JP7328885B2 - MOTION SENSOR SELECTION DEVICE, TARGET DETECTION SYSTEM AND MOTION SENSOR SELECTION METHOD - Google Patents

Description

The present invention mainly relates to a motion sensor selection device for selecting a sensor suitable for detection from a plurality of sensors, and a target detection system including the motion sensor selection device.

Conventionally, a system for detecting an underwater object using a sonobuoy or the like is known. Patent Document 1 discloses a detection system of this kind.

In the underwater vehicle positioning system of Patent Document 1, by arranging a plurality of sonobuoys that receive acoustic signals from pingers mounted on the underwater vehicle in the operational sea area, three or more sonobuoys receive the signals. The position of the underwater vehicle is calculated in real time based on the acoustic signal.

JP-A-2004-245779

Unlike the configuration of the above Patent Document 1, sensors such as a plurality of sound wave generating sonobuoys and sound wave receiving sonobuoys are arranged within the detection range (operational sea area), actively generating sound waves, underwater vehicles, Systems are also used to detect targets such as water vehicles, wrecks, seafloor topography (hereinafter simply "underwater objects"). In order to more accurately detect underwater objects in such a system, there has been a demand for a configuration that automatically selects an appropriate sensor to operate according to the distribution of underwater objects.

In terms of artificial intelligence, which has been developing in recent years, it is possible to automatically select sonobuoys suitable for various situations by building trained models that have learned various situations. However, when artificial intelligence is used, generally, the more complicated the situation to be learned, the longer the learning time required to build a trained model. If it takes a long time to build a trained model, it becomes difficult to operate devices and systems that use the trained model at an early stage.

SUMMARY OF THE INVENTION The present invention has been made in view of the above circumstances, and an object of the present invention is to construct a motion sensor selection model in a short time by machine learning.

The problems to be solved by the present invention are as described above. Next, the means for solving the problems and the effects thereof will be described.

A first aspect of the present invention provides a motion sensor selection device having the following configuration. That is, when a plurality of sensors used for detecting one or more targets existing in a target area are arranged in the target area, the operation sensor selection device selects the sensor to be operated among the plurality of sensors. Select as motion sensor. The motion sensor selection device comprises an input unit, a motion sensor selection model, and a motion sensor selection unit. The input unit receives input of input information, which is information including the target state quantity regarding at least one of position, course and speed regarding each of the targets, regarding at least one of the targets. The motion sensor selection model performs machine learning for selecting at least one sensor to be selected from among the plurality of sensors from base information including the target state quantity in a certain time span for one target. I have been there. The motion sensor selection unit uses the motion sensor selection model to select the sensor to operate from the input information input to the input unit. The motion sensor selection unit decomposes the input information regarding the plurality of targets input to the input unit into information regarding one target, and inputs each decomposed information to the motion sensor selection model. , integrates a plurality of selection results output from the motion sensor selection model, and selects the sensor to be operated as a motion sensor.

This allows the learning data to be limited to situations for only one goal when building the motion sensor selection model in the learning phase. As a result, an explosive increase in the number of learning data can be avoided, and the learning time of the motion sensor selection model can be significantly shortened, so that early operation of the motion sensor selection device can be achieved. In the inference phase, the situations related to multiple goals are decomposed into individual goals and input into the motion sensor selection model. By integrating the output results of the motion sensor selection model, even complex situations where multiple targets appear can be handled. can respond appropriately.

According to a second aspect of the present invention, the following motion sensor selection method is provided. That is, in this operation sensor selection method, when a plurality of sensors used for detecting one or more targets existing in a target area are arranged in the target area, the sensor to be operated among the plurality of sensors is selected. Select as motion sensor. The motion sensor selection method includes a first step, a second step, and a third step. In the first step, with respect to at least one target, input of input information including the target state quantity relating to at least one of position, course and speed relating to each target is received. In the second step, the input information regarding the plurality of targets input in the first step is decomposed into information regarding one target, and each decomposed information is input to a motion sensor selection model. The motion sensor selection model performs machine learning for selecting at least one sensor to be selected from among the plurality of sensors from base information including the target state quantity in a certain time span for one target. This is the trained model that we used. In the third step, a plurality of selection results output from the motion sensor selection model are integrated to select the sensor to operate.

This allows the learning data to be limited to the context of only one goal when building the motion sensor selection model in the learning phase. As a result, an explosive increase in the number of learning data can be avoided, and the learning time of the motion sensor selection model can be significantly shortened, so that early operation of the motion sensor selection device can be achieved. In the inference phase, the situations related to multiple goals are decomposed into individual goals and input into the motion sensor selection model. By integrating the output results of the motion sensor selection model, even complex situations where multiple targets appear can be handled. can respond appropriately.

According to the present invention, a motion sensor selection model can be constructed in a short time by machine learning.

