Frequency determination method and frequency determination device
The method uses machine learning models to predict communication status and interference in wireless train control systems, addressing frequency selection challenges and enhancing reliability by accurately determining available frequencies.
Patent Information
- Application Number
- JP2021020588
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2021-02-12
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2041-02-12
AI Technical Summary
Existing wireless train control systems face challenges in selecting available frequencies with high precision due to interference in the ISM band, particularly when onboard equipment moves, necessitating accurate spectrum sensing to ensure reliable communication.
A method involving machine learning models, such as recurrent neural networks, to estimate communication status and evaluation values based on Signal-to-Noise Ratio (SNR) data, allowing for precise frequency selection by analyzing past data sets and predicting future communication conditions.
Enables high-precision frequency selection by accurately predicting communication status and interference levels, reducing handover delays and ensuring reliable wireless communication in train control systems.
Smart Images

Figure 0007584319000001 
Figure 0007584319000002 
Figure 0007584319000003
Abstract
Description
[Technical field]
[0001] The present invention relates to a frequency determination method for determining a frequency to be used in wireless communication. [Background technology]
[0002] In a wireless train control system known as Communication Based Train Control (CBTC), which uses wireless communication for train control, the use of a frequency band known as the Industrial Scientific and Medical (ISM) band, which does not require a license, for wireless communication between ground-based and on-board trains is being considered. There are various wireless communication standards that use this ISM band, such as wireless LAN (IEEE802.11) and Bluetooth (registered trademark) (IEEE802.15.1), and because they are used for a variety of wireless communication systems, interference is unavoidable.
[0003] In a radio train control system, handover is necessary to switch the radio base station of the communication destination as the train moves, but delays in handover can occur because a frequency with less interference is used as the frequency for radio communication with the new radio base station of the communication destination.Therefore, as a technique for realizing handover without delay, Patent Document 1 discloses a technique in which, for each radio base station, multiple communication frequencies that do not interfere with other radio base stations are determined, the electric field strength of the multiple communication frequencies is periodically checked, and the train performs radio communication with the radio base station at the optimal radio frequency selected based on the electric field strength. [Prior art documents] [Patent documents]
[0004] [Patent Document 1] JP 2012-90328 A Summary of the Invention [Problem to be solved by the invention]
[0005] In wireless communication using the ISM band, there is a technology called cognitive wireless communication that searches for an available wireless communication channel and switches the wireless method to be used. In this cognitive wireless communication, spectrum sensing is required to search for an available wireless communication channel. A wireless communication channel is a frequency. However, when applying cognitive wireless communication to a wireless train control system that requires high reliability, higher accuracy is also required for spectrum sensing. Of course, even in wireless communication other than cognitive wireless communication, if an available frequency can be selected with high accuracy, it is useful as a method for determining a frequency to be used. In particular, in a wireless train control system, since the on-board device moves with the movement of the train, it is necessary to predictively search and select an available frequency at the destination before the movement and determine the frequency to be used.
[0006] The present invention has been made in consideration of the above circumstances, and an object of the present invention is to provide a new technology that can select an available frequency with high accuracy as a technology for determining a frequency to be used in a radio train control system. [Means for solving the problem]
[0007] The first invention for solving the above problem is: acquiring a past data set in which given frequency-specific time series data showing a frequency-specific communication status based on a signal-to-noise ratio (SNR) of a wireless base station installed along a railway line that wirelessly communicates with an on-board device over a predetermined observation period is associated with frequency-specific evaluation value data showing an evaluation value for evaluating the communication status in a subsequent period following the observation period, by frequency; Using the past data set, generate a machine learning model for estimating an evaluation value by inputting the frequency-specific time series data and outputting the frequency-specific evaluation value data; inputting the frequency-specific time series data, in which a given conventional period is set as the observation period, into the machine learning model for estimating evaluation value, to obtain an output of the frequency-specific evaluation value data, in which a future period following the conventional period is set as the subsequent period; and determining a frequency to be used for the wireless communication based on the obtained frequency-specific evaluation value data; The frequency determination method includes the steps of:
[0008] Other inventions include: an acquisition unit (e.g., the past data set acquisition unit 202 in FIG. 10 ) that acquires a past data set that associates given frequency-specific time series data that shows a communication status by frequency based on SNR by a wireless base station installed along a railway line that wirelessly communicates with an on-board device over a predetermined observation period with frequency-specific evaluation value data that shows an evaluation value by frequency that evaluates the communication status in a subsequent period following the observation period; A generation unit (e.g., the evaluation value estimation model generation unit 206 in FIG. 10) that generates a machine learning model for estimating an evaluation value by inputting the frequency-specific time series data and outputting the frequency-specific evaluation value data by using the past data set; a determination unit (e.g., the usage frequency determination unit 210 of FIG. 10 ) that inputs the frequency-specific time series data with a given conventional period as the observation period into the evaluation value estimation machine learning model to obtain an output of the frequency-specific evaluation value data with a future period following the conventional period as the subsequent period, and determines a usage frequency to be used for the wireless communication based on the obtained frequency-specific evaluation value data; A frequency determining device may be configured comprising:
[0009] According to the first invention, etc., a technology that can select an available frequency with high accuracy can be realized as a technology for determining a frequency to be used in a radio train control system. That is, a machine learning model for estimating an evaluation value can be generated by inputting time series data showing a communication status by frequency based on the SNR of the radio base station when the radio base station and the on-board device perform wireless communication over a predetermined observation period, and outputting evaluation value data of the communication status in a subsequent period following the observation period. Then, a frequency to be used by the radio base station for wireless communication can be determined based on the frequency-based evaluation value data of the subsequent period obtained by inputting the time series data of the communication status by frequency in a conventional period into the machine learning model for estimating an evaluation value. As a result, the communication status in a future period following the conventional period can be estimated as an evaluation value from the communication status by frequency in the conventional period, so that a technology for determining a frequency to be used in a radio train control system can be realized as a technology for selecting an available frequency with high accuracy.
