Machine learning device, instruction information output device, machine learning method, instruction information output method, and program

Abstract

To support and achieve operation equivalent to a highly accurate pump dredge by a skilled operator.SOLUTION: A machine learning device 20 learns a relation between sensor data (in other words, state of each part of pump dredge device) detected by a sensor group 12, and operation with optimal construction efficiency by a skilled operator. A sensor data acquisition part 111 acquires time-series sensor data detected by the sensor group 12 according to operation of an operator (second operator) who is more unfamiliar with operation than the above skilled operator. An input part 112 inputs the sensor data acquired by the sensor data acquisition part 111 as an explanation variable, relative to a learned model stored in a storage part 203 of the machine learning device 20. An instruction information output part 113 generates and outputs instruction information according to swing operation amount prediction data obtained as an objective variable from the learned model according to the input of the sensor data.SELECTED DRAWING: Figure 7

Description

Technical Field

[0001] The present invention relates to a technique for performing pump dredging.

Background Art

[0002] In pump dredging work for the purpose of maintaining waterways, berths, etc., an operator operates a swing winch that moves the ladder horizontally and a ladder winch that moves the ladder vertically, and moves a cutter provided at the tip of the ladder along a planned line to perform dredging. For example, Patent Document 1 discloses a technique for controlling a ladder winch and a swing winch based on detection values of a mud content meter, a ladder winch ammeter, and a swing winch ammeter.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In pump dredging work, the efficiency of dredging work varies greatly depending on the skills of the operator, so rich experience and skills of a skilled operator are required. However, due to the retirement of skilled operators due to aging, it has become difficult to secure operators having such experience and skills. In addition, even for young operators, due to the reduction of dredging work and skilled operators, the opportunity to acquire experience and skills from such skilled operators has been decreasing.

[0005] Therefore, an object of the present invention is to support or realize an operation equivalent to high-precision pump dredging by a skilled operator by using various work data during pump dredging work by a skilled operator.

Means for Solving the Problems

[0006] To solve the above problems, the present invention provides a pump dredging device that sucks up seabed sediment excavated by a cutter provided on a ladder through an intake port, The physical quantities indicating the state of at least one of the following: the cutter, the pump communicating with the intake port, the winch driving the rudder, the hull, or the construction state, are power, current, pressure, rotational speed, load, vibration, inclination, flow rate, speed, or construction management values. Time-series sensor data and, By the first operator The ladder Indicates the amount of swing manipulation. An acquisition unit that acquires time-series swing operation amount data, and the acquired chronological The sensor data and the swing operation amount data are correlated on the time axis. , the above Using the data from the data as explanatory variables, The sensor data corresponding to the Training data with swing manipulation amount data as the target variable Generate and the training data This machine learning device is characterized by comprising a machine learning unit that performs machine learning using the machine learning unit, and a storage unit that stores the trained model generated by the machine learning unit.

[0007] The aforementioned sensor data is used by the cutter Indicates the operating status. Cutter data, ladder pump Indicates the operating status. Ladder pump data, ladder winch Indicates the operating status. Ladder winch data, main pump Indicates the operating status. Main pump data, swing winch Indicates the operating status. Swing winch data, The aforementioned hull This indicates the state Hull data, or dredging work This indicates the state Construction Day Ta Even if not, eventually Kao It may include other things.

[0008] In the learning by the aforementioned machine learning unit, the target variable The system may also include a unit for identifying the degree of contribution of the aforementioned sensor data.

[0009] Furthermore, the present invention is The bottom sediment excavated by the cutter on the ladder is sucked in through the intake. In a pump dredging device, in response to the operation of a second operator, The physical quantities indicating the state of at least one of the following: the cutter, the pump communicating with the intake port, the winch driving the rudder, the hull, or the construction state, are power, current, pressure, rotational speed, load, vibration, inclination, flow rate, speed, or construction management values. A sensor data acquisition unit that acquires time-series sensor data, and a machine learning apparatus according to claim 1. generated byAn instruction information output device comprising: an input unit that inputs, as explanatory variables, the sensor data acquired by the sensor data acquisition unit to the learned model; and an instruction information output unit that outputs instruction information corresponding to swing operation amount prediction data obtained as target variables from the learned model in response to the input of the sensor data.

[0010] The instruction information output unit may output, as operation guidance for the ladder, instruction information corresponding to the swing operation amount prediction data to the second operator.

