Optical receiving device and program

Abstract

It precisely tracks the light beam. [Solution] The optical receiving device includes prediction means for performing a prediction process to predict the first position of a sensor irradiated by a light beam when the optical element is in a reference state, using one of a plurality of models, and selection means for performing a selection process to select one model to be used in the prediction process. The plurality of models include a real-time model that predicts the first position based on the history of the second position of the sensor irradiated by the light beam, and N learning models that predict the first position based on input information including the history of the values ​​of each of a plurality of parameters. The selection means selects one or more target models from the plurality of models based on the value of at least one of the plurality of parameters included in the input information, and selects one model to be used in the next prediction process by determining the prediction error between the real-time model and the one or more target models.

Description

Technical Field

[0001] The present disclosure relates to an optical receiver used in a free space optics (FSO) communication system. More specifically, the present disclosure relates to a technique for tracking an optical beam transmitted into space by an optical transmitter in an FSO communication system.

Background Art

[0002] FSO communication is used, for example, for communication between optical communication devices mounted on moving objects such as satellites, or for communication between an optical communication device mounted on a moving object and an optical communication device installed on the ground or the like. In the following description, an optical communication device that transmits an optical beam into space is referred to as an optical transmitter, and an optical communication device that receives the optical beam transmitted by the optical transmitter is referred to as an optical receiver. When at least one of two optical communication devices performing FSO communication moves, the optical receiver performs tracking control of the optical beam transmitted by the optical transmitter.

[0003] Non-Patent Document 1 discloses a tracking configuration of an optical beam in an FSO communication system. Further, Non-Patent Document 2 discloses using a learning model generated by machine learning for tracking an optical beam in an FSO communication system.

Prior Art Documents

Non-Patent Documents

[0004]

Non-Patent Document 1

Non-Patent Document 2

[0005] Light beam tracking includes control to adjust the direction of the optical antenna that focuses the light beam, and control to adjust the attitude of tracking optical elements such as mirrors and lenses so that most of the light beam focused by the optical antenna is input to the receiving section of the optical receiving device. To control the attitude of the tracking optical elements, real-time control such as PID (proportional-integral-derivative) predictive control is used. However, real-time control cannot keep up with rapid changes in the direction or position of incidence of the light beam to the optical receiving device.

[0006] Patent Document 2 discloses the use of a learning model for tracking control, but the number of parameters that affect the direction and position of incidence of the light beam to the optical receiver is large, and the number of input information patterns used as input to the learning model is also enormous, making it difficult to maintain tracking accuracy with a single learning model.

[0007] This disclosure provides a technology for accurately tracking an optical beam. [Means for solving the problem]

[0008] According to one aspect of the present disclosure, the light receiving device includes a tracking mechanism comprising an optical member for tracking a light beam and a sensor for detecting a position illuminated by the light beam that has passed through the optical member; prediction means for performing a prediction process to predict a first position of the sensor illuminated by the light beam when the optical member is in a reference state, using one of a plurality of models; driving means for driving the optical member so that the light beam illuminates a predetermined position of the sensor based on the first position predicted by the prediction means using the one model; and selection means for performing a selection process to select the one model to be used in the prediction process. The system comprises a real-time model that predicts the first position based on the history of the second position of the sensor irradiated by the light beam, and a first to Nth learning model (where N is an integer of 2 or more) that predicts the first position based on input information including the history of the values ​​of each of the multiple parameters. The selection means selects one or more target models from the multiple models based on the value of at least one of the multiple parameters included in the input information before the selection process is started, and selects one of the models to be used in the next prediction process by determining the prediction error between the real-time model and the one or more target models. [Effects of the Invention]

[0009] According to this disclosure, it is possible to track a light beam with high precision. [Brief explanation of the drawing]

[0010] [Figure 1] A diagram showing an example configuration of an optical receiving device. [Figure 2] A diagram showing an example of the tracking unit configuration. [Figure 3] A diagram showing an example of judgment information. [Figure 4] Timing chart for selection and prediction processes. [Modes for carrying out the invention]

[0011] The embodiments will be described in detail below with reference to the attached drawings. Note that the following embodiments do not limit the invention as defined in the claims, and not all combinations of features described in the embodiments are essential to the invention. Two or more of the features described in the embodiments may be combined arbitrarily. Furthermore, identical or similar configurations will be given the same reference numeral, and redundant descriptions will be omitted.

