Method for generating training data, signal type classification system, and program

Abstract

To provide method, system and program for generating high-quality learning data for learning processing in order to implement a high-accuracy signal type classifier.SOLUTION: A method includes: acquiring radio waveform data of radio waves received from an antenna (S11); acquiring feature quantity data indicating features of a time domain or a frequency domain of the radio waveform data, the feature quantity data being data of which the quantity is smaller than the radio waveform data (S12); performing normalization processing on the feature quantity data to acquire normalized feature quantity data, and classifying the normalized feature quantity data into clusters (S13); and labeling the feature quantity data classified by clustering processing to acquire data in which the feature quantity data is associated with the labels, as learning data (S14).SELECTED DRAWING: Figure 5

Description

Technical Field

[0001] The present invention relates to a signal type classification technology for enabling the understanding of the radio wave environment situation.

Background Art

[0002] For the purpose of grasping the operating status and control of various devices, the introduction of IoT (Internet of Things) devices into indoor environments such as factories, hospitals, and commercial facilities is progressing. In such an environment, wireless communication traffic is likely to be concentrated spatially and temporally, and collisions and delays of wireless communication packets are likely to occur. Even in such a situation, stable communication is desired. In addition, in order to grasp the operating status of the entire system, information such as the congestion level and radio wave intensity in the frequency band of the monitoring target is grasped, and it is monitored whether there is a problem with the wireless quality. If a problem occurs, a system that can provide information leading to the identification of the cause is desired. In order to realize such a system, it is important to establish a signal type classification technology for enabling the understanding of the radio wave environment situation.

[0003] Conventionally, technologies for identifying noise sources have been developed. For example, Patent Document 1 discloses a technology for detecting noise (single pulses or burst pulses) generated by lightning discharges.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] In conventional technologies such as those described in Patent Document 1, it is assumed that the nature of the source of the wireless signal or electromagnetic noise is known to some extent, and the frequency spectrum of the wireless signal or electromagnetic noise generated from the source, as well as the time waveform pattern of the signal (or electromagnetic noise) (for example, single pulses or burst-like pulses), can be predicted to some extent, making it possible to identify the source.

[0006] However, in wireless environments where there are numerous sources of wireless signals and electromagnetic noise, and where the frequency spectrum and waveform patterns of these signals and electromagnetic noise cannot be predicted, it is extremely difficult to understand the state of the wireless environment by applying conventional technology.

[0007] Therefore, there is a need for technology that can appropriately understand the state of the wireless environment, even in environments where many wireless signals and electromagnetic noises whose frequency spectra and waveform patterns are difficult to predict are flying around. To realize such technology, it is effective to establish a technology (signal type classification technology) that appropriately classifies the types of wireless signals and electromagnetic noises whose frequency spectra and waveform patterns are difficult to predict. In other words, by using this technology, it is possible to understand what kinds of wireless signals and electromagnetic noises are present in a wireless environment, and as a result, the state of the wireless environment can be appropriately understood.

[0008] Furthermore, in order to realize technology that can appropriately grasp the conditions of the wireless environment, it is desirable to realize a high-precision signal type classifier (signal type classification technology) (for example, a signal type classifier realized by a machine learning model). In order to realize a high-precision signal type classifier (signal type classification technology), it is necessary to prepare high-quality training data to execute efficient learning processing.

[0009] However, there are the following challenges in preparing high-quality training data for the training process necessary to realize a highly accurate signal type classifier (signal type classification technology) (for example, a signal type classifier implemented by a machine learning model).

[0010] Firstly, when using multiple sensors (for example, wireless communication devices or sensor devices with wireless communication capabilities), there is a challenge in labeling training data (teacher data) (assigning correct labels). Even when signals from the same wave source are received by multiple sensors, the received level changes depending on the sensor's position. Also, if the gain levels of each sensor are different, the noise floor rises and the dynamic range changes. In other words, the difference in received levels at each sensor affects the clustering of data acquired by each sensor and the labeling of training data (teacher data), resulting in similar signals and noise not being placed in the same cluster, and consequently, labeling training data (teacher data) becomes difficult.

[0011] Secondly, spurious emissions such as clock leakage and internal noise from peripheral devices (unintended radio waves generated by harmonics, subharmonics, parasitic oscillations, etc., from the radio waves emitted from the transmitter (frequency components not intended in the design of the AC signal)) can alter the spectrum of the signal or noise to be detected, making it impossible to correctly detect the signal or noise to be detected. Furthermore, although spurious emissions do not vary much over time, the condition of peripheral devices can change during long-term monitoring aligned with the production cycle of factories, etc., making it difficult to always consider spurious emissions and acquire the spectrum of the signal or noise to be detected.

[0012] Therefore, in view of the above problems, the present invention aims to realize a method for generating training data that generates high-quality training data for training processing to realize a highly accurate signal type classifier (signal type classification technology) (for example, a signal type classifier realized by a machine learning model). In other words, the present invention aims to realize a method for generating training data that generates common training data (training data) from data acquired from multiple sensors (for example, feature data from spectrum and cepstrum), training data (training data) from which the influence of spurious signals has been eliminated. Furthermore, the present invention aims to realize a signal type classification system that uses a classifier (trained model) acquired by performing training processing on said training data. [Means for solving the problem]

[0013] To solve the above problems, the first invention is a method for generating training data for signal type classification processing, comprising: a radio wave waveform acquisition step; a feature acquisition step; a normalization step; a clustering step; and a training data acquisition step.

[0014] The radio wave waveform acquisition processing step acquires radio wave waveform data of the radio waves received by the antenna.

[0015] The feature acquisition processing step acquires feature data that represents the time domain or frequency domain features of the radio wave waveform data, and which is data with a smaller amount of data than the radio wave waveform data.

[0016] The normalization step involves obtaining normalized feature data by performing a normalization process on the feature data.

[0017] The clustering step classifies the normalized feature data into clusters.

[0018] The training data acquisition step involves assigning labels to the feature data classified by clustering, thereby acquiring data that associates feature data with labels as training data.

[0019] This training data generation method obtains normalized feature data, performs clustering on the obtained normalized feature data, and generates training data. Therefore, even if signals from the same wave source are received by multiple wireless communication devices at different reception levels and gain settings, they can be classified into the same cluster. As a result, this training data generation method can generate highly accurate training data (teaching data) without being affected by differences in reception levels and gain settings.

[0020] The second invention is a method for generating training data for signal type classification processing, comprising: a radio wave waveform acquisition step; a feature acquisition step; a spurious emission removal step; a clustering step; and a training data acquisition step.

[0021] The radio wave waveform acquisition processing step acquires radio wave waveform data of the radio waves received by the antenna.

[0022] The feature acquisition processing step acquires feature data that represents the time domain or frequency domain features of the radio wave waveform data, and which is data with a smaller amount of data than the radio wave waveform data.

[0023] The spurious emission removal step involves obtaining multiple data points from the radio wave waveform data in which the stationary noise component is dominant, taking the time sample average of the multiple obtained data points to obtain time sample average data, and subtracting the obtained time sample average data from the feature data to obtain stationary noise component removed feature data.

[0024] The clustering step classifies the stationary noise component-removed feature data into clusters.

[0025] The learning data acquisition step acquires, as learning data, data in which feature amount data and labels are associated with each other by attaching labels to the feature amount data classified by clustering processing.

[0026] In this learning data generation method, a time sample average spectrum is acquired by taking the time sample average of data mainly including stationary noise (spurious), and feature amount data from which the acquired time sample average data is subtracted is obtained to acquire feature amount data from which stationary noise (spurious) is excluded. Further, based on the data, feature amount data can be acquired. And in this learning data generation method, clustering processing is performed on the feature amount data from which stationary noise (spurious) is excluded to generate learning data, so that high-precision learning data (teacher data) from which the influence of stationary noise (spurious) is excluded can be generated.

[0027] A third invention is the second invention, further including a normalization processing step of acquiring normalized feature amount data obtained by performing normalization processing on the feature amount data from which stationary noise components are removed, and the clustering processing step classifies the normalized feature amount data into clusters.

[0028] Thereby, in this learning data generation method, clustering processing can be performed on the normalized feature amount data obtained by performing normalization processing on the feature amount data from which stationary noise components are removed, so that higher-precision learning data (teacher data) can be generated.

[0029] A fourth invention is any one of the first to third inventions, wherein the feature amount data is data extracted from a cepstrum.

[0030] Thereby, in this learning data generation method, learning data can be generated with the feature amount data being data extracted from a cepstrum.

[0031] The fifth invention is any of the first to third inventions, wherein the feature data is data obtained by extracting or processing the lower-order components of the cepstrum.

