JP5075871B2 - Discriminator - Google Patents

Description

  The present invention relates to a discrimination apparatus that can add a discrimination target.

  In recent years, neural networks that learn using an error back-propagation learning method have been used in various fields such as image recognition, speech recognition, and pattern recognition. Such a neural network can discriminate a discrimination target to which the input data belongs from a plurality of discrimination targets registered in advance based on the input data.

  An underwater detection device is known as one using this neural network (Patent Document 1). Underwater detectors use ultrasound to acquire sediment information about bottom sediments such as mud and rocks in the bottom of the sea, causing habitats for bottom fish, shrimps, crabs, and fishnets. It is used to know rocky places. This bottom sediment information is input to the neural network and input data generated based on the echo signal of the ultrasonic transmission pulse emitted from the transducer of the underwater detector to the bottom of the water, and is output from the neural network. Can be obtained by processing the output data. The obtained output data is notified to the user via a display monitor or the like.

  By the way, there are cases where it is desired to add (register) a new discrimination target in the neural network. For example, in the underwater detection device, there is a case where it is desired to newly add a bottom sediment (user bottom sediment) to be discriminated. In such a case, a method of increasing the number of neurons in the neural network is known as the simplest method.

JP 2008-275351 A

  However, in the method of increasing the number of neurons, it is necessary to change the structure of the neural network when registering a new discrimination target (adding a discrimination target), and the learning time explodes as the number of neurons increases. There is a problem that it increases. In addition, as shown in FIG. 20, in the neural network that outputs the similarity of each registered discrimination target to the input data, the same between the discrimination targets registered in advance before and after changing the structure of the neural network. The ratio of the output value to the input data may change. Furthermore, there is a problem in that it is necessary to change the structure of the neural network again when trying to deregister the registered discrimination target.

  In particular, in the underwater detection device, the bottom sediment that the user wants to identify (user bottom sediment) is often different depending on the water area, season, weather, etc., so the registration of the user bottom sediment (registration of new discrimination target) and the cancellation of the registration Can happen frequently. It is necessary to change the structure of the neural network each time the user bottom sediment is registered and deregistered, and the learning time is enormous. Furthermore, since the ratio of the output value to the same input data between the registered bottom sediments changes, the user may have an illusion that the bottom sediment has changed despite the same input data. There is a problem that the user cannot be accurately grasped.

  The present invention solves the above problems, and the problem is that registration and registration of a new discrimination target can be performed with less learning time, and registration and registration of a new discrimination target are possible. An object of the present invention is to provide a discriminating apparatus in which the ratio of output values between discriminating objects registered in advance is not changed before and after the release of. Furthermore, it is providing the underwater detection apparatus which has this discrimination | determination apparatus.

  The present invention receives an ultrasonic echo signal transmitted into water, inputs input data based on the echo signal to a neural network, and registers a plurality of registered data based on output data output from the neural network. An underwater detection device that generates, every predetermined time, bottom sediment information related to the bottom sediment to which the input data belongs, among a plurality of pre-registered bottom sediments to which the input data belongs In order to determine whether or not the input data belongs to the new bottom sediment when a new bottom sediment is registered by a fixed neural network that outputs the first output data for discriminating and an external operation When a variable neural network that outputs the second output data is added and the registered new bottom sediment is canceled by an external operation, A neural network management unit that deletes a modified neural network; and an output processing unit that can generate the bottom sediment information in which the second output data is prioritized with respect to the first output data. It is characterized by that.

According to the above configuration, the first output data for determining the bottom sediment to which the input data belongs out of the plurality of bottom sediments registered in advance is output from the fixed neural network. When a new bottom sediment is registered by an external operation, a variable neural network is added, and second output data for determining whether or not the input data belongs to the new bottom sediment is the variable neural network. Output from the network. Moreover, it is possible to generate sediment information in which the second output data is prioritized with respect to the first output data. Furthermore, the variable neural network is deleted when the registration of the registered new bottom sediment is released by an external operation.
Thus, by adding and deleting variable neural networks, it is possible to register and deregister new sediments, so there is no need to change the structure of the fixed neural network. The variable neural network is a neural network that determines whether or not it belongs to a new bottom sediment. Since the number of neurons is small, the learning time can be suppressed. Furthermore, it is possible to generate bottom sediment information in which the second output data is given priority over the first output data, which allows the user to recognize whether or not the input data belongs to the new bottom sediment. is there.
Note that the underwater detection device according to the present invention may be any device having a function of determining bottom sediment, and examples thereof include a fish finder, a sounding device, and a sonar.

  Further, in the underwater detection device of the present invention, the variable neural network is generated based on the received echo signal, and is user-registered learning data set as the new bottom sediment learning data by an external operation, and The learning may be performed by fixed learning data that is used when learning the fixed neural network and is set as learning data of sediment other than the new sediment. According to the above configuration, the user registration learning data generated based on the received echo signal and set as new bottom sediment learning data by an external operation, and used for learning of the fixed neural network, Learning is performed using fixed learning data set as learning data for bottom sediment other than bottom sediment. In this way, the learning data used when learning the variable neural network can be acquired without actually confirming with an underwater camera or the like.

  In the underwater detection device of the present invention, the user registration learning data may be configured such that learning data can be added and a part of the user registration learning data can be deleted by an external operation. According to the above configuration, the user registration learning data can be added with learning data and a part of the user registration learning data can be deleted by an external operation. By doing so, it is possible to improve the learning data of the new bottom sediment, and it is possible to accurately generate bottom sediment information about whether or not the input data belongs to the new bottom sediment.

  In the underwater detection device of the present invention, the bottom sediment information may be a value corresponding to the similarity between the input data and the new bottom sediment generated from the second output data, and / or the first sediment. The input data may be made of a value corresponding to the similarity between the input data and each of the plurality of previously registered bottom sediments. According to the above configuration, the bottom sediment information is generated from the second output data, the value corresponding to the similarity between the input data and the new bottom sediment, and / or the first output data. And a value corresponding to the similarity between each of the input data and a plurality of sediments registered in advance. By doing so, the similarity between the input data and the new bottom sediment, and the similarity between the input data and each of the plurality of sediments registered in advance become obvious. In addition, it is obvious at a glance how the similarity between the input data and the new bottom sediment changes within the water area where the bottom sediment is determined, and how the similarity between the input data and each of the multiple registered bottom sediments changes. It becomes.

  The underwater detection device of the present invention further includes a threshold setting unit for setting a threshold, and in the output processing unit, a value corresponding to the similarity between the input data and the new bottom sediment is equal to or greater than the threshold. In this case, the bottom sediment information in which the second output data is prioritized with respect to the first output data is generated, and when the value indicated by the second output data is smaller than the threshold, the first The bottom sediment information may be generated only from the output data. According to said structure, when the value equivalent to the similarity of the said input data and the said new bottom sediment is more than the threshold value set by the threshold value setting part, 2nd output data with respect to 1st output data If the bottom sediment information with priority is generated and is smaller than the threshold value, the bottom sediment information is generated only by the first output data. As described above, whether or not the second output data is used for generating bottom sediment information is determined depending on the magnitude of the value corresponding to the similarity between the input data and the new bottom sediment. It becomes obvious at a glance whether the value corresponding to the similarity is equal to or greater than a set threshold value.

  In the underwater detection device of the present invention, when the value corresponding to the similarity between the input data and the new bottom sediment is not less than the threshold value, the input data is not input to the fixed neural network. A variable neural network output determination unit may be further provided. According to the above configuration, when the value corresponding to the similarity between the input data and the new bottom sediment is equal to or greater than the threshold set by the threshold setting unit, the input data is not input to the fixed neural network. By doing so, when the value corresponding to the similarity between the input data and the new bottom sediment is equal to or greater than the threshold value, the processing in the fixed neural network can be omitted.

  In the underwater detection device of the present invention, the threshold value may be changed by an external operation. According to the above configuration, the threshold value can be changed by an external operation. In this way, the user decides whether or not the second output data should be reflected in the bottom sediment information, and determines the magnitude of the value corresponding to the similarity between the input data and the new bottom sediment according to his / her preference. You can change it accordingly.