1 is an explanatory diagram showing an outline of a target detection system according to a first embodiment of the present invention; FIG. FIG. 4 is a diagram showing the work of constructing a pronunciation target estimation model; FIG. 4 is a diagram for explaining the time required to construct a pronunciation target estimation model; FIG. 10 is a diagram showing an operation of a sound target selection unit selecting a sound wave generator suitable for a multi-target situation using a sound target estimation model; FIG. 4 is a diagram showing the time required for the learning phase and the inference phase in comparison with the reference example in relation to the present embodiment; 4A and 4B are graphs showing the relationship between the accuracy and the number of times of learning for each of the present embodiment and the reference example; The figure which shows the outline of the target detection system which concerns on 2nd Embodiment of this invention. The figure which shows the outline of the target detection system which concerns on 3rd Embodiment of this invention.

Next, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is an explanatory diagram showing an outline of a target detection system 100 according to the first embodiment of the present invention.

A target detection system 100 shown in FIG. 1 is used to identify the existence, position, etc., of an underwater object 1 to be detected, such as an underwater vehicle, a surface vessel, a sunken ship, and seafloor topography. The underwater object 1 can also be called a target or a specific target. The target detection system 100 detects underwater areas that are not visible to the naked eye, for example, by sound waves. This target detection system 100 mainly includes a plurality of sound wave detectors 2, a plurality of sound wave generators (sensors) 3, and a detection control device 4, as shown in FIG.

The sound detector 2 is composed of, for example, a known passive sonobuoy. The sound wave detector 2 receives the reflected wave from the underwater object 1 of the sound wave emitted from the sound wave generator 3 . The sound wave detector 2 performs signal processing on the received reflected waves to generate detection information to be transmitted to the detection control device 4 . The sound wave detector 2 wirelessly transmits the generated detection information to the detection control device 4 via, for example, a communication antenna floating on the water surface.

This detection information includes, for example, the reception time of the reflected wave, the direction of the received reflected wave, the frequency of the received reflected wave, the frequency shift, and the like. The sound wave detector 2 transmits the generated detection information to the detection control device 4 and the like together with identification information that can identify the sound wave detector 2 itself. The identification information of the sound wave detector 2 can be, for example, an identification number.

The sound wave detector 2 may transmit the received reflected waves to a separately provided signal processing device (not shown) instead of the detection information. In this case, the signal processing device performs signal processing on the reflected waves received from the sound detector 2 . This signal processing device may be provided in the detection control device 4 or may be provided separately from the detection control device 4 .

The sound wave generator 3 is composed of, for example, a known active sonobuoy, and generates sound waves of preset specifications such as frequency, pulse waveform, and pulse width (hereinafter simply referred to as “specifications”). The sound wave generator 3 receives a sound generation control instruction from the detection control device 4 via, for example, a communication antenna floating on the water surface. The sound wave generator 3 generates sound waves according to the received sound generation control instructions.

The sound wave detector 2 and the sound wave generator 3 are, for example, sonobuoys, and are laid from a target detector 10 such as an aircraft or ship to a predetermined detection area (target area). When laying the sonobuoy, identification information for uniquely specifying the sonobuoy (sound wave detector 2 and sound wave generator 3) and the laying position are stored in association with each other.

The detection control device 4 is mounted on the target detector 10 . The detection control device 4 selects the sound wave generators 3 for generating sound waves from a plurality of sound wave generators 3 so as to be suitable for the ever-changing detection situation. This detection situation includes at least one of information such as the position, course, speed, etc. of the underwater object 1 in the detection area. There may be one or more underwater objects 1 that exist (or may exist) in the detection area. The detection control device 4 analyzes the detection information received from the sonar detector 2 .

The detection control device (motion sensor selection device) 4 includes, as shown in FIG. Prepare.

The detection control device 4 is configured as a known computer, and includes a CPU, ROM, RAM, HDD, and the like. The HDD stores programs for selecting the sound wave generator 3 to be sounded, analyzing detection information, and the like. This program implements the pronunciation target estimation method (action sensor selection method) of the present invention. By cooperation of the above hardware and software, the detection control device 4 can be operated as the pronunciation target selection unit 5, the radio transmission/reception unit 6, the detection information analysis unit 7, and the input unit 8.

The sound target selection unit 5 is used to select at least one sound wave generator 3 suitable for the detection situation from a plurality of sound wave generators 3 . For this selection, a pronunciation target estimation model (action sensor selection model) 9, which is a pre-constructed trained model, is used. Note that the sound target selection unit 5 may not only select an appropriate sound wave generator 3 but also determine the specifications of the selected sound wave generator 3 . Details of the selection of the sound wave generator 3 will be described later.

The wireless transmission/reception unit 6 uses known wireless communication technology to transmit a sound generation control signal to the sound wave generator 3 selected by the sound generation target selection unit 5 . The sound generation control signal is a control signal for operating the sound wave generator 3 so as to generate sound waves of predetermined specifications. Furthermore, the radio transmitting/receiving section 6 receives detection information obtained by processing the reflected wave by the sound wave detector 2 (or the reflected wave received by the sound wave detector 2 itself).

The detection information analysis unit 7 analyzes the detection information received from the sound wave detector 2 to generate detection target information, which is information about the underwater object 1 and the like, based on the reflected waves received by the sound wave detector 2 . This detection target information includes, for example, the position (coordinates), course, and speed of the detected underwater object 1 . When two or more underwater objects 1 are detected at the same time, the detection target information includes information such as the position of each underwater object 1 .

