Calculation device and noise reduction method

Abstract

To reduce noise contained in measurement data.SOLUTION: A calculation device includes a data input unit that reads measurement data, a condition input unit that reads time-related data of the measurement data, a decomposition unit that decomposes the measurement data into frequency decomposition data which is data for each predetermined frequency, an inference unit that inputs each piece of the frequency decomposition data into respective neural networks subjected to learning in advance to generate each piece of noise-suppressed frequency decomposition data in which noise is suppressed, and an integration unit that integrates each piece of the noise-suppressed frequency decomposition data.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a computing device and a noise reduction method. [Background technology]

[0002] Surveys using acoustic waves are being conducted to explore marine resources and seafloor strata. In these surveys, acoustic waves are emitted toward the seafloor, and signals reflected from the seafloor and other sources are measured by a sensor as vibration data. However, the vibration data contains noise, and measures to deal with this noise are required. Patent Document 1 discloses a method for generating a signal mask from input vibration data, identifying traces contaminated by vibration noise, and identifying the frequency of signals contaminated by vibration noise. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] U.S. Patent No. 9,310,504 Summary of the Invention [Problem to be solved by the invention]

[0004] The invention described in Patent Document 1 leaves room for improvement in terms of noise reduction. [Means for solving the problem]

[0005] A calculation device according to a first aspect of the present invention comprises a data input unit that reads measurement data, a condition input unit that reads data related to time of the measurement data, a decomposition unit that decomposes the measurement data into frequency decomposition data which is data for each predetermined frequency, an inference unit that inputs each of the frequency decomposition data to respective pre-trained neural networks to generate noise-suppressed frequency decomposition data in which noise is suppressed, and an integration unit that integrates each of the noise-suppressed frequency decomposition data. A noise reduction method according to a second aspect of the present invention is a noise reduction method executed by a computer, and includes the steps of reading measurement data, reading data related to time of the measurement data, decomposing the measurement data into frequency resolution data which is data for each predetermined frequency, inputting each of the frequency resolution data into respective pre-trained neural networks to generate noise-suppressed frequency resolution data in which noise is suppressed, and integrating each of the noise-suppressed frequency resolution data. Effect of the Invention

[0006] According to the present invention, it is possible to reduce noise contained in measurement data. [Brief description of the drawings]

[0007] [Figure 1] Calculation device configuration diagram [Diagram 2] Diagram showing acquisition of measurement data through seabed survey [Diagram 3] Diagram showing measurement data [Figure 4] Schematic diagram showing the processing of measurement data by a computing device [Diagram 5] FIG. 1 is a diagram showing an example of a first frequency band data group. [Figure 6] A diagram showing an example of a neural network [Figure 7] 1 is a flowchart showing the processing of the arithmetic unit in the learning phase. [Figure 8] 1 is a flowchart showing the processing of a computing device in an inference phase. [Figure 9] FIG. 13 is a schematic diagram showing measurement data processing in Modification 1. [Figure 10] FIG. 23 shows a learning device according to a seventh modified example. [Figure 11] FIG. 13 shows an inference device according to a seventh modified example. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0008] --Embodiment-- Hereinafter, embodiments of a calculation device and a noise reduction method will be described with reference to FIGS.

[0009] FIG. 1 is a configuration diagram of the arithmetic device 1. The arithmetic device 1 includes a CPU 41 which is a central processing unit, a ROM 42 which is a read-only storage device, a RAM 43 which is a readable and writable storage device, and a non-volatile storage device 44. The CPU 41 performs calculations by expanding a program stored in the ROM 42 into the RAM 43 and executing it. The arithmetic device 1 may be realized by a field programmable gate array (FPGA) which is a rewritable logic circuit or an application specific integrated circuit (ASIC) which is an integrated circuit for a specific application, instead of the combination of the CPU 41, the ROM 42, and the RAM 43. The arithmetic device 1 may also be realized by a combination of different configurations, for example, a combination of the CPU 41, the ROM 42, the RAM 43, and an FPGA, instead of the combination of the CPU 41, the ROM 42, and the RAM 43.

[0010] The RAM 43 includes a reading unit 11, a frequency decomposition unit 12, a channel addition unit 13, a learning unit 14, an inference unit 15, and an integration unit 16, which are functions realized by the programs stored in the ROM 42. The storage device 44 stores frequency conditions 21, measurement conditions 22, measurement data 23, network parameters 24, learning data 27, a neural network 28, and noise reduction measurement data 29.