[0011] The instruction information output unit may output, as an operation command for the ladder, instruction information corresponding to the swing operation amount prediction data to a control device that controls the operation of the ladder.

[0012] The sensor data includes the cutter Indicates the operating status. cutter data, ladder pump data indicating the operating state of the ladder pump, ladder winch data indicating the operating state of the ladder winch, main pump Indicates the operating status. main pump data, swing winch data indicating the operating state of the swing winch, The aforementioned hull This indicates the state hull data, or dredging construction This indicates the state construction data Ta at least any Kao and may include any of them.

[0013] The Detect sensor data When there are a plurality of sensors and the detection periods of each sensor are different, an interpolation processing unit may be provided to perform interpolation processing on each sensor data so that the sensor data of the plurality of sensors has a common detection period.

[0014] Further, the present invention relates to a pump dredging device that sucks sediment excavated by a cutter provided on a ladder from a suction port. In the pump dredging device, :] The physical quantities indicating the state of at least one of the following: the cutter, the pump communicating with the intake port, the winch driving the rudder, the hull, or the construction state, are power, current, pressure, rotational speed, load, vibration, inclination, flow rate, speed, or construction management values. time-series sensor data, and By the first operator the Indicates the amount of swing manipulation. time-series swing operation amount data of the ladder are acquired, and the acquired chronologicalAssociate the sensor data and the swing operation amount data on the time axis , the above Using the sensor data as an explanatory variable, The sensor data corresponding to the teacher data with the swing operation amount data as the target variable is Generate and the training data used to perform a machine learning step of machine learning and a storage step of storing the learned model generated by the machine learning. A machine learning method characterized by comprising these steps.

[0015] Further, the present invention relates to new Obtained during the work sensor data of the first operator as an explanatory variable, input it into the learned model generated by the machine learning method according to claim 9, From the aforementioned trained model the predicted swing operation amount data, which is the output target variable, and The sensor data acquired in the new operation corresponds to the swing operation amount data to A machine learning method characterized by re-learning the learned model using only data whose difference amount is within a threshold as teacher data for re-learning.

[0016] Further, the present invention The soil excavated by the cutter on the ladder is sucked in through the intake port. In a pump dredging device, according to the operation of the second operator, The physical quantities indicating the state of at least one of the following: the cutter, the pump communicating with the intake port, the winch driving the rudder, the hull, or the construction state, are power, current, pressure, rotational speed, load, vibration, inclination, flow rate, speed, or construction management values. [[ID=2%]]a sensor data acquisition step of acquiring time-series sensor data, and according to the machine learning method according to claim 9 or 10 generated For the learned model, an input step of inputting the sensor data acquired in the sensor data acquisition step as an explanatory variable, and an output step of outputting information corresponding to the predicted swing operation amount data obtained as the target variable from the learned model in response to the input of the sensor data. An instruction information output method characterized by comprising these steps.

[0017] Further, the present invention may be a program for causing a computer to execute the above method.

Advantages of the Invention

[0018] According to the present invention, it becomes possible to support or realize an operation equivalent to high-precision pump dredging by a skilled operator.

Brief Description of the Drawings

[0019] [Figure 1] A diagram illustrating the bow section of a pump dredging device according to one embodiment of the present invention. [Figure 2] A block diagram showing the main hardware configuration of the dredging control system 10 and peripheral equipment. [Figure 3] A table illustrating sensor data. [Figure 4] A block diagram showing an example of the hardware configuration of the machine learning device 20. [Figure 5] A block diagram showing an example of the hardware configuration of the control device (instruction information output device) 11. [Figure 6] A block diagram showing an example of the functional configuration of the machine learning device 20. [Figure 7] A block diagram showing an example of the functional configuration of the control device (instruction information output device) 11. [Figure 8] A flowchart illustrating the processing procedure for machine learning operations performed by the machine learning device 20. [Figure 9] A flowchart showing the processing procedure for the instruction information output operation by the control device (instruction information output device) 11. [Modes for carrying out the invention]

