Device for acquiring training data and method for acquiring training data
The learning data acquisition device and method address resource constraints in edge devices by filtering and converting analog signals into low-bit digital data, ensuring effective use in machine learning models with preserved analysis features.
Patent Information
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- TOSHIBA INFORMATION SYSTEMS (JAPAN) CORPORATION
- Filing Date
- 2025-01-16
- Publication Date
- 2026-07-29
AI Technical Summary
Existing technologies for acquiring learning data for machine learning models in edge devices are inefficient due to resource limitations, leading to reduced accuracy and inability to process low-precision data effectively.
A learning data acquisition device and method that utilize a filter to limit the waveform range of analog physical signals from sensors and convert them into low-bit digital signals, ensuring they include necessary features for analysis, using a filter means and an AD converter.
Enables the use of low-precision digital signals as suitable input data for machine learning models in edge devices without excessive load, preserving essential characteristics for analysis.
Smart Images
Figure 2026122858000001_ABST
Abstract
Description
Technical Field
[0002] , , , ,
[0005] ,
[0004] , , , , , ,
[0003] ,
[0006]
[0001] This invention relates to a learning data acquisition device and a learning data acquisition method.
Background Art
[0002] In particular, when obtaining learning data used for creating a machine learning model in an AI mounted on an edge device, it is necessary to provide high-precision data as low-precision data.
[0003] Conventionally, for example, for fault detection by abnormal sounds and weather detection technologies by analysis of electromagnetic waves, etc., high-precision data using a high-precision AD converter is usually used as a learning model with a plurality of observation points. However, in an edge device, resources such as memory and arithmetic circuits for AI are limited, and a method of performing analysis with reduced accuracy internally is applied for optimization of the AI model. Patent Document 2 describes the learning process for an internal combustion engine as follows: When the angle range is changed to the advance side, the injection mode is synchronized with the change in the angle range, thereby allowing for a favorable understanding of the combustion state during transient operation. Furthermore, by variably setting the angle range depending on whether or not the output characteristics of the in-cylinder pressure sensor 26 are learned, it is possible to achieve both appropriate understanding of the combustion state and appropriate learning. Moreover, when the output characteristics of the in-cylinder pressure sensor 26 are learned, the learning process can be performed appropriately by setting the angle range to the advance side compared to when learning is not performed.
[0007] Patent Document 3 discloses a fast magnetic resonance imaging (MRCA) reconstruction method. In this method, a learning unit 22 performs learning using training data. The learning unit 22 performs deep learning to train an MRCNN 21 using training data 50 of k-space data from a high-resolution image 25 and k-space data 15 generated by downsampling the training data 50, optimizing various parameters of the MRCNN 21. In this deep learning step, the training data 50 is downsampled using an undersampling filter 40 that retains all data for frequencies lower than a predetermined spatial frequency and randomly retains data for frequencies higher than a predetermined spatial frequency, thereby generating k-space data 15. Furthermore, in this deep learning step, learning is performed using backpropagation based on the error between the training data 50 and the k-space data corresponding to the high-resolution image 25 reconstructed using the MRCNN 21.
[0008] Next, the data acquisition unit 10 operates according to the above sequence and acquires k-space data 15 based on the nuclear magnetic resonance signal (original signal) observed by magnetic resonance. This acquires k-space data 15 for obtaining a tomographic image of the subject. In this data acquisition, the number of iterations of the sequence or the number of slice positions Z to be observed is determined according to the undersampling filter 40. That is, for low-frequency components in k-space, the data is sampled without decimation, and for high-frequency components, the data is sampled in a randomly decimated state, thereby shortening the data acquisition time (imaging time).
[0009] Furthermore, the data reconstruction unit 20 uses a multi-resolution convolutional neural network (MRCNN21) to reconstruct a high-resolution image 25 from k-space data 15 based on the original signal observed by magnetic resonance. Then, MRCNN21 uses SRCNN35 to reconstruct images obtained by inverse Fourier transform from the k-space data 15, 1 / 2 data, and 1 / 4 data, respectively. Furthermore, MRCNN21 reconstructs the high-resolution image 25 by upsampling the reconstructed data, that is, by matching the data size and superimposing them.