[0011] The measurement data 23 is composed of R pieces of data, namely, the first measurement data 23-1, the second measurement data 23-2, ..., the Rth measurement data 23-R. Hereinafter, the measurement data 23 is also called "fluctuation data". The network parameters 24 are composed of five parameter groups, namely, the first parameter 24-1, the second parameter 24-2, ..., the fifth parameter 24-5. The learning data 27 is composed of pseudo measurement data 27-1 and correct answer data 27-2. The correct answer data 27-2 is, so to speak, a ground truth. The pseudo measurement data 27-1 is, for example, synthetic data, and is generated by adding noise to the correct answer data 27-2. The noise-reduced measurement data 29 is a result of processing the measurement data 23 by a method described later.

[0012] The frequency conditions 21 are frequency divisions created in advance. For example, the frequency conditions 21 store division of frequencies into five frequency bands, 0-2 Hz, 2-4 Hz, 4-8 Hz, 8-16 Hz, and 16-250 Hz, which are equal to or greater than 0 Hz and less than 2 Hz. The measurement conditions 22 store conditions under which the measurement data 23 was measured. For example, the measurement conditions 22 include data such as the number of sensors used in the measurement, the measurement period, and the number of measurements. The measurement data 23 is a collection of data acquired by the sensors. The network parameters 24 are parameters used in the neural network 28 for processing the measurement data 23. The measurement data 23 and the network parameters 24 will be described in detail later. The measurement conditions 22 may be included in the measurement data 23.

[0013] 2 is a diagram showing acquisition of measurement data 23 by seabed survey. In FIG. 2, seabed 820 and boundaries 821-823 of strata of seabed 820 are indicated by solid lines, and sea surface 810 is indicated by dashed lines. Survey vessel 900 is equipped with air gun 910 and multiple hydrophones 920. Air gun 910 emits elastic waves 911 toward seabed 820, and each hydrophone 920 repeatedly records reflected waves 912 from seabed 820 and strata. The number of hydrophones 920 is not particularly limited, but in this embodiment, the number of hydrophones 920 is represented by "R", and the number of measurements by hydrophone 920 is represented by "T".

[0014] There are, for example, 650 hydrophones 920 (R=650), which are arranged in a line along the sea surface 810. These hydrophones 920 are arranged over a distance of several kilometers. In Fig. 1, subnumbers are assigned to identify the multiple hydrophones 920. The subnumbers are 1, 2, 3, ..., 650, starting from the side closest to the research vessel 900. Below, these subnumbers are also referred to as hydrophone IDs.

[0015] The hydrophone 920 receives the elastic waves 911 and the reflected waves 912 and records them as measurement data 23. The hydrophone 920 may include a non-volatile memory for storing the measurement data 23, or may include a communication device for transmitting the measurement data 23 to a recording device (not shown) mounted on the research vessel 900. The measurement data 23 of each hydrophone 920 differs slightly depending on the distance from the research vessel 900. For example, the measurement value of the hydrophone 920-1 closest to the research vessel 900 is the first measurement data 23-1, and the measurement value of the hydrophone 920-650 farthest from the research vessel 900 is the 650th measurement data 23-650.

[0016] The hydrophone 920 records the measured value at each predetermined measurement period. The measured value is, for example, in 255 steps, with the minimum value being "-127" and the maximum value being "+127". When no signal is received, the output is "0".

[0017] Fig. 3 shows the measurement data 23, with Fig. 3(a) showing the time domain and Fig. 3(b) showing the frequency domain. The horizontal axis of Fig. 3(a) is the measured position of the hydrophone 920, in other words the distance from the air gun 910, with the left end being 0 m and the right end being 2500 m. However, since there is a correspondence between the hydrophone ID, which is the identifier of the hydrophone 920, and the distance from the air gun 910, Fig. 3 shows both the distance and the hydrophone ID. From the left, the hydrophone IDs are 1, 2, 3, ..., 650.

[0018] The vertical axis of FIG. 3(a) is time, and indicates the elapsed time from the time when the air gun 910 emitted the elastic wave 911. Time passes from the top to the bottom of FIG. 3(a), with the top end being time 0 seconds and the bottom end being, for example, 3.0 seconds. When the measurement period is 1 ms, 3000 pieces of data are lined up vertically. FIG. 3(a) has T dots lined up vertically and R dots lined up horizontally, and the density of each dot indicates the magnitude of the measurement value of the hydrophone 920. Below, the values ​​on the vertical and horizontal axes of FIG. 3(a) are also referred to as "position data."