[0020] An example of an embodiment for carrying out the present invention will be described. [composition] Figure 1 is a schematic diagram illustrating the bow section of a pump dredging device according to one embodiment of the present invention. The upper part (a) of Figure 1 is a side view of the pump dredging device as seen from the side, and the lower part (b) of Figure 1 is a top view of the pump dredging device as seen from above. This pump dredging device is installed, for example, near the bow of a vessel called a pump dredger. At the shaft support 1 installed on the hull of the pump dredger, the base end (left side in the figure) of the rudder 2 is pivotally supported so that its tip end (right side in the figure) swings vertically (in the Z-axis direction). The tip of the rudder 2 is provided with a cutter 3 for excavating the seabed and an intake port (not shown) for sucking up the seabed sediment excavated by the cutter 3. The rudder 2 extends linearly from the base end on the shaft support 1 side, and the tip of the rudder 2 is suspended from the rudder shear 5 by a wire 4. The rudder 2 swings vertically around the pivot point 1, which is pivotally supported on the hull, i.e., in the XZ plane in the figure, as the wire 4 is paid out and retracted by the rudder winch 6. The tip of the rudder 2 is connected to the swing winch 8 by the swing wire 7. The rudder 2 swings horizontally around its base, i.e., in the XY plane in the figure, as the swing wire 7 is paid out and retracted by the swing winches 8,8. During dredging, the seabed is excavated by the cutter 3 located at the tip of the rudder 2, and the excavated soil is sucked in along with seawater from the intake port at the tip of the rudder 2 using a pump (not shown) and pumped to a designated location such as land via a sand discharge pipe (not shown). In Figure 1, the X-axis is defined as an axis parallel to the overall length of the ship, with its positive direction pointing from the stern to the bow; the Y-axis is defined as an axis parallel to the overall width of the ship, with its positive direction pointing from the port side to the starboard side; and the Z-axis is defined as an axis perpendicular to the X and Y axes, with its positive direction pointing from bottom to top (the same applies hereafter).

[0021] In such a pump dredging device, the operator operates swing winches 8,8 that move the ladder 2 horizontally and a ladder winch 6 that moves it vertically, thereby moving the cutter 3, which is attached to the tip of the ladder 2, underwater along the planned line to perform pump dredging.

[0022] Figure 2 shows the main hardware configuration of the dredging control system 10 and its peripheral devices. The dredging control system 10 is a system in which a control device 11, which functions as an instruction information output device of the present invention, a sensor group 12 including various sensors, a user interface device 13, and a construction management device 14 having construction data are networked by a communication line such as Ethernet or optical fiber. Furthermore, a ladder winch 6 and a swing winch 8 are connected to the control device 11 in a communicative manner, and a machine learning device 20 is also connected to the control device 11 in a communicative manner.

[0023] The sensor group 12 includes, for example, main pump sensors such as an encoder, ammeter, and voltmeter installed on the main pump; rudder pump sensors such as an ammeter and voltmeter installed on the rudder pump; cutter motor sensors such as an ammeter and voltmeter installed on the cutter motor; a GNSS (Global Navigation Satellite System) device and gyro sensor installed on the ship's hull; a tide signal receiving device that receives detection signals from tide gauges installed on the quay, etc.; a rudder depth gauge installed at a predetermined position such as the middle section of the rudder 2; vibration and tilt sensors installed on the ship's hull to acquire vibration data (acceleration data) and tilt data; a draft gauge; a carriage stroke, and various other sensors.

[0024] Figure 3 is a table illustrating the sensor data set detected by the sensor group 12. As illustrated in Figure 3, the sensor data set includes at least one of the following data: cutter data related to the cutter, ladder pump data related to the ladder pump, ladder winch data related to the ladder winch, main pump data related to the main pump, swing winch data related to the swing winch, hull data which is vibration and tilt data related to the hull, and construction data related to dredging work acquired from the construction management device 14. The sensor group 12 includes various sensors for detecting the sensor data set illustrated in Figure 3.

[0025] The user interface device 13 includes, for example, an operating unit such as a keyboard, mouse, microphone, switches, and buttons, and a display unit such as a display, speaker, and LED lamp. The user interface device 13 displays instruction information corresponding to the swing operation amount prediction data obtained as the target variable from a trained model, which will be described later. The operator uses the user interface device 13 to perform operations for pump dredging, such as swing winch operation and ladder winch operation, while referring to the instruction information displayed on the user interface device 13.