The input unit 8 receives input of detection target information output by the detection information analysis unit 7 (first step). Therefore, detection target information is a kind of input information. This detection target information is information about one or a plurality of underwater objects 1 that includes state quantities relating to at least one of the position, course, and speed of each underwater object 1 .

Next, the details of how the sound target selection unit 5 uses the sound target estimation model 9 to select the sound wave generator 3 suitable for the detection situation from the plurality of sound wave generators 3 will be described with reference to FIGS. 2 to 4 and the like. do. This detection situation assumes that at least one underwater object 1 exists. The state of each underwater object 1 can be represented by a state quantity (for example, position, course, speed, etc.). If one or more underwater objects 1 have different positions, courses, speeds, etc., they are treated as different detection situations. In the following description, a detection situation in which only one underwater object 1 exists will be referred to as a "single target situation", and a detection situation in which a plurality of underwater objects 1 exist will be referred to as a "multiple target situation". Also, the state quantity of the underwater object 1 may be referred to as a "target state quantity". Among the detection conditions, the data indicating the single target condition includes the target state quantity data for one underwater object 1 . The data indicating the multiple target states includes target state quantity data for each of the multiple underwater objects 1 .

In the examples of FIGS. 1 to 4, 11 sound wave generators 3 are laid in the detection area. In order to specify each of the 11 sound wave generators 3 in the following description, each sound wave generator is given an underlined number in FIG. In the following description, the sound wave generators 3 are designated as the first sound wave generator 3-1, the second sound wave generator 3-2, . . . may be referred to as the 11th sound wave generator 3-11.

The pronunciation target estimation model 9 of this embodiment is constructed by machine learning (for example, known reinforcement learning). The pronunciation target estimation model 9 is constructed, for example, by learning each individual target situation before being implemented in the pronunciation target selection unit 5 of the detection control device 4 .

In the learning data set (base information) used for this learning, for example, one underwater Considering the case where only the object 1 appears, information indicating target state quantities such as the position, course, speed, etc. of this underwater object 1 in a certain time span is included. That is, a set of learning data corresponds to a set of single target situations (1), (2), . . .

The pronunciation target estimation model 9 calculates each Q value (applicability) when selecting each sound wave generator 3 and its specifications for one single target situation. The Q value means a well-known state action value in Q-learning, which is a type of reinforcement learning.

Briefly, in reinforcement learning, an environment is given to an agent (action subject). The agent optimizes its action by repeating trials of various actions a with respect to the environment. In the environment, an action a and a change in state s corresponding to the action a are defined. When an agent reaches a certain state s as a result of performing an action a, the agent is rewarded by the environment. The value when the agent takes a certain action a in a certain state s is the Q value, which can be expressed as Q(s, a). The Q value is calculated so as to take into consideration not only the immediate reward obtained by the current action a, but also the reward that may be obtained by the next and subsequent actions a. In this embodiment, the agent, in principle, selects the action that maximizes the Q value. Since it can be said that the larger the Q value, the more suitable it is to perform the action a for the state s, the Q value is a kind of fitness. Learning of the pronunciation target estimation model 9 means learning of behavioral value evaluation (specifically, Q value).

In the case of learning the pronunciation target estimation model 9 of the present embodiment, the state s corresponds to a single target situation, and the action a corresponds to selection of the sound wave generator 3 to produce a sound. In the learning phase of the pronunciation target estimation model 9, as a result of the agent uttering any sound wave generator 3 in a single target situation, if the underwater object 1 can be successfully detected by simulation, the agent is rewarded. is given. As shown in FIG. 3, the pronunciation target estimation model 9 is applied to one single target situation multiple times through trial and error ( L times shown in the graph) to learn. However, other conditions can also be used as conditions for ending learning.

The Q value can be calculated using a neural network known as a method of approximating the Q function. A neural network has an input layer, an intermediate layer, and an output layer. Each layer is made up of multiple units that mimic the neurons of the brain. Units are connected between adjacent layers, and a weight indicating ease of information transmission is defined for each connection. In this embodiment, a neural network model having two or more intermediate layers is employed as the pronunciation target estimation model 9 . When the input layer of the neural network is fed with state s, the output layer outputs the respective Q value of action a. In the learning phase, the agent actually tries the action a in the environment based on the output result when the state s is input to the neural network, so that the error in the output of the neural network can be obtained. The weights described above are updated by applying a known error backpropagation method using the obtained errors. Learning can be performed by repeating the above process.

Now consider the time required to complete the learning phase. As shown in FIG. 3, the number of learning data (in other words, the number of single target situations to be learned) when focusing on one underwater object 1 is P, and the single target situations are input to the pronunciation target estimation model 9. Let T be the calculation time required to output the selection result of the sound wave generator 3, and let L be the number of necessary times of learning for one single target situation. In this case, the learning time for learning the pronunciation target estimation model 9 to construct the pronunciation target estimation model 9 is the product (PLT) of the number of single target situations P, the calculation time T, and the number of learning times L. . In this embodiment, since the number P of learning data can be kept small as described above, the pronunciation target estimation model 9 used by the pronunciation target selection unit 5 can be constructed in a short learning time.