[0019] In Figure 3(a), a straight line 951 extending from the upper left to the lower right indicates an elastic wave 911. Of all the hydrophones 920, the hydrophone 920 with a hydrophone ID of "1" is placed closest to the air gun 910, and therefore reacts at the earliest time. The larger the hydrophone ID value, the greater the distance from the air gun 910, and therefore the later the time at which the elastic wave 911 is observed. The elastic wave 911 is a stronger signal than the reflected wave 912, and is therefore displayed in a relatively dark color in Figure 3(a).

[0020] In a seabed survey using a research vessel 900, it is important to measure and analyze the reflected waves 912, not the elastic waves 911. Five reflected waves 912 are shown below the elastic waves 911 in FIG. 3(a), and are shown in a lighter color because their signals are weaker than the elastic waves 911. Of the measurement results of the reflected waves 912, the one shown at the top is indicated by reference number 952. Wave noise 850 exists near IDs 1 to 200 and ID 300. The wave noise 850 hinders recognition of the reflected waves 912, and is therefore removed using a method described below.

[0021] In Fig. 3(b), the horizontal axis indicates frequency and the vertical axis indicates signal strength. In Fig. 3(b), ocean wave noise 850 appears in the region below 10 Hz.

[0022] FIG. 4 is a schematic diagram showing the processing of the measurement data 23 by the arithmetic device 1. First, the reading unit 11 reads the measurement data 23. This measurement data 23 is the graph portion in FIG. 3A, for example, RxT pieces of luminance data. Next, the reading unit 11 reads the measurement conditions 22 and assigns values ​​to the vertical and horizontal axes in FIG. 3A. For example, if the measurement period is described as "1 ms" in the measurement conditions 22, the first point of the measurement data 23 is set to 0 seconds, the second point to 1 ms, ..., and the 1001st point to 1.00 seconds. Also, if the installation interval of the hydrophones 920 is 4 meters, the distance from the air gun 910 is set to 0 m for the first measurement data 23-1, 4 m for the second measurement data 23-2, and 8 m for the third measurement data 23-3. At this point, the data shown in FIG. 3(a), that is, RxT pieces of luminance data and position data, are obtained.

[0023] Next, the frequency decomposition unit 12 refers to the frequency condition 21 and divides the measurement data 23 into data of predetermined frequency bands, for example, five frequency bands of 0-2 Hz, 2-4 Hz, 4-8 Hz, 8-16 Hz, and 16-250 Hz. The divided measurement data 23 is called first frequency band data 23F1, second frequency band data 23F2, third frequency band data 23F3, fourth frequency band data 23F4, and fifth frequency band data 23F5. Hereinafter, the first frequency band data 23F1 to fifth frequency band data 23F5 are also collectively called "frequency decomposition data."

[0024] The frequency decomposition unit 12 extracts a specific frequency signal using, for example, a known bandpass filter. The frequency decomposition unit 12 divides the RxT pieces of measurement data 23 into five frequency bands, and therefore outputs RxTx5 pieces of data.

[0025] Next, the channel adding unit 13 adds channels indicating position data to the measurement data 23 of each frequency band, i.e., a time position channel MT and a measurement position channel MR. The added channels are the same regardless of the frequency band. The channel adding unit 13 adds channels to the first frequency band data 23F1 to generate a first frequency band data set 23F1TR. The channel adding unit 13 similarly generates a second frequency band data set 23F2TR, ..., a fifth frequency band data set 23F5TR.

[0026] Fig. 5 is a diagram showing an example of the first frequency band data group 23F1TR. Here, the first frequency band data group 23F1TR is generated according to the measurement data 23 shown in Fig. 3A. The first frequency band data group 23F1TR is composed of the first frequency band data 23F1, the time position channel MT, and the measurement position channel MR. The first frequency band data 23F1, the time position channel MT, and the measurement position channel MR are all RxT matrices.

[0027] The time position channel MT is a channel that indicates the time value, which is the vertical axis in Fig. 3A. The vertical axis value increases toward the bottom of the figure, and does not change horizontally. Also, since it is difficult to handle decimal points for convenience of calculation, the unit of time is ms and the value of the time position channel MT is an integer value. The measurement period of the measurement data 23 is 1 ms, and an example is shown in which measurements are taken from 0 seconds to 3000 ms.