[0010] The second invention is the first invention, the machine learning model for estimating the evaluation value is a recurrent neural network, The past data set includes a data set of a plurality of frequency-specific time series data, each of which is a period in which the observation period overlaps with another period in which the observation period shifts in sequence, and the frequency-specific evaluation value data corresponding to the frequency-specific time series data (see, for example, FIG. 8 ); This is a frequency determination method.
[0011] According to the second invention, the past data set includes a data set of frequency-specific time series data in each of a plurality of observation periods that are partially overlapped and shifted in sequence, and frequency-specific evaluation value data in a subsequent period following each of the observation periods. Therefore, the past data set is suitable for generating a machine learning model for estimating evaluation values, which is a recurrent neural network.
[0012] The third invention is the first or second invention, Determining the frequency to be used selecting candidate frequencies based on the obtained frequency-specific evaluation value data; determining a current frequency to be used by using a previously determined frequency to be used and the candidate frequency; Including, This is a frequency determination method.
[0013] According to the third invention, a candidate frequency is selected based on the frequency-specific evaluation value data for the subsequent period obtained by using the machine learning model for estimating evaluation value, and the selected candidate frequency and the previous frequency are used to determine the current frequency to be used. Since the previous frequency to be used can be given priority in determining the current frequency to be used, the on-board device can avoid the complication of processing caused by frequently changing the frequency to be used for wireless communication.
[0014] A fourth aspect of the present invention is any one of the first to third aspects of the present invention, Obtaining the past data set includes: inputting data indicating a radio wave environment in a time series into a machine learning model for estimating a communication status, the machine learning model being trained based on teacher data that associates data indicating a radio wave environment of the wireless base station with the communication status by frequency under the radio wave environment, and sequentially acquiring data on the communication status; Among the sequentially acquired communication status data, 1) data corresponding to the observation period is regarded as the frequency-specific time series data, and 2) data corresponding to the subsequent period following the observation period is obtained by calculating the evaluation value by frequency and regards the data as the frequency-specific evaluation value data, thereby acquiring the past data set; Including, This is a frequency determination method.
[0015] According to the fourth aspect of the present invention, it is possible to obtain a past data set based on data showing the radio wave environment of an actual wireless base station. Since the machine learning model for estimating an evaluation value is generated using the past data set, the estimation accuracy of the machine learning model for estimating an evaluation value is improved, and more accurate spectrum sensing can be realized.
[0016] The fifth invention relates to the fourth invention, The machine learning model for estimating a communication state includes a sub-model and a main model, the sub-model is a model trained based on teacher data in which the data indicating the radio wave environment is associated with noise presence / absence information by frequency, the main model is a model trained based on teacher data that associates the data indicating the radio wave environment and the noise presence / absence information with the communication status by frequency, The sequential acquisition of the communication status data includes: inputting data indicating the radio wave environment into the sub-model and acquiring the noise presence / absence information by frequency; inputting the data indicating the radio wave environment and the acquired noise presence / absence information for each frequency into the main model to acquire data on the communication status; Including, This is a frequency determination method.
[0017] According to the fifth invention, the machine learning model for estimating communication conditions includes a sub-model and a main model. The sub-model is a model that inputs data indicating the radio wave environment of a wireless base station and outputs noise presence / absence information by frequency. The main model is a model that inputs data indicating the radio wave environment and the noise presence / absence information by frequency output by the sub-model and outputs the communication conditions by frequency. As a result, the time series data of the communication conditions by frequency obtained using the machine learning model for estimating communication conditions becomes more accurate data that takes into account the presence / absence of noise. As a result, the accuracy of the past data set acquired using the machine learning model for estimating communication conditions is improved, and the estimation accuracy of the machine learning model for estimating an evaluation value generated using this past data set is improved, so that more accurate spectrum sensing can be realized. [Brief description of the drawings]
[0018] [Figure 1] An example of the configuration of a wireless train control system. [Diagram 2]FIG. 1 is a diagram showing the main configuration of a frequency determination device. [Diagram 3] Example output of a machine learning model for estimating communication status. [Figure 4] An example of the configuration of a machine learning model for estimating communication conditions. [Diagram 5] FIG. 1 is a diagram illustrating learning of a machine learning model for estimating communication conditions. [Figure 6] FIG. 1 is an explanatory diagram of a machine learning model for estimating evaluation values. [Figure 7] Example input and output of a machine learning model for estimating rating values. [Figure 8] FIG. 1 is a diagram illustrating learning of a machine learning model for estimating evaluation values. [Figure 9] 13 is a flowchart of a process for determining a frequency to be used. [Figure 10] FIG. 2 is a functional configuration diagram of a frequency determination device. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0019] Hereinafter, an example of a preferred embodiment of the present invention will be described with reference to the drawings. Note that the present invention is not limited to the embodiment described below, and the form to which the present invention can be applied is not limited to the following embodiment. In addition, in the description of the drawings, the same elements are given the same reference numerals.
[0020] [System Configuration] Fig. 1 shows an example of the configuration of a radio train control system 1 according to this embodiment. As shown in Fig. 1, the radio train control system 1 according to this embodiment includes a radio base station 10, a monitoring receiving device 20, and a frequency determining device 30. The radio base station 10 and the monitoring receiving device 20 are installed along the railway line. The frequency determining device 30 is installed in a central control center or the like. Furthermore, each device is communicatively connected via a ground transmission line N.