[0012] Figure 1 shows an example configuration of the optical receiver 200 according to an embodiment. In the following description, it is assumed that the optical receiver 200 is installed on the ground, and the optical transmitter that transmits the optical beam to the optical receiver 200 is mounted on an artificial satellite. The optical receiver 200 receives the optical beam from the optical antenna 100 which focuses the optical beam. The tracking unit 1 controls the state of tracking optical elements such as mirrors and lenses, for example, the position and attitude of the tracking optical elements, so that as much of the optical beam from the optical antenna 100 as possible enters the receiving unit 2. The receiving unit 2 performs processing related to the demodulation of the optical beam.

[0013] Figure 2 shows an example configuration of the tracking unit 1. A light beam is input to the tracking mechanism 30. The tracking mechanism 30 includes tracking optical elements such as lenses and mirrors for tracking the light beam, and a two-dimensional optical sensor that detects the light beam downstream of the tracking optical elements. The two-dimensional optical sensor detects the position on the two-dimensional plane illuminated by the light beam and outputs the detected position on the two-dimensional plane as the detection result. The light beam that has passed through the tracking mechanism 30 is split into a light beam toward the two-dimensional optical sensor and a light beam toward the receiving unit 2. The optical receiving device 200 is configured such that the power of the light beam incident on the receiving unit 2 is maximized when the light beam is irradiated onto a predetermined position on the two-dimensional optical sensor. Tracking of the light beam is performed by controlling the spatial state (position and orientation) of the tracking optical elements so that the light beam continues to irradiate the predetermined position on the two-dimensional optical sensor.

[0014] The drive unit 20 controls the state of the tracking optical element so that the light beam continues to illuminate a predetermined position on the two-dimensional optical sensor, based on the first prediction result output by the prediction unit 10. The first prediction result may indicate, for example, the position on the two-dimensional optical sensor illuminated by the light beam when the tracking optical element is in a predetermined state (hereinafter also referred to as the reference state). The drive unit 20 controls the state of the tracking optical element of the tracking mechanism 30 based on the difference between the position indicated by the first prediction result and the predetermined position on the two-dimensional optical sensor. The tracking mechanism 30 notifies the selection unit 40 and the prediction unit 10 of the position of the two-dimensional optical sensor illuminated by the light beam as an error in the control by the drive unit 20. The error in the control by the drive unit 20 is determined as the difference between the position of the two-dimensional optical sensor illuminated by the light beam and the predetermined position on the two-dimensional optical sensor.

[0015] In the following description, the predetermined position of the 2D optical sensor is defined as the origin in the 2D plane, and therefore, this predetermined position is referred to as the center position of the 2D sensor. In this case, the position of the 2D optical sensor illuminated by the light beam indicates the control error by the drive unit 20. Therefore, in the following description, the position of the 2D optical sensor illuminated by the light beam notified by the tracking mechanism 30 to the selection unit 40 and the prediction unit 10 is referred to as the "control error". Furthermore, the position of the 2D optical sensor illuminated by the light beam when the tracking optical element is in a reference state is referred to as the "uncontrolled illumination position".

[0016] The prediction unit 10 outputs a first prediction result to the drive unit 20, and also outputs the first prediction result, the second prediction result, and the third prediction result to the selection unit 40. In the following description, the first to third prediction results will be collectively referred to simply as "prediction result". In this example, each prediction result indicates an uncontrolled irradiation position.

[0017] The prediction unit 10 has, for example, a PID model that outputs a prediction result by PID (proportional, integral, derivative) calculation. A control error is input to the PID model at every predetermined period. The PID model predicts the non-control irradiation position at the next timing based on the sum of a value proportional to the control error, an integral value over a predetermined past period of the control error, and a derivative value determined based on the control error in the past predetermined period. For example, the control error is input to the PID model every 1 second, the PID model outputs a prediction result every 1 second, the predetermined past period for obtaining the integral value is 10 seconds, and the predetermined past period for obtaining the derivative value is 2 seconds. In this case, the PID model obtains a proportional value based on the latest one control error, obtains an integral value based on the past 10 control errors, and obtains a derivative value based on the past 2 control errors, thereby predicting the non-control irradiation position 1 second later.

