Signal sending device, signal reconstruction processing device, and signal transmission system

Abstract

In the present invention, a sensing unit 10 comprises: a measurement part 12 that compresses a target signal on the basis of an observation matrix Φ to obtain a first signal; a communication part 15 that sends the first signal obtained by the measurement part 12 to an external reconstruction processing unit 20; and a switching signal output part 16 that outputs a switching signal. When the measurement part 12 receives the switching signal, the measurement part 12 compresses the target signal with lower compression than the first signal to obtain a second signal, and a first communication part sends the obtained second signal to the external reconstruction processing unit 20 in place of the first signal. Due to this configuration, provided is a signal transmission technology with which power consumption can be reduced.

Description

Signal transmitting device, signal restoration processing device, and signal transmission system

[0001] The present invention relates to a signal transmission technique that enables low power consumption.

[0002] In recent years, with the spread of the Internet of Things (IoT), there has been a growing demand for sensing devices that sense, for example, the state of a living body and wirelessly transmit the sensed biosignals to a computer (edge ​​computer). Since biosignal sensing devices are typically wearable devices worn on the body, miniaturization and long-term operation are required. Therefore, even with a small battery, it is necessary to continue operating for a long period of time, and there is a strong demand for the development of low-power consumption technologies for sensing devices. Therefore, as described in Non-Patent Documents 1 to 4, compressed sensing technologies have been proposed that can significantly reduce power consumption by compressing electroencephalogram (EEG) signals while sensing them, and their incorporation into many sensing devices is being considered.

[0003] FIG. 6 illustrates the basic process of compressed sensing. (A) illustrates a signal vector x acquired by sensing expressed as a dictionary matrix Ψ and a sparse vector s (i.e., containing many elements that are zero or can be considered to be zero). (B) illustrates how a compressed signal y is obtained from the signal vector x using an observation matrix Φ used to compress the signal. (C) illustrates the relationship between the compressed signal y in (B) and the sparse vector s using the sensing matrix θ when the sensing matrix θ is Φ × Ψ. In FIG. 6, the shading of each square determined by rows and columns conveniently represents the signal level, with white squares representing the zero level. More specifically, in compressed sensing, a signal vector x having n elements is multiplied by an m × n (m < n) observation matrix Φ to obtain a compressed signal y having m elements (see FIG. 6B). Generally, it is impossible to accurately reconstruct x from y and Φ under the condition of m < n. However, when x can be expressed as the product of a dictionary matrix Ψ and a sparse vector s (see FIG. 6A), various algorithms exist for inferring a sparse vector s from a compressed vector y and a sensing matrix θ, and a restored signal ^x can be obtained by multiplying the estimated vector ^s with the dictionary matrix Ψ. Well-known restoration algorithms include BSBL (Block Sparse Bayesian Learning) (Non-Patent Document 5) and OMP (Orthogonal Matching Pursuit). Whether restoration is possible using compressed sensing depends on whether the sensed signal can be converted into a sparse vector. In other words, the existence of a dictionary matrix Ψ that can convert into a vector with high sparsity is important. For example, natural signals, such as biological signals, often have fixed signal patterns, and it is known that they tend to be converted into sparse vectors using a specific dictionary matrix.

[0004] Compressed sensing is an important technology for reducing power consumption, but as mentioned above, it is a type of lossy compression, so maintaining and improving restoration accuracy is required. However, since there is a trade-off between restoration accuracy and restoration time, shortening the restoration time is a challenge 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 it has the issue of long computational time.

[0005] Daisuke Kanemoto, Shun Katsumata, Masao Aihara, and Makoto Ohki, “Framework of Applying Independent Component Analysis After Compressed Sensing for Electroencephalogram Signals,” in Proc. IEEE Biomed. Circuits Syst. Conf. (BioCAS), Oct. 2018, pp. 145-148.Daisuke Kanemoto, Shun Katsumata, Masao Aihara, and Makoto Ohki, “Compressed Sensing Framework Applying Independent Component Analysis after Undersampling for Reconstructing Electroencephalogram Signals,” IEICE Trans. Fundamentals, vol. E103-A, no. 12, pp. 1647-1654, Dec. 2020.Yuki Okabe, Daisuke Kanemoto, Osamu Maida, and Tetsuya Hirose, “Compressed Sensing EEG Measurement Technique with Normally Distributed Sampling Series,” IEICE Trans. Fundamentals, vol.E105-A, no.10,pp. 1429-1433, Oct. 2022.Shun Katsumata, Daisuke Kanemoto, and Makoto Ohki, “Applying Outlier Detection and Independent Component Analysis for Compressed Sensing EEG Measurement Framework,” in Proc. IEEE Biomed. Circuits Syst. Conf. (BioCAS), Oct. 2019, pp. 1-4.Z.Zhang, TP Jung, S. Makeig, and BD Rao, “Compressed sensing of EEG for wireless telemonitoring with low energy consumption and inexpensive hardware,” IEEE Trans. Biomed. Eng., vol.60, no.1, pp. 221-224, Jan. 2013.

[0006] Based on the above background, the inventor proposed a method that enables high-precision and high-speed restoration even when using BSBL as a restoration algorithm by utilizing previously acquired signals as a dictionary (Patent Application No. 2022-28896, filing date February 28, 2022).

[0007] However, when applying this method to low-volume, high-mix products such as IoT devices or user-dependent biosignals, a sensing device that can perform appropriate compression and restoration processing on various detection signals needs to be equipped with an optimal dictionary for each detection signal. Furthermore, even when the measurement target, i.e., the detection signal itself, fluctuates over time, or undergoes gradual changes or transitions, it is necessary to maintain appropriate sensing without sacrificing restoration accuracy, even in the face of such fluctuations, changes, and transitions (hereinafter referred to as "fluctuations") and user dependency. Furthermore, there is a need for further improvements in compressed sensing processing and technology for reducing power consumption when transmitting captured signals to the restoration side.