[0032] This method of generating training data reduces computational complexity because it extracts or processes the low-order components of the cepstrum to obtain feature data. As a result, this method enables the rapid generation of highly accurate training data (with less computation).

[0033] The sixth invention is the first or third invention, wherein the normalization step obtains the maximum value max and minimum value min of the data included in the data set subjected to normalization, and if x is the data value of the data included in the data set subjected to normalization, x' = (x - min) / (max - min) The normalization process is performed by executing the equivalent process.

[0034] This allows this training data generation method to generate training data using normalized feature data obtained through max-min normalization.

[0035] The seventh invention is the first or third invention, wherein the normalization step involves obtaining the maximum value max and minimum value min of the data in the set obtained by excluding data with values ​​smaller than a predetermined threshold from the data set subject to normalization, and if x is the data value of the data included in the data set subject to normalization, x' = (x - min) / (max - min) The normalization process is performed by executing the equivalent process.

[0036] This allows this training data generation method to convert (normalize) data that falls within a range above a predetermined threshold and below a maximum value (max) into data whose possible range is [0,1]. By doing so, this training data generation method eliminates the problem of data without signal or noise being emphasized, which would prevent proper clustering.

[0037] The eighth invention is a signal type classification system that performs signal type classification using a trained model of a signal type classifier obtained by performing a training process using training data generated by a training data generation method according to any of the first to seventh inventions, and comprises a wireless communication device and a host device.

[0038] Wireless communication equipment, A radio wave waveform acquisition process step that acquires radio wave waveform data of radio waves received by an antenna, A feature acquisition process step to acquire feature data which is data that shows the characteristics of the time domain or frequency domain of the radio wave waveform data, and which has a smaller data volume than the radio wave waveform data, A normalization process step to obtain normalized feature data obtained by performing a normalization process on the feature data, A transmission step in which normalized feature data is sent to a host device, Execute this.

[0039] The host device is A receiving step that receives normalized feature data transmitted from a wireless communication device, The normalized feature data acquired by the wireless communication device is input into a trained model of the signal type classifier, and the label acquisition step is to obtain the label of the normalized feature data. Execute this.

[0040] In this signal type classification system 100, the wireless communication device acquires normalized feature data (normalized feature data), and the host device performs signal classification processing (label acquisition) on the acquired normalized feature data. Therefore, even if signals from the same wave source are received by multiple wireless communication devices at different reception levels and set gains, the correct signal type classification result can be obtained for the data subject to signal type classification processing.

[0041] The ninth invention is a signal type classification system that performs signal type classification using a trained model of a signal type classifier obtained by performing a training process using training data generated by a training data generation method which is one of the first to seventh inventions, and comprises a wireless communication device and a host device.

[0042] Wireless communication equipment, A radio wave waveform acquisition process step that acquires radio wave waveform data of radio waves received by an antenna, A feature acquisition process step to acquire feature data which is data that shows the characteristics of the time domain or frequency domain of the radio wave waveform data, and which has a smaller data volume than the radio wave waveform data, The spurious emission removal process involves obtaining multiple data points from radio wave waveform data in which the stationary noise component is dominant, taking the time-sample average of the multiple obtained data points to obtain time-sample average data, and subtracting the obtained time-sample average data from the feature data to obtain feature data with the stationary noise component removed. A transmission step in which steady-state noise component removal feature data is sent to a host device, Execute this.

[0043] The host device is A receiving step of receiving steady-state noise component removal feature data transmitted from a wireless communication device, The process involves inputting stationary noise-removed feature data obtained by a wireless communication device into a trained model of a signal type classifier to obtain labels for the stationary noise-removed feature data, and a label acquisition step. Execute this.

[0044] This signal classification system obtains a time-sample mean spectrum by taking the time-sample mean of data that mainly contains stationary noise (spurious signals). By subtracting the obtained time-sample mean spectrum from the feature data, it obtains feature data from which stationary noise (spurious signals) has been removed. Furthermore, feature data can be obtained based on this data. In this signal classification system, signal classification processing (label acquisition) is performed on the feature data from which stationary noise (spurious signals) has been removed, so the influence of stationary noise (spurious signals) can be appropriately eliminated. For example, even if the status of peripheral devices changes and the state of spurious signals changes, it is possible to obtain the correct signal classification result for the data targeted for signal classification processing.

[0045] The tenth invention is the ninth invention, wherein the host device monitors the confidence level of the classification result by the trained model of the signal type classifier, and if the rate at which the confidence level falls below a predetermined value becomes higher than a predetermined value over a certain period of time, it outputs a command to the wireless communication device to reacquire time sample average data.

[0046] Then, when the wireless communication device receives a command from the host device to reacquire the time sample average data, it executes a time sample average data update step in which it acquires multiple data points from the radio wave waveform data in which the steady noise component is dominant, and acquires time sample average data by taking the time sample average of the multiple acquired data points.

[0047] In this signal type classification system, the accuracy (confidence) of the prediction results is monitored during the prediction process. If the accuracy deteriorates, it is determined that the situation of stationary noise (spurious emissions) has changed, and the time-sample mean spectrum is updated. Spurious emission removal is then performed using the updated time-sample mean spectrum, so that feature data (cepstrum data) free from stationary noise (spurious emissions) can always be obtained. As a result, even if the situation of stationary noise (spurious emissions) changes, this signal type classification system can perform highly accurate prediction processing (signal type classification processing) without being affected by the stationary noise (spurious emissions).

[0048] The eleventh invention is a program for causing a computer to execute a training data generation method which is one of the first to seventh inventions.

[0049] This makes it possible to realize a program that causes a computer to execute a method for generating training data that has the same effect as any of the first to seventh inventions. [Effects of the Invention]

[0050] According to the present invention, it is possible to realize a method for generating training data that generates high-quality training data for training processing to realize a highly accurate signal type classifier (signal type classification technology) (for example, a signal type classifier realized by a machine learning model). Furthermore, it is possible to realize a signal type classification system using a classifier (trained model) obtained by performing training processing with said training data. [Brief explanation of the drawing]

[0051] [Figure 1] A schematic diagram of the signal type classification system 100 according to the first embodiment. [Figure 2] A schematic diagram of the wireless communication device S1_1 according to the first embodiment. [Figure 3] A schematic diagram of the feature acquisition unit 13 and time sample average data acquisition unit 14 of the wireless communication device S1_1 according to the first embodiment. [Figure 4] A sequence diagram showing the sequence of events from the start of the data sampling process in the training data generation process (processing in training data generation mode) executed by the signal type classification system 100. [Figure 5] A flowchart of the training data generation process performed by the signal type classification system 100. [Figure 6] A flowchart of the training data generation process performed by the signal type classification system 100. [Figure 7] A flowchart of the training data generation process performed by the signal type classification system 100. [Figure 8] A diagram illustrating spurious emission elimination processing. [Figure 9] A diagram illustrating spurious emission elimination processing. [Figure 10] A diagram illustrating spurious emission elimination processing. [Figure 11] A diagram illustrating spurious emission elimination processing. [Figure 12] A schematic diagram of the signal type classification system 100 during prediction. [Figure 13] A flowchart of the prediction process (signal type classification process) performed by the signal type classification system 100. [Figure 14] A flowchart of the prediction process (signal type classification process) performed by the signal type classification system 100. [Figure 15] A flowchart of the prediction process (signal type classification process) performed by the signal type classification system 100. [Figure 16] A diagram showing the CPU bus configuration. [Modes for carrying out the invention]

[0052] [First Embodiment] The first embodiment will be described below with reference to the drawings.

[0053] <1.1: Configuration of the signal type classification system> Figure 1 is a schematic diagram of the signal type classification system 100 according to the first embodiment.

[0054] Figure 2 is a schematic diagram of the wireless communication device S1_1 according to the first embodiment.

[0055] Figure 3 is a schematic diagram of the feature acquisition unit 13 and the time sample average data acquisition unit 14 of the wireless communication device S1_1 according to the first embodiment.

[0056] As shown in Figure 1, the signal type classification system 100 comprises N (N: natural number) wireless communication devices S1_1 to S1_N, a host device H1, and N wireless communication devices S1_1 to S1_N (N: natural number).

[0057] (1.1.1: Wireless communication equipment) As shown in Figures 1 and 2, the wireless communication device S1_1 includes an antenna Ant_S1, an RF processing unit 11, an IQ data acquisition unit 12, a feature acquisition unit 13, a time sample average data acquisition unit 14, a first communication interface 15, and a control unit 16 for the wireless communication device.

[0058] Antenna Ant_S1 is an antenna for receiving radio waves (RF signals) radiated (transmitted) from an external source.