[0018] The prediction unit 10 further has N learning models from the first learning model to the Nth learning model. N is an integer of 2 or more. In the following description, the first learning model to the Nth learning model are collectively referred to as "learning models", and the learning models and the PID model are collectively referred to simply as "models". Each learning model outputs a prediction result based on input information. The teacher information used to generate each learning model will be described later.

[0019] The input information includes values for a plurality of parameters that affect the incident direction and incident position of the light beam on the optical receiver 2 over a predetermined past period. For example, the change in the refractive index of the atmosphere, that is, the incident direction of the light beam on the optical receiver 200 can vary due to scintillation. Also, the incident direction of the light beam on the optical receiver 2 can vary due to scattering by suspended matter in the atmosphere. Since the refractive index of the atmosphere and the suspended matter depend on weather, temperature, humidity, atmospheric pressure, wind speed, the position of the artificial satellite carrying the optical transmitter, the amount of dust in the atmosphere, etc., these parameters can be used as input information. Furthermore, temperature changes and vibrations of the optical receiver 200 and the optical transmitter can also affect the incident direction of the light beam on the optical receiver 200. Therefore, these parameters can also be used as input information.

[0020] The optical receiving device 200 collects input information by a collection unit (not shown). Parameters such as weather, temperature, humidity, atmospheric pressure, wind speed, and the amount of dust in the atmosphere can be obtained from a server device used by a provider that provides weather information and the like to provide weather information. Parameters such as the temperature, temperature change, and vibration of the optical receiving device 200 can be obtained by measurement by the optical receiving device 200. Regarding the position of the artificial satellite equipped with the optical transmitting device, it can be determined based on the orbital information of the artificial satellite. Parameters such as the temperature, temperature change, and vibration of the optical transmitting device can be measured by the optical transmitting device and notified to the optical receiving device.

[0021] Each learning model outputs a prediction result at the next timing based on the history of the values of each parameter for a predetermined period in the past. For example, assume that the values of each parameter are sampled at a 1-second cycle, 10 parameters are used, the predetermined period in the past is 10 seconds, and the prediction result is output every 1 second. In this case, the learning model outputs the prediction result 1 second later based on 10 values for each of the 10 parameters, that is, a total of 100 values.

[0022] The input information is also input to the selection unit 40. The selection unit 40 classifies the input information into any one of group #1 (G#1) to group #N (G#N) based on the values of one or more parameters included in the input information. Note that the classification information indicating the correspondence between the values of one or more parameters used for classification and the groups is determined in advance and set in the selection unit 40. As an example, when classifying the input information based only on the weather, the classification information indicates the correspondence between each weather such as sunny, rainy, foggy, etc. and the group. Also, when classifying based on the weather and temperature, the classification information indicates the correspondence between each combination of the weather such as sunny, rainy, foggy, etc. and the temperature range and the group. As an example, the classification information indicates that when the weather is sunny and the temperature is within the first range, it is classified into G#1, when the weather is rainy and the temperature is within the first range, it is classified into G#2, and when the weather is sunny and the temperature is within the second range, it is classified into G#3, etc.

[0023] The selection unit 40 notifies the prediction unit 10 of the group to which the input information belongs and the model that should output the second prediction result based on the judgment information shown in Figure 3. In this example, the third prediction result is always output by the PID model. As shown in Figure 3, the judgment information indicates the correspondence between the group of input information and the model that should output the second prediction result. According to the example shown in Figure 2, if the input information belongs to G#1, the first learning model is used to output the second prediction result; if the input information belongs to G#2, the third learning model is used to output the second prediction result; and if the input information belongs to G#3, the PID model is used to output the second prediction result. The method for creating the judgment information will be described later.

[0024] Furthermore, the selection unit 40 performs a selection process based on the prediction errors of the second and third prediction results notified by the prediction unit 10, selecting the model that should output the first prediction result from the model that output the second prediction result and the PID model that output the third prediction result, and notifying the prediction unit 10 of this selection.