[0008] The present invention has been made in view of the above, and has an object to provide a signal transmission device, a signal restoration processing device, and a signal transmission system that enable low power consumption in compressed sensing processing and external transmission.

[0009] The signal transmission device of the present invention comprises a measurement unit that compresses a target signal based on an observation matrix and captures a first signal, a first communication unit that transmits the first signal captured by the measurement unit to the outside, and a switching signal output unit that outputs a switching signal, wherein when the measurement unit receives the switching signal, it compresses the target signal at a lower compression rate than the first signal and captures a second signal, and the first communication unit transmits the captured second signal to the outside.

[0010] According to the present invention, the measurement unit compresses the target signal via an observation matrix, captures it as a first signal, and transmits it to the outside via the first communication unit. The measurement unit also compresses the target signal at a lower compression rate than the first signal and captures the second signal. Upon receiving a switching signal from the switching signal output unit, the first communication unit transmits the second signal to the outside. Therefore, the measurement unit and the first communication unit perform compression to the first signal and external transmission of the compressed signal during normal operation, enabling low-power operation. Furthermore, although power consumption temporarily increases due to the low-compression process and transmission process when capturing a usable second signal, for example, at a timing related to updating the restoration dictionary matrix, the device can operate with low power consumption during other normal operations. Therefore, the overall operation can be performed with low power consumption while maintaining, for example, the restoration accuracy of the detection signal. Note that compression at a lower compression rate than the first signal includes no compression.

[0011] Furthermore, the signal restoration device according to the present invention includes a second communication unit that receives a signal from a first communication unit of the signal transmission device; a dictionary memory that records a dictionary matrix; a restoration algorithm unit that uses the observation matrix and the dictionary matrix to determine an inferred signal from the first signal among the signals received by the second communication unit, and determines a product of the determined inferred signal and the dictionary matrix to derive a restored signal; and a dictionary generation unit that generates a new dictionary matrix from the second signal among the signals received by the second communication unit, and stores the new dictionary matrix in the dictionary memory.

[0012] According to the present invention, the restoration algorithm unit calculates a predicted signal from a first signal using an observation matrix and a dictionary matrix, and then calculates the product of the calculated predicted signal and the dictionary matrix to derive a restored signal. Meanwhile, the dictionary generation unit generates a new dictionary matrix from a second signal and stores it in a dictionary memory. In this way, high restoration accuracy for the restored signal is maintained by generating and updating the dictionary matrix from the second signal.

[0013] In addition, the signal transmission system of the present invention comprises the signal transmitting device and the signal restoration processing device connected by a communication unit, and this configuration enables low power consumption while maintaining high restoration accuracy.

[0014] According to the present invention, it is possible to provide a signal transmitting device and a signal transmission system that can reduce power consumption, and a signal restoration processing device that can maintain the restoration accuracy of a detected signal regardless of changes in the target over time.

[0015] 1 is a block diagram showing a first embodiment of a compressed sensing system. It is a block diagram functionally illustrating a portion of the configuration of the compressed sensing system of FIG. 1, where (A) is a diagram of the sensing unit side and (B) is a diagram of the restoration processing unit side. It is a flowchart showing an example of the procedure for dictionary update processing executed by the control unit of the restoration processing unit. It is a diagram briefly explaining the relationship between the discretized detection signal x and the dictionary matrix Ψ, and more specifically, it shows an example of the configuration of the dictionary matrix Ψ having a set number of columns formed from the discretized detection signal x, where (A) is a diagram of a dictionary consisting of 20 columns (set number of columns) when sorted, and (B) is a diagram of a smaller dictionary consisting of 11 columns (set number of columns) selected from the 20 columns of the discretized detection signal x in (A) taking into consideration the frequency of appearance, etc. It is a block diagram showing a second embodiment of a sensing unit. 1A and 1B are diagrams illustrating the basic processing of compressed sensing, in which (A) represents a signal vector x acquired by sensing using a dictionary matrix Ψ and a sparse vector s (i.e., containing many elements that are 0 level or can be regarded as 0 level), (B) is a diagram illustrating how a compressed signal y is obtained from the signal vector x using an observation matrix Φ used to compress the signal, and (C) is a diagram illustrating the relationship between the compressed signal y in (B) and the sparse vector s using the sensing matrix θ when the sensing matrix θ = Φ × Ψ.

[0034]

[0035]

[0036]

[0037]

[0038]

[0039]

[0040]

[0041]

[0042]

[0043]

[0044]

[0045]

[0046]

[0047]

[0048]

[0049]

[0050]

[0051]

[0052]

[0053]

[0054]

[0055]

[0056]

[0057]

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067]

[0068]

[0069]

[0070]

[0071]

[0072]

[0073]

[0074]

[0075]

[0076]

[0077]

[0078]

[0079]

[0080]

[0081]

[0082]

[0083]

[0084] [008

[0016] 1 is a block diagram showing a first embodiment of a compressed sensing system 1 applied as a signal transmission device according to the present invention. The compressed sensing system 1 includes a sensing unit 10 that is attached to a living body, for example, to measure the state of a measurement target in a wearable manner, and a restoration processing unit 20 that incorporates a processor (such as a CPU). First, an overview of the sensing unit 10 and the restoration processing unit 20 will be described.

[0017] The sensing unit 10 includes a sensor 11, a measurement unit 12 having an analog circuit 13 equipped with an amplifier and filter for preprocessing and a compression unit 14 consisting of, for example, an ADC (Analog Digital Converter), and a communication unit 15 including an antenna capable of short-range wireless communication, such as Bluetooth (registered trademark) or Wi-Fi, between the sensing unit 10 and the restoration processing unit 20. The sensing unit 10 also includes a switching signal output unit 16 that outputs a switching signal to the measurement unit 12 to switch modes (described below), and a setting unit 17 for mode switching, which is used as needed. Although not shown, the sensing unit 10 includes a built-in power source, such as a battery (e.g., a secondary battery) or a power generation element. The built-in battery or power generation element enables the sensing unit 10 to perform remote sensing with respect to the restoration processing unit 20. Functionally, the sensor 11 may be included in the measurement unit 12.