[0028] The measurement position channel MR is a channel that indicates the distance value, which is the horizontal axis in Figure 3A. The value on the horizontal axis increases toward the right in the figure, and there is no change in the vertical direction. Also, since it is difficult to handle decimal points for convenience of calculation, the unit of distance is set to meters and the value of the measurement position channel MR is set to an integer value. Figure 5 shows an example in which the hydrophones 920 are installed at intervals of 4 m, and 650 hydrophones are present. Returning to Figure 4, the explanation will be continued.

[0029] When the first frequency band data group 23F1TR to which the time position channel MT and the measurement position channel MR are added is generated, the inference unit 15 performs inference. Specifically, the inference unit 15 generates a first network 28-1 based on the neural network 28 and the network parameters 24, and outputs a first output 25F1 using the first frequency band data group 23F1TR as an input. Similarly, the inference unit 15 outputs a second output 25F2 to a fifth output 25F5. Hereinafter, the first output 25F1 to the fifth output 25F5 are also collectively referred to as "noise suppression frequency decomposition data".

[0030] Neural network 28 may be a neural network common to first network 29-1 to fifth network 29-5. Neural network 28 may also be an individual neural network corresponding to each of first network 29-1 to fifth network 29-5. First parameter 24-1 to fifth parameter 24-5 are parameters used in first network 29-1 to fifth network 29-5.

[0031] FIG. 6 is a diagram showing an example of the neural network 28. The neural network 28 includes seven sub-blocks 281 to 287 and three attention blocks 289. The seven sub-blocks 281 to 287 have the same configuration, and are configured to sequentially apply a two-dimensional convolution layer, a two-dimensional batch normalization layer, an activation layer, a two-dimensional convolution layer, a two-dimensional batch normalization layer, and an activation layer. The parameters used in the seven sub-blocks 281 to 287 may be the same or different. The configuration of the neural network 28 shown in FIG. 6 is merely an example, and any configuration may be used. Returning to FIG. 4, the description will be continued.

[0032] The integrating unit 16 sums the first output 25F1 to the fifth output 25F5, adds a header, and writes the noise reduced measurement data 29 to the storage device 44. For example, the integrating unit 16 writes the noise reduced measurement data 29 to the storage device 44 in the SEG-Y format. However, the integrating unit 16 may write the noise reduced measurement data 29 to the storage device 44 in the SEG-D format or another format.

[0033] 7 is a flowchart showing the processing of the arithmetic device 1 in the learning phase. In the learning phase, the network parameters 24 are updated using the learning data 27. In step S301, the reading unit 11 reads the pseudo measurement data 27-1. In step S302, the frequency decomposition unit 12 reads the frequency conditions 21, and in step S303, decomposes the pseudo measurement data 27-1 into frequency bands according to the frequency conditions 21. In the following step S304, the channel addition unit 13 generates a time position channel MT and a measurement position channel MR, and the process proceeds to step S305.

[0034] In step S305, the learning unit 14 inputs the pseudo measurement data 27-1, the time position channel MT, and the measurement position channel MR to the neural network 28 for each frequency band. In the following step S306, the learning unit 14 calculates the loss. In the following step S307, the learning unit 14 judges whether the loss calculated in step S306 is below a predetermined threshold or the number of epochs exceeds a specified value. If the learning unit 14 judges that the loss is below the predetermined threshold or the number of epochs exceeds a specified value, the process proceeds to step S308, where the network parameters 24 are stored in the storage device 44, and the process shown in FIG. 7 is terminated. If the learning unit 14 judges that the loss is not below the predetermined threshold and the number of epochs does not exceed a specified value, the process proceeds to step S309.

[0035] In step S309, the learning unit 14 calculates a convergence condition. In the following step S310, the learning unit 14 updates the network parameters 24 and returns to step S305. As described above, the learning unit 14 ends the process shown in FIG. 7 at least when the number of epochs, which is the number of times the negative determination is made in step S307 and the process returns to step S305, exceeds a specified value.

[0036] 8 is a flowchart showing the processing of the calculation device 1 in the inference phase. First, in step S321, the reading unit 11 reads the measurement data 23. In step S322, the frequency decomposition unit 12 reads the frequency conditions 21, and in step S323, the frequency decomposition unit 12 decomposes the measurement data 23 into frequency bands according to the frequency conditions 21. In the following step S324, the channel addition unit 13 generates a time position channel MT and a measurement position channel MR, and the process proceeds to step S325.