[0021] The wireless communication between the wireless base station 10 and the on-board device 82 mounted on a train 80 traveling on the tracks is wireless communication using ISM (Industrial Scientific and Medical) bands such as the 2.4 GHz band, 5.7 GHz band, and 920 MHz band, and employs cognitive wireless communication that searches for frequencies not used by other communication systems and switches between wireless communication standards and wireless channels to perform wireless communication.
[0022] A plurality of radio base stations 10 are installed along the railway line so that their communication ranges partially overlap each other, so that on-board equipment 82 of trains 80 running on the track can perform continuous wireless communication with the frequency determination device 30. Furthermore, the radio base station 10 can establish wireless channels with a predetermined number of trains 80, and performs wireless communication with the on-board equipment 82 of each train 80 through a wireless channel determined based on an instruction from the frequency determination device 30.
[0023] The monitor receiving device 20 is a device installed to obtain data indicating the radio wave environment of the corresponding radio base station 10, and is installed near the corresponding radio base station 10 to receive radio signals around the radio base station 10 and output them as data indicating the radio wave environment to the frequency determining device 30. Although two radio base stations 10 and two monitor receiving devices 20 are shown in Fig. 1, in reality, a plurality of them are installed along the railway line.
[0024] The frequency determination device 30 estimates the communication state around the radio base station 10 based on data indicating the radio wave environment input from the corresponding monitor receiving device 20 for each radio base station 10. Then, based on the estimated communication state, dynamically assigns a radio channel related to cognitive radio communication between the radio base station 10 and the on-board device 82 of the train 80. For example, at the time of handover of the radio base station 10, an available frequency is determined as the radio channel to be assigned based on the future communication state around the next radio base station 10 estimated in real time. Then, the determined radio channel is notified to the on-board device 82 of the train 80 and the radio base station 10. A radio channel is a term in the technical field of communication, and means a frequency, and more precisely, corresponds to a frequency band. Therefore, determining a radio channel to be used corresponds to determining a frequency to be used. Hereinafter, a frequency is appropriately referred to as a channel.
[0025] [Frequency determination device] Fig. 2 is a diagram showing the main components of a frequency determination device 30 related to the determination of a radio channel to be used in this embodiment. As shown in Fig. 2, the frequency determination device 30 has a machine learning model 40 for estimating communication conditions, a machine learning model 50 for estimating evaluation values, and a usage frequency determination unit 210 as main components related to the determination of a radio channel to be used. The machine learning model 40 for estimating communication conditions and the machine learning model 50 for estimating evaluation values are generated for each radio base station 10. In the following, for the sake of simplicity, a description will be given of one radio base station 10, but the same applies to each of the other radio base stations 10.
[0026] (A) Machine learning model for estimating communication status The machine learning model 40 for estimating communication conditions is a machine learning model for performing so-called spectrum sensing, which estimates the communication conditions by frequency around the wireless base station 10 corresponding to the monitoring receiving device 20 based on data indicating the radio wave environment input from the monitoring receiving device 20.
[0027] In this embodiment, the machine learning model 40 for estimating communication conditions is realized by a convolutional neural network (CNN), but may be another neural network. The machine learning model 40 for estimating communication conditions is a machine learning model trained using teacher data as described later, and a weighting coefficient is determined. Since the monitoring receiving device 20 is installed near the wireless base station 10, the wireless signal received by the monitoring receiving device 20 can be regarded as the wireless signal received by the wireless base station 10. This can significantly reduce the processing time required to estimate the communication conditions, and can realize estimation of the communication conditions in near real time or in real time.
[0028] Then, the machine learning model 40 for estimating communication status outputs the communication status by frequency around the monitoring receiving device 20 when the wireless signal input as data indicating the radio wave environment is received by the monitoring receiving device 20. This communication status by frequency becomes the communication status around the corresponding wireless base station 10. The communication status by frequency is, for example, the communication status by frequency in units of wireless channels. The communication status is classified into multiple stages based on the SNR (Signal to Noise Ratio) when the wireless base station 10 performs wireless communication in the wireless environment. In this embodiment, the communication status is classified into four stages of interference levels "0" to "3", but may be three stages or less, or may be five stages or more. The higher (larger) the interference level is, the higher the possibility that wireless communication is being performed at the frequency (= it is being used). For the wireless train control system 1, the fact that wireless communication is being performed means that it is being used by another wireless communication system, which becomes noise. Therefore, the higher (larger) the interference level is, the more noise there is.
[0029] FIG. 3 is an example of an output from the machine learning model 40 for estimating communication conditions. In FIG. 3, the horizontal axis represents time t, and the vertical axis represents an interference level, which is a communication condition, and an output example of the communication conditions for each of three frequencies (wireless channels) f1, f2, and f3 is shown. As shown in FIG. 3, the machine learning model 40 for estimating communication conditions inputs data indicating the radio wave environment in a time series from the monitor receiving device 20, and outputs the communication conditions by frequency in a time series. Note that, in FIG. 3, the communication conditions output by the machine learning model 40 for estimating communication conditions are shown as a graph that is continuous in time, but in reality, the processing performed by the frequency determining device 30 is digital processing, and becomes discrete time series data (frequency-specific time series data) at a predetermined time interval.
[0030] Fig. 4 is a diagram showing a configuration of the machine learning model for estimating communication conditions 40. As shown in Fig. 4, the machine learning model for estimating communication conditions 40 includes a sub-model 42, a noise power estimation unit 44, and a main model 46. The sub-model 42 and the main model 46 are each realized by a convolutional neural network, but may be other neural networks.