[0059] The RF processing unit 11 receives an external RF signal via the antenna Ant_S1, performs RF processing for reception (RF demodulation, AD conversion, etc.) on the received RF signal, and obtains the RF-processed signal Sig0 (RF demodulated signal Sig0 (e.g., baseband OFDM signal)). The RF processing unit 11 then outputs the RF-processed signal Sig0 to the IQ data acquisition unit 12. The RF processing unit 11 receives a control signal CTL1 from the wireless communication control unit 16 and processes the signal according to the control signal CTL1. The RF processing unit 11 also sets the frequency band to be processed for RF processing (e.g., center frequency fo, frequency band Δf) according to the control signal CTL1, and obtains the RF demodulated signal Sig0 of the RF signal in the frequency range set above by performing RF processing for reception.

[0060] The IQ data acquisition unit 12 receives the control signal CTL2 output from the wireless communication control unit 16 and the signal Sig0 output from the RF processing unit 11. In accordance with the control signal CTL2, the IQ data acquisition unit 12 acquires data of the I component signal (common-mode component signal) (I component data) and data of the Q component signal (orthogonal component signal) (Q component data) from the signal Sig0 output from the RF processing unit 11, and outputs the acquired data as data D1 to the feature acquisition unit 13.

[0061] The feature acquisition unit 13 receives the control signal CTL3 output from the wireless communication control unit 16 and the data D1 output from the IQ data acquisition unit 12. The feature acquisition unit 13 performs a feature acquisition process on the data D1 output from the IQ data acquisition unit 12 according to the control signal CTL3 and acquires a predetermined feature (for example, a spectrum or cepstrum). The feature acquisition unit 13 then outputs the data containing the acquired feature as data D2 to the first communication interface 15. When the predetermined feature is a cepstrum, the feature acquisition unit 13 is configured as shown in Figure 2. That is, the feature acquisition unit 13 includes a frame processing unit 131, a cepstrum calculation unit 132, and a normalization processing unit 133.

[0062] The frame processing unit 131 receives data D1 output from the IQ data acquisition unit 12 and performs frame processing on data D1. Specifically, the frame processing unit 131 stores (stores and holds) data for the number of samples to be subjected to the Discrete Fourier Transform (DFT) from data D1, which is time-series data. Then, once the frame processing unit 131 has acquired the number of samples to be subjected to the Discrete Fourier Transform, it outputs this data as data D11 to the cepstrum calculation unit 132.

[0063] As shown in Figure 3, the cepstrum calculation unit 132 includes a DFT processing unit 1321, a subtractor 1322, a log calculation unit 1323, and an inverse DFT processing unit 1324.

[0064] The DFT processing unit 1321 receives the data D11 output from the frame processing unit 131 and performs a Discrete Fourier Transform (DFT) on the data D11. The DFT processing unit 1321 outputs the data after the Discrete Fourier Transform as data D12 to the subtractor 1322.

[0065] The subtractor 1322 receives data D12 output from the DFT processing unit 1321 and data Df_ave (time sample average spectrum) output from the time sample average data acquisition unit 14. The subtractor 1322 subtracts data Df_ave (time sample average spectrum) from data D12 (frequency domain data). Then, the subtractor 1322 outputs the data after the subtraction process as data D13 to the log calculation unit 1323.

[0066] The Log calculation unit 1323 receives the data D13 output from the subtractor 1322 and performs a logarithmic operation (a process to calculate the logarithm) on the data D13. Then, the Log calculation unit 1323 outputs the processed data as data D14 to the inverse DFT processing unit 1324.

[0067] The inverse DFT processing unit 1324 receives data D14 output from the log calculation unit 1323 and performs an inverse discrete Fourier transform on data D14 to obtain cepstrum data (cepstrum sequence). Then, the inverse DFT processing unit 1324 extracts a sequence of predetermined dimensions (for example, 16 dimensions) from the obtained cepstrum sequence, excluding the lower 0th dimension, to obtain feature data (cepstrum data). The inverse DFT processing unit 1324 then outputs the data obtained above as data D15 to the normalization processing unit 133.

[0068] The normalization processing unit 133 receives the data D15 output from the inverse DFT processing unit 1324 of the cepstrum calculation unit 132 and performs normalization processing (for example, max-min normalization) on the data D15. Then, the normalization processing unit 133 outputs the normalized data as data D2 to the first communication interface 15.

[0069] The time sample average data acquisition unit 14 receives the control signal CTL4 output from the wireless communication control unit 16 and performs time sample average data acquisition processing according to the control signal CTL4. As shown in Figure 3, the time sample average data acquisition unit 14 includes a spectrum acquisition unit 141, a time sample average spectrum calculation unit 142, and a subtraction unit 143.

[0070] The spectrum acquisition unit 141 receives data D1 output from the IQ data acquisition unit 12, performs frequency transformation processing (for example, Fourier transform) on data D1 (time-series data), and acquires the spectrum (frequency domain data) corresponding to data D1 (time-series data). The spectrum acquisition unit 141 then outputs the data including the acquired spectrum as data Df1 to the time-sample average spectrum calculation unit 142 and the subtraction unit 143.

[0071] The time-sample-averaged spectrum calculation unit 142 receives data Df1 output from the spectrum acquisition unit 141 and data D_H transmitted from the host device H1 (data transmitted from the host device H1 received via the first communication interface 15 and output from the first communication interface 15 to the time-sample-averaged data acquisition unit 14). Based on data D_H, the time-sample-averaged spectrum acquisition unit 14 acquires the time-sample-averaged spectrum and outputs the data including the acquired time-sample-averaged spectrum as data Df_ave to the subtraction unit 143 and the subtractor 1322 of the cepstrum calculation unit 132 of the feature acquisition unit 13.

[0072] The subtraction unit 143 receives data Df1 output from the spectrum acquisition unit 141 and data Df_ave output from the time sample average spectrum calculation unit 142. The subtraction unit 143 subtracts data Df_ave from data Df1 and outputs the resulting data as data Df_m to the first communication interface 15 (data Df_m is transmitted to the host device H1 via the first communication interface 15).

[0073] The first communication interface 15 is a communication interface for transmitting and receiving data with an external device, for example, via a wired network (including a communication path compliant with a predetermined serial bus standard (e.g., USB or PCI Express)) or a wireless network. The first communication interface 15 is a communication interface for communicating commands, data, etc., between the wireless communication device S1_1 and the host device H1 (for example, a communication interface for USB communication).

[0074] The first communication interface 15 receives data D2 output from the feature acquisition unit 13 and data Df_rm output from the time sample average data acquisition unit 14, converts the input data into a format that can be communicated via a wired or wireless network, and transmits it to the host device H1.

[0075] Furthermore, the first communication interface 15 outputs the command Cmd1 received from the host device H1 to the wireless communication control unit 16. Also, the first communication interface 15 outputs the data D_H received from the host device H1 to the time sample average data acquisition unit 14.

[0076] The wireless communication control unit 16 is a functional unit for controlling each functional unit of the wireless communication device S1_1. It generates a predetermined control signal according to a command Cmd1 received from the host device H1 via the first communication interface 15, and controls each functional unit by outputting the generated control signal to each functional unit. For example, the wireless communication control unit 16 generates a control signal CTL1 for controlling the RF processing unit 11 and outputs the control signal CTL1 to the RF processing unit 11. The wireless communication control unit 16 also generates a control signal CTL2 for controlling the IQ data acquisition unit 12 and outputs the control signal CTL2 to the IQ data acquisition unit 12. The wireless communication control unit 16 also generates a control signal CTL3 for controlling the feature acquisition unit 13 and outputs the control signal CTL3 to the feature acquisition unit 13. The wireless communication control unit 16 also generates a control signal CTL4 for controlling the time sample average data acquisition unit 14 and outputs the control signal CTL4 to the time sample average data acquisition unit 14.

[0077] The wireless communication devices S1_2 to S1_N each have the same configuration as described above (the same configuration as wireless communication device S1_1). The signal type classification system 100 may contain one or more wireless communication devices.

[0078] (1.1.2: Host device) As shown in Figure 1, the host device H1 includes a second communication interface 21, a host device control unit 22, a ROM 23, a RAM 24, a buffer for acquired data 25, a storage unit 26, a third communication interface 27, and a bus Bus_H1. The second communication interface 21, the host device control unit 22, the ROM 23, the RAM 24, the buffer for acquired data 25, the storage unit 26, and the third communication interface 27 are connected to the bus Bus_H1, and can send and receive data, commands, etc., to and from each other via the bus Bus_H1. Note that some or all of the above-mentioned functional units of the host device H1 do not necessarily have to be connected by bus, and may be directly connected.

