary to block some unnecessary signals.

[0021] The compression unit 13 samples the EEG signal output from the analog circuit 12 at a predetermined frequency, for example, 200 Hz, with 3 seconds as one frame. The compression unit 13 performs sensing, i.e., compression processing, while further thinning out the signal to be sampled. As shown in FIG. 2(A), the compression unit 13 has a calculation module 131 that performs a multiplication operation, and performs a calculation to obtain the product of the sampled EEG signal x and the observation matrix Φ used for compression, to calculate the compressed signal y as y=Φx. The calculation module 131 may be configured as hardware or as a compression algorithm made of software. In addition, there is no limitation on the type of the observation matrix Φ, and it is possible to use random undersampling, matrices according to Gaussian distribution or Bernoulli distribution, etc., but when random undersampling is used, the multiplication operation of the signal vector x and the observation matrix Φ in the compression unit 13 can be replaced with an A / D conversion circuit that simply thins out the sampling points. The wireless transmitting section 14 transmits the compressed signal y to the measurement signal processing unit 20 side every time the compressed signal y is acquired, for example, every frame (every 3 seconds).

[0022] The sampling operation of each frame may be continuous, or may be started by a specific signal in the EEG signal as a trigger. The frame length is preferably set to include a time span that includes a periodic tendency in the EEG signal.

[0023] Here, the dictionary matrix Ψ that gives sparsity to the signal x will be described. A possible case in which the dictionary matrix Ψ can convert the signal x into a sparse vector s having sparsity is when each column vector constituting the dictionary matrix Ψ has a correlation with the signal x. For example, a signal having a series of waveforms with relatively similar characteristics, such as a biological signal, does not change significantly from moment to moment in terms of signal characteristics, and therefore can be considered as a signal having a certain correlation between the extracted vectors. Based on this knowledge, the inventors have noticed that if Ψ, whose column vectors are signals acquired in the past (past signals), is applied as a dictionary matrix, it may be possible to generate a sparse vector s exhibiting the required sparsity.

[0024] Returning to Fig. 1, the measurement signal processing unit 20 will be described. The wireless receiving unit 21 continuously receives the compressed signal y transmitted from the wireless transmitting unit 14 for each frame. The received compressed signal y is temporarily stored in the storage unit 25. Note that, when the restoration time is faster than the frame reception cycle and streaming reception is possible, the compressed signal y may be directly input to the restoration unit 22 for processing.

[0025] The restoration unit 22 includes a restoration algorithm execution module 221 and a calculation module 222 that performs a multiplication operation. The restoration unit 22 executes a process of deriving a restored signal ^x corresponding to an original signal x from a compressed signal y according to a flowchart showing an example of a procedure of a restoration process shown in FIG. 3, which is executed by the control unit 24. The restoration unit 22 first specifies the compressed signal y of a frame to be processed (step S1), and inputs the compressed signal y and a sensing matrix θ (=Φ×Ψ) to the restoration algorithm execution module 221, which employs, for example, the BSBL algorithm, to calculate a guess vector ^s (step S3). Next, the calculation module 222 inputs the calculated guess vector ^s and the dictionary matrix Ψ to derive a restored signal ^x corresponding to a detection signal x (step S5). Next, the presence or absence of a frame to be processed is monitored, and if there is a next frame, the same process is repeated, and if there is not, this flow is terminated (step S7). The restored signal ^x derived in step S5 is output to the analysis unit 23, or is stored in the storage unit 25 or the like.

[0026] The analysis unit 23 performs information processing on the restored signal ^x according to the purpose of analysis. For example, when the detection target is an electroencephalogram signal, the analysis is expected to be applied not only to the healthcare and medical fields but also to a wide range of fields, including diagnosis of dementia, epilepsy, sleep disorders, Alzheimer's, and other brain diseases, measurement of concentration, and judgment of comfort and discomfort. The analysis can also be performed by another analysis device using the restored signal.