[0025] Figure 4 is an explanatory diagram of the process of changing the model that outputs the first prediction result. The number n (where n is an integer from 1 to N) within the rectangle indicates that the nth learning model is outputting the corresponding prediction result, and the letter P within the rectangle indicates that the PID model is outputting the corresponding prediction result.

[0026] The process in Figure 4 includes a selection process that runs for a period of T1 with a period of T2, and a prediction process that runs for a period of T2 with a period of T2. The prediction process outputs the first prediction result. The selection process selects the model to be used in the next prediction process.

[0027] As shown in Figure 4, at time T#0, the selection unit 40 notified the prediction unit 10 that the first learning model should be used to output the first prediction result. As a result, the prediction unit 10 uses the first learning model to output the first prediction result in the prediction process from time T#0.

[0028] The selection unit 40 determines the learning model that outputs the second prediction result based on the group to which the input information immediately preceding time T#1 belongs and the judgment information, and notifies the prediction unit 10 of the determined learning model. Between time T#1 and time T#2, the prediction unit 10 outputs the second prediction result by inputting the input information to the learning model notified by the selection unit 40. According to Figure 4, the group of input information immediately preceding time T#1 is G#1. According to the judgment information shown in Figure 3, the learning model corresponding to G#1 is the first learning model. Therefore, between time T#1 and time T#2, the prediction unit 10 outputs the second prediction result by inputting the input information to the first learning model. The prediction unit 10 also outputs a third prediction result using the PID model between time T#1 and time T#2.

[0029] The selection unit 40 determines the evaluation value of the prediction error between the second and third prediction results at time T#1 to time T#2. The evaluation value of the prediction error may be the sum of the squares or absolute values ​​of the prediction error at each timing between time T#1 and time T#2. The selection unit 40 receives the first prediction result from the prediction unit 10 and the control error of the first prediction result from the tracking mechanism 30. Therefore, the selection unit 40 can determine the actual uncontrolled irradiation position at time T#1 to time T#2. Thus, the selection unit 40 can determine the error of the prediction result at each timing between the second and third prediction results at time T#1 to time T#2.

[0030] Furthermore, the tracking mechanism 30 can be configured to detect uncontrolled irradiation positions in addition to control errors. For example, by splitting the light beam upstream of the tracking optical element, directing one beam into the tracking optical element and the other into another two-dimensional optical sensor for detecting the "uncontrolled irradiation position," the tracking mechanism 30 can detect the actual uncontrolled irradiation position. If the tracking mechanism 30 can detect the actual uncontrolled irradiation position, the selection unit 40 can determine the error in the prediction results at each timing of the second prediction result and the third prediction result based on the actual uncontrolled irradiation position detected by the tracking mechanism 30. In this case, it is not necessary to output the first prediction result and the control error to the selection unit 40.

[0031] The selection unit 40 determines that the model that outputs the prediction result with the smaller evaluation value of the prediction error between time T#1 and time T#2 will be used for the prediction processing from time T#2, and notifies the prediction unit 10. In the example in Figure 4, since the evaluation value of the prediction error of the second prediction result is smaller than the evaluation value of the prediction error of the third prediction result, the selection unit 40 selects the first learning model as the model to be used for the prediction processing from time T#2. Therefore, the prediction unit 10 outputs the first prediction result using the first learning model between time T#2 and time T#4.

[0032] The selection unit 40 determines that the group of input information immediately preceding time T#3 is G#2, and therefore notifies the prediction unit 10 of the third learning model as the learning model that outputs the second prediction result based on the judgment information shown in Figure 3. Accordingly, the prediction unit 10 outputs the second prediction result by inputting the input information to the third learning model between time T#3 and time T#4. The prediction unit 10 also outputs the third prediction result using the PID model between time T#3 and time T#4.

[0033] The selection unit 40 determines the model to be used for prediction processing from time T#4 by determining the evaluation value of the prediction error of the second prediction result and the third prediction result between time T#3 and time T#4. In the example in Figure 4, since the evaluation value of the prediction error of the second prediction result is smaller than the evaluation value of the prediction error of the third prediction result, the selection unit 40 selects the third learning model as the model to be used for prediction processing from time T#4. Therefore, the prediction unit 10 outputs the first prediction result using the third learning model between time T#4 and time T#6.