[0079] The second communication interface 21 is a communication interface for transmitting and receiving data with an external device via, for example, a wired or wireless network (including a communication path compliant with a predetermined serial bus standard (e.g., USB or PCI Express)). Furthermore, the second communication interface 21 is a communication interface for communicating commands, data, etc., between the wireless communication device S1_1 and the host device H1 (for example, a communication interface for USB communication).

[0080] The second communication interface 21 receives data transmitted from the wireless communication devices S1_1 to S1_N. The second communication interface 21 then outputs the received data to the data acquisition buffer 25. The second communication interface 21 also generates transmission data for sending the command Cmd1 generated by the host device control unit 22 to a destination specified by the host device control unit 22 (one or more of the wireless communication devices S1_1 to S1_N), and transmits the said transmission data to the destination.

[0081] The host device control unit 22 is a functional unit for controlling each functional unit of the host device H1. It generates control signals for controlling each functional unit and outputs the generated control signals to the corresponding functional unit to control it. The host device control unit 22 also generates a command Cmd1 for controlling one or more of the wireless communication devices S1_1 to S1_N, and transmits the generated command Cmd1 to the corresponding (specified destination) wireless communication device via the second communication interface 21.

[0082] Furthermore, the host device control unit 22 has a mode switching function that switches between (1) learning data generation mode, (2) learning mode, and (3) prediction mode. The host device control unit 22 sets the operating mode of the host device H1 to one of the learning data generation mode, learning mode, or prediction mode.

[0083] The host device control unit 22 controls each functional unit so that (1) in the learning data generation mode, the host device H1 executes the learning data generation process (the process of generating learning data). The host device control unit 22 also controls each functional unit so that (2) in the learning mode, the host device H1 executes the learning process (the learning process of a classifier that classifies signal types from feature data). The host device control unit 22 also controls each functional unit so that (3) in the prediction mode, the host device H1 executes the prediction process (for example, the signal type classification process).

[0084] Furthermore, the host device control unit 22 reads data (instructions, codes, etc.) stored in the ROM 23 and executes programs, etc., stored in the ROM. In addition, the host device control unit 22 stores predetermined data, codes, etc., in the RAM 24 and executes predetermined processing (programs, etc.) by reading the data, codes, etc., stored in the RAM 24 at predetermined timings.

[0085] Furthermore, in learning data generation mode, the host device control unit 22 performs clustering on the data acquired from the wireless communication device S1_k (k: natural number, 1 ≤ k ≤ N), and transmits the data including the result Rst_cls as data D_H to the wireless communication device S1_k that acquired the data subject to clustering via the second communication interface 21. Also, in learning data generation mode, the host device control unit 22 receives data Df_m acquired by the wireless communication device S1_k (k: natural number, 1 ≤ k ≤ N) via the second communication interface 21.

[0086] Furthermore, in prediction mode, the host device control unit 22 monitors the steady-state noise, and if it determines that the steady-state noise conditions have changed, it generates data Det_chng indicating that the steady-state noise conditions have changed, and transmits the data containing this data Det_chng as data D_H to the wireless communication device S1_k (k: natural number, 1 ≤ k ≤ N) via the second communication interface 21.

[0087] ROM23 is a read-only memory that stores programs, libraries, instructions, code, etc., executed by the host device H1. The data stored in ROM23 is read by the host device control unit 22.

[0088] RAM24 is a random access memory that stores data, commands, codes, etc., for executing predetermined processes on the host device H1. The data, commands, codes, etc., stored in RAM24 are read out by the host device control unit 22.

[0089] The acquired data buffer 25 is a buffer for storing the transmission data from the wireless communication devices S1_1 to S1_N received via the second communication interface 21. The acquired data buffer 25 may include, for example, the same number of buffers as the number of wireless communication devices (e.g., buffers Buf_1 to Buf_N in Figure 1) to store and manage data separately for each wireless communication device S1_1 to S1_N. Alternatively, the acquired data buffer 25 may consist of a single buffer, within which the storage area is divided into the same number of memory areas as the number of wireless communication devices for memory management.

[0090] The memory unit 26 is a device for storing data. For example, learning data acquired by the host device H1 is stored in the memory unit 26.

[0091] The third communication interface 27 is a communication interface for transmitting and receiving data with an external device, for example, via a wired or wireless network (e.g., LAN). The host device H1 communicates data with an externally located server, for example, via the third communication interface 27.

[0092] <1.2: Operation of the signal type classification system> The operation of the signal type classification system 100 configured as described above will be explained below with reference to the diagrams.

[0093] Figure 4 is a sequence diagram showing the sequence of events from the start of the data sampling process in the training data generation process (processing in training data generation mode) executed by the signal type classification system 100.

[0094] Figures 5 to 7 are flowcharts of the training data generation process performed by the signal type classification system 100.

[0095] Figures 8 to 11 are diagrams illustrating the spurious emission elimination process.

[0096] In the following explanation, for the sake of clarity, we will describe an example where two wireless communication devices S1_1 and S1_2, and a host device H1 are installed in a confined space (for example, inside a factory), with N=2, and (1) wireless communication device S1_1 acquires data of radio signals and electromagnetic noise in the 920MHz band, and (2) wireless communication device S1_2 acquires data of signals and electromagnetic noise in the 920MHz band.

[0097] Furthermore, the operation of the signal type classification system 100 will be explained below, divided into (1) processing in learning data generation mode (training data generation processing), (2) processing in learning mode (learning processing), and (3) processing in prediction mode (prediction processing).

[0098] (1.2.1: Processing in training data generation mode (training data generation process)) First, we will explain the processing (training data generation process) of the signal type classification system 100 in training data generation mode.

[0099] (Step S11): In step S11 (see Figure 5), data sampling (the process of sampling data with each wireless communication device) is performed. Specifically, the following processes are performed. The data sampling process will be explained with reference to the sequence diagram in Figure 4.

[0100] (Step SS1): As shown in Figure 4, in step SS1, the host device H1 performs an initialization process. Specifically, the process is executed as follows.

[0101] The host control unit 22 of the host device H1 generates a command Cmd1(fo=920MHz,Δf=15MHz) to the wireless communication device S1_1 to obtain data of wireless signals and electromagnetic noise in a frequency band (920MHz band) with a center frequency fo=920MHz and a frequency band Δf=15MHz, and transmits data including the command Cmd1(fo=920MHz,Δf=15MHz) to the wireless communication device S1_1, which is the recipient of the command, via the second communication interface 21.

[0102] Furthermore, the host control unit 22 of the host device H1 generates a command Cmd1(fo=2.4GHz,Δf=15MHz) to the wireless communication device S1_2 to acquire data of wireless signals and electromagnetic noise in a frequency band (2.4GHz band) with a center frequency fo=920MHz and a frequency band Δf=15MHz, and transmits data including the command Cmd1(fo=2.4GHz,Δf=15MHz) to the wireless communication device S1_2, which is the recipient of the command, via the second communication interface 21.

[0103] (Step SS1.1): The wireless communication device S1_1 receives communication data including the command Cmd1 (fo=920MHz, Δf=15MHz) transmitted from the host device H1. The first communication interface 15 of the wireless communication device S1_1 extracts the command Cmd1 (fo=920MHz, Δf=15MHz) transmitted from the host device H1 from the received communication data and outputs the extracted command Cmd1 (fo=920MHz, Δf=15MHz) to the control unit 16 for the wireless communication device.

[0104] The control unit 16 for the wireless communication device S1_1 generates a control signal CTL1 instructing the RF processing unit 11 to receive wireless signals (and electromagnetic noise) in the frequency band of fo=920MHz, Δf=15MHz in accordance with the command Cmd1 (fo=920MHz, Δf=15MHz), and outputs the control signal CTL1 to the RF processing unit 11.

[0105] The RF processing unit 11 of the wireless communication device S1_1 sets the receivable frequency band to a frequency band of fo=920MHz, Δf=15MHz according to the control signal CTL1.

[0106] After performing the above frequency band setting process, the wireless communication control unit 16 of the wireless communication device S1_1 transmits an Ack signal to the host device H1 via the first communication interface 15, indicating that the processing corresponding to the command Cmd1 received from the host device H1 has been completed.

[0107] (Step SS1.2): The wireless communication device S1_2 receives communication data including the command Cmd1 (fo=920MHz, Δf=15MHz) transmitted from the host device H1. The first communication interface 15 of the wireless communication device S1_2 extracts the command Cmd1 (fo=920MHz, Δf=15MHz) transmitted from the host device H1 from the received communication data and outputs the extracted command Cmd1 (fo=2.4GHz, Δf=15MHz) to the wireless communication control unit 16.

[0108] The control unit 16 for the wireless communication device S1_2 generates a control signal CTL1 instructing the RF processing unit 11 to receive wireless signals (and electromagnetic noise) in the frequency band of fo=920MHz, Δf=15MHz in accordance with the command Cmd1 (fo=920MHz, Δf=15MHz), and outputs the control signal CTL1 to the RF processing unit 11.