[0018] The restoration processing unit 20 includes a communication unit 21 including an antenna for transmitting and receiving signals to and from the communication unit 15, a restoration algorithm unit 22 for generating a restored signal from a compressed signal received via the communication unit 21, and a dictionary memory 23 for storing various information applied to the restoration process, such as a dictionary matrix Ψ and a sensing matrix θ (=Φ×Ψ). The observation matrix Φ is a transformation matrix applied in compressed sensing. The restoration processing unit 20 also includes a control unit 24 for updating the dictionary matrix Ψ and an operation unit 25 for manually instructing mode switching from the outside, as needed. The control unit 24 includes a processor (e.g., a CPU) and a memory unit 24a for storing a control program for selectively switching between compressed sensing mode and dictionary generation mode. By executing the control program, the control unit 24 functions as a dictionary generation unit 241 and a mode switching processing unit 242, which will be described later.

[0019] Returning to the sensing unit 10, the sensor 11 can be adapted to various detection targets, and in this embodiment, the sensor 11 has electrodes for measuring biosignals, such as electroencephalogram (EEG) signals (electrical signals). The sensor 11 may be, for example, a pair of electrodes attached to the human body, or a predetermined pair of electrodes pre-mounted in a distributed manner on headgear worn on the head and placed in contact with the scalp.

[0020] In this embodiment, the analog circuit 13 and compression unit 14 of the measurement unit 12 are switched between a compressed sensing mode and a dictionary generation mode (described later) when a switching signal from the switching signal output unit 16 is input to the operation switching control terminals 13a, 14a. Note that the circuit state of the measurement unit 12 when operating in the compressed sensing mode is conveniently referred to as a first circuit 112 and includes the analog circuit 13 and the compression unit 14, while the circuit state when operating in the dictionary generation mode is conveniently referred to as a second circuit 112' and includes the analog circuit 13' and the compression unit 14'.

[0021] As will be described later, when the switching signal output unit 16 receives a switching instruction signal from the restoration processing unit 20 or from the setting unit 17, it outputs a switching signal to switch the mode to the dictionary generation mode. Note that the setting unit 17 includes, for example, a timer, and when it measures the elapsed time from the start of operation of the sensing unit 10 or a predetermined repetitive time, it outputs a switching instruction to the switching signal output unit 16 and also outputs it to the mode switching processing unit 242 via the communication unit 15. In this way, even when a switching instruction signal is issued on the sensing unit 10 side, the sensing unit 10 and the restoration processing unit 20 can work together to execute the dictionary generation mode.

[0022] Here, the compressed sensing mode is a normal operation mode of the sensing unit 10 in a state in which power consumption is suppressed. In the first embodiment, the S / N ratio decreases due to circuit noise generated in the analog circuit 13, which is driven with low power consumption. However, in the compressed sensing mode, the compressor 14 compresses and acquires data at low sampling rate, and the amount of data transmitted to the restoration processing unit 20 is reduced by the compression, thereby suppressing power consumption in the preprocessing, compression processing, and transmission processing. Note that, depending on the application, the analog circuit 13 may employ a circuit with a high S / N ratio even in the compressed sensing mode.

[0023] Next, the dictionary matrix Ψ that imparts sparsity to the detection signal x will be described. When the detection signal x can be expressed as a highly sparsity sparse vector s using the dictionary matrix Ψ, there may be a high correlation between each column vector constituting the dictionary matrix Ψ and the detection signal x. For example, a signal with a series of waveforms with relatively similar characteristics, such as a biological signal, does not change significantly from moment to moment, and therefore can be considered a signal with a certain correlation between the extracted vectors. On the other hand, when the biological signal fluctuates over time, the correlation with the dictionary matrix Ψ gradually decreases, resulting in a deterioration in sparsity, and as a result, the accuracy (restoration accuracy) of the restored signal ^x decreases. Therefore, the dictionary generation mode is switched to at an appropriate time, and the most recent biological signal is acquired as a high-quality signal to be used in the dictionary, and the dictionary matrix Ψ is updated to maintain the accuracy of the restored signal ^x. The dictionary generation mode is a mode in which a predetermined number of columns of biological signals are temporarily acquired at an appropriate time with high sampling. More specifically, in the first embodiment, the dictionary generation mode operates the analog circuit 13' at high power consumption to suppress circuit noise, increase the S / N ratio, and expand the signal band. Furthermore, typically, high sampling operation is performed without compression, and while power consumption increases due to the transmission of more data to the restoration processing unit 20, data used for the dictionary is acquired with high accuracy.

[0024] There are several methods for changing the S / N ratio of the analog circuit 13, for example, by adjusting the current flowing in the circuit. More specifically, switching between the compressed sensing mode and the dictionary generation mode in the sensing unit 10 according to the first embodiment is performed by a switching signal from the switching signal output unit 16. In the first embodiment, the circuit configuration of the analog circuit 13 having an amplifier and a filter is almost shared, and the operating mode is switched by a switching signal input to the control terminal 13 a, thereby preventing the circuit configuration from becoming larger.

[0025] There are various methods for adjusting the current flowing to the amplifier in analog circuit 13, and it is possible to employ, for example, a mode in which the value of the current flowing to the amplifier is controlled by employing a control circuit (not shown), or a mode in which the magnitude of the current value is adjusted by a binary switching signal corresponding to the mode output from switching signal output unit 16 to control terminal 13a. The amplifier in analog circuit 13 executes the compressed sensing mode with high noise but low power consumption (first power), and the amplifier in analog circuit 13' executes the dictionary generation mode with high power consumption (second power) but low noise.