[0034] The selection unit 40 determines that the group of input information immediately preceding time T#5 is G#2, and therefore notifies the prediction unit 10 of the third learning model as the learning model that will output the second prediction result based on the judgment information shown in Figure 3. Accordingly, the prediction unit 10 outputs the second prediction result by inputting the input information to the third learning model between time T#5 and time T#6. The prediction unit 10 also outputs the third prediction result using the PID model between time T#5 and time T#6.

[0035] The selection unit 40 determines the model to be used for prediction processing from time T#6 by determining the evaluation value of the prediction error of the second prediction result and the third prediction result between time T#5 and time T#6. In the example in Figure 4, since the evaluation value of the prediction error of the third prediction result is smaller than the evaluation value of the prediction error of the second prediction result, the selection unit 40 selects the PID model as the model to be used for prediction processing from time T#6. Therefore, the prediction unit 10 outputs the first prediction result using the PID model between time T#6 and time T#8.

[0036] The selection unit 40 determines that the learning model that outputs the second prediction result is a PID model based on the judgment information shown in Figure 3, since the group of input information immediately preceding time T#8 is G#3. Since the learning model that outputs the third prediction result is also a PID model, the selection unit 40 determines that the model to be used in the prediction process from time T#8 is a PID model and notifies the prediction unit 10. In this case, there is no need to determine the prediction error. Therefore, the prediction unit 10 outputs the first prediction result using the PID model from time T#8 until the period T2 has elapsed.

[0037] Next, we will explain the generation of the learning model. In order to generate the learning model, we collect training information in advance. The training information includes input information and the correct label for that input information. The correct label for the input information may be the uncontrolled irradiation position that the learning model should output when that input information is obtained. Next, the training information is classified into sets 1 through N based on its input information. The classification criteria are the same as the classification information set in the selection unit 40. In other words, the input information included in the training information of the nth set (where n is an integer from 1 to N) belongs to G#n. The nth learning model is generated by performing machine learning based on the training information of the nth set. In the following explanation, the nth learning model generated with training information that includes input information belonging to G#n, that is, training information classified into the nth set, will also be referred to as the "learning model based on G#n".

[0038] Next, we will explain the generation of judgment information. The nth learning model is generated using training information that includes input information belonging to G#n. Therefore, it is generally assumed that the nth learning model will have the smallest prediction error when input information belonging to G#n is input to each of the first to nth learning models. However, as mentioned above, the number of parameters that should be used as input information is large, and when the input information is grouped by combinations of values ​​for each of the multiple parameters that should be used as input information, the number of groups N becomes enormous. For example, if the number of parameters is 10, and one parameter is grouped into a range of 10 values, the number of groups N becomes 10 10 However, 10 10 Generating a single learning model is practically impossible. Therefore, in this embodiment, the number of learning models to be generated is made a realistic number by grouping the input information based on some of the parameters used as input information.

[0039] However, by grouping input information based on some of the multiple parameters used as input information, the correlation between input information belonging to the same group decreases. Therefore, even for input information belonging to G#n, there are cases where prediction using a different model than the nth-th learning model results in a lower error than prediction using the nth-th learning model.

[0040] Therefore, in this embodiment, verification information is used to create determination information that shows the correspondence between the group to which the input information belongs and the model. The verification information is information that shows the time-series data of the uncontrolled irradiation position and the temporal changes of the input information during that time, and can be obtained by prior measurement. The verification information is also classified into sets 1 to N based on the group to which the input information belongs.

[0041] The prediction results obtained by using the input information of the validation information included in the nth set as input to the first to the nth learning models are compared with the uncontrolled irradiation positions indicated by the validation information to obtain an evaluation value of the prediction error of the first to the nth learning models. Similarly, for the PID model, the initial value of the error is appropriately set and the output of the PID model is compared with the uncontrolled irradiation positions indicated by the validation information to obtain an evaluation value of the prediction error of the PID model. The evaluation value of the prediction error is the same as the evaluation value obtained by the selection unit 40 in the selection process.