[0109] The RF processing unit 11 of the wireless communication device S1_2 sets the receivable frequency band to a frequency band of fo=920MHz, Δf=15MHz according to the control signal CTL1.

[0110] After performing the above frequency band setting process, the wireless communication control unit 16 of the wireless communication device S1_2 transmits an Ack signal to the host device H1 via the first communication interface 15, indicating that the processing corresponding to the command Cmd1 received from the host device H1 has been completed.

[0111] When the host device H1 receives an Ack signal from the wireless communication devices S1_1 to S1_2 in response to the transmitted command Cmd1, it determines that the wireless communication devices S1_1 to S1_2 have completed the setup to comply with the transmitted command Cmd1, and proceeds to step SS2.

[0112] (Steps SS2, SS3): In step SS2, the host device H1 executes a data acquisition instruction process. Specifically, the process is executed as follows:

[0113] The host control unit 22 of the host device H1 generates a command Cmd1 (Data_get_start) to instruct each of the wireless communication devices S1_1 to S1_2 to start acquiring data. The host control unit 22 then transmits the communication data, including the generated command Cmd1 (Data_get_start), to each of the wireless communication devices S1_1 to S1_2 via the second communication interface 21.

[0114] The wireless communication device S1_1 receives communication data including the command Cmd1(Data_get_start) transmitted from the host device H1. The first communication interface 15 of the wireless communication device S1_1 extracts the command Cmd1(Data_get_start) transmitted from the host device H1 from the received communication data and outputs the extracted command Cmd1(Data_get_start) to the wireless communication device control unit 16.

[0115] The control unit 16 for the wireless communication device S1_1 generates control signals CTL1 to CTL2 to execute (start) the data acquisition process according to the command Cmd1 (Data_get_start), and outputs the generated control signal CTL1 to the RF processing unit 11 and the generated control signal CTL2 to the IQ data acquisition unit 12.

[0116] The RF processing unit 11 and the IQ data acquisition unit 12 of the wireless communication device S1_1 execute (start) data acquisition processing according to the control signals CTL1 to CTL2 input from the wireless communication device control unit 16 (step SS3).

[0117] Furthermore, the same processing as described above is performed in wireless communication devices S1_2 and S1_3, and the data acquisition process (data sampling process) is executed (started) (step SS3).

[0118] (Step S12): In step S12 (see Figures 5 and 6), spurious emission removal (steady-state noise removal) is performed. Specifically, the following processes are performed.

[0119] (Step S121): In step S121 (see Figure 6), loop processing (loop 1 processing) is started. Loop processing (loop 1 processing) is executed individually for each wireless communication device (each of wireless communication devices S1_1 to S1_2).

[0120] (Step S122A): In step S122A, frequency conversion processing (spectrum acquisition processing) is performed. Specifically, the following processes are performed.

[0121] The spectrum acquisition unit 141 of the time sample average data acquisition unit 14 of the wireless communication device S1_1 performs frequency transformation processing (for example, Fourier transform) on the data D1 (time series data) acquired by the IQ data acquisition unit 12 to acquire the spectrum (frequency domain data) corresponding to the data D1 (time series data). The spectrum acquisition unit 141 then outputs the data including the acquired spectrum as data Df1 to the time sample average spectrum calculation unit 142 and the subtraction unit 143.

[0122] (Step S122B): In step S122B, a clustering process for spurious emission detection is performed. Specifically, the following processes are executed. Note that the spurious emission detection clustering process may be executed in parallel with the frequency conversion process (step S122A).

[0123] The feature acquisition unit 13 of the wireless communication device S1_1 performs a feature acquisition process on the data D1 (time-series data) acquired by the IQ data acquisition unit 12. Specifically, the feature acquisition unit 13 accumulates (acquires) data from data D1 for the number of samples to be subjected to the discrete Fourier transform, performs a discrete Fourier transform on this data, and obtains data D12. Then, it performs a logarithmic operation on the acquired data D12, and further performs an inverse discrete Fourier transform on the result to obtain data D15. Furthermore, the feature acquisition unit 13 performs a normalization process on data D15 to obtain data D2.

[0124] Then, the wireless communication device S1_1 transmits the data D2 acquired above to the host device H1.

[0125] The host device H1 performs clustering on the data D2 received from the wireless communication device S1_1, and assigns a cluster number to each sample data (data acquired at the sample time) of data D2. Then, the host device H1 transmits data containing the cluster number information assigned to each sample data of data D2 (clustering result Rst_cls) as data D_H to the wireless communication device S1_1.

[0126] (Step S123): In step S123, the time sample mean spectrum calculation process is performed. Specifically, the following processes are performed.

[0127] The time-sample average spectrum calculation unit 142 of the time-sample average data acquisition unit 14 of the wireless communication device S1_1 calculates (acquires) the time-sample average spectrum for the data Df1 output from the spectrum acquisition unit 141 based on the data D_H received from the host device H1, and acquires the time-sample average spectrum (data Df_ave).

[0128] (Step S124): In step S124, a time-sample average spectrum subtraction process is performed. Specifically, the following process is performed.

[0129] The subtraction unit 143 of the time sample average data acquisition unit 14 of the wireless communication device S1_1 receives data Df1 output from the spectrum acquisition unit 141 and data Df_ave output from the time sample average spectrum calculation unit 142. The subtraction unit 143 subtracts data Df_ave from data Df1 and outputs the resulting data as data Df_m to the first communication interface 15 (data Df_m is transmitted to the host device H1 via the first communication interface 15).

[0130] Here, we will explain the spurious emission elimination process (steps S122A, 122B, 123, and 124) with specific examples.

[0131] Figure 8 shows the spectrum of data acquired by the wireless communication device S1_1 from time 0 seconds to 1 second (corresponding to data Df1) (an example). In Figure 8, below the spectrum, the class numbers acquired by the clustering process described above (processing for classifying into three clusters (classes)) are shown. The clustering process described above can be performed using, for example, the technology disclosed in Japanese Patent Application No. 2020-169784.

[0132] In the spectrum shown in Figure 8, the noise extending along the time axis is stationary noise (spurious signals), and by eliminating this stationary noise, it becomes possible to obtain high-quality training data. On the other hand, in the spectrum of Figure 8, the data in the region labeled "short-period pulse noise," the data in the region labeled "communication signals (wireless signals)" (narrowband signals), and the data in the region labeled "impulsive noise" are data that should be subjected to signal type classification processing.

[0133] As can be seen from Figure 8, the sample data determined to be cluster number "0" does not contain short-period pulse noise, communication signals (wireless signals), or impulsive noise, but does contain stationary noise (spurious signals). Therefore, by obtaining the time-averaged data of the sample data determined to be cluster number "0" and subtracting this time-averaged data from each sample data, it is possible to obtain data from which stationary noise (spurious signals) has been eliminated.

[0134] The graph on the right side of Figure 9 shows the time-averaged data (time-sample-averaged spectrum) for the sample data that was determined to be cluster number "0" in the spectrum of Figure 8.

[0135] Furthermore, the lower part of Figure 10 shows the spectrum obtained by subtracting the time-sample-averaged spectrum from each sample data in the spectrum of Figure 8 (spectrum after spurious emission removal). From the lower part of Figure 10, it can be seen that the steady-state noise (spurious emissions) has been properly removed.

[0136] Furthermore, in Figure 11, in the spectrum of Figure 8, (1) Data that has been determined to have cluster number "0" (data that should be labeled "no signal / noise"), (2) Data that has been determined to have cluster number "1" (data that should be labeled as "narrowband signal"), and (3) Data that has been determined to have cluster number "2" (data that should be labeled as "short-period pulse noise"), The image shows the data before spurious emission elimination (upper data) and the data after spurious emission elimination (lower data).

[0137] Figure 11 shows that steady-state noise (spurious signals) is properly eliminated from all types of data.

[0138] As described above, by performing spurious emission elimination processing in the signal type classification system 100, it is possible to obtain data from which steady-state noise (spurious emissions) has been appropriately eliminated.

[0139] (Step S125): In step S125, it is determined whether the termination condition for loop 1 processing is met. If the termination condition for loop 1 processing is not met (i.e., processing for each wireless communication device is not completed), the process returns to step S121. On the other hand, if the termination condition for loop 1 processing is met, the process proceeds to step S13. In this embodiment, it is determined that the termination condition for loop 1 processing is met once processing for wireless communication devices S1_1 and S1_2 is completed.

[0140] (Step S13): In step S13 (see Figures 5 and 7), normalized clustering is performed. Specifically, the following processes are executed.