[0027] Next, the relationship between the dictionary matrix Ψ and past signals will be described with reference to the simplified diagram of Fig. 4. Fig. 4 shows an example of the configuration of a dictionary matrix, where (A) is a diagram of a dictionary having 20 columns after sorting, and (B) is a diagram of a smaller dictionary in which 11 columns have been selected from the 20 columns in (A) taking into account the frequency of occurrence. The dictionary matrix Ψ shown in Fig. 4(A) p In the case where a signal similar to the detection target is obtained, past signals are arranged as column vectors to generate a dictionary matrix Ψ. For example, here, 20 column vectors x1 to x20 are used to generate a dictionary matrix Ψ pThe 20 column vectors may be arranged in the order in which they were obtained at the time of sampling, but in this example, the average frequency of the column vectors is calculated (for example, using the MATLAB (registered trademark) function meanfreq) and the average frequencies are sorted in ascending order and rearranged. In this way, by rearranging the average frequencies in descending order, that is, by placing signals with high correlation in adjacent columns, a sparse vector s with higher sparseness (with clearer sparseness and denseness) can be obtained. p A dictionary matrix Ψ that allows the conversion to p has been created.

[0028] In addition, the frequency of occurrence of the average frequency of the column vectors is calculated as a histogram, and as shown in FIG. 4(B), column vectors having an average frequency with a high frequency of occurrence are selectively adopted, that is, signals with high correlation are selected, so that a smaller dictionary Ψ ps A sparse vector s ps It is preferable to enable conversion to . It is also possible to arrange the signals according to the average frequency with a high frequency of occurrence. Alternatively, a dictionary with a higher correlation can be created by selecting frames with a high frequency of occurrence by sorting a larger number of frames.

[0029] Next, a test for verifying the restoration accuracy will be described.

[0030] <Test> (1) Test procedure Of the data used in the test, 500 frames of EEG signals (FP1-F7, sampling frequency: 200Hz) from three subjects were used as test signals, divided into frames every 3 seconds (600 samples / frame) from the CHB-MIT Scalp EEG Database (see https: / / physionet.org / content / chbmit / 1.0.0). The past signals used for the dictionary matrix were also obtained from the CHB-MIT Scalp EEG Database, but were obtained from a different subject than the test signals. All restoration algorithms used BSBL, and the observation matrix Φ used random undersampling. In this test, compression ratios of 7 to 15 were used.

[0031] (2) Dictionary matrix contents Comparative example: (K-SVD dictionary) A dictionary (5,000 columns) created using the K-SVD dictionary learning algorithm (see Non-Patent Documents 2 and 3). Created using a feature extraction method for high-dimensional data using past signals. Example 1: (Sorted Dictionary) EEGs are sorted in order of highest correlation and used as a dictionary (10,000 columns). In this test, the average frequency of each EEG is calculated, and the EEG signals are sorted in order of average frequency to create a dictionary. Example 2: (3-6Hz dictionary) Only EEG signals with an average frequency of 3-6Hz are selected from the sorted dictionary of Example 1 and made into a dictionary (number of columns: 4,140) Example 3: (4-5 Hz dictionary) Only EEG signals with an average frequency of 4-5 Hz were selected from the sorted dictionary of Example 1 and made into a dictionary (1,517 columns).

[0032] In Examples 1, 2, and 3, the MATLAB (registered trademark) function meanfreq was used to calculate the average frequency.

[0033] (3) Results The test results are shown in Fig. 5 and Fig. 6. Fig. 5 is a chart of the test results showing the average value of NMSE, which is an index of restoration error in each dictionary, versus the compression ratio. NMSE (normalized mean square error) indicates the difference between the original signal x and the restored signal ^x, and the lower the value, the higher the restoration accuracy. As is clear from Fig. 5, the sorted dictionary of "Example 1" of the present invention can suppress NMSE to about 0.05 overall, compared to the K-SVD dictionary of "Comparative Example", which is a conventional method, and it was found to be effective in improving restoration accuracy. In addition, although the NMSE of each of the dictionaries of "Example 2" and "Example 3" is somewhat inferior to the dictionary of "Example 1", the NMSE showed high accuracy compared to the K-SVD dictionary of "Comparative Example". The compression ratio (=n / m) on the horizontal axis of Figs. 5 and 6 corresponds to the thinning rate of the signal in one frame, and a high compression ratio means that a high compression is applied. In addition, the 1,500 frame average in FIG. 5 refers to the average number of detection frames for three subjects (500 frames / subject).

[0034] Figure 6 is a chart showing the test results showing the restoration time (in this test, the average of 150 frames of the restoration time for one frame) when using the K-SVD dictionary, sort dictionary, 3-6Hz dictionary, and 4-5Hz dictionary against the compression ratio. Although there was no significant relationship between the compression ratio and the restoration time for any dictionary, it was confirmed that the restoration time became shorter as the number of columns in the dictionary matrix became smaller. Note that the calculation time was measured using a PC equipped with an Intel(R) Core(TM) i7-8565U CPU and 16GB RAM, and the software used was MATLAB(registered trademark) 2021a.