[0026] Similarly to the amplifier, the filter in the analog circuit 13 is configured to increase the S / N ratio in dictionary generation mode depending on the power consumption depending on the mode. Furthermore, in the case of the filter, a binary switching signal depending on the mode can be used to control, for example, the transconductance in the filter circuit with voltage, or to switch the functional parts of the L and C elements that make up the filter. In this way, adjusting the S / N ratio and adjusting the bandpass band of the filter—for example, setting a first passband in compressed sensing mode, and a second passband wider than the first passband in dictionary creation mode—can acquire signals with a wider frequency component, thereby enabling the generation of a highly accurate dictionary. Note that mode switching for the amplifier and filter of the analog circuit 13 is not limited to both, and may be performed on at least one of them.

[0027] Even if the update frequency in the dictionary generation mode differs depending on the target, the dictionary generation mode is short compared to the overall operating period of the compressed sensing mode, so even if the dictionary generation mode is used intermittently, the increase in overall power consumption does not pose a problem. Furthermore, in the dictionary generation mode, the dictionary matrix Ψ is generated using a low-noise, high-precision signal, and this dictionary matrix Ψ is used to perform restoration processing on the detection signal x, so even if noise is mixed into the detection signal, restoration is unlikely to occur, thereby achieving a noise reduction effect in the compressed sensing mode.

[0028] Next, the compression unit 14 selectively executes a dictionary generation mode in which the target signal output from the analog circuit 13 is highly sampled at a predetermined frequency, for example, 200 Hz, which is equivalent to the Nyquist sampling rate in the case of an electroencephalogram signal, with 3 seconds per frame, and a compressed sensing mode in which the target signal is low-sampled while being thinned out to a predetermined ratio, for example, 1 / 10.

[0029] 2A, the compressor 14 has a calculation module 140 that performs a product calculation, and in compressed sensing mode, performs a calculation to multiply the sampled detection signal x by the observation matrix Φ used for compression to realize a compressed signal y as y = Φx. Note that compressed sensing may include a mode in which an ADC or the like that constitutes the compressor 14 performs a sampling operation while thinning out at a preset timing based on the observation matrix Φ.

[0030] In the dictionary generation mode, the compressor 14' outputs, in response to a switching signal, the discretized detection signal xi (see FIG. 2A) that has been discretized by uncompressed high sampling, for example, by applying a unit matrix U (a square matrix whose diagonal elements are 1 and the rest are zero) instead of the observation matrix Φ, as a dictionary update signal. Note that the high sampling in the dictionary generation mode is not limited to the uncompressed sensing described above, and may be compressed sampling with a compression ratio lower than that in the compressed sensing mode.

[0031] The calculation module 140 is not limited to a hardware module such as an ADC, but may be a software module configured using a compression algorithm. In the compressed sensing mode, information is compressed by using, as the observation matrix Φ, a matrix that realizes random undersampling, a Bernoulli matrix, or a matrix composed of Gaussian random elements. The communication unit 15 transmits the compressed signal y to the restoration processing unit 20 every time the compressed signal y is acquired, for example, every frame (3 seconds).

[0032] The sampling operation for each frame may be continuous or may be triggered by a specific signal in the EEG signal. Preferably, the frame length is set to include a time span that includes a periodic trend in the EEG signal.

[0033] Next, a more detailed description will be given of the restoration processing unit 20. The communication unit 21 receives signals from the sensing unit 10 and transmits a switching instruction signal from the restoration processing unit 20 to the switching signal output unit 16. In the compressed sensing mode, the switch SW21 guides the received compressed signal y to the restoration algorithm unit 22, while in the dictionary generation mode, it guides the received discretized detection signal xi to the dictionary generation unit 241 as a dictionary update signal.

[0034] Next, the restoration process in compressed sensing mode will be described. As shown in FIG. 2B , the restoration algorithm unit 22 includes a restoration algorithm execution module 221 and an arithmetic module 222 that performs a multiplication operation. When a compressed signal y of a frame to be processed is input to the restoration algorithm unit 22, the restoration algorithm execution module 221, which employs, for example, the BSBL algorithm, inputs the compressed signal y and a sensing matrix θ (=Φ×Ψ) to calculate a sparse estimated vector ^s. If the detection signal x can be expressed as the product of a dictionary matrix Ψ and a sparse vector s (i.e., containing many elements that are at level 0 or can be considered to be at level 0), various algorithms exist for estimating the sparse vector s from the compressed signal y and the sensing matrix θ. Examples of well-known restoration algorithms include BSBL (Block Sparse Bayesian Learning) and OMP (Orthogonal Matching Pursuit).

[0035] Next, the calculation module 222 receives the calculated estimated vector ^s and the dictionary matrix Ψ, derives a restored signal ^x corresponding to the detected signal x, and outputs it to an external device or an image display (not shown). The derived restored signal ^x is sent to an analysis unit (not shown) as needed, or stored in the dictionary memory 23 or another memory (not shown). Furthermore, if the detection target is an EEG signal, for example, the analysis is expected to be applicable not only to the healthcare and medical fields but also to a wide range of fields, including diagnosing dementia, epilepsy, sleep disorders, Alzheimer's, and other brain diseases, measuring concentration, and assessing comfort and discomfort.

[0036] Next, in the control unit 24, the dictionary generation unit 241 executes a function of generating a new dictionary matrix Ψ. The mode switching processing unit 242 outputs a switching instruction to switch the mode to the dictionary generation mode. In this embodiment, the switching instruction is issued when it is determined that a predetermined condition has been met, or may be issued without any condition as necessary. The switching instruction signal output from the mode switching processing unit 242 is guided to the switching signal output unit 16 via the communication units 21 and 15. As a result, when a switching instruction signal is issued from the restoration processing unit 20, the sensing unit 10 and the restoration processing unit 20 can be linked to execute the dictionary generation mode.