[0042] The judgment information is created by selecting the model with the smallest evaluation value as the model to use for input information belonging to G#n. The judgment information can also be generated by having all models make predictions in the configuration shown in Figure 2, and then determining which model produced the prediction result with the lowest evaluation value of the prediction error, and which group the input information belongs to at that time. In other words, the judgment information is generated based on prior measurements of the prediction errors of each of the multiple models. Prior measurements include measuring the prediction error using verification information collected in advance, and measuring the prediction error while measuring the input information and uncontrolled irradiation position in the actual environment. In any case, the judgment information can be generated before the operation of the optical receiver 200 and stored in the selection unit 40 before the start of operation of the optical receiver 200.

[0043] Based on the above, the nth training model is generated using the nth set of training information. The nth set of training information includes input information classified as G#n. Then, using validation information, the model to be used for the input information classified as G#n is selected from the first training model to the nth training model and the PID model to generate judgment information. In actual operation, the selection unit 40 determines the evaluation value of the prediction error of the model indicated by the judgment information and the evaluation value of the prediction error of the PID model, and selects the model to be used in the next prediction process based on the judgment result. With this configuration, accurate tracking control can be performed while suppressing prediction errors.

[0044] Furthermore, during operation, the optical receiver 200 collects the input information used for prediction and the uncontrolled irradiation position in order to generate training information used for retraining the learning model. In addition, it also collects verification information for updating the judgment information. The optical receiver 2 is configured to be able to update each model and judgment information.

[0045] Furthermore, the selection unit 40 calculates the proportion of times a PID model was selected in the selection process, even though the learning model indicated by the judgment information was the nth learning model. In other words, it calculates the proportion of times a PID model was selected in these selection processes relative to the total number of times the nth learning model was used for selection processes. If the number of times the nth learning model was used for selection processes exceeds a predetermined number, and the proportion exceeds a predetermined threshold, the selection unit 40 determines that the prediction performance of the nth learning model is low. In this case, the selection unit 40 may inform the user that the nth learning model needs to be retrained. Also, when the nth learning model is retrained and updated, the judgment information may also be updated using the verification information.

[0046] In this embodiment, the decision information indicated one of the first to Nth learning models and the PID model as the model that outputs the second prediction result. However, it is also possible to exclude the PID model from the decision information and indicate one of the first to Nth learning models. This is based on the fact that the PID model is always used in the selection process.

[0047] Furthermore, in this embodiment, the selection process selects the model to be used in the next prediction process from the model indicated by the judgment information and the PID model. However, it is also possible to configure the selection process to select from the model indicated by the judgment information, the PID model, and the learning model based on the group to which the input information belongs. For example, if the input information belongs to G#2 and the judgment information indicates a third learning model as the learning model corresponding to the input information, the selection process can measure the prediction errors of the third learning model indicated by the judgment information, the PID model, and the second learning model, which is a learning model based on G#2 to which the input information belongs, and select one model. Moreover, if the judgment information does not exclude the PID model as the model that outputs the second prediction result, the model to be used in the next prediction process can be configured to select from the model indicated by the judgment information and the learning model based on the group to which the input information belongs. Furthermore, it is also possible to configure the selection process to select the model to be used in the next prediction process from the learning model and the PID model based on the group to which the input information belongs, without using judgment information.

[0048] Furthermore, although the embodiment was described using a PID model as an example of a real-time model, any real-time model that predicts the uncontrolled irradiation position at the next timing based on the history of control errors, the history of uncontrolled irradiation positions, etc., can be used instead of the PID model.

[0049] As described above, the input information includes values ​​for each of the multiple parameters over a predetermined past period. The grouping of the input information is performed based on the values ​​of one or more of the multiple parameters. Therefore, changes in the values ​​of one or more parameters within the predetermined past period may result in the values ​​of one or more parameters spanning multiple groups during that period. For example, when classifying input information based solely on weather, it is possible that the weather was sunny in the first half of the predetermined past period and rainy in the second half. In this case, the selection unit 40 can be configured to group based on the latest input information value within the predetermined past period.