In the inference phase of the pronunciation target estimation model 9, when a single target situation is given, the sound wave generator 3 with the largest Q value among the plurality of sound wave generators 3 is selected as the sound wave generator 3 to be sounded. Output.

Here, as the detection situation given to the sound target selection unit 5 via the input unit 8, there are cases where only one underwater object 1 appears, and cases where a plurality of underwater objects 1 appear.

First, consider the case where the detection situation input to the pronunciation target selection unit 5 is the single target situation. In this case, the pronunciation target selection unit 5 inputs the single target situation as it is to the pronunciation target estimation model 9, and outputs the selection result output by the pronunciation target estimation model 9 as it is as the sound wave generator 3 to be sounded.

Next, let us consider a case where the detection situation input to the sound target selection unit 5 is a multi-target situation, specifically, a case where a plurality of underwater objects 1 to be detected appear. In this case, as shown in FIG. 4, the pronunciation target selection unit 5 decomposes the input multiple target situations into single target situations corresponding to the number of underwater objects 1 . When a plurality of single target situations are obtained, the pronunciation target selection unit 5 inputs each of the single target situations to the pronunciation target estimation model 9 (second step). In the example of FIG. 4, a multiple target situation in which five underwater objects 1 appear is decomposed into five single target situations. The single target situation (1), the single target situation (2), .

For each of the input single target situations (1) to (5), the pronunciation target estimation model 9 selects, in the same manner as described above, the sound wave generation with the largest Q value among the candidates for the plurality of sound wave generators 3. Output body 3 as an object to be pronounced. In the example of FIG. 4, in single target situation (1) and single target situation (2), the fifth sound wave generator 3-5 is output as the object to be sounded. Similarly, for the single target situations (3) to (5), the target sound wave generator 3 to be sounded is output from the sound target estimation model 9. FIG. Hereinafter, the sound wave generator 3 output from the pronunciation target estimation model 9 as being suitable for operating with respect to each of the single target situations (1) to (5) may be referred to as a single motion compatible target.

The pronunciation target selection unit 5 integrates a plurality of selection results output by the pronunciation target estimation model 9, and outputs the sound wave generator 3 that is comprehensively most suitable for pronunciation in a multi-target situation (third process). Hereinafter, the sound wave generator 3 output from the sound generation target selection unit 5 as being suitable for multiple target situations may be referred to as a motion matching target (motion sensor). For example, the number of times the sound wave generator 3 is output from the sound target estimation model 9 as the target to be sounded is counted for each sound wave generator 3. 3 as the object to be pronounced in the input multi-goal situation. FIG. 4 shows an example in which the fifth sound wave generator 3-5 is output as the object to be pronounced in the multi-target situation input to the pronunciation object selection unit 5 by the above-described majority method. there is

However, the method of summarizing the selection results output by the pronunciation target estimation model 9 for a plurality of single target situations is not limited to the above. For example, the pronunciation target estimation model 9 outputs to the pronunciation target selection unit 5 the Q value when each sound wave generator 3 is selected for each single target situation obtained by decomposing a plurality of target situations. can be considered. In this case, the sound generation target selection unit 5 calculates the sum of the Q values for each sound wave generator 3, and selects the sound wave generator 3 that maximizes this sum (sum of suitability) in the input multi-target situation. Select as a target.

Specifically, in the example of FIG. 4, the pronunciation target estimation model 9 causes the first to eleventh sound wave generators 3-1 to 3-11 to produce sounds for the single target situation (1). The Q value for each case is calculated and output to the pronunciation target selection unit 5 . As the Q value, the output value of the neural network described above may be used as it is. The output of the Q value in the pronunciation target estimation model 9 is similarly performed for each of the remaining single target situations (2) to (5). The sound target selection unit 5 totalizes the obtained Q values for each of the 11 sound wave generators 3-1 to 3-11. In this way, 11 sum totals of Q values are obtained. When selecting targets to be pronounced in a multi-target situation, the pronunciation target selection unit 5 selects, from among the 11 sound wave generators 3, the sound wave generators 3 having the maximum sum of Q values. It is selected as an object to be pronounced in the target situation.

The detection control device 4 transmits a sound generation control signal to the sound wave generator 3 selected by the sound generation target selection unit 5 as described above, via the wireless transmission/reception unit 6 . The sound wave generator 3 selected by the sound generation target selection unit 5 generates sound waves according to the sound generation control signal received from the sound generation target selection unit 5 via the communication antenna. The sound wave detector 2 receives the reflected sound of the sound waves emitted from the sound wave generator 3 from the underwater object 1 and the like, and transmits the detection information obtained by signal processing to the detection information analysis unit 7 via the communication antenna. Send. The detection information analysis unit 7 analyzes the detection information and updates the target candidate list. The detection status reflecting the latest target candidate list is input to the pronunciation target selection unit 5 .