[0037] Here, the predetermined condition for generating a switching signal preferably refers to a case where the restoration accuracy of the restored signal ^x has decreased and the dictionary matrix Ψ needs to be updated. The deterioration of the restoration accuracy of the restored signal ^x can be determined by various methods. For example, the deterioration can be determined based on a deterioration in the sparsity of the estimated vector ^s, a deterioration (dispersion) in the histogram of the spectrum of the restored signal ^x, or a decrease in the correlation between the restored signal ^x and the dictionary matrix Ψ. These determinations may also include a method of comparing the current value with a threshold, or a method of determining the magnitude of the difference between two values ​​over time. In this embodiment, the mode switching processing unit 242 sequentially acquires the estimated vector ^s output from the restoration algorithm execution module 221 for each restoration and calculates a value representing the sparsity. When the difference between the previous and current values, i.e., the amount of deterioration, exceeds a threshold, the mode switching processing unit 242 outputs a switching instruction to the dictionary generation unit 241 and the switching signal output unit 16.

[0038] The following cases may also include cases where there are no conditions for outputting a switching instruction: That is, when a timer (not shown) in the control unit 24 is used to measure a predetermined elapsed time from the start of a sensing operation for a predetermined observation target, or when an external operation is received from the operation unit 25, the mode switching processing unit 242 outputs a switching instruction.

[0039] FIG. 3 is a flowchart showing an example of the dictionary update process executed by the control unit 24. A predetermined timing during each restoration process is monitored as a determination timing. When the determination timing is met (Yes in step S1), the mode switching processing unit 242 determines whether to update the dictionary matrix based on, for example, the difference between the sparsity value of the previous and current inferred vector ^s and a threshold (step S3). That is, if the difference indicating degradation exceeds the threshold, the mode switching processing unit 242 determines that the dictionary matrix Ψ should be updated (Yes in step S5). On the other hand, if the difference indicating degradation does not exceed the threshold, the mode switching processing unit 242 terminates the process, assuming that updating is not necessary (No in step S5). When the mode switching processing unit 242 determines that the dictionary matrix Ψ should be updated in step S5, it outputs an instruction to switch to the dictionary generation mode, switches the sensing unit 10 to operate in a high sampling state, acquires the discretized detection signals xi (see FIG. 2A) for a predetermined number of columns (multiple frames), and switches the signal path of the restoration processing unit 20 so that the received discretized detection signals xi are input to the dictionary generation unit 241 (step S7).

[0040] Next, the dictionary generation unit 241 generates a dictionary matrix Ψ with a preset number of columns based on the input discretized detection signals x for multiple frames (step S9), and transfers the generated dictionary matrix Ψ to the dictionary memory 23 to store it as an updated matrix (step S11). Then, the mode switching processing unit 242 executes a switching process to return the mode to the compressed sensing mode (step S13). Note that if it is not necessary to determine whether an update is necessary, for example, if the timer has counted a predetermined time, the process may proceed immediately to step S7.

[0041] In step S7, in the dictionary generation mode, the dictionary generation unit 241 continuously acquires a predetermined number of columns of the discretized detection signal x for one frame at a high sampling rate. Then, in step S9, the dictionary generation unit 241 forms a matrix by arranging one frame's worth of data as column vectors in the order of acquisition, based on the dictionary generation program. Alternatively, the dictionary generation unit 241 forms a new matrix by performing a predetermined sorting (rearrangement) in the column direction to further increase correlation. Alternatively, the dictionary generation unit 241 forms a matrix by selectively thinning out the data in consideration of factors such as frequency of occurrence. The number of columns in the dictionary matrix Ψ is equal to the number of rows in the sparse vector s. Therefore, in the above-described mode in which selective thinning is performed in consideration of factors such as frequency of occurrence, a large number of frames of the discretized detection signal x can be acquired in advance, including the number of columns to be thinned out. In this case, the number of frames of the discretized detection signal x may be acquired over a predetermined period of time, or may be acquired over a predetermined period of time that is adjustable as needed (i.e., a predetermined number of columns in which the number of acquired frames can vary). For example, the mode switching processing unit 242 may determine the degree of degradation in the sparsity of the estimated vector ^s, the degradation (dispersion) of the histogram of the spectrum of the restored signal ^x, and the degree of degradation in the correlation between the restored signal ^x and the dictionary matrix Ψ, as described above, in accordance with the degree of degradation in the restoration accuracy of the restored signal ^x, and adjust the acquisition time, i.e., the number of columns to be acquired, in accordance with the determined degree of degradation.

[0042] 4A and 4B are diagrams for briefly explaining the relationship between the discretized detection signal x and the dictionary matrix Ψ. More specifically, they show an example of the configuration of a dictionary matrix Ψ having a set number of columns formed from the discretized detection signal x, in which (A) is a diagram of a dictionary consisting of 20 columns (the set number of columns) after sorting, and (B) is a diagram of a smaller dictionary consisting of 11 columns (the set number of columns) selected from the 20 columns of the discretized detection signal x in (A) taking into account the frequency of appearance, etc.

[0043] These are executed by the dictionary generation unit 241 based on the dictionary generation program. The dictionary matrix Ψ shown in FIG. ｐ is generated by arranging the discretized detection signals x of one frame as a column vector, assuming that signals of similar detection targets are obtained. For example, here, the dictionary matrix Ψ is generated by arranging the discretized detection signals x of one frame as a column vector, with 20 column vectors x1 to x20.ｐ The 20 column vectors may be arranged in the order obtained in the dictionary generation mode, but in this example, the average frequencies of the column vectors are calculated (for example, MATLAB (registered trademark) function meanfreq) and the average frequencies are sorted and rearranged in the high-low direction. In this way, by rearranging the average frequencies in the high-low direction, i.e., by placing highly correlated signals in adjacent columns, a sparse vector s with higher sparseness (with clearer sparseness and density) can be obtained. ｐ A dictionary matrix Ψ that allows the conversion to ｐ has been created.