  The underwater detection device of the present invention further includes an output unit (such as a display monitor or a speaker) that outputs the bottom sediment information, and there are a plurality of output forms of the bottom sediment information that are output to the output unit. It may be made changeable by an external operation. According to the above configuration, there are a plurality of output forms for causing the output unit to output bottom sediment information, which can be changed by an external operation. By doing so, the user can change the output form according to his / her preference.

  Further, the present invention inputs input data to a neural network, and generates discrimination information related to a discrimination target to which the input data belongs among a plurality of registered discrimination targets based on the output data output from the neural network. A discriminating device, which includes a fixed neural network that outputs first output data for discriminating a discriminating target to which the input data belongs among a plurality of discriminating targets registered in advance, and a new by an external operation When a discrimination target is registered, a variable neural network that outputs second output data for determining whether or not the input data belongs to the new discrimination target is added, and a new discrimination target is obtained by an external operation. A neural network tube that deletes the second neural network when the registration is canceled And parts, characterized in that it comprises a possible output processing unit configured to generate the said determination information and the second output data is given priority over the first output data.

According to said structure, the 1st output data for discriminating the discrimination | determination object to which input data belong among the some discrimination | determination object registered beforehand is output from a fixed neural network. When a new discrimination target is newly registered by an external operation, a variable neural network is added, and second output data for discriminating whether or not the input data belongs to the new discrimination target is the variable neural network. Output from the network. In addition, it is possible to generate discrimination information in which the second output data has priority over the first output data. Furthermore, the variable neural network is deleted when the registration of the registered new discrimination target is canceled by an external operation.
In this way, by adding and deleting variable neural networks, new discrimination targets can be registered and deregistered, so there is no need to change the structure of the fixed neural network. The variable neural network is a neural network that determines whether or not it belongs to a new determination target. Since the number of neurons is small, the learning time can be reduced. Furthermore, since discrimination information in which the second output data is prioritized over the first output data is generated, it is possible to easily determine whether or not the input data belongs to a new discrimination target.

  The present invention makes it possible to register a new discrimination target and cancel the registration without changing the structure of a neural network that discriminates a discrimination target to which input data belongs among a plurality of discrimination targets registered in advance.

It is explanatory drawing explaining the underwater detection apparatus concerning one Embodiment of this invention. It is a functional block diagram of the underwater detection apparatus concerning one Embodiment of this invention. It is an external structure of the display apparatus of the underwater detection apparatus concerning one Embodiment of this invention. It is a block diagram which shows the electric constitution of the underwater detection apparatus concerning one Embodiment of this invention. It is a figure explaining the display form of the bottom sediment information displayed on the display monitor of the underwater detection apparatus concerning one Embodiment of this invention, (a) is a 1st bottom sediment display form, (b) is a 1st ratio. It is a figure explaining a display form. It is a figure which shows the output data output from the neural network of the underwater detection apparatus concerning one Embodiment of this invention, (a) is the output data output from a fixed neural network, (b) is output from a variable neural network. It is a figure which shows output data. It is a figure explaining the display form of the bottom sediment information displayed on the display monitor of the underwater detection apparatus concerning one Embodiment of this invention, (a) is a 1st bottom sediment display form, (b) is a 1st ratio. It is a figure explaining a display form. It is a figure explaining the display form of the bottom sediment information displayed on the display monitor of the underwater detection apparatus concerning one Embodiment of this invention, (a) is a 2nd bottom sediment display form, (b) is a 2nd ratio. It is a figure explaining a display form. It is a figure explaining the display form of the bottom sediment information displayed on the display monitor of the underwater detection apparatus concerning one Embodiment of this invention, (a) is a 2nd bottom sediment display form, (b) is a 2nd ratio. It is a figure explaining a display form. It is a figure explaining the display form of the bottom sediment information displayed on the display monitor of the underwater detection apparatus concerning one Embodiment of this invention, (a) is a 3rd bottom sediment display form, (b) is a 3rd ratio. It is a figure explaining a display form. It is a figure explaining the display form of the bottom sediment information displayed on the display monitor of the underwater detection apparatus concerning one Embodiment of this invention, (a) is a 3rd bottom sediment display form, (b) is a 3rd ratio. It is a figure explaining a display form. It is a figure explaining one modification of the display form of the bottom sediment information displayed on the display monitor of the underwater detection device concerning one embodiment of the present invention. It is a flowchart figure showing the flow of operation | movement of the main process in the underwater detection apparatus which concerns on one Embodiment of this invention. It is a flowchart figure showing the flow of operation of the operation button processing in the underwater detection device concerning one embodiment of the present invention. It is a flowchart figure showing the flow of operation of user bottom sediment addition processing in the underwater detection device concerning one embodiment of the present invention. It is a flowchart figure showing the flow of operation | movement of the user bottom sediment correction process in the underwater detection apparatus which concerns on one Embodiment of this invention. It is a flowchart figure showing the flow of operation | movement of the learning process in the underwater detection apparatus which concerns on one Embodiment of this invention. It is a figure which shows the screen displayed on the display monitor in the user bottom sediment correction process of the underwater detection apparatus which concerns on one Embodiment of this invention. It is a figure which shows the modification of the underwater detection apparatus which concerns on one Embodiment of this invention. It is explanatory drawing explaining the conventional neural network.

  Hereinafter, an underwater detection device according to an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is an explanatory diagram for explaining the underwater detection device 1.

(Schematic configuration of the underwater detection device 1)
As shown in FIG. 1, the underwater detection device 1 is mounted on a ship 5. A vibrator 11 is provided on the bottom of the ship 5 on which the underwater detection device 1 is mounted. The vibrator 11 emits a transmission pulse that is an ultrasonic wave toward the bottom of the water, and the transmission pulse is generated at the bottom of the water. A reflected echo signal is received.

  The ship 5 has a detection device main body 90 installed in the wheelhouse. The detection device main body 90 includes an input data generation unit 6, a variable neural network unit 7 (hereinafter, variable NN unit 7), a fixed neural network. 26 (hereinafter, fixed NN 26), output control unit 8, and operation input unit 70.

  The input data generation unit 6 calculates a feature amount based on the echo signal received by the transducer 11 and generates input data to be input to a later-described variable neural network 40 (hereinafter, variable NN40) or fixed NN26. .

  The fixed NN 26 generates first output data for determining a bottom sediment to which the input data belongs among a plurality of bottom sediments registered in advance.

  The variable NN unit 7 generates second output data for determining whether or not the input data belongs to the new bottom sediment when a new bottom sediment is registered by the user via the operation input unit 70. The variable NN 40 to be added is added before the fixed NN 26. Then, according to the second output data of the variable NN 40, it is determined whether or not the input data is output to the fixed NN 26. Only when it is determined that the input data is output, the input data is output to the fixed NN 26. .

  Further, when a new bottom sediment is registered by the user via the operation input unit 70, the variable NN unit 7 deletes the variable NN 40 added to the previous stage of the fixed NN 26.

  The output control unit 8 generates bottom sediment information 55 based on the first output data output from the fixed NN 26 and / or the second output data output from the variable NN 40, and the bottom sediment information 55 is generated. The information is output to the display monitor 50 provided in the detection device main body 90 so that the user can recognize it.

(Detailed configuration of the underwater detection device 1)
FIG. 2 is a block diagram showing a configuration of the underwater detection device according to the embodiment of the present invention. The underwater detection device 1 has a function of detecting a school of fish and a function of determining the bottom sediment of the seabed. Since the function of detecting a school of fish is the same as that of the conventional one, the description thereof is omitted.

  The vibrator 11 emits a transmission pulse, which is an ultrasonic wave having two frequencies (here, 50 kHz and 200 kHz), toward the bottom of the water, and receives an echo signal in which the transmission pulse is reflected from the bottom of the water. When the transmission pulse width is 1 ms and the frequency is 50 kHz, the transmission pulse is a sine wave signal for 50 cycles.

  The input data generation unit 6 includes a transmission / reception unit 12, a frequency conversion unit 13, an A / D converter 14, a detection unit 15, an amplitude data storage unit 16, a depth calculation unit 17, a transmission pulse width calculation unit 18, and an amplitude data extraction unit 19 , A gain correction unit 20, a smoothing unit 21, a normalization unit 22, and a feature amount calculation unit 23.