In this way, the sound target selection unit 5 selects the sound wave to be sounded according to the detection situation that changes at any time due to the movement (movement, appearance/disappearance, etc.) of the underwater object 1, the sound wave detector 2, and the sound wave generator 3. The generator 3 can be selected at any time.

In the target detection system 100 of the present embodiment, the sound target selection unit 5 can also select the sound wave generators 3 that are second and subsequent ones, instead of the sound wave generators 3 that are most suitable for the detection situation, depending on the situation. .

For example, it is assumed that the sound wave generator 3, which is judged to be most suitable for sound generation by the sound target selection unit 5, has sounded consecutively more than a predetermined number of times most recently in response to the input single target situation. In this case, the sound target selection unit 5 outputs the second most suitable sound wave generator 3 as a target to be sounded in the input single target situation.

Similarly, when multiple target situations are input, the sound target selection unit 5 excludes the sound wave generators 3 that have recently made sounds continuously for a predetermined number of times or more, and selects the sound waves to be produced in the multiple target situations. Select generator 3.

In this way, if there is a sound wave generator 3 that satisfies the condition (operation exclusion condition) that it has been sounding continuously for a predetermined number of times or more most recently, then the selected sound wave generator 3 will may be different from the sound wave generator 3 selected the most times in , or the sound wave generator 3 with the largest sum of Q values.

By making such a selection, it is possible to prevent one sound wave generator 3 from continuously generating sound waves many times. As a result, the effect of, for example, heat generated by the operation of the sound wave generator 3 can be reduced, and the function of the sound wave generator 3 can be favorably maintained.

Advantages of the pronunciation target selection unit 5 of the present embodiment will now be described in detail with reference to FIGS. 5 and 6 from the viewpoint of time required for processing.

FIG. 5 shows this embodiment in which only a single target situation is learned and a multi-target situation in addition to a single target situation in order to realize a target detection system that can handle cases where a maximum of two underwater objects 1 appear. A comparison with a reference example of learning is shown.

First, consider the learning phase. If the number of learning data when only one underwater object 1 appears is P, the number of learning data when two underwater objects 1 appear simultaneously is P squared. In this embodiment, only P is the number of learning data because only single target situations are learned. On the other hand, in the reference example, since the multi-goal situation is learned in addition to the single goal situation, the number of learning data is P+ ^P2 .

Therefore, in the pronunciation target estimation model 9, if the time required to output the selection result after the input of a single target situation or multiple target situations is T, and the number of learning times required to converge the output is L, In the reference example, the learning time required to construct the trained model is (P+P ² )LT. Thus, in the reference example, the learning time is longer than the learning time PLT for constructing the pronunciation target estimation model 9 used by the pronunciation target selection unit 5 of the present invention.

This time, the explanation is given with a maximum of two cases, but when considering a maximum of three, four, . Generally speaking, when the target maximum number of underwater objects 1 is N, the learning time required in the present embodiment and the reference example follows the following formula (1).

Next, consider the inference phase. In this embodiment, when two underwater objects 1 appear, two single target situations obtained by decomposing a plurality of target situations are input to the sound target estimation model 9 to obtain the selection result of the sound wave generator 3. There is a need. Therefore, simply thinking, it takes at most twice as much time as the reference example. On the other hand, in the reference example, even when two underwater objects 1 appear, it is possible to obtain the selection result of the sound wave generator 3 suitable for producing sound from a comprehensive point of view in the same amount of time as in the case of one underwater object.

ただし、学習済モデルの複雑さ及び探知制御装置４の処理能力にもよるが、推論に必要な時間は学習時間と比較すると大幅に短く、かつ並列計算も可能であるため、本実施形態が大きく不利になることはないと考えられる。 Assuming that the maximum number of target underwater objects 1 is N, the inference time required in this embodiment and the reference example follows the following equation (2).

However, although it depends on the complexity of the learned model and the processing power of the detection control device 4, the time required for inference is significantly shorter than the learning time, and parallel computation is also possible. It is not considered to be disadvantageous.

The table at the bottom of FIG. 5 shows the results of an experiment conducted by the inventor regarding the learning time, inference time, and accuracy of the reference example and the present embodiment when N=2, P=36, and L=1000. Are listed. In this experiment, in the configuration of the present embodiment, the sum of the Q values for each sound wave generator 3 was used as a criterion for selecting the sound wave generator 3 to be sounded in a multi-target situation.

As shown in this table, in the present embodiment, the learning time can be greatly reduced while keeping the increase in the time required for inference within a range (0.1 to 0.2 seconds) that does not pose a problem. Therefore, system operation can be started early. In addition, in this embodiment, the selection accuracy of this embodiment is equal to or higher than that of the reference example, and in terms of the selection accuracy of the sound wave generator 3 to be sounded, this embodiment has performance equal to or higher than that of the reference example. It was confirmed that

FIG. 6 shows transitions of the selection accuracy according to the number of times of learning for each of the reference example and the present embodiment. With a sufficient number of times of learning, the accuracy converged to about 75% in the reference example, and the accuracy converged to about 82% in the present embodiment. According to the graph of FIG. 6, it can be seen that the present embodiment achieves good selection accuracy with a smaller number of times of learning than the reference example.