[0044] In addition, the occurrence frequency of the average frequency of the column vectors is calculated as a histogram, and as shown in FIG. 4B, column vectors having an average frequency with a high occurrence frequency are selectively adopted, that is, signals with high correlation are selected, thereby making it possible to obtain a smaller dictionary Ψ ｐｓ However, the sparse vector s ｐｓ In this case, the signals may be arranged according to the average frequency of the signals that appear most frequently. Alternatively, a dictionary with a higher correlation can be created by selecting frames that appear most frequently from a larger number of frames through a sorting process.

[0045] Next, a second embodiment of the sensing unit 10 will be described using FIG. 5 . The sensing unit 10 shown in FIG. 1 electrically controls operation in each mode based on a switching signal. In contrast, the measurement unit 120 shown in FIG. 5 is configured such that switches SW11 and SW12 are switched in response to a switching instruction signal to selectively operate a first circuit 121 for the compressed sensing mode and a second circuit 122 for the dictionary generation mode, which are provided in parallel. With this configuration, the analog circuit 131 and the compression unit 141, and the analog circuit 132 and the compression unit 142, are configured with circuit elements corresponding to each mode and are designed based on the supplied power (the first and second powers) and the bandpass bandwidths (the first and second passbands). This allows for operation similar to that of the measurement unit 12 shown in FIG. 1 , and also allows for the characteristics of each circuit to be individually designed to suit the mode, even if the sampling ratios of the two modes differ significantly.

[0046] The present invention can include the following aspects.

[0047] (1) The signals used to update the dictionary matrix are not limited to signals from the same object, but may be signals from objects having similar characteristics.

[0048] (2) In the above embodiment, biological signals have been used as examples of targets. However, other signals with sparseness, such as meteorological and environmental information (including cases where the signal becomes sparse when converted using a dictionary matrix), and characteristics of current values, voltage values, and other physical quantities acquired by IoT sensors and IoT devices in factories, particularly smart factories, for monitoring the operating status of various facilities, may also be included.

[0049] (3) In the above embodiment, an example was described in which filters with a band-pass band as the passband were used as the filters provided in the analog circuits 13, 13′, 131, and 132. However, the present invention is not limited to such band-pass filters, and various filters with a designed passband, such as high-pass filters, low-pass filters, and notch filters, can be used.

[0050] (4) In the above embodiment, the compressed signal obtained in the compressed sensing mode and the signal obtained while being compressed at a low compression ratio are selectively transmitted to the restoration processing unit 20. However, both sensing signals may be simultaneously acquired and transmitted to the restoration processing unit 20. A known time division or frequency division method may be employed as a method for simultaneous transmission. Regardless of whether simultaneous transmission is performed, identification information such as a flag indicating the type of signal may be added to the header of the transmission signal sequence when transmitting signals bidirectionally from the sensing unit 10 and the restoration processing unit 20. Furthermore, the communication units 15 and 21 may each have a configuration for transmitting and receiving both sensing signals.

[0051] (5) In the above embodiment, an example was described in which a signal acquired by low-compression sampling was used to update a dictionary matrix. However, the present invention is not limited to this application and may be applied to other applications. For example, by transmitting both a compressed signal acquired in compressed sensing mode and a low-compression or uncompressed signal acquired at the same time to the decompression side, and then comparing a signal restored from the compressed signal acquired in compressed sensing mode with the low-compression or uncompressed signal on the decompression side, it is possible to evaluate the quality of the signal acquired in the currently implemented compressed sensing mode. One possible evaluation method is to utilize the difference in waveform between the restored signal and the signal acquired by uncompressed sensing. Furthermore, the evaluation results can be used to evaluate whether the currently created dictionary is sufficiently usable in compressed sensing mode, the level of the circuit characteristics set on the sensing side, and the level of the compression ratio set.

[0052] (6) Furthermore, the present invention provides a sensing unit used for small-lot, high-mix production such as IoT with a dictionary generation mode function, thereby enabling the generation of dictionary matrices corresponding to signals from various targets, thereby providing a highly versatile sensing unit or system.

[0053] (7) In the above-described embodiment, the compressors 14, 14′, 141, and 142 that perform compressed sensing employ a method of reducing and compressing information by thinning out the information using an ADC. However, the present invention is not limited to such a compressed sensing method. For example, in addition to a method of compressing at a random undersampling rate, a compressed sensing method that compresses at a low sampling rate before or after AD conversion can also be employed (see FIG. 7 ).

[0054] FIG. 7 is a circuit diagram illustrating the relationship between compressed sensing methods applicable to the present invention and information transmitted from the reconstruction processing unit 20 to the sensing unit 10 in response to the update of the dictionary matrix Ψ. (A) shows a compressed sensing method utilizing random undersampling, (B) shows an analog domain compressed sensing method, and (C) shows compressed sensing systems 1A, 1B, and 1C using a digital domain compressed sampling method. The information transmitted from the reconstruction processing unit 20 to the sensing unit 10 includes triggers for mode switching, as well as instructions for improving the S / N ratio and changing the circuit passband, and information related to the update of the dictionary matrix Ψ. For example, the dictionary matrix Ψ and the observation matrix Φ are related to each other, and the sensing matrix θ, which is their product, significantly affects the signal reconstruction accuracy. In other words, once the dictionary matrix Ψ is created, the reconstruction processing unit 20 can determine the observation matrix Φ that will improve reconstruction accuracy, and the more appropriate observation matrix Φ is transmitted to the sensing unit 10.