  The transmission / reception unit 12 supplies a transmission drive signal having a pulse width corresponding to a transmission pulse width control signal input from a transmission pulse width calculation unit 18 (described later) to the transducer 11, and generates an echo signal output from the transducer 11. On the other hand, band limiting and signal amplification are performed.

  The frequency converter 13 converts an echo signal of 50 kHz or 200 kHz into an intermediate frequency signal. By the conversion of the intermediate frequency signal, the subsequent signal processing for the echo signals of two frequencies is made common. The A / D converter 14 converts the intermediate frequency signal into a digital signal. The detector 15 detects the digital signal output from the A / D converter 14 and outputs amplitude data. The amplitude data is stored in the amplitude data storage unit 16 and is read out from the amplitude data storage unit 16 during arithmetic processing by the depth calculation unit 17 and the amplitude data extraction unit 19.

  The depth calculation unit 17 obtains the seabed depth from the amplitude data. Since the calculation of the seabed depth is the same as the conventional method, the description thereof is omitted. The transmission pulse width calculation unit 18 calculates a transmission pulse width proportional to the seabed depth based on the seabed depth calculated by the depth calculation unit 17, and sends a transmission pulse width control signal corresponding to the transmission pulse width to the transmission / reception unit 12. send. It should be noted that the transmission pulse width may be exactly proportional to the seabed depth or may be only approximately proportional. The amplitude data extraction unit 19 extracts amplitude data used for calculation of a feature amount in a feature amount calculation unit 23 described later from input amplitude data at a time interval proportional to the seabed depth.

  The gain correction unit 20 performs TVG (Time Variable Gain) processing on the amplitude data extracted by the amplitude data extraction unit 19. By this TVG processing, the attenuation amount of the received signal level due to the propagation loss of the ultrasonic signal is corrected. In FIG. 2, the gain correction unit 10 is arranged at the subsequent stage of the amplitude data extraction unit 19, but may be arranged at the previous stage.

  The smoothing unit 21 smoothes the amplitude data after gain correction. This smoothing suppresses fluctuations in the amplitude data value due to slight fluctuations of the hull, noise, etc., and improves the stability of the discrimination result.

  The normalizing unit 22 normalizes the smoothed amplitude data with their maximum values. By normalizing the amplitude data, the bottom sediment discrimination result is not affected by the variation in the reception level due to the secular change of the sensitivity of the vibrator 11 or the bubbles near the sea surface. In the present embodiment, a transmission pulse having a pulse width proportional to the seabed depth is emitted from the transducer 11, amplitude data is extracted from a predetermined range of the echo signal at a time interval proportional to the seabed depth, and the extracted amplitude data is extracted. TVG processing is performed, and amplitude data after TVG processing is further normalized. As a result, the normalized amplitude data is the same regardless of the seabed depth if the bottom sediment is the same. Therefore, in the underwater detection device 1 of the present embodiment that acquires bottom sediment information based on the feature amount calculated from the normalized amplitude data, the bottom sediment information does not depend on the seabed depth. That is, if the bottom sediment is the same, the bottom sediment information is the same even if the seabed depth is different.

  The feature amount calculation unit 23 calculates a feature amount from the normalized amplitude data, and generates input data that is input to the fixed NN 26, the variable NN 40, and the like. The feature amount is calculated for each processing cycle of 50 kHz and 200 kHz. Further, in the present embodiment, the input data is formed based on the feature amount obtained in the processing cycle of 50 kHz and the feature amount obtained in the processing of 200 kHz as one set of feature amount sets. 19 input signals.

  The fixed NN 26 generates first output data for discriminating the bottom sediment to which the input data output from the feature amount calculation unit 23 belongs among a plurality of bottom sediments registered in advance. In this embodiment, four types of sediments {rock, stone, sand, mud} are registered in advance, and the first output data is a value corresponding to the similarity between the input data and each sediment. .

  Here, the configuration of the fixed NN 26 will be described. The fixed NN 26 is a hierarchical neural network composed of three layers: an input layer, an intermediate layer, and an output layer. The input layer is composed of 19 input neurons corresponding to 19 input signals as input data. The output layer is composed of four output neurons each corresponding to four types of sediments registered in advance. Each output neuron outputs a value corresponding to the similarity between the input data and the corresponding sediment as an output signal. The first output data consists of the four output signals.

  Next, learning of the fixed NN 26 will be described. First, at each location where the bottom sediment is rock, stone, sand, or mud, a large amount of input data is collected using the underwater detector 1 and the bottom sediment type is actually confirmed by an underwater camera or the like. The Next, the synapse coupling coefficient is determined by the back propagation method from the collected input data and the confirmed sediment type by a personal computer or the like. The synapse coupling coefficient determined by the error back propagation method is stored in the ROM 82 described later. The collected input data is stored in the ROM 82 as fixed learning data, which is learning data for sediment other than the new sediment.

  The variable NN unit 7 includes a variable neural network determination unit 24 (hereinafter, variable NN determination unit 24), a variable neural network output determination unit 25 (hereinafter, variable NN output determination unit 25), and a neural network management unit 27 (hereinafter, NN management). Unit 27), variable NN 40, threshold setting unit 28, and learning data acquisition unit 31.

  The variable NN determination unit 24 determines whether or not the variable NN 40 is registered based on the registration signal output from the NN management unit 27. When it is determined that the variable NN 40 is registered, the input data output from the feature amount calculation unit 23 is output to the variable NN 40.

  The variable NN output determination unit 25 determines whether or not to output the input data output from the feature amount calculation unit 23 to the fixed NN 26 based on a threshold signal output from the threshold setting unit 28 described later. When it is determined that the input data is output to the fixed NN 26, the input data is output to the fixed NN 26. Specifically, a threshold signal indicating that the value corresponding to the similarity between the input data and the new bottom sediment is equal to or greater than a threshold set by the user via the operation input unit 70 is input from the threshold setting unit 28. If it is determined that the input data is not output. On the other hand, when a threshold signal that is smaller than the threshold is input from the threshold setting unit 28, it is determined that the input data is output. When a threshold signal indicating that the mode is not the threshold mode is received from the threshold setting unit 28, the input data is always output to the fixed NN 26. Here, the threshold mode is a mode for determining whether to output to the fixed NN 26 according to a value corresponding to the similarity between the input data and the new bottom sediment.

  The NN management unit 27 stores the input data output from the feature amount calculation unit 23 and the input data output from the learning data acquisition unit 31. Further, the NN management unit 27 receives from the operation input unit 70 an NN operation signal indicating registration of a new bottom sediment as a new discrimination target or cancellation of registration of a new bottom sediment. It should be noted that the NN operation signal indicating registration of a new sediment includes designation data for designating input data (user registration learning data) to be learned by the variable NN 40 among the input data stored in the NN management unit 27. ing.

  Further, when the NN management unit 27 receives an NN operation signal indicating registration of a new bottom sediment, the NN management unit 27 sets input data specified by the specified data as user registration learning data, and sets the user registration learning data and the fixed learning data. To make the variable NN 40 learn. At the same time, a registration signal indicating that the variable NN 40 has been added is output to the variable NN determination unit 24. On the other hand, when an NN operation signal indicating cancellation of registration of new sediment is received, the NN management unit 27 outputs a registration signal indicating that the variable NN 40 has been deleted to the variable NN determination unit 24.

  The variable NN 40 generates second output data for determining whether or not the input data output from the feature amount calculation unit 23 belongs to the new bottom sediment.

  Here, the configuration of the variable NN 40 will be described. The variable NN 40 is a hierarchical neural network including three layers: an input layer, an intermediate layer, and an output layer. Similar to the fixed neural network 26, the input layer is composed of 19 input neurons corresponding to 19 input signals as input data. The output layer is composed of two output neurons each corresponding to a new bottom sediment and a bottom sediment other than the new bottom sediment. Each output neuron outputs a value corresponding to the similarity to the corresponding bottom sediment as an output signal. The second output data consists of the two output signals.