[0141] Steps S1311 to S1331 are performed in radio transmitter S1_1, steps S1312 to S1332 are performed in radio transmitter S1_2, ..., and steps S131N to S133N are performed in radio transmitter S1_N. Steps S1311 to S1331, steps S1312 to S1332, and steps S131N to S133N are performed in parallel.

[0142] (Step S1311): In step S1311, the first framing process is performed. Specifically, the following processes are performed.

[0143] The frame processing unit 131 of the feature acquisition unit 13 of the wireless communication device S1_1 receives data D1 output from the IQ data acquisition unit 12 and performs frame processing on data D1. Specifically, the frame processing unit 131 stores (stores and holds) data from data D1, which is time-series data, for the number of samples to be subjected to the discrete Fourier transform. Then, once the frame processing unit 131 has acquired the number of samples to be subjected to the discrete Fourier transform, it outputs this data as data D11 to the cepstrum calculation unit 132.

[0144] Furthermore, the data processed in the first framing process is assumed to be time-synchronized with the data processed in the second framing process 1312 to the Nth framing process. In other words, the data processed in the first framing process and the data processed in the second framing process 1312 to the Nth framing process are data acquired during the same period.

[0145] (Step S1321): In step S1321, the first cepstrum calculation process is executed. Specifically, the following processes are performed.

[0146] The DFT processing unit 1321 of the cepstrum calculation unit 132 of the feature acquisition unit 13 of the wireless communication device S1_1 receives the data D11 output from the frame processing unit 131 and performs a Discrete Fourier Transform (DFT) on the data D11. The DFT processing unit 1321 outputs the data after the Discrete Fourier Transform as data D12 to the subtractor 1322.

[0147] The subtractor 1322 receives data D12 output from the DFT processing unit 1321 and data Df_ave (time-sample-averaged spectrum) output from the time-sample-averaged data acquisition unit 14. The subtractor 1322 subtracts data Df_ave (time-sample-averaged spectrum) from data D12 (frequency-domain data). This makes it possible to obtain a spectrum (frequency-domain data) from which steady-state noise (spurious signals) has been eliminated.

[0148] The subtractor 1322 then outputs the data after subtraction (the spectrum (frequency domain data) with stationary noise (spurious emissions) removed) as data D13 to the log calculation unit 1323.

[0149] The Log calculation unit 1323 receives the data D13 output from the subtractor 1322 and performs a logarithmic operation (a process to calculate the logarithm) on the data D13. Then, the Log calculation unit 1323 outputs the processed data as data D14 to the inverse DFT processing unit 1324.

[0150] The inverse DFT processing unit 1324 receives data D14 output from the log calculation unit 1323 and performs an inverse discrete Fourier transform on data D14 to obtain cepstrum data (cepstrum sequence). Then, the inverse DFT processing unit 1324 extracts a sequence of predetermined dimensions (for example, 16 dimensions) from the obtained cepstrum sequence, excluding the lower 0th dimension, to obtain feature data (cepstrum data). The inverse DFT processing unit 1324 then outputs the data obtained above as data D15 to the normalization processing unit 133.

[0151] (Step S1331): In step S1331, the first normalization process is performed. Specifically, the following process is performed.

[0152] The normalization processing unit 133 receives the data D15 output from the inverse DFT processing unit 1324 of the cepstrum calculation unit 132 and performs normalization processing (for example, max-min normalization) on the data D15. In other words, the normalization processing unit 133 obtains the maximum value max and the minimum value min from the set of sample data (sample data included in data D15) as the target of processing. Then, if the value of each sample data included in data D15 is x, x' = (x - min) / (max - min) Normalization is performed by executing a process equivalent to the above. As a result, the range of possible values ​​for the normalized data x' becomes [0,1].

[0153] Furthermore, in the normalization process described above, data with values ​​smaller than a predetermined threshold may be excluded. This allows data that falls within the range of a predetermined threshold and a maximum value (max) to be converted (normalized) into data whose possible range is [0,1]. By doing so, data without signal or noise is not emphasized, preventing incorrect clustering.

[0154] The normalization processing unit 133 then transmits the normalized data as data D2 to the host device H1 via the first communication interface 15.

[0155] (Steps S1312~S1332, ..., Steps S131N~S133N): In steps S1312 to S1332, ..., and steps S131N to S133N, the same processing as in steps S1312 to S1332 is performed in wireless communication devices S1_2, ..., and S1_N, respectively.

[0156] (Step S134): In step S134, data integration processing is performed. Specifically, the following processes are performed.

[0157] The host device H1 integrates the normalized data (normalized cepstrum data) acquired by the radio communication devices S1_1, S1_2, ..., and S1_N and transmitted to the host device H1. This makes it possible to obtain aggregate data (integrated data) of the normalized data (normalized cepstrum data) acquired by the radio communication devices S1_1, S1_2, ..., and S1_N during the same period.

[0158] (Step S135): In step S135, the clustering process is performed. Specifically, the following processes are performed.

[0159] The host device H1 performs clustering on the integrated data acquired in step S134 (a collection of normalized data (normalized cepstrum data) acquired by wireless communication devices S1_1, S1_2, ..., and S1_N during the same period). For this clustering process, for example, the technology disclosed in Japanese Patent Application No. 2020-169784 is used. This allows for appropriate clustering of the integrated data acquired in step S134 (a collection of normalized data (normalized cepstrum data) acquired by wireless communication devices S1_1, S1_2, ..., and S1_N during the same period), and a cluster number is assigned to each cluster.

[0160] Furthermore, in the signal type classification system 100, as described above, clustering is performed on the data after normalization, so that signal and noise data output from the same wave source are not classified into different clusters. For example, even if signal and noise data output from the same wave source is acquired by multiple wireless communication devices, if normalization is not performed, the values ​​and ranges of the acquired data will differ due to differences in the reception levels of each wireless communication device, which may lead to the classification of signal and noise data output from the same wave source into different clusters. In contrast, in the signal type classification system 100, clustering is performed on the data after normalization, so that signal and noise data output from the same wave source can be classified into the same cluster with extremely high accuracy.

[0161] Furthermore, for example, even when acquiring signal and noise data output from the same wave source using multiple wireless communication devices, if normalization processing is not performed, the acquired data values ​​and ranges may differ due to the position and gain settings of each wireless communication device, and changes thereto, or the noise floor and dynamic range may differ or change. In such cases, signal and noise data output from the same wave source may be classified into different clusters. In contrast, the signal type classification system 100 performs clustering processing on the data after normalization processing, so even in the cases described above, signal and noise data output from the same wave source can be classified into the same cluster with extremely high accuracy.

[0162] As a result, the signal type classification system 100 can perform highly accurate clustering processing, drastically reducing the amount of data that is incorrectly classified into an inappropriate cluster.

[0163] (Step S14): In step S14, a labeling process is performed. Specifically, a label (for example, the name of the data) is manually assigned to the clustering result data obtained by the normalized clustering process, that is, the data consisting of the cepstrum after normalization and the cluster number indicating the clustering result. At this time, the labeling work may be performed while referring to the spectrum data Df_m after spurious emission elimination (for example, corresponding to the lower data in Figure 10 (spectrum after spurious emission elimination)) transmitted from each wireless communication device to the host device H1. For example, by displaying the spectrum data Df_m after spurious emission elimination and the clustering result on a display device (not shown) with the time axis aligned, it is possible to easily distinguish between short-period pulse noise, narrowband signals, impulsive noise, etc., and perform the labeling work efficiently.

[0164] As a result, the signal type classification system 100 can obtain high-quality training data, which includes feature data (cepstrum data after normalization) and its labels (ground truth labels).

[0165] (1.2.2: Processing in learning mode (learning process)) Next, we will explain the processing (learning process) of the signal type classification system 100 in learning mode.

[0166] In the training data generation mode, the training data generated is used to train a classifier that classifies signal types from feature data (cepstrum data after normalization) using a classification model such as a neural network (for example, a machine learning model (e.g., a support vector machine model) or a deep learning model (e.g., a neural network model)). This allows for the acquisition of a trained model of the classifier that classifies signal types from feature data. This training process may be performed on the host device H1 or on a high-performance computer such as a server.

[0167] The trained model of the classifier that classifies signal types from feature data, obtained through the above learning process, is built into the ROM23 or RAM24 of the host device H1. In other words, the program, code, and parameter data for executing the processing using the trained model of the classifier that classifies signal types from feature data, obtained through the above learning process, are stored in the ROM23 or RAM24 of the host device H1.

[0168] (1.2.3: Processing in prediction mode (prediction processing)) Next, we will explain the processing (prediction processing) of the signal type classification system 100 during prediction.

[0169] Figure 12 is a schematic diagram of the signal type classification system 100 during prediction.