[0055] In FIG. 7A , the amplifier / filter unit 13A included in the measurement unit 12A performs preprocessing similar to that shown in FIG. 1 , and the ADC 14A1 performs compressed sampling. Then, when the dictionary generation unit 241 updates the dictionary matrix Ψ, it creates an observation matrix Φ corresponding to the updated content and transmits the created observation matrix Φ to the ADC 14A1 as transmission information D1, causing it to be updated and written. Therefore, even if the sensed signal contains some noise or the signal bandwidth is narrowed, high-accuracy restoration is possible. In other words, the quality of the restored signal can be guaranteed even if the circuit characteristics are degraded or power consumption is reduced. The restoration processing unit 20 transmits adjustment information indicating the extent to which the circuit characteristics can be degraded as transmission information D2 to the corresponding circuit unit of the sensing unit 10 and sets it. Furthermore, in the Digital 14A2 and the communication unit 15, if accuracy can be guaranteed even if the amount of information used for processing and transmission is reduced, as described above, the restoration processing unit 20 transmits instructions to the corresponding circuit unit of the sensing unit 10 as transmission information D2 to reduce the amount of information handled. By reducing the amount of information handled in this way, further reductions in power consumption can be expected.

[0056] In FIG. 7B, the amplifier, filter, and analog compression calculation unit 134B included in the measurement unit 12B performs preprocessing and compression sampling similar to those in FIG. 1. More specifically, the amplifier, filter, and analog compression calculation unit 134B performs compression calculation on the preprocessed analog signal. Various forms of the observation matrix Φ can be used in the analog domain, including Gaussian random matrices and Bernoulli matrices. When the restoration processing unit 20 updates the dictionary matrix Ψ, it creates an observation matrix Φ that improves restoration accuracy and transmits it as transmission information D1 to the amplifier, filter, and analog compression calculation unit 134B, which updates and writes it. By transmitting the observation matrix Φ to the compression unit on the sensing unit 10 side, appropriate compression becomes possible. In FIG. 7B, the ADC 14B1 performs converter operation, and the Digital 14B2 and communication unit 15 perform the same functions as those in FIG. 7A.

[0057] In FIG. 7C, Digital (including digital compression calculation) 14C2, connected downstream of ADC 14C1 included in measurement unit 12C, performs compression sampling on the digital signal. Note that various forms of observation matrix Φ can be used in the digital domain, such as Gaussian random matrix and Bernoulli matrix. When the restoration processing unit 20 updates the dictionary matrix Ψ, it creates an observation matrix Φ that improves restoration accuracy and transmits it as transmission information D1 to Digital (including digital compression calculation) 14C2, which updates and writes it. By transmitting the observation matrix Φ to the compression unit on the sensing unit 10 side, appropriate compression becomes possible. Note that in FIG. 7C, the amplifier / filter unit 13C performs preprocessing, and the ADC 14C1 performs converter operation.

[0058] Furthermore, a mixed compression method such as those shown in Figures 7(A) to 7(C) may be adopted, and even in such a mixed method, by basically transmitting the same information as each method to the sensing unit 10, it is possible to improve the restoration accuracy and narrow down the circuit characteristics (reducing power consumption).

[0059] As described above, the signal transmission device according to the present invention includes a measurement unit that compresses a target signal based on an observation matrix and captures a first signal, a first communication unit that transmits the first signal captured by the measurement unit to the outside, and a switching signal output unit that outputs a switching signal, and it is preferable that, upon receiving the switching signal, the measurement unit compresses the target signal at a lower compression rate than the first signal and captures a second signal, and the first communication unit transmits the captured second signal to the outside.

[0060] According to the present invention, the measurement unit compresses the target signal via an observation matrix, captures it as a first signal, and transmits it to the outside via the first communication unit. The measurement unit also compresses the target signal at a lower compression rate than the first signal and captures the second signal. Upon receiving a switching signal from the switching signal output unit, the first communication unit transmits the second signal to the outside. Therefore, the measurement unit and the first communication unit perform compression to the first signal and external transmission of the compressed signal during normal operation, enabling low-power operation. Furthermore, although power consumption temporarily increases due to the low-compression process and transmission process when capturing a usable second signal, for example, at a timing related to updating the restoration dictionary matrix, the device can operate with low power consumption during other normal operations. Therefore, the overall operation can be performed with low power consumption while maintaining, for example, the restoration accuracy of the detection signal. Note that compression at a lower compression rate than the first signal includes no compression.

[0061] Furthermore, it is preferable that the measurement unit, upon receiving the switching signal, acquires the second signal instead of the first signal. With this configuration, when the second signal is acquired, the acquisition and transmission of the first signal are not performed, resulting in operation with low power consumption.

[0062] Preferably, the measurement unit includes an analog circuit that performs preprocessing on the target signal, and the analog circuit operates by being supplied with a first power when capturing the first signal, and by being supplied with a second power greater than the first power when capturing the second signal. With this configuration, the second signal requires the second power when capturing, resulting in high power consumption during this period, but the second signal is captured as a low-noise, high-precision signal, which is suitable for use externally in, for example, updating a dictionary matrix.

[0063] Preferably, the measurement unit includes an analog circuit that performs preprocessing on the target signal, and the analog circuit is set to a first passband when capturing the first signal, and is set to a second passband wider than the first passband when capturing the second signal. With this configuration, the wider second passband is set when capturing the second signal, which results in high power consumption during capture, but the second signal is detected as a signal containing frequency components within the second passband, making it suitable for applications such as updating a dictionary matrix with high correlation.

[0064] Preferably, the first communication unit outputs the switching instruction received from an external device to the switching signal output unit, and the switching signal output unit outputs the switching signal for a predetermined period of time upon receiving the switching instruction. With this configuration, a switching instruction can be given from an external device via the first communication unit.

[0065] Preferably, the device further includes a setting unit that outputs a switching instruction to the switching signal output unit under predetermined conditions, and the switching signal output unit outputs the switching signal for a predetermined period of time upon receiving the switching instruction. With this configuration, the setting unit can set the switching instruction at an appropriate timing, for example.