[0035] (4) Discussion As shown in Figs. 5 and 6, the 3-6Hz dictionary of "Example 2" was found to have improved NMSE at all compression ratios compared to the K-SVD dictionary of "Comparative Example", and furthermore, the average restoration time was shortened by about 25%. Furthermore, the 4-5Hz dictionary of "Example 3" shows that 3 seconds of compressed EEG data could be restored in about 1 second, which is less than 3 seconds. This means that streaming reception, i.e., restoration in almost real time, is possible by achieving a higher restoration accuracy than the K-SVD dictionary of "Comparative Example" (see Fig. 5) while enabling a significant reduction in restoration time (restoring 3 seconds of EEG signal in about 1 second), and it was confirmed that the dictionary has high versatility. In addition, in the case of the 3-6Hz dictionary of "Example 2", since the restoration time is close to 3 seconds, streaming reception can be expected by slightly reducing the number of columns.

[0036] Next, a second embodiment of the present invention will be described with reference to FIG. 7. When an artifact (noise) is mixed into an electroencephalogram signal, the sparsity of the electroencephalogram signal is impaired accordingly, which may affect the restoration accuracy. Therefore, as shown in FIG. 7, a signal measurement system 1A shown in the second embodiment is provided with a random undersampling A / D conversion circuit 13A that thins out sampling points instead of the compression section 13 in the first embodiment on the sensing unit 10A side, and an artifact removal section 26 is provided in front of the restoration section 22 on the measurement signal processing unit 20A side. As is well known, the random undersampling A / D conversion circuit 13A performs thinning sampling, i.e., compression, at a frequency or timing lower than twice the maximum frequency of the signal.

[0037] When an artifact is mixed into an electroencephalogram signal, a spectrum corresponding to the artifact may appear, for example, in the frequency component domain. The artifact removal unit 26 separates the compressed signal received by the wireless receiving unit 21 into signals of multiple independent components, and further performs outlier detection processing as necessary to effectively remove the artifact, and reconstructs and outputs the signal after removal (see Patent Document 1). By removing noise as much as possible and increasing sparsity, the restoration unit 22 performs restoration processing on the signal, enabling restoration with higher accuracy.

[0038] The present invention can include the following aspects.

[0039] (1) In the above embodiment, the restoration algorithm is described as BSBL, but the restoration algorithm is not limited to BSBL, and other restoration algorithms such as ADMM (Alternating Direction Method of Multipliers), CoSaMP (Compressive Sampling Matching Pursuit), and other algorithms can be applied as long as a certain degree of restoration accuracy can be expected. In addition, the OMP restoration algorithm can also be applied depending on the application.

[0040] (2) Past signals to be applied to the dictionary matrix may be past signals detected from a third party or the person himself / herself, or may be classified by gender, age group, etc. The past signals may include the same type of target signal, i.e., a common target, as well as the same target. The past signals may contain noise and artifacts, or may be signals from which noise and artifacts have been removed. Also, a large number of columns of past signals may be acquired, and a dictionary with the required number of columns may be selected from among them each time a dictionary is created according to correlation conditions, or a dictionary created in advance may be selected.

[0041] (3) In the above embodiment, the past signals are sorted by the average frequency, but instead of or in addition to this, sorting may be performed by conditions other than frequency. For example, since correlation can be evaluated by the similarity of signal waveforms, sorting may be performed by the correlation coefficient with the reference signal.

[0042] (4) In the above embodiment, the target signal is described as a biosignal (detection signal) such as an electroencephalogram signal, but other signals that have a certain degree of periodicity or waveform similarity in whole or in part can be applied as past signals. For example, in addition to electroencephalogram signals, the present invention can be applied to various biosignals such as myoelectricity, signals with sparseness such as weather and environmental information (including cases where the signal has sparseness as a result of conversion using a dictionary matrix), and for monitoring the operation status of various facilities, for example, in a factory, particularly in a smart factory, to measure the characteristics of current values, voltage values, and other physical quantities acquired by IoT sensors and IoT devices. In this way, the present invention can be applied to compression and restoration processing of information in the natural world, that is, acquireable physical quantity data. Furthermore, the present invention can be applied to artificially created signals.