[0026] Figure 4 shows the hardware configuration of the machine learning device 20. Physically, the machine learning device 20 is configured as a computer device including a processor 2001, memory 2002, storage 2003, communication device 2004, and a bus connecting them. Each of these devices operates on power supplied from a power supply (not shown). The hardware configuration of the machine learning device 20 may include one or more of the devices shown in Figure 4, or it may be configured without some of the devices. Alternatively, some devices may be attached externally to the machine learning device 20.

[0027] Each function in the machine learning device 20 is realized by loading predetermined software (programs) onto hardware such as the processor 2001 and memory 2002, which allows the processor 2001 to perform calculations, control communication by the communication device 2004, acquire data transmitted from other devices, output data calculated based on the acquired data, and control at least one of data reading and writing in the memory 2002 and storage 2003.

[0028] The processor 2001 reads programs (program code), software modules, data, etc., from at least one of the storage 2003 and the communication device 2004 into the memory 2002, and performs various processes according to these.

[0029] Memory 2002 is a computer-readable recording medium and may consist of at least one of the following: ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), RAM (Random Access Memory), etc. Memory 2002 may also be called a register, cache, main memory, etc. Memory 2002 can store executable programs (program code), software modules, etc., for carrying out the method according to this embodiment.

[0030] Storage 2003 is a computer-readable recording medium and may consist of at least one of the following: optical discs such as CD-ROM (Compact Disc ROM), compact discs, digital multipurpose discs, and Blu-ray® discs; magneto-optical discs such as hard disk drives, solid-state drives, and flexible discs; smart cards; flash memory (e.g., cards, sticks, key drives); floppy® discs; and magnetic strips. Storage 2003 may also be called an auxiliary storage device.

[0031] Communication device 2004 is hardware (transceiver / receiver device) for communicating between a computer and other devices via at least one of wired or wireless connections, and is also referred to as a network device, network controller, network card, communication module, etc.

[0032] Figure 5 shows the hardware configuration of the control device 11. Similar to the machine learning device 20, the control device 11 is physically configured as a computer device including a processor 1101, memory 1102, storage 1103, communication device 1104, and a bus connecting these components. Each function in the control device 11 is realized by loading predetermined software (programs) onto hardware such as the processor 1101 and memory 1102, which allows the processor 1101 to perform calculations, control communication by the communication device 1104, and control at least one of data reading and writing in the memory 1102 and storage 1103. The control device 11 may also be implemented as a computer device called a PLC (Programmable Logic Controller).

[0033] Figure 6 shows an example of the functional configuration of the machine learning device 20. In the machine learning device 20, the acquisition unit 201 acquires from the control device 11 a time-series sensor data set (see Figure 3) detected by the sensor group 12 in accordance with the operation of the pump dredging device, and time-series swing operation amount data of the ladder 2 in accordance with the operation of the skilled operator (first operator). Specifically, the swing operation amount data refers to the operation direction in the XY direction, the swing direction, and the swing speed of the ladder 2.

[0034] The machine learning unit 202 associates the sensor data acquired by the acquisition unit 201 with the swing operation amount data on a time axis. Here, it is assumed that a skilled operator, referring to the numerical values ​​of each device displayed on the control panel of the pump dredging device, performs an operation that optimizes construction efficiency, for example, 1 second later. Therefore, the machine learning unit 202 associates the swing operation amount data, which indicates the content of the operation performed by the skilled operator at a predetermined time elapsed (for example, 1 second later) from the detection timing of the sensor data installed on each device, with the sensor data. Then, the machine learning unit 202 generates training data with the sensor data as explanatory variables and the swing operation amount data corresponding to that sensor data (i.e., the swing operation amount data detected at a predetermined time elapsed (for example, 1 second later) from the detection timing of that sensor data) as the objective variable, and generates a trained model by performing so-called supervised machine learning using this training data.

[0035] The memory unit 203 stores the trained model generated by the machine learning unit 202. The output unit 204 outputs data as the target variable to the control device 11.

[0036] As described above, the machine learning device 20 learns the relationship between the sensor data set detected by the sensor group 12 (i.e., the status of each part of the pump dredging device and the pump dredging vessel) and the operation performed by a skilled operator to optimize construction efficiency.

[0037] Figure 7 shows an example of the functional configuration of the control device 11. In the control device 11 of Figure 7, the sensor data acquisition unit 111 acquires a time-series sensor data set detected by the sensor group 12 in response to the operation of an operator (second operator) who is less familiar with the operation than the skilled operator mentioned above.