As described above, in this embodiment, a plurality of sound wave generators 3 used for detecting one or more underwater objects 1 existing in the detection area are arranged in the detection area. The detection control device 4 selects the sound wave generator 3 to be sounded from among the plurality of sound wave generators 3 as an operation adaptation target. The detection control device 4 includes an input unit 8 , a pronunciation target estimation model 9 , and a pronunciation target selection unit 5 . The input unit 8 receives an input of detection target information, which is information including a target state quantity relating to at least one of the position, course and speed of each underwater object 1 for at least one underwater object 1 . The pronunciation target estimation model 9 selects at least one sound wave generator 3 to be selected from a plurality of sound wave generators 3 from a learning data set containing a target state quantity for one underwater object 1 in a certain time span. We performed machine learning to The sound target selection unit 5 uses the sound target estimation model 9 to select the sound wave generator 3 to operate from the detection target information input to the input unit 8 . The pronunciation target selection unit 5 decomposes the detection target information regarding the plurality of underwater objects 1 input to the input unit 8 into information regarding one underwater object 1, and applies each of the decomposed information to the pronunciation target estimation model 9. A plurality of selection results input and output from the pronunciation target estimation model 9 are integrated, and the sound wave generator 3 to be operated is selected as an operation adaptation target.

As a result, when constructing the pronunciation target estimation model 9 in the learning phase, the learning data can be limited to the situation (single target situation) regarding one underwater object 1 . As a result, an explosive increase in the number of learning data is avoided, the learning time of the pronunciation target estimation model 9 is remarkably shortened, and early operation of the pronunciation target selection unit 5 (and thus the detection control device 4) is attempted. be able to. In the inference phase, multiple target situations regarding a plurality of underwater objects 1 are decomposed into single target situations (for each underwater object 1) and input to the pronunciation target estimation model 9, and the output result of the pronunciation target estimation model 9 is used. By integrating, even in a complicated situation where a plurality of underwater objects 1 appear, the sound wave generator 3 to be sounded can be appropriately selected.

Further, in the present embodiment, the sound target selection unit 5 further determines the specifications of the selected sound wave generator 3 .

Accordingly, the specifications of the sound wave generated by the sound wave generator 3 (for example, frequency, pulse waveform, pulse width, etc.) can be appropriately determined according to the situation regarding the underwater object 1 .

In addition, in the present embodiment, the detection control device 4 responds to the detection target information regarding the plurality of underwater objects 1 input to the input unit 8 (detection target information regarding multiple target situations), and the pronunciation target estimation model 9 outputs Among the plurality of selection results, the sound wave generator 3 that has been selected the most times is output as an operation matching target.

As a result, the sound wave generator 3 suitable for multiple target situations can be selected by simple processing.

However, this embodiment can also be modified as follows. That is, the output of the sound target estimation model 9 includes each Q value when each of the plurality of sound wave generators 3 is selected with respect to the target state quantity for one underwater object 1 . The pronunciation target selection unit 5 decomposes the detection target information regarding the plurality of underwater objects 1 input to the input unit 8 (detection target information regarding multiple target situations) into information regarding one underwater object 1, is input to the pronunciation object estimation model 9, and the sum of the Q values of each sound wave generator 3 included in the output of the pronunciation object estimation model 9 is obtained for each sound wave generator 3, and the sum of the Q values The sound wave generator 3 with the maximum is output as an operation-adaptive target.

As a result, it is possible to accurately select the sound wave generator 3 suitable for multiple target situations.

Further, when some of the plurality of sound wave generators 3 satisfy a preset operation exclusion condition, the detection control device 4 of the present embodiment excludes the sound wave generators 3 and selects an operation compatible target. do.

As a result, it is possible to adjust the selection result of the pronunciation target selection unit 5 according to the actual operational situation.

Further, the target detection system 100 of this embodiment uses the detection control device 4 and a plurality of sound wave generators 3, and further uses the sound wave detector 2 to detect a detection area in which the plurality of sound wave generators 3 are arranged. Detect underwater object 1 inside.

As a result, the underwater object 1 can be accurately detected according to the conditions within the detection area.

Although the preferred embodiments of the present invention have been described above, the above configuration can be modified, for example, as follows.

The sound target selection unit 5 selects the sound wave generator 3 suitable for detecting the target from among the plurality of sound wave generator 3 candidates. etc. may be determined.

The pronunciation target estimation model 9 is not limited to a neural network, and can employ decision rules, for example.

The learning method of the pronunciation target estimation model 9 is not limited to reinforcement learning, and genetic machine learning, for example, can be employed.

For example, the same pronunciation target estimation model 9 may operate in parallel in the inference phase, for example, by configuring the pronunciation target selection unit 5 from a plurality of computers (hardware). In this case, the reasoning time in multi-goal situations can be shortened overall.

The single target state may include, for example, tidal current information, wind information, water temperature information, etc., in addition to the target state quantity.