[0010] In the invention described in Reference 3, a tomographic image of a subject is obtained from a degraded tomographic image of the subject using high-speed magnetic resonance imaging (MRI). High-speed MRI restores the tomographic image of the subject with high resolution.
[0011] Furthermore, Patent Document 4 discloses a method for reducing the amount of data in an image data frame by masking or dropping a predetermined number of bits from the least significant bit of a digital image sensor data word, and Patent Documents 5 and 6 disclose techniques related to resolution control.
[0012] The inventions described in Patent Documents 1 to 3 above involve complex processing methods and are by no means suitable for edge devices. Furthermore, while the inventions described in Patent Documents 4 to 6 can be described as techniques for reducing data accuracy, they are unsuitable as training data because they are not created by extracting the necessary features for use in training data or analysis. [Prior art documents] [Patent Documents]
[0013] [Patent Document 1] International Publication No. 2021-025092 [Patent Document 2] Japanese Patent Publication No. 2008-138581 [Patent Document 3] Japanese Patent Publication No. 2020-146449 [Patent Document 4] Special Publication No. 2006-502673 [Patent Document 5] Japanese Patent Application Publication No. 8-331543 [Patent Document 6] Japanese Patent Publication No. 2014-135992 [Overview of the project] [Problems that the invention aims to solve]
[0014] The present invention has been made in view of the above-mentioned circumstances, and its purpose is to provide a learning data acquisition device and a learning data acquisition method that are suitable for using data obtained from sensors as learning data in edge devices and the like, without losing necessary features. [Means for solving the problem]
[0015] A learning data acquisition device according to an embodiment of the present invention is a learning data acquisition device that obtains learning data to be used to create a machine learning model for detecting the occurrence of an event, by obtaining a digital signal created from analog physical signals obtained from sensors provided at one or more locations, and is characterized by comprising: a filter means that limits the waveform range of the analog physical signal obtained from the sensor to the waveform range of the analog physical signal obtained from the sensor from the waveform range of the analog physical signal when the event does not occur to the waveform range of the analog physical signal when the event occurs; and an AD converter that converts the output signal of the filter means into a digital signal of a predetermined number of bits.
[0016] A learning data acquisition method according to an embodiment of the present invention is a learning data acquisition method that obtains learning data to be used for creating a machine learning model that detects the occurrence of an event using a digital signal created from analog physical signals obtained from sensors provided at one or more locations, and is characterized in that the filter means is equipped with a range limiting step of inputting the analog physical signals obtained from the sensors provided at one or more locations and limiting the waveform range of the analog physical signals obtained from the sensors to the range of the waveform of the analog physical signals obtained from the sensors from the waveform range of the analog physical signals when the event does not occur to the waveform range of the analog physical signals when the event occurs, and an AD conversion step of converting the output signal of the filter means into a digital signal of a predetermined number of bits using an AD converter. [Brief explanation of the drawing]
[0017] [Figure 1] A functional block diagram of an event detection system configured using a conventional learning data acquisition device. [Figure 2] A functional block diagram of an event occurrence detection system configured using a learning data acquisition device according to the first embodiment of the present invention. [Figure 3] A functional block diagram of an event occurrence detection system configured using a learning data acquisition device according to the second embodiment of the present invention. [Modes for carrying out the invention]
[0018] Referring to the accompanying drawings below, a learning data acquisition device and a learning data acquisition method according to an embodiment of the present invention will be described. In each figure, the same components are denoted by the same reference numerals, and overlapping descriptions are omitted. FIG. 2 shows a functional block diagram of an event occurrence detection system configured using the learning data acquisition device according to the first embodiment of the present invention.
[0019] This embodiment is a learning data acquisition device 200 that obtains learning data used for creating a machine learning model for detecting event occurrence by using a digital signal created from an analog physical signal obtained from a sensor 101 provided at one location. The learning data acquisition device 200 includes a filter means 201 and an AD converter 202.