[0038] The input unit 112 inputs the sensor data acquired by the sensor data acquisition unit 111 via the acquisition unit 201 as explanatory variables to the trained model stored in the memory unit 203 of the machine learning device 20.

[0039] The instruction information output unit 113 receives instruction information corresponding to the swing operation amount prediction data obtained as the target variable from the trained model via the output unit 204 in response to the input of the sensor data group, generates and outputs the instruction information. The swing operation amount prediction data obtained as the target variable from the trained model corresponds to an operation equivalent to high-precision pump dredging by a skilled operator, and can be said to correspond to the optimal operation for an operator unfamiliar with the operation (second operator). The instruction information corresponding to this swing operation amount prediction data is, for example, if the swing operation amount prediction data obtained as the target variable indicates the operating direction and operating speed of the ladder 2, then the instruction information is characters, numbers, arrows, or images instructing the operator to operate the ladder in that direction and operating speed, and these are displayed on the user interface device 13. In other words, the instruction information output unit 113 outputs instruction information corresponding to the swing operation amount prediction data as operation guidance for the ladder 2 to the operator (second operator) during operation. By operating the ladder 2 according to or referring to this instruction information, the operator (second operator) can achieve an operation equivalent to that performed by a skilled operator (first operator).

[0040] [Operation] Next, the operation of this embodiment will be described. First, with reference to Figure 8, the machine learning operation by the machine learning device 20 will be described. In this machine learning operation, a skilled operator (first operator) operates the pump dredging device. The acquisition unit 201 of the machine learning device 20 acquires a time-series sensor data group detected by the sensor group 12 in response to the operation of the skilled operator (first operator), and time-series swing operation amount data of the ladder 2 corresponding to that operation, via the input unit 112 from the sensor data acquisition unit 111 of the control device 11 (step S10).

[0041] Next, the machine learning unit 202 associates the sensor data acquired by the acquisition unit 201 with the swing operation amount data on the time axis, and then generates training data with the sensor data set as explanatory variables and the corresponding swing operation amount data as the target variable (step S11). Then, the machine learning unit 202 generates a trained model by performing supervised machine learning using this training data (step S12). This trained model is stored in the storage unit 203 (step S13).

[0042] Next, with reference to Figure 9, the instruction information output operation by the control device 11 will be explained. In this instruction information output operation, an operator less familiar with the operation than a skilled operator (a second operator) operates the pump dredging device. The sensor data acquisition unit 111 of the control device 11 acquires a time-series sensor data group detected by the sensor group 12 in response to the operation of the operator (the second operator) (step S20).

[0043] Next, the input unit 112 of the control device 11 inputs the sensor data acquired by the sensor data acquisition unit 111 via the acquisition unit 201 as explanatory variables to the trained model stored in the memory unit 203 of the machine learning device 20 (step S21).

[0044] Then, the instruction information output unit 113 obtains swing operation amount prediction data from the trained model as the target variable via the output unit 204 in response to the input of the sensor data group (step S22), generates instruction information corresponding to this swing operation amount prediction data, and outputs (displays) it to the user interface device 13 (step S23). The operator (second operator) operates the ladder 2 according to or by referring to this instruction information.

[0045] According to the embodiments described above, it is possible to support operations equivalent to high-precision pump dredging performed by a skilled operator.

[0046] [Differentiation] The present invention is not limited to the embodiments described above. The embodiments described above may be modified as follows.

[0047] [Example 1] In this embodiment, the machine learning unit 202 generated training data by associating the swing operation amount data detected at a predetermined time elapsed (for example, 1 second) from the detection timing of the sensor data group with the sensor data. This predetermined time corresponds to the time difference between when a skilled operator grasps the state of the pump dredging device and when they perform their own operation, and is a value that can be arbitrarily changed. Furthermore, if there are multiple sensors included in the sensor group 12 and the detection cycles of each sensor are different, the control device 11 may be equipped with an interpolation processing unit that performs interpolation processing between preceding and succeeding sensor data on the time axis so that the sensor data of the multiple sensors have a common detection cycle. In addition, if multiple sensors of the same type are installed at the same location, the average, maximum, or minimum value of the sensor data of the multiple sensors may be used.