When the sound wave generators 3 suitable for multiple target situations are selected by a majority method, and there are a plurality of sound wave generators 3 selected by the pronunciation target estimation model 9 the maximum number of times, for example, Q A larger value may be selected.

A value other than the Q value may be used as the degree of suitability (degree of suitability) for sounding the sound wave generator 3 to the single target situation.

The action exclusion condition is not limited to the fact that the sound has been pronounced continuously for a predetermined number of times or more in the recent past, and other conditions may be used.

A trained model such as the pronunciation target estimation model 9 can also be applied to the operation of other devices besides the sound wave generator 3 .

The sound target selection unit 5 may use the sound target estimation model 9 to select two or more sound wave generators 3 to operate from the detection target information input to the input unit 8 .

All or part of the detection control device 4 may be deployed on a ship or ground equipment other than the target detector 10 such as an aircraft.

The target detection system 100 can be applied not only to detection of the underwater object 1 but also to detection of other objects.

Next, another embodiment using the target detection system 100 of the present invention will be briefly described with reference to FIGS. 7 and 8 and the like. In the following description, descriptions that are the same as or similar to those of the first embodiment may be omitted.

A target detection system 100 of the second embodiment shown in FIG. 7 detects an airborne object 1a such as an aerial vehicle, a living creature, or a cumulonimbus cloud, using, for example, a radar 3a as a sensor. A plurality of radars 3a are installed. Each radar 3a is arranged on the ground.

The detection control device 4 of the present embodiment is mounted, for example, on ground equipment (not shown). However, it is not limited to this, and a part or the whole of the detection control device 4 may be mounted on an aerial vehicle or the like. The detection control device 4 selects from a plurality of radars 3a a radar 3a for generating radio waves, which are detection signals, so as to be suitable for ever-changing detection conditions. Furthermore, the detection control device 4 determines the specifications of the selected radar 3a (for example, waveform, irradiation time, irradiation direction, frequency, etc.).

The detection situation assumes the presence of at least one airborne object 1a. In the following description, a detection situation in which only one airborne object 1a exists will be referred to as a "single target situation", and a detection situation in which a plurality of airborne objects 1a exist will be referred to as a "multiple target situation". The state of each airborne object 1a can be represented by a state quantity (for example, position, course, speed, etc.).

The detection control device 4 of the present embodiment includes an unillustrated motion radar estimation model (motion sensor selection model) instead of the sound object estimation model 9 . In the learning data set used for learning by the motion radar estimation model, for example, only one airborne object 1a appears in a detection area in which four (but not limited to) radars 3a are installed as shown in FIG. Considering the case of the airborne object 1a, it includes the target state quantities such as the position, course, speed, etc. of the airborne object 1a in a certain time span. In this way, the detection situation learned by the motion radar estimation model is only the single target situation.

The detection control device 4 of this embodiment includes an unillustrated movement radar selection section (motion sensor selection section) instead of the sound target selection section 5 . Instead of the sound wave generator 3, the operating radar selection unit selects the radar 3a to which radio waves are to be emitted from a plurality of radars 3a installed within the detection area. At this time, if the detection target information input to the input unit 8 relates to multiple target situations, the operating radar selection unit selects the detection target information as information on one airborne object 1a (corresponding to a single target situation). information) and inputs it to a motion radar estimation model, integrates a plurality of selection results output from the motion radar estimation model, and selects the radar 3a to be operated as a motion adaptation target.

As a result, the radar 3a suitable for the movement (movement, appearance/disappearance) of the airborne object 1a and its operation (irradiation direction, irradiation time, radio wave waveform, frequency, etc.) can be determined. The airborne object 1a can be searched and tracked more accurately. In addition, since the learning time of the motion radar estimation model is short, the operation of the system can be started early.

A target detection system 100 according to the third embodiment shown in FIG. 8 is configured as a monitoring system that monitors a monitoring target 1b such as a person (for example, a suspicious person) or a vehicle using a camera 3b as a sensor, for example.

The detection control device 4 of this embodiment is provided, for example, in a monitoring room as shown in FIG. However, it is not limited to this, and a part or the whole of the detection control device 4 may be mounted on a vehicle or the like. The detection control device 4 includes a monitoring device 41 operated by a monitoring staff 40, as shown in FIG. The monitoring device 41 can select one or more of the images captured by the cameras 3 b and display them on the display 42 . Normally, all of the cameras 3b are constantly shooting for monitoring purposes. However, when the image captured by the camera 3b is displayed on the display 42, the camera 3b is substantially operating in relation to the monitoring device 41, and otherwise the camera 3b is substantially stopped. It can be said that

The detection control device 4 displays an image on the display 42 so as to suit the detection situation regarding the object to be monitored 1b such as a suspicious person instructed by the surveillance staff 40 (or determined based on the image captured by the camera 3b). select the camera 3b to be used. Furthermore, the detection control device 4 may determine the specifications of the selected camera 3b (for example, the orientation of the camera 3b, zooming, etc.).

The detection situation is based on the assumption that at least one monitored object 1b exists. In the following description, a detection situation in which there is only one monitoring target 1b is called a "single target situation", and a detection situation in which a plurality of monitoring targets 1b exist is sometimes called a "multiple target situation". The state of each monitored object 1b can be represented by state quantities (for example, position, course, speed, etc.).