[0043] (5) In the above embodiment, the target signal is described as a one-dimensional signal, but the present invention is not limited to this and can be applied to two-dimensional signals such as image signals, and even higher-dimensional signals such as three-dimensional signals, and a highly versatile device can be proposed. In this case, the multidimensional signal can be rearranged into one dimension and treated as a detection signal, and similar compression and restoration processing can be performed.

[0044] (6) In Fig. 1 and Fig. 7, for the sake of illustration, the sensing unit 10 (10A) is shown as having the sensor 11, the compression section 13, and the wireless transmission section 14 integrated together, but instead, each section may be separated depending on the application. For example, one or more sensors 11 may be connected by wire to each of the child devices and a parent device having the compression section 13 and the wireless transmission section 14 integrated together to function as a sensing unit.

[0045] (7) In this embodiment, the measurement signal processing unit 20 is configured as a computer such as a personal computer with a built-in processor (CPU, etc.), but it can be applied to devices with built-in computers such as personal computers (PCs), server devices, smartphones, tablet devices, and other portable information processing devices. Also, a wired system can be applied instead of a wireless system depending on the application.

[0046] As described above, the signal restoration device according to the present invention includes a receiving unit that receives a compressed signal obtained by compressing a target signal using an observation matrix, and a restoration unit that restores the received compressed signal. It is preferable that the restoration unit includes the observation matrix and a dictionary matrix in which past signals that are the same type of signal as the target signal and are previously acquired multiple times and are arranged in each column, and inputs the received compressed signal, the observation matrix, and the dictionary matrix to a restoration algorithm execution module to obtain a guess vector, and inputs the obtained guess vector and the dictionary matrix to a calculation module to obtain a product, thereby deriving a restored signal corresponding to the target signal.

[0047] Moreover, it is preferable that the signal compression device according to the present invention further comprises a compression unit that converts the target signal into a compressed signal compressed using an observation matrix, and a transmission unit that transmits the compressed signal to the signal restoration device.

[0048] Moreover, the signal processing system according to the present invention includes a compression unit that converts a target signal into a compressed signal compressed using an observation matrix, and a restoration unit that restores the compressed signal compressed by the compression unit. It is preferable that the restoration unit includes the observation matrix and a dictionary matrix in which past signals that are the same type of signal as the target signal and are previously acquired multiple times and are arranged in each column, and inputs the received compressed signal, the observation matrix, and the dictionary matrix to a restoration algorithm execution module to obtain a guess vector, and inputs the obtained guess vector and the dictionary matrix to a calculation module to obtain a product, thereby deriving a restored signal corresponding to the target signal.

[0049] According to these inventions, by creating a dictionary matrix in which past signals, which are the same type of signal as the target signal and are previously acquired multiple times, are arranged in each column, a dictionary containing data that is correlated or highly correlated with the target signal can be created, and by using this dictionary, it is possible to convert the acquired target signal into a sparse vector. Moreover, since the dictionary matrix has high correlation, even a dictionary with a small number of columns can maintain the required restoration accuracy in addition to shortening the restoration time. Note that the target signal can include detection signals that can be detected by a sensor, physical quantity data containing various information existing in the natural world, and even artificially created signals.

[0050] In addition, it is preferable that the dictionary matrix arranges highly correlated signals in adjacent columns for the multiple past signals. With this configuration, by arranging highly correlated signals in order in the column direction so that they are adjacent to each other, it is possible to convert the signals into vectors with larger sparse chunks and high sparsity, and restoration accuracy is maintained.

[0051] In addition, it is preferable that the dictionary matrix is ​​arranged by selecting highly correlated signals from among the multiple past signals. With this configuration, since highly correlated signals are adopted, higher restoration accuracy is maintained.

[0052] In addition, it is preferable that the dictionary matrix is ​​arranged in descending order of average frequency for the multiple past signals. With this configuration, signals with similar average frequencies are arranged in adjacent columns, making it possible to convert the signal into a vector with larger sparse chunks and higher sparsity, thereby maintaining restoration accuracy.