[0031] The sub-model 42 is a machine learning model for estimating the presence or absence of noise for each frequency based on data indicating the radio wave environment input from the monitoring receiver 20. For example, the presence or absence of noise can be classified based on the interference level, which is the communication state, into a communication state with an interference level of "1 to 3" as "noise present" and a communication state with an interference level of "0" as "no noise present". The sub-model 42 receives the power spectrum for each frequency of the radio signal received by the monitoring receiver 20 as data indicating the radio wave environment. In this embodiment, N pieces of data obtained by normalizing the amplitude and phase of the radio signal received by the monitoring receiver 20 and performing discrete Fourier transform processing are input as the power spectrum for each frequency. In addition, since the estimation by the sub-model 42 is repeatedly performed at a predetermined time interval, the presence or absence of noise for each frequency output by the sub-model 42 becomes discrete time series data.
[0032] The noise power estimation unit 44 calculates an estimate of the noise power Pn contained in the wireless signal received by the monitoring receiver 20 based on the noise presence / absence information by frequency output from the sub-model 42. The sub-model 42 receives data indicating the radio wave environment from the monitoring receiver 20 and outputs the presence / absence of noise by frequency in a time series. Therefore, the noise power estimation unit 44 calculates, as an estimate of the noise power Pn, an average value of a predetermined number of signal powers selected in ascending order from the signal powers of the wireless signal received from the monitoring receiver 20 as data indicating the radio wave environment when the noise presence / absence output by the sub-model 42 is classified as "no noise" by frequency. Note that the calculation by the noise power estimation unit 44 does not need to be performed at the same time interval as the time interval at which the sub-model 42 and the main model 46 perform the estimation, but may be performed at a time interval longer than the time interval.
[0033] The main model 46 is a machine learning model for estimating a communication situation by frequency based on data indicating the radio wave environment input from the monitor receiving device 20 and information on the presence or absence of noise by frequency input from the sub-model 42. The main model 46 receives the signal power (time average of the squared amplitude) of the wireless signal received by the monitor receiving device 20 as data indicating the radio wave environment. In addition, the main model 46 receives the noise power Pn by frequency output from the noise power estimation unit 44 as information on the presence or absence of noise by frequency. The main model 46 estimates and outputs an interference level "0 to 3" classified into four stages as the communication situation by frequency based on the signal power normalized by the noise power Pn for each frequency. The estimation by the main model 46 is repeated at a predetermined time interval like the sub-model 42, so that the communication situation by frequency output by the main model 46 becomes discrete time series data. The time series data of the communication situation by frequency output by the main model 46 becomes the output of the machine learning model 40 for estimating communication situation.
[0034] In estimating a communication situation using the machine learning model 40 for estimating a communication situation, first, data indicating the radio wave environment based on the wireless signal received by the monitoring receiving device 20 is input to the sub-model 42, and frequency-specific noise presence / absence information included in the input data indicating the radio wave environment is output from the sub-model 42. Next, this noise presence / absence information is input to a noise power estimation unit 44, and noise power Pn is output from the noise power estimation unit 44. Then, this noise power Pn and the data indicating the radio wave environment based on the wireless signal received by the monitoring receiving device 20 are input to the main model 46, and the main model 46 outputs a frequency-specific communication situation that matches the input data indicating the radio wave environment.
[0035] FIG. 5 is a diagram for explaining learning (machine learning) of the machine learning model 40 for estimating communication conditions. As shown in FIG. 5, the machine learning model 40 for estimating communication conditions is trained using a plurality of teacher data 312 prepared in advance. Specifically, the machine learning model 40 for estimating communication conditions includes a sub-model 42 and a main model 46, the sub-model 42 is trained using teacher data 312a for the sub-model, and the main model 46 is trained using teacher data 312b for the main model. The teacher data 312a for the sub-model is data that associates the power spectrum of a wireless signal (specifically, N pieces of data obtained by normalizing the amplitude and phase of a wireless signal and performing a discrete Fourier transform process) which is data indicating the radio wave environment of the wireless base station 10, with noise presence / absence information by frequency based on the SNR when the wireless base station 10 performs wireless communication in the radio wave environment. The sub-model 42 is generated by performing learning using data indicating the radio wave environment (power spectrum) as input and noise presence / absence information by frequency as output.
[0036] The teacher data 312b for the main model is data that associates the signal power of a wireless signal (specifically, the signal power normalized by the noise power), which is data indicating the radio wave environment of the wireless base station 10, with communication status information by frequency based on the SNR when the wireless base station 10 performs wireless communication in the radio wave environment. The main model 46 is generated by learning using the data (signal power) indicating the radio wave environment as input and the communication status information by frequency as output. In addition, the type of learning is deep learning in this embodiment, but any other type of machine learning may be used.
[0037] The teacher data 312 can be generated by using actual wireless signal data, or by simulation calculation using a computer, for example. Specifically, parameters such as a wireless communication standard and wireless propagation characteristics are set, a transmission signal for transmitting arbitrary transmission data according to the set wireless communication standard is generated, and a wireless signal that is assumed to be received by the wireless base station 10 when the transmission signal is transmitted through a wireless channel with the set wireless propagation characteristics is obtained. Then, the power spectrum (N pieces of data obtained by normalizing the amplitude and phase of the wireless signal and performing discrete Fourier transform processing) of the obtained wireless signal and the signal power (signal power normalized by noise power) of the wireless signal are calculated as data indicating the radio wave environment. Furthermore, the SNR of the transmission signal is calculated for each frequency from the power spectrum of the wireless signal and the transmission signal to calculate noise presence / absence information for each frequency and communication status information for each frequency. Then, the data indicating the radio wave environment (power spectrum) and the noise presence / absence information for each frequency are used as teacher data 312 for one sub-model, and the data indicating the radio wave environment (signal power of the wireless signal normalized by noise power) and communication status information for each frequency are used as teacher data 312 for one main model. By setting various parameters such as wireless communication standards and wireless propagation characteristics, a large number of training data 312 can be generated.