[0037] In step S325, inference unit 15 inputs measurement data 23, time position channel MT, and measurement position channel MR to neural networks for each frequency band, for example, first network 29-1 to fifth network 29-5 shown in Fig. 4. In the following steps S326 to S328, integration unit 16 combines the outputs of each neural network, adds a header, and writes noise-reduced measurement data 29 to storage device 44, thereby completing the process shown in Fig. 8.

[0038] According to the above-described embodiment, the following advantageous effects can be obtained. (1) The arithmetic device 1 includes a reading unit 11 that reads measurement data 23, a reading unit 12 that reads measurement conditions 22 that are data related to the time of the measurement data 23, and a frequency resolution unit 12 that resolves the measurement data 23 into frequency resolution data such as first frequency band data 23F1 that is data for each predetermined frequency. The arithmetic device 1 includes an inference unit 15 that inputs each of the frequency resolution data to each of pre-trained neural networks 28 to generate noise-suppressed frequency resolution data such as a first output 25F1 in which noise is suppressed, and an integration unit 16 that integrates each of the noise-suppressed frequency resolution data. Therefore, the noise included in the measurement data 23 that is fluctuation data can be reduced.

[0039] (2) The measurement data 23 is a measurement value measured at a plurality of points at a plurality of times as shown in Fig. 2 and Fig. 3. A channel adding unit 13 is provided which creates a time position channel MT indicating the time and a measurement position channel MR indicating the position. The first frequency band data group 23F1TR shown in Fig. 5 and the like, i.e., the frequency resolved data, the time position channel MT, and the measurement position channel MR, are input to the neural network 28. Therefore, by inputting the time and measurement position data of the measurement data 23 to the neural network 28, the noise contained in the measurement data 23 can be accurately reduced.

[0040] (3) The neural network 28 includes an attention mechanism, which is expected to lead to faster convergence and improved accuracy.

[0041] (4) A learning unit 14 is provided that updates the network parameters 24 of the neural network 28 using the created ground truth and pseudo-measurements created by adding noise to the ground truth. Therefore, the learning process, i.e., the updating of the network parameters 24, can be performed without using real data.

[0042] (Variation 1) 9 is a schematic diagram showing the processing of the measurement data 23 in the first modification. In the above-mentioned embodiment, the calculation device 1 includes the channel addition unit 13. However, the calculation device 1 does not need to include the channel addition unit 13. In this case, the first frequency band data 23F1 obtained by the frequency decomposition unit 12 decomposing the measurement data 23 and the like are input directly to the neural network 28. When this modification is adopted, it is necessary to change the configuration of the neural network 28 in accordance with the reduction in the number of channels and then execute the learning phase again.

[0043] (Variation 2) In the above embodiment, the channel adding unit 13 adds the time position channel MT and the measurement position channel MR. However, the channel adding unit 13 may add only one of the time position channel MT and the measurement position channel MR. When this modification is adopted, it is necessary to change the configuration of the neural network 28 in accordance with the reduction in the number of channels, and then execute the learning phase again.

[0044] (Variation 3) In the above-described embodiment, the frequency bands are divided into five bands according to the frequency condition 21. However, the frequency bands may be divided into four or less bands, or six or more bands. The lower limit of the frequency in the lowest frequency band does not have to be 0, and the upper limit of the frequency in the highest frequency band may be unlimited, in other words, infinity. Furthermore, the frequency bands do not necessarily have to be continuous. For example, the frequency bands may be divided into four bands: 0-2 Hz, 2-5 Hz, 8-16 Hz, and 16-250 Hz.

[0045] (Variation 4) In the above embodiment, the neural network 28 has an attention mechanism. However, the attention mechanism is not an essential component of the neural network 28, and the neural network 28 does not have to include the attention mechanism.

[0046] (Variation 5) In the above-described embodiment, the pseudo measurement data 27-1 created by the synthesis process constitutes the learning data 27. However, the learning data 27 may be composed of actual measurement data and ground truth data obtained by removing noise from the measurement data 23 by some means. In this case, the noise removal is performed by a human being, for example.