The detection control device 4 of this embodiment includes an unillustrated image selection estimation model (motion sensor selection model) in place of the pronunciation target estimation model 9 . The learning data set used for learning by the video selection estimation model is only the single target situation. In the learning phase of the image selection estimation model, for example, as a result of the agent displaying the image of any camera 3b on the display 42 in the single target situation, the face of the suspicious person who is the surveillance object 1b can be captured on the display 42. If successful, the agent is rewarded.

The detection control device 4 of this embodiment includes an unillustrated image selection unit (motion sensor selection unit) instead of the sound target selection unit 5 . Instead of the sound wave generator 3, the image selection unit selects the camera 3b whose photographed image is to be displayed on the display 42 from a plurality of cameras 3b installed within the detection area. At this time, if the detection target information input to the input unit 8 is related to the situation of multiple targets, the video selection unit selects the detection target information as information related to one monitoring target 1b (information corresponding to a single target situation). ) and input it to the image selection estimation model, and integrate a plurality of selection results output from the image selection estimation model, and select the camera 3b that should be the image source as the motion matching target.

As a result, the surveillance staff 40 can easily and accurately search and track the surveillance target 1b such as a suspicious person using the target detection system 100 of the present embodiment. In addition, since the learning time of the video selection estimation model is short, system operation can be started early.

The surveillance staff 40 can refer to the contents displayed on the display 42 to operate the crime prevention gate 3c or give instructions to the guards near the surveillance object 1b. The operation of the security gate 3c can also be automated by machine learning.

1 underwater object (target)
2 sound detector 3 sound generator (sensor)
4 Detection control device (movement sensor selection device)
5 Sound target selection section (motion sensor selection section)
6 radio transceiver unit 7 detection information analysis unit 8 input unit 9 pronunciation target estimation model (motion sensor selection model)
10 target detector 100 target detection system

Claims (7)

A motion sensor selection device, wherein a plurality of sensors used for detecting one or more targets existing in a target region are arranged in the target region, and the sensor to be operated from among the plurality of sensors is selected as a motion sensor. ,
an input unit for receiving an input of input information, which is information including a target state quantity relating to at least one of position, course and speed relating to each target, with respect to at least one target;
a motion sensor selection model that performs machine learning for selecting at least one sensor to be selected from among the plurality of sensors from base information including the target state quantity in a certain time span for one target; and ,
a motion sensor selection unit that selects the sensor to operate from the input information input to the input unit using the motion sensor selection model;
with
The motion sensor selection unit decomposes the input information regarding the plurality of targets input to the input unit into information regarding one target, and inputs each decomposed information to the motion sensor selection model. 1. A motion sensor selection device, which integrates a plurality of selection results output from said motion sensor selection model and selects said sensor to be operated as a motion sensor. The motion sensor selection device of claim 1, comprising:
The motion sensor selection device, wherein the motion sensor selection unit further determines specifications of the selected motion sensor. 3. A motion sensor selection device according to claim 1 or 2,
The motion sensor selection unit selects the target with the largest number of times from among the plurality of selection results output by the motion sensor selection model for the input information about the plurality of targets input to the input unit. A motion sensor selection device characterized by outputting a sensor as the motion sensor. 3. A motion sensor selection device according to claim 1 or 2,
The output of the motion sensor selection model includes each degree of fitness when each of the plurality of sensors is selected with respect to the target state quantity for one of the targets,
The motion sensor selection unit decomposes the input information regarding the plurality of targets input to the input unit into information regarding one target, and inputs each decomposed information to the motion sensor selection model. and determining, for each of the sensors, the sum of the degrees of goodness of fit obtained by adding the degrees of goodness of each of the sensors included in the output of the motion sensor selection model, and selecting the sensor having the maximum sum of goodness of fit as the motion sensor. A motion sensor selection device characterized by outputting. A motion sensor selection device according to any one of claims 1 to 4,
A motion sensor selection device, wherein when some of the plurality of sensors satisfy a preset motion exclusion condition, the motion sensor is selected by excluding the sensor. A specific target is detected in a detection area in which a plurality of sensors are arranged using the motion sensor selection device according to any one of claims 1 to 5 and a plurality of sensors. target detection system. A motion sensor selection method in which a plurality of sensors used for detecting one or more targets existing in a target region are arranged in the target region, and one of the plurality of sensors to be operated is selected as a motion sensor. ,
a first step of receiving an input of input information regarding at least one of the targets, which is information including a target state quantity regarding at least one of position, course and speed regarding each of the targets;
The input information regarding the plurality of targets input in the first step is decomposed into information regarding one target, and each decomposed information is input to a motion sensor selection model, and the motion sensor selection model is , a trained model that has undergone machine learning for selecting at least one sensor to be selected from among the plurality of sensors from base information including the target state quantity in a certain time span for one target a second step;
a third step of integrating a plurality of selection results output from the motion sensor selection model and selecting the sensor to operate;
A motion sensor selection method, comprising:

Images