[0038] (B) Machine learning model for estimating evaluation value The machine learning model for estimating evaluation value 50 is a machine learning model for estimating an evaluation value of a communication situation by frequency in a future period around the corresponding wireless base station 10, based on time series data of the communication situation by frequency estimated by the machine learning model for estimating communication situation 40. The machine learning model for estimating evaluation value 50 is realized by a recurrent neural network (RNN), but may be another neural network. The machine learning model for estimating evaluation value 50 is a machine learning model trained using a past data set as described later, and a weighting coefficient is determined.
[0039] FIG. 6 is a schematic diagram of the machine learning model 50 for estimating an evaluation value. As shown in FIG. 6, the machine learning model 50 for estimating an evaluation value receives as input time series data of the communication status by frequency for a predetermined conventional period t0 to t1 going back from a predetermined time (for example, the current time) t1, out of the time series data of the communication status by frequency output by the machine learning model 40 for estimating a communication status. Since the time series data of the communication status is discrete time series data, n pieces of data (x1, x2, . . . , x n-1 ,x n ) is input. Then, the machine learning model 50 for estimating evaluation value outputs an evaluation value B of the communication status by frequency for the future period t1-t2 following the conventional period t0-t1. The evaluation value B is a value indicating the communication status in the entire period from the future period t1-t2. In this embodiment, the evaluation value B is indicated by being classified into six levels from "0 to 5", but it may be five levels or less, or seven levels or more. The larger the evaluation value B, the higher the possibility that wireless communication is being performed at the frequency (= it is being used). For the wireless train control system 1, the fact that wireless communication is being performed means that it is being used by another wireless communication system, which becomes noise. Therefore, the larger the evaluation value B, the more noise there is.
[0040] FIG. 7 shows an example of input and output of the machine learning model 50 for estimating evaluation values. In FIG. 7, the horizontal axis represents time t, and k frequencies (f1, f2, . . . , f k7 shows input data and output data for each frequency (radio channel). The input data is time series data of the communication status in the conventional period t0 to t1. The output data is the evaluation value B of the communication status in the future period t1 to t2 following the conventional period t0 to t1. For the input, the vertical axis represents the interference level, which is the communication status. As shown in FIG. 7, the machine learning model for estimating evaluation value 50 inputs the communication status (interference level) by frequency in the conventional period t0 to t1 from the machine learning model for estimating communication status 40 in time series, and outputs the evaluation value B of the future communication status by frequency in the future period t1 to t2. Note that, although FIG. 7 shows the communication status input to the machine learning model for estimating evaluation value 50 as a graph that is continuous in time, in reality, it is discrete time series data (time series data by frequency) at a predetermined time interval.
[0041] FIG. 8 is a diagram for explaining the learning of the machine learning model 50 for estimating an evaluation value. As shown in FIG. 8, the learning of the machine learning model 50 for estimating an evaluation value is performed using an acquired past data set 314. The past data set 314 is data in which 1) frequency-specific time series data showing the communication status by frequency based on the SNR by the wireless base station 10 in a time series manner over a predetermined observation period, and 2) frequency-specific evaluation value data showing the evaluation value evaluating the communication status in the subsequent period following the observation period, by frequency, are associated with each other. The observation period in the past data set 314 is a period of the same length as the conventional period t0 to t1 of the data input to the machine learning model 50 for estimating an evaluation value. The subsequent period is a period of the same length as the future period t1 to t2 of the data output from the machine learning model 50 for estimating an evaluation value.
[0042] As shown in FIG. 8, the past data set 314 can be acquired based on the time series data of the communication status by frequency obtained from the data of the actual wireless signal. In FIG. 8, as the time series data of the communication status for one frequency (wireless channel), discrete time series data in which the communication status is set as an interference level of "0 to 3" is illustrated. For this time series data of the communication status, a plurality of observation periods that overlap with each other are determined by shifting the data by one piece of data, and a subsequent period following the observation period is determined for each observation period. Then, the time series data of the communication status in the observation period is associated with the evaluation value B based on the time series data in the subsequent period following the observation period, to form one data set. This makes it possible to acquire the past data set 314 including a data set of each of a plurality of frequency-specific time series data that are periods that overlap with each other but are shifted in sequence, and frequency-specific evaluation value data in the subsequent period following the observation period.
[0043] The evaluation value B is calculated as the proportion of the interference level, which is the communication state during the entire subsequent period, that is equal to or higher than a predetermined level (e.g., equal to or higher than "3"). Specifically, as shown in the calculation table of FIG. 8, the evaluation value estimation machine learning model 50 is calculated as the proportion of the number of pieces of data with an interference level equal to or higher than a predetermined level (e.g., equal to or higher than "3") to the number of pieces of data during the entire subsequent period. The evaluation value estimation machine learning model 50 is generated by learning using the frequency-specific time series data during the observation period in the past data set 314 as input and the evaluation value B of the communication state during the subsequent period as output.
[0044] (C) Use frequency determination section The usage frequency determination unit 210 determines the frequency (wireless channel) that the wireless base station 10 will use for wireless communication, based on the time series data of communication conditions by frequency for the previous period t0 to t1 input from the machine learning model for estimating communication conditions 40 and the frequency-specific evaluation value data for the future period t1 to t2 input from the machine learning model for estimating evaluation value 50.
[0045] 9 is a flowchart for explaining the determination of the use frequency by the use frequency determination unit 210. As shown in FIG. 9, the use frequency determination unit 210 first selects, as a candidate frequency, a frequency having a predetermined number (for example, two) stages of the evaluation value B from the smallest (lowest) stage based on the evaluation value B of the communication situation by frequency for the future period t1 to t2 output from the evaluation value estimation machine learning model 50 (step S1). In this embodiment, since the evaluation value B is classified into six stages of "0 to 5", the "two" stages from the smallest stage means that the frequency having the evaluation value B of "0" or "1" is selected as the candidate frequency.