[0066] Furthermore, it is preferable that the signal restoration device according to the present invention comprises: a second communication unit that receives a signal from a first communication unit of the signal transmission device; a dictionary memory that records a dictionary matrix; a restoration algorithm unit that uses the observation matrix and the dictionary matrix to determine an inferred signal from the first signal among the signals received by the second communication unit, and determines a product of the determined inferred signal and the dictionary matrix to derive a restored signal; and a dictionary generation unit that generates a new dictionary matrix from the second signal among the signals received by the second communication unit, and stores the dictionary matrix in the dictionary memory.

[0067] According to the present invention, the restoration algorithm unit calculates a predicted signal from a first signal using an observation matrix and a dictionary matrix, and then calculates the product of the calculated predicted signal and the dictionary matrix to derive a restored signal. Meanwhile, the dictionary generation unit generates a new dictionary matrix from a second signal and stores it in a dictionary memory. In this way, high restoration accuracy for the restored signal is maintained by generating and updating the dictionary matrix from the second signal.

[0068] In addition, the present invention preferably includes a mode switching processing unit that outputs a switching instruction, and the second communication unit transmits the switching instruction to an external device. According to this configuration, the switching instruction is transmitted to an external device, for example, a signal transmitting device, via the second communication unit, and is used in the acquisition process of the first and second signals.

[0069] Furthermore, it is preferable that the mode switching processing unit outputs the switching instruction when the restoration accuracy of the restored signal is reduced. With this configuration, since fluctuations over time of the target signal or the like lead to a reduction in restoration accuracy, the switching instruction is output at an appropriate timing and the dictionary matrix is ​​updated, thereby improving correlation and maintaining restoration accuracy.

[0070] Preferably, the mode switching processor outputs the switching instruction when a predetermined time has elapsed since the previous dictionary generation. With this configuration, the switching instruction is output periodically or at predetermined intervals to update the dictionary matrix, thereby improving correlation and maintaining restoration accuracy.

[0071] Preferably, the mode switching processor outputs a switching instruction in response to an external operation. With this configuration, the switching instruction is output from the outside at any timing, for example, by manual operation, and the dictionary matrix is ​​updated, thereby improving correlation and maintaining restoration accuracy.

[0072] Furthermore, it is preferable that the signal transmission system according to the present invention comprises the signal transmitting device and the signal restoration processing device connected by a communication unit, and this configuration enables low power consumption while maintaining high restoration accuracy.

[0073] 1, 1A, 1B, 1C Compressed sensing system 10 Sensing unit (sensing device) 11 Sensor 12, 120, 12A, 12B, 12C Measurement unit 112, 121 First circuit 112', 122 Second circuit 13, 13', 131, 132 Analog circuit 14, 14', 141, 142 Compression unit 15 Communication unit (first communication unit) 21 Communication unit (second communication unit) 16 Switching signal output unit 20 Restoration processing unit (restoration processing device) 22 Restoration algorithm unit 23 Dictionary memory 24 Control unit 241 Dictionary generation unit 242 Mode switching processing unit SW21, SW11, SW12 Switch

Claims

1. a measurement unit that compresses a target signal based on an observation matrix and captures a first signal; a first communication unit that transmits the first signal captured by the measurement unit to an outside; a switching signal output unit that outputs a switching signal, When the measurement unit receives the switching signal, the measurement unit compresses the target signal uncompressed or at a lower compression rate than the first signal and captures a second signal; the first communication unit transmits the captured second signal to an outside; The measurement unit includes an analog circuit that performs preprocessing on the target signal, and the analog circuit sets a first S / N ratio when capturing the first signal, and sets a second S / N ratio higher than the first S / N ratio when capturing the second signal.

2. 2. The signal transmitting device according to claim 1, wherein the measuring unit receives the second signal in place of the first signal upon receiving the switching signal.

3. 2. The signal transmitting device according to claim 1, wherein the analog circuit operates by being supplied with a first power when receiving the first signal, and operates by being supplied with a second power greater than the first power when receiving the second signal.

4. 2. The signal transmission device according to claim 1, wherein the analog circuit is set to a first passband when receiving the first signal, and is set to a second passband wider than the first passband when receiving the second signal.

5. the first communication unit outputs a switching instruction received from an external device to the switching signal output unit; The signal transmitting device according to claim 1 , wherein the switching signal output unit outputs the switching signal for a predetermined period of time upon receiving the switching instruction.

6. a setting unit that outputs a switching instruction to the switching signal output unit under a predetermined condition; The signal transmitting device according to claim 1 , wherein the switching signal output unit outputs the switching signal for a predetermined period of time upon receiving the switching instruction.

7. a second communication unit that receives a signal from the first communication unit of the signal transmission device according to any one of claims 1 to 6; a dictionary memory for recording a dictionary matrix; a restoration algorithm unit that calculates a predicted signal from the first signal among the signals received by the second communication unit using the observation matrix and the dictionary matrix, and calculates a product of the calculated predicted signal and the dictionary matrix to derive a restored signal; a dictionary generation unit that generates a new dictionary matrix from the second signal among the signals received by the second communication unit and stores the new dictionary matrix in the dictionary memory.

8. a mode switching processing unit that outputs a switching instruction; The signal restoration processing device according to claim 7 , wherein the second communication unit transmits the switching instruction to an external device.

9. The signal restoration processing device according to claim 8 , wherein the mode switching processing unit outputs the switching instruction when the restoration accuracy of the restored signal is reduced.

10. The signal restoration processing device according to claim 8 , wherein the mode switching processing unit outputs the switching instruction when a predetermined time has elapsed since the previous dictionary generation.

11. The signal restoration processing device according to claim 8 , wherein the mode switching processing unit receives an external operation and outputs a switching instruction.

12. A signal transmission system comprising the signal transmitting device according to any one of claims 1 to 6 and the signal restoration processing device according to claim 7.