[0050] Furthermore, in this embodiment, the model that outputs the second prediction result was selected based on the group of input information immediately before the selection process. However, it is also possible to configure the system to select the model that outputs the second prediction result in the selection process based on the group of input information at any time before the selection process. Alternatively, it is also possible to configure the system to determine one group based on the input information over a predetermined period before the selection process, and then select the model that outputs the second prediction result in the selection process based on the determined group. For example, the selection unit 40 may determine and count the groups of input information at each time point in the predetermined period before the selection process, and determine the group with the highest count as the one group described above.

[0051] Furthermore, while the nth learning model is generated based on the nth set of training information, that is, training information containing input information belonging to G#n, the set of training information used to generate the nth learning model may be classified by a different criterion than the criterion used to group the input information. This is based on creating judgment information by predetermining the model corresponding to the group of input information based on validation information, and then selecting the model to be used in the prediction process from the model indicated by the judgment information and the real-time model in the selection process. In this case, the number of input information groups and the number of learning models may differ.

[0052] As described above, according to this embodiment, the prediction unit 10 predicts the uncontrolled irradiation position (first position), which is the position of the 2D optical sensor irradiated by the light beam when the tracking optical member is in a reference state, using one of the multiple models in the prediction process. The multiple models include a real-time model that predicts the first position based on the history of the second position of the 2D optical sensor irradiated by the light beam, and N learning models (N is an integer of 2 or more) that predict the first position based on input information including the history of the values ​​of each of the multiple parameters. The second position corresponds to the control error in the embodiment. The selection unit 40 performs a selection process to select one model to be used in the next prediction process and notifies the prediction unit 10 of the selected model. The selection unit 40 selects one or more target models from the multiple models based on the value of at least one parameter among the multiple parameters included in the input information before starting the selection process, and selects one model to be used in the next prediction process by determining the prediction error between the real-time model and the one or more target models. This configuration enables accurate tracking of the light beam.

[0053] In the above example, the optical receiver 200 was assumed to be installed on the ground and the optical transmitter mounted on a moving object. However, the above embodiment can also be applied when the optical transmitter is installed on the ground and the optical receiver 200 is mounted on a moving object, or when both the optical receiver 200 and the optical transmitter are mounted on a moving object. Furthermore, fluctuations in the light beam due to scintillation and airborne particles can occur even when both the optical receiver 200 and the optical transmitter are installed on the ground. Therefore, the above embodiment is also applicable when both the optical receiver 200 and the optical transmitter are installed on the ground. In addition, the optical receiver 200 is also applicable to FSO communication, which propagates an optical beam underwater. Therefore, the optical receiver 200 may be fixedly installed in the water or installed on a moving object that moves underwater. The parameters used for input information are determined according to the installation configuration of the optical transmitter and the optical receiver 200.

[0054] The selection unit 40 and prediction unit 10 of the tracking unit 1 shown in Figure 2 can be realized by having one or more processors provided in the optical receiver 200 execute an appropriate program. In other words, an optical receiver comprising a tracking mechanism 30, a drive unit 20, a receiving unit 2, one or more processors, and one or more memory devices can be made to function as the optical receiver 200 by having one or more processors execute an appropriate program. The program can be stored in one or more memory devices, which are non-temporary computer-readable storage media. Therefore, according to this embodiment, a program for functioning as the optical receiver 200 and a computer-readable storage medium are provided.

[0055] With the above configuration, the tracking of the light beam can be performed with high precision. Therefore, it becomes possible to contribute to Goal 9 of the United Nations-led Sustainable Development Goals (SDGs), "Build resilient infrastructure, promote sustainable industrialization and foster innovation."

[0056] The invention is not limited to the embodiments described above, and various modifications and changes are possible within the scope of the gist of the invention. [Explanation of Symbols]

[0057] 10: Prediction unit, 20: Drive unit, 30: Tracking mechanism, 40: Selection unit

Claims

1. A tracking mechanism including an optical member for tracking a light beam and a sensor for detecting the position irradiated by the light beam that has passed through the optical member, Prediction means that performs prediction processing to predict the first position of the sensor irradiated by the light beam when the optical element is in a reference state, using one of multiple models, A driving means that drives the optical member so that the light beam illuminates a predetermined position of the sensor based on the first position predicted by the prediction means using the one model, Selection means for performing a selection process to select one of the models to be used in the prediction process, Equipped with, The plurality of models include a real-time model that predicts the first position based on the history of the second position of the sensor irradiated by the light beam, and a first to Nth learning model (where N is an integer of 2 or more) that predicts the first position based on input information including the history of the values ​​of each of the plurality of parameters. The selection means selects one or more target models from the plurality of models based on the value of at least one of the plurality of parameters included in the input information before the selection process is started, and selects one model to be used in the next prediction process by determining the prediction error between the real-time model and the one or more target models.