[0046] Next, it is determined whether the selected candidate frequencies include a frequency used last time (immediately before), and if so (step S3: YES), the frequency used last time (immediately before) is subsequently determined as the frequency used this time (step S5). If the candidate frequencies do not include a frequency used last time (step S3: NO), the frequency used this time is determined based on the time series data of the communication status by frequency input from the machine learning model 40 for estimating communication status (step S7). Specifically, for each frequency, the ratio of the interference level, which is the communication status, in the entire subsequent period that is a predetermined level or higher (e.g., "2" or higher), that is, the ratio of the number of data items whose interference level is a predetermined level or higher (e.g., "2" or higher) to the number of data items in the entire subsequent period, is calculated. Then, the frequency with the smallest calculated ratio is determined as the frequency used this time.
[0047] [Generating machine learning models] As described above, the sub-model 42 and main model 46 included in the machine learning model 40 for estimating communication conditions, and the machine learning model 50 for estimating evaluation values are generated by determining weighting coefficients of the neural network using teacher data (teacher data 312a for the sub-model 42, teacher data 312b for the main model 46, and past data set 314 for the machine learning model 50 for estimating evaluation values). In this embodiment, a stochastic gradient descent method is used as a method for determining the weighting coefficients in order to generate a machine learning model with higher accuracy.
[0048] Specifically, two sets of teacher data with different data contents are prepared as teacher data used to generate a machine learning model. Then, first, a weighting coefficient Wn of the machine learning model is obtained by stochastic gradient descent from an appropriate initial point IPn (initial value of the weighting coefficient) using one of the teacher data. Next, the estimation accuracy is obtained for the machine learning model to which the obtained weighting coefficient Wn is applied, using the other teacher data. The estimation accuracy is, for example, the mean square value of the difference between the estimated value, which is the output when the input data of the other teacher data is input to the machine learning model, and the output data associated with the input data. If this estimation accuracy is equal to or greater than a predetermined threshold, the obtained weighting coefficient Wn is adopted, and if it is less than the threshold, the initial point IPn is changed and the same process is repeated again.
[0049] [Function configuration] Fig. 10 is a functional configuration diagram of the frequency determining device 30. According to Fig. 10, the frequency determining device 30 includes an operation unit 102, a display unit 104, a sound output unit 106, a communication unit 108, a processing unit 200, and a storage unit 300, and can be configured as a kind of computer system.
[0050] The operation unit 102 is realized by an input device such as a keyboard, a mouse, a touch panel, various switches, etc., and outputs an operation signal corresponding to the operation performed to the processing unit 200. The display unit 104 is realized by a display device such as a liquid crystal display or a touch panel, etc., and performs various displays corresponding to display signals from the processing unit 200. The sound output unit 106 is realized by an audio output device such as a speaker, etc., and performs various sound outputs corresponding to audio signals from the processing unit 200. The communication unit 108 is realized by a wireless or wired communication device, and is connected to a terrestrial transmission line N to communicate with various external devices such as the wireless base station 10 and the monitor receiving device 20.
[0051] The processing unit 200 is a processor realized by an arithmetic device or arithmetic circuit such as a CPU (Central Processing Unit) or an FPGA (Field Programmable Gate Array), and performs overall control of the frequency determination device 30 based on programs and data stored in the storage unit 300, input data from the operation unit 102, and the like. The processing unit 200 also has, as functional processing blocks, a past data set acquisition unit 202, a communication status estimation model generation unit 204, an evaluation value estimation model generation unit 206, a radio wave environment data acquisition unit 208, and a usage frequency determination unit 210. Each of these functional units possessed by the processing unit 200 can be realized in software by the processing unit 200 executing a program, or can be realized by a dedicated arithmetic circuit. In this embodiment, the former software realization will be described.
[0052] The past data set acquisition unit 202 acquires a past data set 314 in which given frequency-specific time series data showing, in a time series manner over a predetermined observation period, the communication status by frequency based on the SNR by the wireless base station 10 installed along the railway line that performs wireless communication with the on-board device 82, and frequency-specific evaluation value data showing, by frequency, evaluation values evaluating the communication status in a subsequent period following the observation period. The past data set 314 includes a data set of each of a plurality of frequency-specific time series data whose observation periods overlap partially but are shifted in sequence thereafter, and frequency-specific evaluation value data corresponding to the frequency-specific time series data (see FIG. 8).
[0053] Furthermore, the past data set acquiring unit 202 inputs data indicating a radio wave environment in a time series to the machine learning model 40 for estimating communication conditions to sequentially acquire data on the communication conditions. The machine learning model 40 for estimating communication conditions is a machine learning model trained based on teacher data 312 that associates data indicating a radio wave environment of the wireless base station 10 with communication conditions by frequency under the radio wave environment. The past data set acquiring unit 202 acquires a past data set 314 by 1) treating data corresponding to an observation period as time-series data by frequency, and 2) treating data obtained by calculating evaluation values by frequency for data corresponding to a period following the observation period as frequency-specific evaluation value data, among the sequentially acquired data on the communication conditions.
[0054] Furthermore, the past data set acquisition unit 202 inputs data indicating the radio wave environment to the sub-model 42 included in the machine learning model 40 for estimating communication conditions to acquire noise presence / absence information by frequency. Then, the data indicating the radio wave environment and the acquired noise presence / absence information by frequency are input to the main model 46 included in the machine learning model 40 for estimating communication conditions to acquire data on the communication conditions, thereby sequentially acquiring data on the communication conditions.