[0020] The filter means 201 limits the waveform range of the analog physical signal obtained from the sensor 101 from the waveform range of the analog physical signal when the event does not occur to the waveform range of the analog physical signal when the event occurs. The AD converter 202 converts the output signal of the filter means 201 into a digital signal of a predetermined number of bits.
[0021] The AD converter 202 can be configured to convert the signal into a digital signal of a predetermined number of bits from 4 bits to 8 bits. Also, the waveform range is not defined only by the signal level. That is, when the frequency of the analog physical signal obtained from the sensor 101 varies from the waveform range when the event does not occur to the waveform range when the event occurs, the filter means 201 can use a band-pass filter, a high-pass filter, or a low-pass filter.
[0022] The digital signal created as described above is given to an event occurrence device 300 composed of a microcomputer or the like as an edge device, and is used as learning data for creating a machine learning model for detecting event occurrence.
[0023] The generated digital signal is filtered by the filter means 201 to limit the waveform range of the analog physical signal obtained from the sensor 101 from the waveform range of the analog physical signal when no event occurs to the waveform range of the analog physical signal when the event occurs. Therefore, even if the low-precision (low-bit count) digital signal generated above is used as input data for a learning model in an edge device such as a microcomputer, the problem of excessive load and inability to process it will not occur.
[0024] Furthermore, the generated digital signal, filtered by the filter means 201, has a waveform range of the analog physical signal obtained from the sensor 101 that ranges from the waveform range of the analog physical signal when no event occurs to the waveform range of the analog physical signal when the event occurs, and includes the characteristic range necessary for analysis. Therefore, even though it is a low-precision (low-bit) digital signal, it can be said to be suitable data as input data for a learning model.
[0025] Figure 3 shows a functional block diagram of an event occurrence detection system configured using a learning data acquisition device according to the second embodiment of the present invention. This system uses microphones 11 and 12 as sensors to capture audio signals as analog physical signals. The outputs of microphones 11 and 12 are captured and amplified by an amplifier configured with operational amplifiers 13 and 14 at its core, and then supplied to a filtering means 210.
[0026] The filter means 210 is configured as a differential amplifier that receives two sets of analog physical signals, which are audio signals, obtained from microphones 11 and 12, which are sensors installed at two locations, and outputs a spike-shaped waveform when noise occurs. With a filter means 210 configured in this way, training data can be obtained to create a machine learning model that recognizes the occurrence of a malfunction due to abnormal noise as "when the event occurred".
[0027] This filtering means 210 limits the waveform range of the audio signal, which is an analog physical signal obtained from the microphones 11 and 12, to the range of the analog physical signal when the event does not occur, from the waveform range of the analog physical signal when the event does occur.
[0028] The AD converter 220 converts the output signal of the filter means 210 into a predetermined number of bits of digital signal. The AD converter 220 can convert the signal into a predetermined number of bits, from 4 bits to 8 bits.
[0029] The above-mentioned filter means 210 and AD converter 220 constitute a training data acquisition device, and the output of the AD converter 220 is output to the microcomputer 230, which is an edge device, as input data for the training model.
[0030] In this embodiment as well, the filter means 210 limits the waveform range of the analog physical signals obtained from the microphones 11 and 12, which are sensors, from the waveform range of the analog physical signals when no event occurs to the waveform range of the analog physical signals when the event occurs. Therefore, even if the low-precision (low-bit) digital signals created above are used as input data for the learning model in the microcomputer 230, the problem of high load and inability to process them does not occur.
[0031] Furthermore, the digital signal created by this embodiment, filtered by the filter means 210, has a waveform range of the analog physical signal obtained from the microphones 11 and 12 that extends from the waveform range of the analog physical signal when no event occurs to the waveform range of the analog physical signal when the event occurs, and includes the characteristic range necessary for analysis. Therefore, even though it is a low-precision (low-bit) digital signal, it can be said to be suitable data as input data for a learning model.