[0053] In addition, it is preferable that the dictionary matrix is ​​arranged by selecting signals having a highly frequent average frequency from among the multiple past signals. With this configuration, signals having a highly frequent average frequency, i.e., signals with high correlation, are adopted, thereby maintaining higher restoration accuracy. [Explanation of symbols]

[0054] 1. Signal measurement system (signal processing system) 10 Sensing unit (signal compression device) 11 Sensors 13 Compression section 131 Calculation Module 14 Radio transmitter 20 Measurement signal processing unit (signal restoration device) 21 Radio receiving unit (receiving unit) 22 Restoration Department 221 Restoration Algorithm Execution Module 222 Calculation Module 23 Analysis Department 24 Control Unit 25 Memory section

Claims

1. a receiving unit for receiving a compressed signal obtained by compressing a target signal using an observation matrix; a restoration unit that restores the received compressed signal, the restoration unit comprises the observation matrix and a dictionary matrix in which past signals which are the same type of signal as the target signal and which are previously acquired multiple times are arranged in each column, and the dictionary matrix arranges signals which have a high correlation with the multiple past signals in adjacent columns, and the signal restoration device is characterized in that the signal restoration device inputs the received compressed signal, the observation matrix, and the dictionary matrix to a restoration algorithm execution module to obtain a predicted vector, and inputs the obtained predicted vector and the dictionary matrix to a calculation module to obtain a product, thereby deriving a restored signal corresponding to the target signal.

2. A receiving unit for receiving a compressed signal obtained by compressing a target signal using an observation matrix; a restoration unit that restores the received compressed signal, the restoration unit comprises the observation matrix and a dictionary matrix in which past signals that are the same type of signal as the target signal and that are previously acquired multiple times are arranged in each column, and the dictionary matrix is ​​arranged by selecting and arranging highly correlated signals from the multiple past signals, and the signal restoration device is characterized in that the signal restoration device inputs the received compressed signal, the observation matrix, and the dictionary matrix to a restoration algorithm execution module to obtain a predicted vector, and inputs the obtained predicted vector and the dictionary matrix to a calculation module to obtain a product, thereby deriving a restored signal corresponding to the target signal.

3. A receiving unit for receiving a compressed signal obtained by compressing a target signal using an observation matrix; a restoration unit that restores the received compressed signal, the restoration unit comprises the observation matrix and a dictionary matrix in which past signals that are the same type of signal as the target signal and that are previously acquired multiple times are arranged in each column, and the dictionary matrix arranges the multiple past signals in descending order of average frequency, and the signal restoration device is characterized in that the received compressed signal, the observation matrix, and the dictionary matrix are input to a restoration algorithm execution module to obtain a predicted vector, and the obtained predicted vector and the dictionary matrix are input to a calculation module to obtain a product, thereby deriving a restored signal corresponding to the target signal.

4. A receiving unit for receiving a compressed signal obtained by compressing a target signal using an observation matrix; a restoration unit that restores the received compressed signal, the restoration unit comprises the observation matrix and a dictionary matrix in which past signals of the same type as the target signal and which are previously acquired multiple times are arranged in each column, and the dictionary matrix is ​​arranged by selecting signals having an average frequency with a high occurrence frequency from the multiple past signals, and the signal restoration device is characterized in that the received compressed signal, the observation matrix, and the dictionary matrix are input to a restoration algorithm execution module to obtain a guess vector, and the obtained guess vector and the dictionary matrix are input to a calculation module to obtain a product, thereby deriving a restored signal corresponding to the target signal.

5. a compression unit that converts the target signal into a compressed signal by using an observation matrix; A signal compression device comprising: a transmission unit for transmitting the compressed signal to the signal restoration device according to any one of claims 1 to 4.

6. A compression unit that converts a target signal into a compressed signal by using an observation matrix; a restoration unit that restores the compressed signal compressed by the compression unit, the restoration unit comprises the observation matrix and a dictionary matrix in which past signals of the same type as the target signal and which are previously acquired multiple times are arranged in each column, and the dictionary matrix arranges signals with high correlation with the multiple past signals in adjacent columns, and the received compressed signal, the observation matrix, and the dictionary matrix are input to a restoration algorithm execution module to obtain a predicted vector, and the obtained predicted vector and the dictionary matrix are input to a calculation module to obtain a product, thereby deriving a restored signal corresponding to the target signal.

7. A compression unit that converts a target signal into a compressed signal by using an observation matrix; a restoration unit that restores the compressed signal compressed by the compression unit, the restoration unit comprises the observation matrix and a dictionary matrix in which past signals that are the same type of signal as the target signal and that are previously acquired multiple times are arranged in each column, and the dictionary matrix is ​​arranged by selecting and arranging highly correlated signals from the multiple past signals, and the signal processing system is characterized in that the received compressed signal, the observation matrix, and the dictionary matrix are input to a restoration algorithm execution module to obtain a predicted vector, and the obtained predicted vector and the dictionary matrix are input to a calculation module to obtain a product, thereby deriving a restored signal corresponding to the target signal.

Images