[0170] As shown in Figure 12, it is assumed that the trained model of the classifier that classifies signal types from feature data, obtained through the above training process, is stored in the ROM23 of the host device H1. In other words, the program, code, and parameter data for executing the processing using the trained model of the classifier that classifies signal types from feature data, obtained through the above training process, are stored in the ROM23 of the host device H1.

[0171] Figures 13 to 15 are flowcharts of the prediction process (signal type classification process) performed by the signal type classification system 100.

[0172] (Step S21): In step S21 (see Figure 13), the process of acquiring inspection data is executed. Specifically, similar to step S11 in the learning data generation mode, the initialization process and the data acquisition instruction process are executed, and the process of acquiring inspection data (data to be used for signal type classification processing) using the wireless communication device is started (executed).

[0173] (Step S22): In step S22 (see Figures 13 and 14), spurious emission monitoring is performed. Specifically, the following processes are executed.

[0174] (Step S221): In step S221, a determination process is performed to determine whether time sample average data (time sample average spectrum data) has been acquired (acquired and stored) in the wireless communication device that acquired the inspection data (data subject to signal type classification processing). If the determination result shows that time sample average data (time sample average spectrum data) has not been acquired in the wireless communication device, the process proceeds to step S223. On the other hand, if time sample average data (time sample average spectrum data) has been acquired (stored and stored), the process proceeds to step S222.

[0175] Furthermore, once the time-sampled average spectrum data is acquired, it will be stored in the wireless communication device that acquired it.

[0176] (Step S222): In step S222, a determination is made as to whether the proportion of samples whose posterior probability value (confidence of the predicted result data (classification result)) falls below a certain level is greater than the threshold Th1. The trained classifier model will output the posterior probability value for the result data after the prediction process has been performed on the input data (sample). In other words, the trained classifier model will perform the prediction process on the input data (sample), and if it is classified into a cluster (class) to which a predetermined label has been assigned, it will output the probability that it can be judged that the input data (sample) is correctly classified into that cluster (class) (confidence of the prediction result) as the posterior probability value.

[0177] Based on the above determination, if the proportion of samples whose posterior probability value (confidence level of the predicted result data (classification result)) falls below a certain level is greater than threshold Th1, the process proceeds to step S223. On the other hand, if the proportion of samples whose posterior probability value (confidence level of the predicted result data (classification result)) falls below a certain level is not greater than threshold Th1, the process proceeds to step S23.

[0178] (Step S223): In step S223, spurious emission removal is performed. In step S222, if the proportion of samples whose posterior probability value (confidence level of the predicted result data (classification result)) falls below a certain level is determined to be greater than the threshold Th1, it can be inferred that the prediction accuracy has decreased and the situation of stationary noise (spurious emissions) has changed. Therefore, it is necessary to recalculate the time-sample mean spectrum of the stationary noise (spurious emissions) and update the time-sample mean spectrum data used in the spurious emission removal process.

[0179] Therefore, in step S223, the same process as in step S12 (spurious emission removal process) of the training data generation process is executed. This allows the time-sample average spectrum of the stationary noise (spurious emissions) to be recalculated and the time-sample average spectrum data used in the spurious emission removal process to be updated. By performing the spurious emission removal process using the updated time-sample average spectrum data, the accuracy of the spurious emission removal process can be maintained (maintained with high accuracy). In other words, this allows the system to respond to changes in the stationary noise (spurious emissions) situation and maintain the prediction accuracy of the signal type classification process with high accuracy.

[0180] (Step S23 (Steps S231-S233)): In step S23 (see Figures 13 and 15), the normalized feature data acquisition process is performed. Specifically, the wireless communication device that acquired the test data (data to be processed) performs the framing process (step S231), the cepstrum calculation process (step S232), and the normalization process (step S233). The framing process (step S231), the cepstrum calculation process (step S232), and the normalization process (step S233) are the same processes as the first framing process (step S1311), the first cepstrum calculation process (S1321), and the first normalization process (step S1331) during the training data generation process, respectively.

[0181] Then, the normalized feature data (normalized cepstrum data) obtained by executing the normalized feature data acquisition process on the wireless communication device that acquired the inspection data (data to be processed) is transmitted from the wireless communication device to the host device H1.

[0182] (Step S24): In step S24, signal type classification processing is performed. Specifically, the host device H1 inputs the normalized feature data (normalized cepstrum data) (normalized feature data stored in the data acquisition buffer 25) obtained from the inspection data (data to be processed) into the trained model of the signal type classifier built on the host device H1, and obtains the data output from the trained model of the signal type classifier. The data output from the trained model of the signal type classifier is data with labels assigned to the input feature data (normalized feature data), and the type of wireless signal or electromagnetic noise from which the input feature data (normalized feature data) was obtained can be identified (the signal type can be classified) by the assigned labels.

[0183] (Step S25): In step S25, it is determined whether or not to continue the prediction process (signal type classification process). If the prediction process (signal type classification process) is to be continued, the process returns to step S21, and steps S21 to S24 are executed in the same manner as described above. On the other hand, if the prediction process (signal type classification process) is not to be continued, the prediction process (signal type classification process) is terminated.

[0184] Summary As described above, the signal type classification system 100 acquires normalized feature data (normalized feature data) from each wireless communication device, performs clustering on the acquired normalized feature data, and generates training data. Therefore, even if signals from the same wave source are received by multiple wireless communication devices at different reception levels and set gains, they can be classified into the same cluster. As a result, the signal type classification system 100 can generate highly accurate training data (teacher data) without being affected by differences in reception levels and set gains.

[0185] Furthermore, in the signal type classification system 100, a time-sample average spectrum is obtained by taking the time-sample average of the spectrum mainly containing stationary noise (spurious signals) for each wireless communication device. By subtracting the obtained time-sample average spectrum from the feature data (frequency domain data), feature data (frequency domain data) from which stationary noise (spurious signals) has been removed is obtained. Furthermore, feature data (cepstrum data) can be obtained based on this data. The signal type classification system 100 then performs clustering processing using the feature data (cepstrum data) from which stationary noise (spurious signals) has been removed to generate training data, thereby generating highly accurate training data (teaching data) that eliminates the influence of stationary noise (spurious signals).

[0186] Furthermore, the signal type classification system 100 monitors the accuracy (confidence) of the prediction processing results during the prediction process. If the accuracy deteriorates, it determines that the situation of stationary noise (spurious emissions) has changed, updates the time-sample mean spectrum, and performs spurious emission removal processing using the updated time-sample mean spectrum. As a result, it is always possible to obtain feature data (cepstrum data) that has had stationary noise (spurious emissions) removed. Consequently, even if the situation of stationary noise (spurious emissions) changes, the signal type classification system 100 can perform highly accurate prediction processing (signal type classification processing) without being affected by the stationary noise (spurious emissions).

[0187] [Other embodiments] In the above embodiment, a case was described in which the signal type classification system 100 performs both normalization processing and spurious emission elimination processing. However, the system is not limited to this, and it may perform either normalization processing or spurious emission elimination processing. In this case, the signal type classification system 100 may be configured in a way that omits the functional units for performing normalization processing and spurious emission elimination processing.

[0188] Furthermore, in the above embodiment, the spurious emission elimination process was described in which the time-sample average spectrum of data (frequency domain data) mainly containing stationary noise components (spurious components) is obtained, and this time-sample average spectrum is subtracted from the frequency domain data (data D12) in the cepstrum calculation unit. However, the method is not limited to this. For example, in the spurious emission elimination process, the time-sample average spectrum of data (frequency domain data) mainly containing stationary noise components (spurious components) may be obtained, the time-sample average spectrum may be inversely Fourier transformed to obtain time domain data, and the obtained time domain data may be subtracted from the time series data (data D1) (time domain data) to eliminate the stationary noise components (spurious components). In this case, the feature acquisition process (process to obtain a normalized cepstrum) may be performed on the data obtained by subtracting the acquired time domain data from the time series data (data D1) (time domain data).

[0189] Furthermore, in the above embodiment, the clustering process disclosed in Japanese Patent Application No. 2020-169784 (clustering process using a machine learning model) was described as a method for detecting data (frequency domain data) that mainly contains steady-state noise components (spurious components) in the spurious emission elimination process, but the invention is not limited to this. For example, the acquired spectrum may be displayed as a two-dimensional image, and the two-dimensional image may be analyzed (image analysis methods may be used, or image analysis may be performed by processing with a machine learning model or a deep learning model) to detect data (frequency domain sample data) that mainly contains steady-state noise components (spurious components), or data (frequency domain sample data) that does not contain (or can be determined to have a low probability of containing) short-period pulse noise, narrowband signals, radio signals, or pulsed noise, and the data (frequency domain sample data) that mainly contains steady-state noise components (spurious components) may be detected by taking the time sample average of the detected data.