[0032] In this embodiment, the filter means 210 may be configured using a differential amplifier that receives two time-series analog physical signals obtained from a sensor installed at one location and outputs a spike-shaped waveform when noise occurs. Alternatively, analog physical signals obtained from sensors installed at three or more locations may be input, and two different combinations of these signals may be directed to the filter means 210 using differential amplifiers to obtain training data to be used in creating a machine learning model.
[0033] As described above, each embodiment is suitable for use when data obtained from a sensor is used as training data in an edge device or the like, and provides a training data acquisition device and training data acquisition method that do not lose necessary features.
[0034] Furthermore, if the filtering means of this embodiment ensures that the waveform range of the analog physical signal obtained from the sensor falls between the waveform range of the analog physical signal when no event occurs and the waveform range of the analog physical signal when the event occurs, then the digital data after digital conversion will include the characteristic range necessary for analysis, and therefore, a limiter circuit can be used as the filtering means. [Explanation of Symbols]
[0035] 11, 12 Microphone 13, 14 Op-amps 101 Sensor 200 Training data acquisition device 201, 210 Filtering means 202, 220 AD converters 230 Microcomputers 300 Event Occurrence Detection Device
Claims
1. In a learning data acquisition device that obtains learning data to be used in creating a machine learning model for detecting the occurrence of an event, using digital signals created from analog physical signals obtained from sensors installed at one or more locations, A filter means that limits the waveform range of the analog physical signal obtained from the sensor from the waveform range of the analog physical signal when the event does not occur to the waveform range of the analog physical signal when the event occurs, A converter that converts the output signal of the filter means into a digital signal of a predetermined number of bits. A learning data acquisition device characterized by comprising the following:
2. The learning data acquisition device according to claim 1, characterized in that the filtering means is configured with a differential amplifier that receives two time-series analog physical signals obtained in time series from a sensor provided at one location, or two sets of analog physical signals obtained from sensors provided at two locations, and outputs a spike-shaped waveform when noise occurs.
3. The learning data acquisition device according to claim 1, characterized in that the filtering means is a limiter circuit.
4. The learning data acquisition device according to claim 1, characterized in that the filtering means is a bandpass filter.
5. The learning data acquisition device according to claim 1, characterized in that the filtering means is a high-pass filter.
6. The learning data acquisition device according to claim 1, characterized in that the filtering means is a low-pass filter.
7. The learning data acquisition device according to claim 1, characterized in that the AD converter converts the data into a digital signal of a predetermined number of bits, from 4 bits to 8 bits.
8. A method for acquiring training data to be used to create a machine learning model for detecting the occurrence of an event, using digital signals created from analog physical signals obtained from sensors installed at one or more locations, A range limiting step involves inputting the analog physical signals obtained from sensors provided at one or more locations to a filter means and limiting the waveform range of the analog physical signals obtained from the sensors from the waveform range of the analog physical signals when the event does not occur to the waveform range of the analog physical signals when the event occurs. The AD conversion step involves using an AD converter to convert the output signal of the filter means into a digital signal of a predetermined number of bits. A method for acquiring learning data, characterized by comprising the following:
9. The learning data acquisition method according to claim 8, characterized in that the filtering means inputs two time-series analog physical signals obtained in a time series from a sensor provided at one location, or two sets of analog physical signals obtained from sensors provided at two locations, to a differential amplifier, and outputs a spike-shaped waveform when noise occurs.
10. The learning data acquisition method according to claim 8, characterized in that the filtering means uses a limiter circuit.
11. The method for acquiring training data according to claim 8, characterized in that the filtering means uses a bandpass filter.
12. The method for acquiring training data according to claim 8, characterized in that the filtering means uses a high-pass filter.
13. The method for acquiring training data according to claim 8, characterized in that the filtering means uses a low-pass filter.
14. The learning data acquisition method according to claim 8, characterized in that the AD converter converts the data into a digital signal of a predetermined number of bits, from 4 bits to 8 bits.