  Next, learning of the variable NN 40 will be described. First, user registration learning data is set as bottom sediment information of a new bottom sediment, and fixed learning data stored in the ROM 82 is set as bottom sediment information other than the new bottom sediment. If the data amount of the user registration learning data is different from the data amount of the fixed learning data, the data amount is increased by duplicating a part of the data having a smaller data amount, and the data amount of the user registration learning data And the amount of fixed learning data is set to be the same. Next, a synapse coupling coefficient is determined from the user registration learning data and the fixed learning data by the error back propagation method using the main control circuit 80 described later. The synapse coupling coefficient determined by the error back propagation method and the user registration learning data are stored in the RAM 83 described later. By doing so, the variable NN 40 can be made to learn without actually checking the bottom sediment type with an underwater camera or the like. In addition, since the number of output neurons in the output layer is as small as two, the learning time required for registering a new sediment can be reduced.

  The threshold setting unit 28 determines whether or not the value corresponding to the similarity between the input data and the new bottom sediment, which is the second output data, is equal to or greater than the threshold set by the user via the operation input unit 70. The threshold value signal indicating the determination result is output to the variable NN output determination unit 25 and the output processing unit 29. When the first output data of the fixed NN 26 is always used when generating bottom sediment information in the output processing unit 29 described later, a variable NN output determination unit 25 generates a threshold signal indicating that the mode is not the threshold mode. And output to the output processing unit 29.

  The output control unit 8 includes an output processing unit 29, a display control unit 51, an audio control unit 61, a display monitor 50, and a speaker 60.

  Based on the threshold signal output from the threshold setting unit 28, the first output data output from the fixed NN 26, and / or the second output data output from the variable NN 40, the output processing unit 29 Is generated. Specifically, when the output processing unit 29 receives from the threshold setting unit 28 a threshold signal indicating a determination result that is equal to or greater than the threshold, the bottom sediment giving priority to the second output data over the first output data. Generate information. On the other hand, when a threshold signal indicating a determination result that is smaller than the threshold is received, sediment information is generated only from the first output data. When a threshold signal indicating that the threshold mode is not set is received, and variable NN 40 is registered, bottom sediment information is generated using the first output data and the second output data, and variable NN 40 is generated. Is not registered, bottom sediment information is generated only from the first output data.

  The display control unit 51 converts the sediment information generated by the output processing unit 29 into display data, and causes the display monitor 50 to display the display data. The voice control unit 61 generates voice data based on predetermined bottom sediment information and outputs the voice data from the speaker 60. When the underwater detection device 1 does not include the display monitor 50 or the speaker 60, display data may be displayed on a monitor of another device, or audio data may be output from a speaker of another device. The operation input unit 70 detects user operations such as a threshold change operation, a display mode (output mode) change operation, and a new sediment registration operation.

  The learning data acquisition unit 31 can acquire learning data of a new bottom sediment from the outside. For example, new bottom sediment learning data acquired by another underwater detection device 1 is acquired via wireless communication, or new bottom sediment learning data stored in an external memory such as a USB (Universal Serial Bus) memory. Have been able to get.

  Here, the “fixed neural network” is a neural network that outputs output data for determining a discrimination target to which input data belongs among a plurality of discrimination targets registered before registering a new discrimination target. is there. Also. A “variable neural network” is a neural network that outputs output data for determining whether input data belongs to a new discrimination target or a plurality of discrimination targets registered before registering the new discrimination target. is there.

  “Fixed learning data” is data used for learning of a fixed neural network. “User registration learning data” is data about a new discrimination target.

(Operation of underwater detector)
Next, the operation of the underwater detection device 1 will be described. The transmission / reception unit 12 supplies the transducer 11 with a transmission drive signal having a pulse width corresponding to the transmission pulse width control signal input from the transmission pulse width calculation unit 18. As a result, an ultrasonic transmission pulse of 50 kHz or 200 kHz is emitted (transmitted) from the vibrator 11 toward the seabed. When the transmission pulse width is 1 ms and the transmission frequency is 50 kHz, the transmission pulse is a sine wave signal for 50 cycles.

  The transmission pulse emitted from the transducer 11 is reflected by the seabed, and a signal including a seabed echo is received by the transducer 11. Band limiting and signal amplification for the received signal are performed by the transmitting / receiving unit 12, and the received signal is converted into an intermediate frequency signal by the frequency converting unit 13. By this frequency conversion, subsequent signal processing for the received signals of two frequencies is made common. The intermediate frequency signal (frequency-converted received signal) is converted into a digital signal by the A / D converter 14, further detected by the detector 15, and amplitude data is output from the detector 15. The amplitude data is stored in the amplitude data storage unit 16 and is read out from the amplitude data storage unit 16 during arithmetic processing by the depth calculation unit 17 and the amplitude data extraction unit 19.

  The depth calculation unit 17 obtains the seabed depth by a conventional method. The transmission pulse width calculation unit 18 calculates a transmission pulse width proportional to the seabed depth based on the seabed depth calculated by the depth calculation unit 17, and sends a transmission pulse width control signal corresponding to the transmission pulse width to the transmission / reception unit 12. send. For example, when the seabed depth is 50 m, 100 m, and 200 m, the transmission pulse width is 0.5 ms, 1 ms, and 2 ms, respectively. Further, when the seabed depth is 70 m to 90 m and 90 m to 110 m, the transmission pulse width may be set to 0.8 ms and 1 ms, respectively, and may be proportional to the seabed depth stepwise. That is, the transmission pulse width may be exactly proportional to the seabed depth, or may be only approximately proportional. When the transmission drive signal described above is supplied to the vibrator 11 from the transmission / reception unit 12 that has received the transmission pulse width control signal, a transmission pulse having a pulse width proportional to the seabed depth is emitted from the vibrator 11.

  The amplitude data extraction unit 19 extracts amplitude data from a predetermined range of the amplitude data series output from the detection unit 15, that is, from a predetermined range of the echo signal at a time interval proportional to the seabed depth. Here, the series of amplitude data is a numerical sequence generated by arranging a plurality of amplitude data obtained by transmitting a transmission pulse once in the order of times when they occurred.

  The gain correction unit 20 performs TVG (Time Variable Gain) processing on the amplitude data extracted by the amplitude data extraction unit 19. By this TVG processing, the attenuation amount of the received signal level due to the propagation loss of the ultrasonic signal is corrected. In FIG. 2, the gain correction unit 10 is disposed at the subsequent stage of the amplitude data extraction unit 19, but may be disposed at the preceding stage. The smoothing unit 21 smoothes the amplitude data after gain correction. Specifically, amplitude data obtained by a plurality of processing cycles of the same frequency (50 kHz or 200 kHz) is averaged for each data number. The value output from the smoothing unit 21 in the k-th processing cycle is, for example, an average value of amplitude data obtained in five processing cycles from the (k−4) th to the k-th processing cycle. The smoothing process suppresses fluctuations in the amplitude data value due to slight fluctuations in the hull or noise, and improves the stability of the bottom sediment discrimination result.

  The normalizing unit 22 normalizes the smoothed amplitude data with their maximum values. By normalizing the amplitude data, the bottom sediment discrimination result is not affected by the variation in the reception level due to the secular change of the sensitivity of the vibrator 11 or the bubbles near the sea surface. In the present embodiment, a transmission pulse having a pulse width proportional to the seabed depth is emitted from the transducer 11, amplitude data is extracted from a predetermined range of the echo signal at a time interval proportional to the seabed depth, and the extracted amplitude data is extracted. TVG processing is performed, and amplitude data after TVG processing is further normalized. As a result, the normalized amplitude data is the same regardless of the seabed depth if the bottom sediment is the same. Therefore, in the present invention in which the bottom sediment is determined based on the feature amount calculated from the normalized amplitude data, the bottom sediment determination result does not depend on the seabed depth. That is, if the bottom sediment is the same, the bottom sediment discrimination result is the same even if the seabed depth is different.

  The feature amount calculation unit 23 calculates a total value of amplitude data in each segment (hereinafter referred to as a feature amount) for a plurality of sections (hereinafter referred to as segments) in the normalized amplitude data series. The feature amount is calculated for each processing cycle of 50 kHz and 200 kHz. In the present embodiment, the feature amount is calculated by summing the amplitude data in each segment. However, the moment value calculated using the standard deviation of the amplitude data value in each segment or from the amplitude data in each segment is used. The feature amount may be calculated by using it. Input data is formed based on the calculated feature amount. In the present embodiment, 19 input signals formed on the basis of one set of feature values obtained in the processing cycles of 50 kHz and 200 kHz are used as input data.