[0055] The communication status estimation model generation unit 204 uses the teacher data 312 to input data indicating a radio wave environment and generate a machine learning model 40 for estimating a communication status, which outputs a communication status by frequency under the radio wave environment. The machine learning model 40 for estimating a communication status includes a sub-model 42 and a main model 46. The sub-model 42 is a machine learning model trained based on teacher data 312a in which data indicating a radio wave environment is associated with noise presence / absence information by frequency. The main model 46 is a machine learning model trained based on teacher data 312b in which data indicating a radio wave environment and noise presence / absence information are associated with a communication status by frequency.
[0056] The rating value estimation model generation unit 206 uses the past data set 314 acquired by the past data set acquisition unit 202 to input frequency-specific time series data and generate a rating value estimation machine learning model 50 that outputs frequency-specific rating value data (see FIG. 8). The rating value estimation machine learning model 50 is, for example, a recurrent neural network.
[0057] The radio wave environment data acquisition unit 208 acquires data indicating the radio wave environment of the wireless base station 10. Specifically, the unit acquires N pieces of data obtained by normalizing the amplitude and phase and performing discrete Fourier transform processing as the power spectrum of a wireless signal received by the monitoring receiver 20 installed near the wireless base station 10, as data indicating the radio wave environment of the wireless base station 10. The unit also acquires the signal power of the received wireless signal as data indicating the radio wave environment.
[0058] The usage frequency determination unit 210 inputs frequency-specific time series data with a given conventional period as an observation period to the machine learning model 50 for estimating evaluation value, thereby obtaining output of frequency-specific evaluation value data with a future period following the conventional period as a subsequent period. Then, the usage frequency to be used for wireless communication is determined based on the obtained frequency-specific evaluation value data. Specifically, a candidate frequency is selected based on the obtained frequency-specific evaluation value data, and the current usage frequency is determined using the previous usage frequency determined last time and the candidate frequency (see FIG. 9). In addition, the frequency-specific time series data (time series data of communication conditions by frequency) during the observation period input to the machine learning model 50 for estimating evaluation value is data of communication conditions by frequency that is output in time series by inputting data indicating the radio wave environment acquired by the radio wave environment data acquisition unit 208 to the machine learning model 40 for estimating communication conditions in time series.
[0059] The storage unit 300 is realized by an IC (Integrated Circuit) memory such as a ROM (Read Only Memory) or a RAM (Random Access Memory) or a storage device such as a hard disk, and stores programs, data, etc. for the processing unit 200 to comprehensively control the frequency determination device 30. The storage unit 300 is also used as a working area for the processing unit 200, and temporarily stores the results of calculations performed by the processing unit 200, input data from the operation unit 102, etc. In this embodiment, a frequency determination program 302 and radio base station information 310 are stored in the storage unit 300.
[0060] The radio base station information 310 is generated for each radio base station 10, and stores machine learning model data 311 for estimating communication conditions, teacher data 312 used to generate the machine learning model 40 for estimating communication conditions, machine learning model data 313 for estimating evaluation value, a past dataset 314 acquired by the past dataset acquisition unit 202 and used to generate the machine learning model 50 for estimating evaluation value, radio wave environment data 315, frequency-specific communication conditions accumulation data 316, and frequency-specific evaluation value accumulation data 317 in association with the identification information (base station ID) of the corresponding radio base station 10.
[0061] The machine learning model data for estimating communication conditions 311 is data defining the machine learning model for estimating communication conditions 40 for the corresponding wireless base station 10, specifically, data of weighting coefficients. This data is generated by the model generating unit for estimating communication conditions 204. The teacher data 312 is data used for generating the machine learning model for estimating communication conditions 40 by the model generating unit for estimating communication conditions 204. The machine learning model data for estimating evaluation value 313 is data defining the machine learning model for estimating evaluation value 50 for the corresponding wireless base station 10, specifically, data of weighting coefficients. This data is generated by the model generating unit for estimating evaluation value 206. The past data set 314 is data acquired by the past data set acquiring unit 202, and is data used for generating the machine learning model for estimating evaluation value 50 by the model generating unit for estimating evaluation value 206. The radio wave environment data 315 is data indicating the radio wave environment of the corresponding wireless base station 10 acquired by the radio wave environment data acquiring unit 208. The frequency-specific communication status accumulated data 316 is time-series accumulated data of the communication status by frequency of the corresponding wireless base station 10 obtained using the machine learning model for estimating communication status 40. The frequency-specific evaluation value accumulated data 317 is time-series accumulated data of the evaluation value B by frequency of the corresponding wireless base station 10 obtained using the machine learning model for estimating evaluation value 50.
[0062] [Effects] According to the present embodiment, a technique for determining a frequency to be used in a radio train control system, which can select an available frequency with high accuracy, can be realized. That is, a machine learning model for estimating an evaluation value 50 can be generated, which inputs time series data showing a communication status by frequency based on the SNR of the radio base station 10 when the radio base station 10 and the on-board device 82 perform radio communication over a predetermined observation period, and outputs evaluation value data of the communication status in a subsequent period following the observation period. Then, a frequency to be used by the radio base station 10 for radio communication can be determined based on the frequency-specific evaluation value data of the subsequent period obtained by inputting the time series data of the communication status by frequency in a conventional period to the machine learning model for estimating an evaluation value 50. As a result, the communication status in a future period following the conventional period can be estimated as an evaluation value from the communication status by frequency in the conventional period, so that a technique for determining a frequency to be used in a radio train control system, which can select an available frequency with high accuracy, can be realized.
[0063] Incidentally, the applicable embodiments of the present invention are not limited to the above-described embodiments, and can of course be modified as appropriate without departing from the spirit of the present invention.