[0190] Furthermore, in the above embodiment, the normalization clustering process (see Figure 7) was described in the case where the normalization process is performed after the cepstrum calculation process. However, the normalization clustering process may be performed after the normalization process. In this case, the feature acquisition unit 13 may be configured such that the normalization processing unit 133 is located after the frame processing unit 131, and the cepstrum calculation unit 132 is located after the normalization processing unit 133, and the normalization clustering process may be performed with this configuration.

[0191] In other words, in normalized clustering, linearity is maintained even when time waveform and spectrum data are normalized, so the execution order of the cepstrum calculation process and the normalization process can be swapped.

[0192] Furthermore, although the above embodiment describes the case where the normalized feature data acquisition process (see Figure 15) performs the cepstrum calculation process followed by the normalization process, it is not limited to this, and the normalized feature data acquisition process may also perform the normalization process followed by the cepstrum calculation process. In the normalized feature data acquisition process, linearity is maintained even when time waveform and spectrum data are normalized, so the execution order of the cepstrum calculation process and the normalization process can be reversed.

[0193] Furthermore, in the above embodiment, the wireless communication device used in the signal type classification system may be implemented using an SDR (software defined radio). In this case, some or all of the functional parts of the wireless communication device may be implemented using, for example, an FPGA.

[0194] Furthermore, while the above embodiment described the case where data obtained by extracting the low-dimensional portion of cepstrum data is used as feature data, the method is not limited to this. For example, if there is data that reflects the time-domain signal characteristics of wireless signals and / or electromagnetic noise, or the spectral characteristics in the frequency domain, other data may be adopted as feature data and the feature data acquisition process may be performed.

[0195] Furthermore, in the signal type classification system, wireless communication device, and host device described in the above embodiments, each block may be individually integrated into a single chip using semiconductor devices such as LSIs, or it may be integrated into a single chip that includes some or all of the blocks.

[0196] Although we have used the term LSI here, depending on the degree of integration, they may also be called IC, system LSI, super LSI, or ultra LSI.

[0197] Furthermore, the method of integrated circuit implementation is not limited to LSIs; it may also be implemented using dedicated circuits or general-purpose processors. After LSI manufacturing, FPGAs (Field Programmable Gate Arrays) that can be programmed, or reconfigurable processors that allow for the reconfiguration of the connections and settings of circuit cells inside the LSI, may also be used.

[0198] Furthermore, some or all of the processing of each functional block in each of the above embodiments may be implemented by a program. And some or all of the processing of each functional block in each of the above embodiments is performed by the central processing unit (CPU) in a computer. The programs for each of these processes are stored in a storage device such as a hard disk or ROM, and are read from the ROM or RAM and executed.

[0199] Furthermore, each of the processes in the above embodiments may be implemented by hardware, or by software (including cases where it is implemented together with an OS (operating system), middleware, or a predetermined library). Moreover, it may be implemented by a hybrid process of software and hardware.

[0200] Furthermore, for example, when each functional part of the above embodiment (including modified examples) is implemented by software, the hardware configuration shown in Figure 16 (for example, a hardware configuration in which a CPU (which may be a GPU), ROM, RAM, input unit, output unit, etc. are connected by a bus) may be used to implement each functional part by software processing.

[0201] Furthermore, when each functional part of the above embodiment (modified version) is implemented by software, the software may be implemented using a single computer having the hardware configuration shown in Figure 16, or it may be implemented using distributed processing with multiple computers.

[0202] Furthermore, the execution order of the processing method in the above embodiments is not necessarily limited to the description of the embodiments, and the execution order can be changed without departing from the spirit of the invention. Also, in the processing method in the above embodiments, some steps may be executed in parallel with other steps without departing from the spirit of the invention.

[0203] A computer program that causes a computer to execute the method described above, and a computer-readable recording medium on which such program is recorded, are included in the scope of the present invention. Examples of computer-readable recording media include flexible disks, hard disks, CD-ROMs, MOs, DVDs, DVD-ROMs, DVD-RAMs, high-capacity DVDs, next-generation DVDs, and semiconductor memory.

[0204] The above-mentioned computer program is not limited to one recorded on the above-mentioned recording medium, but may also be transmitted via telecommunications lines, wireless or wired communication lines, networks such as the Internet, etc.

[0205] Furthermore, in the above embodiments (including modified examples), "same" is a concept that includes being approximately the same. "Simultaneous" is a concept that includes being approximately simultaneous. "Agreement" is a concept that includes being approximately agreeing.

[0206] It should be noted that the specific configuration of the present invention is not limited to the embodiments described above, and various changes and modifications are possible without departing from the spirit of the invention. [Explanation of Symbols]

[0207] 100. Signal Type Classification System S1_1~S1_N Wireless communication device 13 Feature acquisition unit 14-hour sample average data acquisition unit H1 Host Device

Claims

1. A radio wave waveform acquisition process step of acquiring radio wave waveform data of radio waves received by an antenna using a computer, A feature acquisition process step involves using a computer to acquire feature data that represents the time domain or frequency domain characteristics of the radio wave waveform data, and which has a smaller data volume than the radio wave waveform data; A spurious emission removal process step involves using a computer to acquire multiple data points from the radio wave waveform data in which a stationary noise component is dominant, taking the time sample average of the acquired data to obtain time sample average data, and subtracting the acquired time sample average data from the feature data to obtain stationary noise component removal feature data. A clustering process step in which the stationary noise component removal feature data is classified into clusters using a computer, A training data acquisition step involves using a computer to assign labels to the feature data classified by the clustering process, thereby acquiring data that associates the feature data with the labels as training data. A method for generating training data for signal type classification processing, comprising the following:

2. The present invention further comprises a normalization processing step of obtaining normalized feature data obtained by performing a normalization process on the steady-state noise component removal feature data using a computer, The clustering processing step involves classifying the normalized feature data into clusters. The method for generating training data according to claim 1.

3. The aforementioned feature data is data extracted from the cepstrum. The method for generating training data according to claim 1 or 2.

4. The aforementioned feature data is data obtained by extracting or processing the lower-order components of the cepstrum. A method for generating training data according to any one of claims 1 to 3.

5. The normalization step obtains the maximum value max and minimum value min of the data included in the data set subject to normalization, and if x is the data value of the data included in the data set subject to normalization, x'=(x-min) / (max-min) The normalization process is performed by executing a process equivalent to the above. The method for generating training data according to claim 2.

6. The normalization step involves obtaining the maximum value max and minimum value min of the data in the set obtained by excluding data with values ​​smaller than a predetermined threshold from the data set subject to normalization. If x is the data value of the data in the data set subject to normalization, x'=(x-min) / (max-min) The normalization process is performed by executing a process equivalent to the above. The method for generating training data according to claim 2.

7. A signal type classification system that performs signal type classification using a trained model of a signal type classifier obtained by performing a training process using training data generated by the training data generation method described in any one of claims 1 to 6, wherein Wireless communication equipment, The host device and Equipped with, The aforementioned wireless communication device is A radio wave waveform acquisition process step that acquires radio wave waveform data of radio waves received by an antenna, A feature acquisition process step to acquire feature data which is data that shows the characteristics of the time domain or frequency domain of the radio wave waveform data, and which has a smaller data volume than the radio wave waveform data; A spurious emission removal process step involves obtaining multiple data points from the aforementioned radio wave waveform data in which a steady-state noise component is dominant, obtaining time-sample average data by taking the time-sample average of the multiple obtained data points, and obtaining steady-state noise component-removed feature data by subtracting the obtained time-sample average data from the feature data. A transmission step of transmitting the aforementioned steady-state noise component removal feature data to a host device, Execute, The aforementioned host device, A receiving step of receiving the steady-state noise component removal feature data transmitted from the wireless communication device, A label acquisition step involves inputting the stationary noise component removal feature data acquired by the wireless communication device into the trained model of the signal type classifier to obtain a label for the stationary noise component removal feature data. Execute Signal type classification system.

8. The aforementioned host device, The reliability of the classification result obtained by the trained model of the signal type classifier is monitored, and if the percentage of times the reliability falls below a predetermined value during a certain period exceeds a predetermined value, a command is output to the wireless communication device to reacquire the time sample average data. The aforementioned wireless communication device is When the host device receives a command to reacquire the time-sample averaged data, the time-sample averaged data update step is executed, which involves acquiring multiple data points from the radio wave waveform data in which the steady-state noise component is dominant, and then taking the time-sample average of the multiple acquired data points to acquire the time-sample averaged data. The signal type classification system according to claim 7.

9. A program for causing a computer to execute any of the training data generation methods described in claims 1 to 6.

Images