  When the input data is input from the feature amount calculation unit 23, the variable NN determination unit 24 determines whether the variable NN 40 is registered based on the registration signal output from the NN management unit 27. When it is determined that the variable NN 40 is registered, the input data is output to the registered variable NN 40.

  The variable NN output determination unit 25 determines whether or not to output the input data to the fixed NN 26 based on the threshold signal output from the threshold setting unit 28 when input data is input from the feature amount calculation unit 23. To do. Specifically, when the threshold signal is equal to or greater than the threshold value, the input data is not output to the fixed NN 26. On the other hand, when the second output data is smaller than the threshold value, the input data is output to the fixed NN 26. In the case of a threshold signal indicating that the mode is not the threshold mode, the input data is output to the fixed NN 26.

  The fixed NN 26 is a hierarchical neural network, and generates first output data based on input data output from the feature amount calculation unit 23 or input data output from the NN management unit 27. The output data consists of four output signals, and the synapse coupling coefficient is determined so that each output signal has a value corresponding to the similarity to the sediment type {rock, stone, sand, mud}. . The synapse coupling coefficient is recorded in the ROM 82 in advance.

  That is, the fixed NN 26 generates a numerical value (sediment information) indicating a similarity to a previously registered bottom sediment type based on input data calculated from an echo signal from the seabed that determines the bottom sediment. The number of sediment types registered in advance is not limited to four, and the sediment type is not limited to {rock, stone, sand, mud}.

  When the input data is input from the feature amount calculation unit 23 or when the input data is input from the learning data acquisition unit 31, the NN management unit 27 converts the input data into characteristics of the input data (when the input data is acquired). Information, such as time, water temperature, acquisition method, and water area). Further, when the NN management unit 27 receives an NN operation signal indicating registration of a new bottom sediment, the NN management unit 27 sets input data indicated by designation data included in the NN operation signal as user registration learning data, and the user registration The variable NN 40 is trained using the learning data and the fixed learning data. At the same time, a registration signal indicating that the variable NN 40 is registered is output to the variable NN determination unit 24. On the other hand, when an NN operation signal indicating cancellation of registration of a new sediment is received, a registration signal indicating that registration of the variable NN 40 is canceled is output to the variable NN determination unit 24, and the variable NN 40 is deleted.

  The variable NN 40 is a hierarchical neural network, and generates second output data based on the input data output from the feature amount calculation unit 23. The second output data is composed of two output signals. One output signal has a value corresponding to the similarity between the input data and the new bottom sediment, and the other output signal has a value corresponding to the similarity between the input data and the bottom sediment other than the new bottom sediment. In addition, the synaptic coupling coefficient is determined. The synapse coupling coefficient is recorded in the RAM 83.

  When the second output data is input from the variable NN 40, the threshold setting unit 28 has a value corresponding to the similarity between the input data and the new bottom sediment equal to or greater than the threshold set by the user via the operation input unit 70. And a threshold signal indicating the determination result is output to the variable NN output determination unit 25 and the output processing unit 29.

  Based on the threshold signal output from the threshold setting unit 28, the first output data output from the fixed NN 26, and / or the second output data output from the variable NN 40, the output processing unit 29 Is generated. Specifically, when the output processing unit 29 receives a threshold signal from the threshold setting unit 28, the output processing unit 29 generates bottom sediment information giving priority to the second output data with respect to the first output data, and receives the threshold signal. If not, bottom sediment information is generated only from the first output data. Further, when an output form is selected via the operation input unit 70, the output processing unit 29 generates bottom sediment information based on the selected output form.

  The display control unit 51 converts the sediment information generated by the output processing unit 29 into display data, and causes the display monitor 50 to display the display data. The voice control unit 61 generates voice data based on predetermined bottom sediment information and outputs the voice data from the speaker 60.

(External structure of the detector main body 90)
Next, with reference to FIG. 3, the external structure of the detection device main body 90 of the underwater detection device 1 in the present embodiment will be described.

  As shown in FIG. 3, the detection device main body 90 is provided with a display monitor 50 and an operation input unit 70 having various operation buttons to be operated by the user. The operation input unit can perform operations such as changing the display form of the bottom sediment information displayed on the display monitor 50, determining the threshold value of the second output data output from the variable NN 40, and registering the new sediment (variable NN 40). I have to.

(Electric configuration of underwater detector)
Next, the electrical configuration of the underwater detection device 1 according to the present embodiment will be described with reference to FIG. The underwater detection device 1 in the present embodiment includes a main control circuit 80 and a peripheral device that is electrically connected to the main control circuit. The main control circuit 80 includes a CPU 81 that performs a control operation in accordance with a preset program, a ROM 82, and a RAM 83.

The CPU 81 includes the vibrator 11, an operation input unit 70 for inputting operations such as various instructions, a display control unit 51 for performing display control of the display monitor 50, a sound control unit 61 for performing sound control of the speaker 60, and a variable NN 40. A learning data acquisition unit 31 for acquiring input data to be learned is connected.
The ROM 82 stores various programs, a synapse coupling coefficient of the fixed NN 26, fixed learning data, and the like. The RAM 83 stores the amplitude data, the feature amount, the input data, the synapse coupling coefficient of the variable NN 40, information on the created user sediment (new sediment), and the like.

(Display format of the display monitor)
Next, the display form of the display monitor in the present embodiment will be described with reference to FIGS.

(First display form)
FIG. 5 is a diagram for explaining a display form (hereinafter referred to as a first display form) when bottom sediment information is generated using only the first output data of the fixed NN 26 at an arbitrary time.

FIG. 5 is an image of the display monitor 50 when the ship 5 is traveled while generating the bottom sediment information 55. The horizontal direction in the figure represents time, but when the ship speed of the ship 5 is clear, it can also be represented as a distance. Image display monitor 50 is displayed so as to move from the right end t ₀ in FIG direction of the arrow A, and reaches the left end portion t _n, it is displayed so as to sequentially disappear. Therefore, the image displayed at the lower part of the right end t ₀ is the latest information and becomes more past information as it approaches the left end t _n . The upper half of the image on the display monitor 50 is the same image as that displayed by a normal underwater detection device. 53 is an image directly attributable to the transmission drive signal, and 54 is an image of the seabed. 55 is an image of bottom sediment information, and a basic display image 56 representing each bottom sediment can be identified (for example, colored). More specifically, 56a is a basic display image representing mud, 56b sand, 56c stone, and 56d rock. In addition, the matter about the image of the display monitor 50 is a matter applied to the second display form and the third display form described with reference to FIGS. 8 and 10.

  In the present embodiment, as shown in FIG. 5 (a), a display form (hereinafter referred to as the first bottom) that displays only the bottom sediment having the highest value corresponding to the similarity among the bottom sediments registered in advance. Quality display form) and a display form (hereinafter referred to as a first ratio display) that displays each value corresponding to the similarity of the bottom sediment registered in advance according to the ratio to the whole as shown in FIG. 5 (b). Two types of first display forms.

In the first bottom sediment display form shown in FIG. 5 (a), for example, the value corresponding to the similarity of each bottom sediment output from the fixed NN 26 is 100% as shown in FIG. 6 (a). If mud is 60%, sand is 20%, stone is 10%, and rock is 10%, the value corresponding to the similarity of mud is the highest, so as shown in FIG. Only 56a, which is a basic display image of mud, is displayed below the right end t ₀ of the display monitor 50.

  Further, the ratio display form shown in FIG. 5B displays a value corresponding to the similarity of each sediment output from the fixed NN 26 so that the user can recognize it. In the present embodiment, as shown in FIG. 7B, the vertical width of each basic display image 56 is set to a length corresponding to the ratio of the value corresponding to the similarity of each sediment to the whole. , So that the user can recognize it. For example, if the mud is 60%, the sand is 20%, the stone is 10%, and the rock is 10%, the ratio of the vertical width is 60: 20: 10: 10. Is assigned to each basic display image 56 so as to have the ratio.