[0064] (A) Number of frequencies used For example, the frequency determination device 30 may determine a plurality of frequencies as the frequencies to be used. In this case, if there is a previously used frequency among the candidate frequencies, the frequency to be used this time is determined by the frequency to be used last time, and further, based on the time-series data of the communication conditions by frequency output from the machine learning model for estimating communication conditions 40, the frequency ratio of frequencies with an interference level equal to or higher than a predetermined level is calculated for each frequency, and the frequencies are selected in ascending order of the calculated ratios as the frequencies to be used, thereby determining the plurality of frequencies to be used (see FIG. 9).
[0065] (B) Monitor receiver 20 In the above embodiment, the monitoring receivers 20 are installed near the corresponding wireless base stations 10 in one-to-one correspondence with the wireless base stations 10, but they may be installed at any positions. In this case, for example, the communication conditions around the wireless base station 10 are estimated based on the wireless signals received by the monitoring receiver 20 installed at the position closest to the installation position of the wireless base station 10.
[0066] Alternatively, the wireless base station 10 may also serve as the monitoring receiving device 20. In this case, the frequency determining device 30 estimates the communication conditions around the wireless base station 10 based on the radio signal received by the wireless base station 10. [Explanation of symbols]
[0067] 1. Radio train control system 10...Radio base station 20...Monitor receiver 30...Frequency determination device 40…Machine learning model for estimating communication status 42…Submodel 44…Noise power estimation section 46…Main model 50…Machine learning model for estimating evaluation value 200... Processing section 202…Past Data Set Acquisition Section 204...Communication status estimation model generation unit 206...Evaluation value estimation model generation unit 208…Radio wave environment data acquisition section 210...Used frequency determination section 300...Storage section 302...Frequency determination program 310...Radio base station information 311…Machine learning model data for estimating communication status 312(312a,312b)…Teacher data 313…Machine learning model data for estimating evaluation values 314...Past Data Sets 315…Radio wave environment data 316…Accumulated communication status data by frequency 317…Frequency-specific evaluation value accumulation data 80…Train 82...Onboard equipment
Claims
1. acquiring a past data set in which given frequency-specific time series data showing frequency-specific communication conditions based on SNR (Signal to Noise Ratio) by wireless base stations installed along railway lines that perform wireless communication with on-board equipment is associated with frequency-specific evaluation value data showing evaluation values by frequency of the communication conditions in a subsequent period following the observation period; Using the past data set, generate a machine learning model for estimating an evaluation value by inputting the frequency-specific time series data and outputting the frequency-specific evaluation value data; inputting the frequency-specific time series data, in which a given conventional period is set as the observation period, into the machine learning model for estimating evaluation value, to obtain an output of the frequency-specific evaluation value data, in which a future period following the conventional period is set as the subsequent period; and determining a frequency to be used for the wireless communication based on the obtained frequency-specific evaluation value data; A frequency determination method including:
2. the machine learning model for estimating the evaluation value is a recurrent neural network, The past data set includes a data set of a plurality of frequency-specific time series data, each of which is a period in which the observation period overlaps with another period in which the observation period shifts in sequence, and the frequency-specific evaluation value data corresponding to the frequency-specific time series data, The frequency determination method according to claim 1 .
3. Determining the frequency to be used selecting candidate frequencies based on the obtained frequency-specific evaluation value data; determining a current frequency to be used by using a previously determined frequency to be used and the candidate frequency; Including, The frequency determination method according to claim 1 or 2.
4. Obtaining the past data set includes: inputting data indicating the radio wave environment in a time series into a machine learning model for estimating a communication status, the machine learning model being trained based on teacher data that associates data indicating a radio wave environment of the wireless base station with the communication status by frequency under the radio wave environment, and sequentially acquiring data on the communication status; Among the sequentially acquired communication status data, 1) data corresponding to the observation period is regarded as the frequency-specific time series data, and 2) data corresponding to the subsequent period following the observation period is obtained by calculating the evaluation value by frequency and regards the data as the frequency-specific evaluation value data, thereby acquiring the past data set; Including, A frequency determination method according to any one of claims 1 to 3.
5. The machine learning model for estimating a communication state includes a sub-model and a main model, the sub-model is a model trained based on teacher data in which the data indicating the radio wave environment is associated with noise presence / absence information by frequency, the main model is a model trained based on teacher data that associates the data indicating the radio wave environment and the noise presence / absence information with the communication status by frequency, The sequential acquisition of the communication status data includes: inputting data indicating the radio wave environment into the sub-model and acquiring the noise presence / absence information by frequency; inputting the data indicating the radio wave environment and the acquired noise presence / absence information for each frequency into the main model to acquire data on the communication status; Including, The frequency determination method according to claim 4.
6. an acquisition unit that acquires a past data set that associates given frequency-specific time series data, which shows a communication status by frequency based on an SNR (Signal to Noise Ratio) by a wireless base station installed along a railway line that wirelessly communicates with an on-board device, over a predetermined observation period, with frequency-specific evaluation value data, which shows an evaluation value by frequency that evaluates the communication status in a subsequent period following the observation period; a generation unit that generates a machine learning model for estimating an evaluation value by inputting the frequency-specific time series data and outputting the frequency-specific evaluation value data, using the past data set; a determination unit that inputs the frequency-specific time series data, in which a given conventional period is set as the observation period, into the machine learning model for estimating an evaluation value, thereby obtaining an output of the frequency-specific evaluation value data, in which a future period following the conventional period is set as the subsequent period, and determines a frequency to be used for the wireless communication based on the obtained frequency-specific evaluation value data; A frequency determining device comprising:
Citation Information
Patent Citations
A method for cognitive radio anti-interference intelligent decision based on reinforcement learning
CN108712748A
Radio control method and radio control system
JP2012090328A
Cognitive radio systems, and their resource allocation devices and methods
JP2015518356A
Management device, program for making computer execute, and computer-readable recording medium recording program
JP2018142957A