[0047] (Variation 6) In the above-described embodiment, the neural network 28 is created in advance, and only the parameters are determined by the learning unit 14. However, the hyperparameters and network depth of the neural network 28 may be determined by the learning unit 14. In this case, the learning unit 14 may use techniques such as Bayesian optimization, random sampling, genetic algorithms, and Latin hypercube sampling. The learning unit 14 may use any one of the above-described four techniques, or a combination of two or more of them.

[0048] (Variation 7) In the above-described embodiment, the arithmetic device 1 performs learning and inference. However, learning and inference may be performed by separate devices. For example, the functions of the arithmetic device 1 may be implemented separately in a learning device 1A and an inference device 1B.

[0049] Fig. 10 is a diagram showing a learning device 1A in Modification 7, and Fig. 11 is a diagram showing an inference device 1B in Modification 7. Learning device 1A has a configuration in which inference unit 15 and integrating unit 16 are deleted from the configuration of calculation device 1, and measurement data 23 and noise-reduced measurement data 29 are deleted from storage device 44. Inference device 1B has a configuration in which learning unit 14 is deleted from the configuration of calculation device 1, and learning data 27 is deleted from storage device 44.

[0050] Learning device 1A generates network parameters 24 in the same manner as in the embodiment. Network parameters 24 generated by learning device 1A are copied to inference device 1B. Inference device 1B performs inference using network parameters 24 on measurement data 23 in the same manner as in the embodiment, and generates noise-reduced measurement data 29.

[0051] In each of the above-mentioned embodiments and modifications, the configuration of the functional blocks is merely an example. Some functional configurations shown as separate functional blocks may be integrated, or a configuration shown in one functional block diagram may be divided into two or more functions. In addition, some of the functions of each functional block may be provided by other functional blocks.

[0052] In the above-described embodiments and modifications, the program in the arithmetic device 1 is stored in the ROM 42, but the program may be stored in the storage device 44. The arithmetic device 1 may also be provided with an input / output interface (not shown), and the program may be read from another device via the input / output interface and a medium available to the arithmetic device 1 when necessary. The medium here refers to, for example, a storage medium that is detachable from the input / output interface, or a communication medium, i.e., a wired, wireless, optical, or other network, or a carrier wave or digital signal that propagates through the network. Some or all of the functions realized by the program may be realized by a hardware circuit or an FPGA.

[0053] The above-mentioned embodiments and modifications may be combined with each other. Although various embodiments and modifications have been described above, the present invention is not limited to these. Other aspects that are conceivable within the scope of the technical idea of ​​the present invention are also included in the scope of the present invention. [Explanation of symbols]

[0054] 1: Arithmetic device 11: Reading section 12: Frequency decomposition section 13: Channel Addition Section 14: Learning Department 15: Reasoning part 16: Integration Department 21: Frequency conditions 22: Measurement conditions 23: Measurement data 23F1: First frequency band data 23F1TR: First frequency band data set 24: Network parameters 27: Training data 27-1: Pseudo-measurement data 27-2: Correct data 28:Neural Networks 29: Noise reduction measurement data 289: Attention Block 920: Hydrophone MR: Measurement position channel MT: Time position channel

Claims

1. a data input section for reading measurement data; a condition input unit for reading data relating to time of the measurement data; A decomposition unit that decomposes the measurement data into frequency decomposition data, which is data for each predetermined frequency; an inference unit that inputs each of the frequency resolution data to a pre-trained neural network to generate noise-suppressed frequency resolution data; An integration unit that integrates each of the noise suppression frequency decomposition data.

2. 2. The computing device according to claim 1, the measurement data are measurements taken at multiple points in time, a channel adding unit that creates a time position channel indicating the time and a measurement position channel indicating the point, A computing device, the neural network receiving the frequency resolved data, the time position channel, and the measurement position channel.

3. 2. The computing device according to claim 1, The neural network includes an attention mechanism.

4. 2. The computing device according to claim 1, The arithmetic device further includes a learning unit that updates parameters of the neural network using the created ground truth and pseudo-measured values ​​created by adding noise to the ground truth.

5. 1. A computer implemented method for noise reduction, comprising: Reading the measurement data, Reading data relating to time of the measurement data; Decomposing the measurement data into frequency decomposition data, which is data for each predetermined frequency; inputting each of the frequency decomposition data into a respective neural network that has been trained in advance to generate noise-suppressed frequency decomposition data; and combining each of the noise-suppressed frequency decomposition data.

Images