(Second display form)
Next, a display form for generating sediment information using only one of the first output data of the fixed NN 26 and the second output data of the variable NN 40 at an arbitrary time (hereinafter referred to as the second output data). Display mode) will be described. The second display form indicates the user bottom when the value corresponding to the similarity of the user bottom (new bottom) output from the variable NN 40 is larger than the threshold set by the threshold setting unit 28. The user bottom sediment display image 57 is displayed below the right end t ₀ of the display monitor 50. On the other hand, if it is smaller than the threshold value, the basic display image 56 is displayed below the right end t ₀ of the display monitor 50 based on the value output from the fixed NN 26 and corresponding to the similarity of each sediment. Note that the user bottom sediment display image 57 is made distinguishable from the basic display image 56 (for example, colored). Hereinafter, when the value corresponding to the similarity between the user bottom sediment and the bottom sediment other than the user bottom sediment output from the variable NN 40 is 100% as shown in FIG. 6B, The case where the user bottom sediment (new bottom sediment) is 60% and the bottom sediment other than the user bottom sediment is 40% will be described.

  In the present embodiment, the second display mode includes a second bottom sediment display mode and a second ratio display mode.

Fig.8 (a) has shown the 2nd bottom sediment display form. As shown in FIG. 9A, in the second bottom sediment display form, when the threshold set by the threshold setting unit 28 is 50%, the value corresponding to the similarity of the user bottom sediment is larger. The user bottom sediment display image 57 is displayed below the right end t ₀ of the display monitor 50. On the other hand, when the threshold set by the threshold setting unit 28 is 80%, the value corresponding to the similarity of the user bottom sediment is smaller, so the value corresponding to the similarity among the sediments registered in advance. Only the bottom sediment with the highest _value is displayed below the right end t ₀ of the display monitor 50.

FIG. 8B shows a second ratio display form. As shown in FIG. 9B, in the second ratio display form, when the threshold set by the threshold setting unit 28 is 50%, the value corresponding to the similarity of the user bottom sediment is larger, so that the user A bottom display image 57 is displayed below the right end t ₀ of the display monitor 50. On the other hand, when the threshold set by the threshold setting unit 28 is 80%, the value corresponding to the similarity of the user bottom sediment is smaller, so each value corresponding to the similarity of the bottom sediment registered in advance is set. In accordance with the ratio with respect to the whole, it is displayed below the right end portion t ₀ of the display monitor 50.

(Third display form)
Next, a display form for generating sediment information using the output data of both the first output data of the fixed NN 26 and the second output data of the variable NN 40 at an arbitrary time (hereinafter referred to as a third display form). Will be described.

  In the present embodiment, the third display mode includes a third sediment display mode and a third ratio display mode.

  FIG. 10A shows a third bottom sediment display form. In the third bottom sediment display form, first, the length in the vertical direction of the bottom sediment information 55 is set according to the ratio of the value corresponding to the similarity of the bottom sediment other than the user bottom sediment and the bottom sediment to the whole. The user bottom display image 57 and the basic display image 56 are assigned. For example, as shown in FIG. 6 (b), when the whole is 100%, if the user bottom sediment is 60% and the bottom sediment other than the user bottom sediment is 40%, as shown in FIG. 11 (a), The ratio of the length in the vertical direction between the user bottom display image 57 and the basic display image 56 is 60:40. Next, the length of the basic display image 56 in the vertical direction is assigned to each basic display image 56 according to the ratio of the value corresponding to the similarity of each bottom sediment registered in advance to the whole.

  FIG. 10B shows a third ratio display form. In the third ratio display form, as shown in FIG. 11B, when the value corresponding to the similarity of the user sediment output from the variable NN 40 is equal to or larger than the threshold set by the threshold setting unit 28, When the display form is the same as the third bottom sediment display form and is smaller than the threshold value, the first ratio display form is adopted.

  Although the display form of the display monitor in the present embodiment has been described above, the present invention is not limited to this. For example, as shown in FIG. 12, two bottom sediments registered in advance may be handled as one bottom sediment and displayed as a basic mixed display image 58. FIG. 12 shows a display form when mud and sand, which are sediments registered in advance, are handled as the same sediment.

  Also, as shown in FIG. 12, the bottom sediment registered in advance and the user bottom sediment may be handled as one type of bottom sediment and displayed as a user mixed display image 59. FIG. 12 shows a display form when the user bottom sediment and the previously registered bottom sediment are handled as the same bottom sediment.

  In addition, a plurality of display forms may be displayed on the display monitor 50.

  The various display forms described above can be selected by the user via the operation input unit 70. By doing so, the user can change the output form (display form) according to his / her preference.

(Operation of the CPU 81 of the underwater detection device 1)
Next, the operation of the underwater detection device 1 according to the present embodiment will be described.

(Main process)
The main process of the underwater detection device 1 will be described. FIG. 13 is a flowchart showing the operation flow of the underwater detection device 1 in the main process.

  First, the CPU 81 determines whether or not an underwater detection start operation has been performed by the user via the operation input unit 70 (S1). If it is determined that the underwater detection start operation has not been performed (S1: NO), the process returns to S1. On the other hand, when it is determined that an underwater detection start operation has been performed (S1: YES), the seabed depth is calculated, a transmission pulse having a pulse width corresponding to the seabed depth is transmitted toward the seabed, and an echo signal is received ( S2).

  Next, the CPU 81 generates input data based on the received echo signal (S3), and determines whether or not the variable NN 40 is registered (added) (S4). If it is determined that the variable NN 40 is not registered (S4: NO), the process proceeds to S8 described later. On the other hand, when it is determined that the variable NN 40 is registered (S4: YES), input data is input to the variable NN 40 (S5), and it is determined whether or not the threshold mode is set (S6). If it is determined that the threshold mode is not set (S6: NO), the process proceeds to S8 described later. On the other hand, if it is determined that the threshold mode is set (S6: YES), it is determined whether the second output data output from the variable NN 40 is equal to or greater than the threshold (S7). If it is determined that it is not equal to or greater than the threshold (S7: NO), the process proceeds to S8 described later. On the other hand, if it is determined that the threshold value is exceeded (S7: YES), the display monitor 50 and the speaker 60 are controlled to notify the user that the bottom sediment is a new bottom sediment (S9).

  In the process of S8, the CPU 81 inputs input data to the fixed NN 26. After the process of S8 and the process of S9, the CPU 81 outputs the bottom sediment information based on the second output data of the variable NN 40 and / or the first output data of the fixed NN 26 and the selected display form. The bottom sediment information is generated and displayed on the display monitor 50 (S10).

  Next, the CPU 81 stores the input data in the RAM 83 in association with the characteristics of the input data (information such as the time when the input data was acquired, the water temperature, the acquisition method, and the water area) (S11). Next, the CPU 81 determines whether or not the operation button of the operation input unit 70 has been operated (S12). If it is determined that it has not been operated (S12: NO), the process returns to S2. On the other hand, when it is determined that the operation button has been operated (S12: YES), it is determined whether or not the operation is an underwater detection end operation (S13). If it is determined that the operation is to end underwater detection (S13: YES), this process ends. On the other hand, if it is determined that the operation is not an underwater detection end operation (S13: NO), an operation button process described later is performed (S14), and the process returns to S2.

(Operation button processing)
The operation button process of the underwater detection device 1 will be described. FIG. 14 is a flowchart showing the operation flow of the underwater detection device 1 in the operation button process.

  First, the CPU 81 determines whether or not the operation of the operation button of the operation input unit 70 is a display form change operation (S20). If it is determined that the operation is a display mode changing operation (S20: YES), the display mode to be displayed on the display monitor 50 is changed to the selected display mode (S21), and this process is terminated. On the other hand, when it is determined that the operation is not a display form changing operation (S20: NO), it is determined whether the operation is a user sediment addition operation (S22). If it is determined that the operation is a user bottom sediment addition operation (S22: YES), a user bottom sediment addition process described later is performed (S23), and this process is terminated.

  On the other hand, when it is determined that it is not a user bottom sediment addition operation (S22: NO), the CPU 81 determines whether it is a user bottom sediment deletion operation (S24). If it is determined that the operation is a user bottom sediment deletion operation (S24: YES), the variable NN 40 is deleted, and the learning data of the variable NN 40 is replaced with the characteristics of the learning data (user bottom sediment name, registration, and deregistration operations). The information is stored in the RAM 83 in association with the date and time information), and the process is terminated.