2. The period of the aforementioned selection process is the first period, The optical receiving device according to claim 1, wherein the selection means performs the selection process each time a second period longer than the first period has elapsed to select one model to be used in the next prediction process.

3. The optical receiving device according to claim 1, wherein the selection means determines the prediction error of the first position predicted by the real-time model and the one or more target models, respectively, based on the first position predicted by one model in the prediction process performed in parallel with the selection process and the second position of the sensor irradiated by the light beam.

4. The system further comprises another sensor for measuring the first position, The optical receiving device according to claim 1, wherein the selection means determines the prediction error of the first position predicted by the real-time model and the one or more target models, based on the first position measured by the other sensor.

5. The optical receiving device according to claim 1, wherein the selection means determines the group to which the input information belongs based on the value of at least one of the plurality of parameters included in the input information, and selects one of the one or more target models based on determination information indicating the correspondence between the group and one target model.

6. The input information is classified into one of the first to N groups based on the value of at least one of the multiple parameters. The nth learning model (where n is an integer from 1 to N) is generated by machine learning based on input information classified into the nth group. The optical receiving device according to claim 5, wherein the determination information is generated based on prior measurement of the prediction error of each of the plurality of models.

7. The optical receiving device according to claim 5, wherein the determination information indicates one of the first learning model to the N learning models and the real-time model as a target model corresponding to one group.

8. The optical receiving device according to claim 5, wherein the determination information indicates one of the first to N learning models as a target model corresponding to one group.

9. The input information is classified into one of the first to N groups based on the value of at least one of the multiple parameters. The nth learning model (where n is an integer from 1 to N) is generated by machine learning based on input information classified into the nth group. The optical receiving device according to claim 5, wherein the selection means determines the group to which the input information belongs based on the value of at least one of the plurality of parameters included in the input information, and selects a learning model generated by machine learning based on the input information classified into the determined group as one of the one or more target models.

10. The input information is classified into one of the first to N groups based on the value of at least one of the multiple parameters. The nth learning model (where n is an integer from 1 to N) is generated by machine learning based on input information classified into the nth group. The optical receiving device according to claim 1, wherein the selection means determines the group to which the input information belongs based on the value of at least one of the plurality of parameters included in the input information, and selects a learning model generated by machine learning based on the input information classified into the determined group as one of the one or more target models.

11. The aforementioned one or more target models are one target model, The optical receiving device according to claim 1, wherein the selection means, when the one target model selected based on the value of at least one of the plurality of parameters included in the input information before the selection process is started is the real-time model, selects the real-time model as the one model to be used in the next prediction process without determining the prediction error.

12. The aforementioned one or more target models are one target model, The optical receiving device according to claim 1, wherein the selection means, in a selection process in which a first target model among the plurality of models is selected as one target model, notifies that the proportion of real-time models selected as one model to be used in the next prediction process exceeds a threshold.

13. The optical receiving device according to claim 1, wherein the real-time model predicts the first position based on the sum of a value proportional to the difference between the second position and the predetermined position, the integral of the difference, and the derivative of the difference.

14. The optical receiver according to claim 1, wherein the plurality of parameters include one or more of the following: temperature, humidity, atmospheric pressure, wind speed, amount of dust in the atmosphere, temperature of the optical receiver, temperature of the optical transmitter that transmits the light beam to the optical receiver, and the position of the moving part of the optical receiver and the optical transmitter.

15. A tracking mechanism including an optical member for tracking a light beam and a sensor for detecting the position irradiated by the light beam that has passed through the optical member, A driving means for driving the optical element, One or more processors, A program that, when executed on one or more processors of an optical receiver equipped with, causes the optical receiver to function as an optical receiver according to any one of claims 1 to 14.

Images