  On the other hand, when it is determined that the operation is not a user bottom sediment deletion operation (S24: NO), the CPU 81 determines whether the operation is a user bottom sediment correction operation (S26). If it is determined that the operation is a user bottom sediment correction operation (S26: YES), a user bottom sediment correction process described later is performed (S27), and this process is terminated.

  On the other hand, if it is determined that the operation is not a user sediment removal operation (S26: NO), the CPU 81 determines whether or not the operation is a threshold value changing operation (S28). If it is determined that the operation is not a threshold change operation (S28: NO), this process is terminated. On the other hand, if it is determined that the operation is a threshold value changing operation, the threshold value stored in the RAM 83 is changed to the threshold value selected by the user, and this process is terminated.

(User sediment additional processing)
The user bottom sediment addition process of the underwater detection device 1 will be described. FIG. 15 is a flowchart showing an operation flow of the underwater detection device 1 in the user sediment addition process.

  First, the CPU 81 determines whether or not to create a new user sediment based on a user operation of the operation input unit 70 (S40). If it is determined that a new user sediment is to be created (S40: YES), a learning process described later is performed (S41). Next, the CPU 81 creates user bottom sediment information including information such as the synapse coupling coefficient obtained by the learning process, the learning data used for learning of the variable NN 40, the name of the user bottom sediment, and the user bottom sediment information. Is stored in the RAM 83 (S42). Next, the CPU 81 adds the variable NN 40 having the synapse coupling coefficient obtained by the learning process to the preceding stage of the fixed NN 26 (S45), and ends this process.

  On the other hand, if it is determined that a user bottom sediment is not newly created (S40: NO), the CPU 81 stores a list of previously created user bottom sediment information (names of user bottom sediments) stored in the RAM 83. A list or the like) is displayed on the display monitor 50 (S43). Next, the CPU 81 determines whether or not user bottom sediment information has been selected by the user via the operation input unit 70 (S44). If it is determined that it has not been selected (S44: NO), the process returns to S44. On the other hand, if it is determined that it has been selected (S44: YES), the variable NN 40 having the synapse coupling coefficient of the selected user sediment information stored in the RAM 83 is added to the preceding stage of the fixed NN 26 (S45). ), This process is terminated.

(User bottom sediment correction processing)
The user bottom sediment correction process of the underwater detection device 1 will be described. FIG. 16 is a flowchart showing an operation flow of the underwater detection device 1 in the user bottom sediment correction process.

  First, the CPU 81 displays user bottom sediment information created in the past on the display monitor 50 (S60). Next, the CPU 81 determines whether or not user bottom sediment information is selected by the user via the operation input unit 70 (S61). If it is determined that no user sediment has been selected (S61), the process returns to S61. On the other hand, if it is determined that the user bottom has been selected, it is determined whether or not an additional learning operation has been performed by the user via the operation input unit 70 (S62). If it is determined that an additional learning operation has been performed (S62: YES), a learning process described later is performed (S67), and the synapse coupling coefficient of the user sediment information stored in the RAM 83 and the learning of user registration learning data are performed. The data information is updated (S68), and this process is terminated.

  On the other hand, when it is determined that it is not an additional learning operation (S62: NO), as shown in FIG. 18, a list of learning data of user registration learning data of the selected user bottom sediment (user bottom sediment information) Is displayed (S63). Next, the CPU 81 determines whether or not learning data to be deleted is selected by the user via the operation input unit 70 (S64). If it is determined that the learning data to be deleted has not been selected (S64: NO), the process returns to S63. On the other hand, when it is determined that the learning data to be deleted is selected (S64: YES), the selected learning data is deleted from the user registration learning data (S64). Next, the CPU 81 calculates the synapse coupling coefficient of the variable NN 40 by the error back propagation method using the user registration learning data and the fixed learning data stored in the ROM 82 (S65), and the user base stored in the RAM 83 is calculated. The synapse coupling coefficient of the quality and the learning data information of the user registration learning data are updated (S68), and this process is terminated.

(Learning process)
The learning process of the underwater detection device 1 will be described. FIG. 17 is a flowchart showing the operation flow of the underwater detection device 1 in the learning process.

  First, the CPU 81 determines whether or not to learn using the input data stored in the RAM 83 when learning the variable NN 40 based on the user's operation of the operation input unit 70 (S80). If it is determined that the learning is performed using the input data stored in the ROM 82 (S80: YES), a list of characteristics of the input data stored in the RAM 83 is displayed (S81). Next, the CPU 81 determines whether or not input data has been selected by the user via the operation input unit 70 (S82). If it is determined that it has not been selected (S82: NO), the process returns to S82. On the other hand, if it is determined that the input data is selected, the input data is set as user registration learning data, and the synapse coupling of the variable NN40 is performed by the error back propagation method using the user registration learning data and the fixed learning data stored in the ROM 82. A coefficient is calculated (S90), and this process is terminated.

  On the other hand, if it is determined that learning is not performed using the input data stored in the RAM 83 (S80: NO), the CPU 81 has performed a learning data acquisition start operation by the user via the operation input unit 70. It is determined whether or not (S83). If it is determined that the learning data acquisition start operation has not been performed (S83: NO), the process returns to S83. On the other hand, when it is determined that the learning data acquisition start operation has been performed (S83: YES), the seabed depth is calculated, a transmission pulse having a pulse width corresponding to the seabed depth is transmitted toward the seabed, and an echo signal is received. (S84).

  Next, the CPU 81 generates input data based on the received echo signal (S85), and inputs the input data to the fixed NN 26 (S86). Next, the CPU 81 generates bottom sediment information using the first output data output from the fixed NN 26, and displays the bottom sediment information on the display monitor 50 (S87). Next, the CPU 81 stores the input data in the RAM 83 (S88).

  Next, the CPU 81 determines whether or not a learning data acquisition end operation has been performed by the user via the operation input unit 70 (S89). If it is determined that the learning data acquisition end operation has not been performed (S89: NO), the process returns to S84.

  On the other hand, if it is determined that the learning data acquisition end operation has been performed (S89: YES), the input data acquired in this learning process is set as user registration learning data, and the user registration learning data and the fixed data stored in the ROM 82 are stored. The learning data is used to calculate the synapse coupling coefficient of the variable NN 40 by the error back propagation method (S90), and the process is terminated.

(Outline of this embodiment)
As described above, the underwater detection device 1 of the present embodiment receives an echo signal of an ultrasonic wave transmitted into the water, inputs input data based on the echo signal to the neural network, and is output from the neural network. An underwater detection device that generates, every predetermined time, bottom sediment information 55 relating to bottom sediment to which the input data belongs among a plurality of registered bottom sediments based on output data. Among the bottom sediments, when the new bottom sediment is registered by the fixed NN 26 that outputs the first output data for determining the bottom sediment to which the input data belongs and the operation from the outside, the input data becomes the new bottom sediment. When the variable NN40 for outputting the second output data for determining whether or not the registered new bottom sediment is canceled by an external operation, An NN management unit 27 that deletes the variable NN 40; an output processing unit 29 that can generate bottom sediment information 55 in which the second output data is prioritized with respect to the first output data; and bottom sediment information 55 Display monitor 50 and speaker 60 are provided.

According to said structure, the 1st output data for discriminating the bottom sediment to which input data belong among the several registered bottom sediments is output from fixed NN26. When a new bottom sediment is newly registered by an external operation, the variable NN 40 is added, and second output data for determining whether or not the input data belongs to the new bottom sediment is received from the variable NN 40. Is output. The variable NN 40 is a neural network that determines whether or not it belongs to a new sediment, and the number of neurons can be reduced, so that the learning time can be suppressed. Further, it is possible to generate bottom sediment information 55 in which the second output data has priority over the first output data, and the bottom sediment information 55 is output from the display monitor 50 or the speaker 60. Further, the variable NN 40 is deleted when the registration of the registered new bottom sediment is canceled by an external operation.
In this way, by adding and deleting the variable NN 40, it is possible to register and deregister new sediments, so there is no need to change the structure of the fixed NN 26. Further, since the bottom sediment information 55 in which the second output data is prioritized with respect to the first output data is output from the display monitor 50 and the speaker 60, the user determines whether or not the input data belongs to the new bottom sediment. Can be easily recognized.

  Further, in the underwater detection device 1 of the present invention, the variable NN 40 is generated based on the received echo signal, and when the user registration learning data set as the new bottom sediment and the fixed NN 26 are learned by an external operation. And is learned by registered learning data set as bottom sediment other than new bottom sediment. According to the above configuration, the user registration learning data generated based on the received echo signal and set as a new bottom sediment by an external operation and the fixed NN 26 are used for learning, and other than the new bottom sediment. It is learned by fixed learning data set as bottom sediment. By doing in this way, the learning data used when learning the variable NN 40 can be acquired without actually confirming with an underwater camera or the like.

  In the underwater detection device 1 of the present invention, the user registration learning data can be added with learning data and a part of the user registration learning data can be deleted by an external operation. According to the above configuration, the user registration learning data can be added with learning data and a part of the user registration learning data can be deleted by an external operation. By doing so, it is possible to improve the learning data of the new bottom sediment, and it is possible to accurately generate the bottom sediment information 55 as to whether or not the input data belongs to the new bottom sediment. .

  Moreover, in the underwater detection device 1 of the present invention, the bottom sediment information 55 is generated from the second output data, and corresponds to a value corresponding to the similarity between the input data and the new bottom sediment, and / or the first output data. From the input data and a value corresponding to the similarity between each of the plurality of sediments registered in advance. According to the above configuration, the bottom sediment information 55 is generated from the value corresponding to the similarity between the input data and the new bottom sediment generated from the second output data, and / or from the first output data. , Consisting of a value corresponding to the degree of similarity between the input data and each of a plurality of sediments registered in advance. By doing so, the similarity between the input data and the new bottom sediment, and the similarity between the input data and each of the plurality of sediments registered in advance become obvious. In addition, it is obvious at a glance how the similarity between the input data and the new bottom sediment changes within the water area where the bottom sediment is determined, and how the similarity between the input data and each of the multiple registered bottom sediments changes. It becomes.

  In addition, the underwater detection device 1 of the present invention further includes a threshold setting unit 28 for setting a threshold, and in the output processing unit 29, a value corresponding to the similarity between the input data and the new bottom sediment is equal to or greater than the threshold. The bottom sediment information 55 in which the second output data is prioritized with respect to the first output data is generated. If the bottom sediment information 55 is smaller than the threshold value, the bottom sediment information is generated only by the first output data. According to the above configuration, when the value corresponding to the similarity between the input data and the new bottom sediment is not less than the threshold set by the threshold setting unit 28, the second output data is compared to the first output data. If the priority sediment information 55 is generated and is smaller than the threshold value, the sediment information 55 is generated only from the first output data. As described above, whether or not the second output data is used for generating the bottom sediment information 55 is determined by the magnitude of the value corresponding to the similarity between the input data and the new bottom sediment. It becomes obvious at a glance whether or not the value corresponding to the similarity to is greater than or equal to the set threshold value.

  Further, in the underwater detection device 1 of the present invention, a variable NN output determination unit that prevents input data from being input to the fixed NN 26 when the value corresponding to the similarity between the input data and the new bottom sediment is equal to or greater than a threshold value. 25 is further provided. According to the above configuration, when the value corresponding to the similarity between the input data and the new bottom sediment is equal to or larger than the threshold set by the threshold setting unit 28, the input data is not input to the fixed NN 26. By doing so, when the value corresponding to the similarity between the input data and the new bottom sediment is equal to or greater than the threshold value, the processing in the fixed NN 26 can be omitted.

  Further, in the underwater detection device 1 of the present invention, the threshold value can be changed by an external operation. According to the above configuration, the user determines the magnitude of the value corresponding to the similarity between the input data and the new bottom sediment that determines whether or not the second output data is reflected in the generation of the bottom sediment information 55. You can change it according to your own preference.

  Further, in the underwater detection device 1 of the present invention, there are a plurality of output forms for outputting bottom sediment information to the display monitor 50 and the speaker 60, and the output forms can be changed by an external operation. According to said structure, the user can change an output form according to own liking.

(Modification of this embodiment)
The embodiments of the present invention have been described above, but only specific examples have been illustrated, and the present invention is not particularly limited. Specific configurations and the like can be appropriately changed in design. Further, the actions and effects described in the embodiments of the invention only list the most preferable actions and effects resulting from the present invention, and the actions and effects according to the present invention are described in the embodiments of the present invention. It is not limited to what was done.

  For example, in the present embodiment, only one variable NN unit 7 (variable NN40) is provided in front of the fixed NN 26, but a plurality of variable NN units 7 (variable NN40) are provided as shown in FIG. May be enabled. This makes it possible to add a plurality of types of new bottom sediments as discrimination targets. In this case, the fixed learning data used for learning the variable NN 40 of each variable NN unit 7 is the learning data of the fixed NN 26 and the user registration learning data of the variable NN provided in the subsequent stage. Further, by changing the threshold value set by the threshold value setting unit 28 of each variable NN unit 7, it is possible to change the output data of the variable NN 40 that is prioritized when generating bottom sediment information.

DESCRIPTION OF SYMBOLS 1 Underwater detection apparatus 25 Variable neural network output determination part 26 Fixed neural network 27 Neural network management part 28 Threshold setting part 29 Output processing part 50 Display monitor (output part)
60 Speaker (output unit)

Claims (9)

The ultrasonic echo signal transmitted into the water is received, input data based on the echo signal is input to the neural network, and a plurality of registered bottom sediments are output based on the output data output from the neural network. Among them, an underwater detection device that generates bottom sediment information about the bottom sediment to which the input data belongs, every predetermined time,
A fixed neural network that outputs first output data for determining a bottom sediment to which the input data belongs among a plurality of bottom sediments registered in advance;
When a new bottom sediment is registered by an external operation, a variable neural network that outputs second output data for determining whether or not the input data belongs to the new bottom sediment is added. When the registration of the registered new bottom sediment is canceled by the operation of, the neural network management unit that deletes the variable neural network,
An underwater detection device comprising: an output processing unit capable of generating the bottom sediment information in which the second output data is prioritized with respect to the first output data.   The variable neural network is generated based on the received echo signal, and is used for learning the user registration learning data set as the new bottom sediment learning data and learning the fixed neural network by an external operation. The underwater detection device according to claim 1, wherein learning is performed using fixed learning data set as learning data for bottom sediment other than the new bottom sediment.   The underwater detection device according to claim 2, wherein the user registration learning data is configured such that learning data can be added and a part of the user registration learning data can be deleted by an external operation. .   The bottom sediment information is generated from the second output data, the value corresponding to the similarity between the input data and the new bottom sediment, and / or the input generated from the first output data. The underwater detection device according to any one of claims 1 to 3, wherein the underwater detection device includes a value corresponding to a degree of similarity between the data and each of the plurality of bottom sediments registered in advance. A threshold setting unit for setting the threshold;
The output processing unit gives priority to the second output data over the first output data when a value corresponding to the similarity between the input data and the new bottom sediment is equal to or greater than the threshold. The bottom sediment information is generated, and when the value indicated by the second output data is smaller than the threshold value, the bottom sediment information is generated only by the first output data. Underwater detection device.   A variable neural network output determination unit that prevents the input data from being input to the fixed neural network when a value corresponding to the similarity between the input data and the new bottom sediment is equal to or greater than the threshold; The underwater detection device according to claim 5.   The underwater detection device according to claim 5, wherein the threshold value is changeable by an external operation. Provided with an output unit (display monitor, speaker, etc.) for outputting the bottom sediment information,
The output form of the bottom sediment information to be output to the output unit is plural, and the output form can be changed by an operation from the outside. The underwater detection device described. A discrimination device that inputs input data to a neural network and generates discrimination information related to a discrimination target to which the input data belongs among a plurality of registered discrimination targets based on output data output from the neural network,
A fixed neural network that outputs first output data for determining a determination target to which the input data belongs among a plurality of determination targets registered in advance;
When a new discrimination target is registered by an external operation, a variable neural network that outputs second output data for judging whether or not the input data belongs to the new discrimination target is added. A neural network management unit for deleting the second neural network when registration of a new discrimination target is canceled by the operation of
And an output processing unit capable of generating the discrimination information in which the second output data is prioritized with respect to the first output data.

Images