[0048] [Differentiation 2] In machine learning, it is possible to calculate the extent to which explanatory variables contribute to the target variable. Therefore, the machine learning device 20 may be equipped with an identification unit that identifies the degree of contribution of each sensor data from the sensor data set used as explanatory variables to the manipulated variable data, and outputs this degree of contribution. This makes it possible to identify meaningful sensor data as explanatory variables and to output a more accurate target variable by reviewing the weighting. Furthermore, by displaying the degree of contribution, operators during operation can use the machine learning's decision-making basis as a reference to make decisions about their operations.

[0049] [Difference 3] In the above embodiment, the instruction information output unit 113 output instruction information corresponding to the swing operation amount prediction data to the operator (second operator) during operation as guidance for operating the ladder 2. However, this instruction information may also be used for automatic control. Specifically, the pump dredging device is equipped with a control device that controls the operation of each part, such as the ladder 2, according to operation commands. Furthermore, the instruction information output unit 113 of the control device 11 outputs instruction information corresponding to the swing operation amount prediction data to the control device as an operation command for the ladder 2. The control device controls the operation of the ladder 2 according to this operation command. By using the results of machine learning for automatic control in this way, it becomes possible to achieve operation equivalent to high-precision pump dredging by a skilled operator.

[0050] [Differentiation Example 4] In machine learning, a trained model can be repeatedly retrained to create a more accurate model. Therefore, the machine learning device 20 may input sensor data corresponding to a new operation by a skilled operator (first operator) as explanatory variables into the trained model, compare the outputted swing manipulation amount prediction data (the target variable) with the swing manipulation amount data from the new operation, and retrain the trained model using only the data where the difference between the swing manipulation amount data from the new operation and the swing manipulation amount prediction data (the target variable) is within a threshold. The reason for using only the data where the difference between the swing manipulation amount data corresponding to the new operation and the swing manipulation amount prediction data (the target variable) is within a threshold is to exclude operation amount data corresponding to outlier values ​​(so-called bug data) that should not be used as training data, as even skilled operators may occasionally perform inefficient operations.

[0051] Furthermore, the present invention can also be implemented as a machine learning method or a method for outputting instruction information. The present invention may also be a program for executing these methods. [Explanation of symbols]

[0052] 1: Axis support, 2: Ladder, 3: Cutter, 4: Wire, 5: Ladder shear, 6: Ladder winch, 7: Swing wire, 8: Swing winch, 10: Dredging control system, 11: Control device, 12: Sensor group, 13: User interface device, 14: Construction management device, 111: Sensor data acquisition unit, 112: Input unit, 113: Instruction information output unit, 1101: Processor, 1102: Memory, 1103: Storage, 1104: Communication device, 20: Machine learning device, 201: Acquisition unit, 202: Machine learning unit, 203: Memory unit, 204: Output unit, 2001: Processor, 2002: Memory, 2003: Storage, 2004: Communication device.

Claims

1. In a pump dredging device that sucks up seabed sediment excavated by a cutter installed on a rudder through an intake port, an acquisition unit acquires time-series sensor data indicating at least one of the following physical quantities representing the state of the cutter, the pump communicating with the intake port, the winch driving the rudder, the hull, or the construction state: power, current, pressure, rotational speed, load, vibration, inclination, flow rate, speed, or construction management value, and time-series swing operation amount data indicating the amount of swing operation of the rudder by a first operator. A machine learning unit that associates the acquired time-series sensor data and the swing operation amount data on the time axis, generates training data with the sensor data as explanatory variables and the swing operation amount data corresponding to the sensor data as the target variable, and performs machine learning using the training data, A storage unit for storing the trained model generated by the machine learning unit and A machine learning device characterized by having the following features.

2. The sensor data includes at least one of the following: cutter data indicating the operating state of the cutter, rudder pump data indicating the operating state of the rudder pump, rudder winch data indicating the operating state of the rudder winch, main pump data indicating the operating state of the main pump, swing winch data indicating the operating state of the swing winch, hull data indicating the state of the hull, or construction data indicating the state of the dredging work. The machine learning device according to claim 1, characterized in that it is a machine learning device.

3. The machine learning unit comprises a unit that identifies the degree of contribution of the sensor data to the target variable during learning. The machine learning apparatus according to claim 2, characterized in that it is as described above.

4. A pump dredging device that sucks up seabed sediment excavated by a cutter provided on a ladder from an intake port, comprising: a sensor data acquisition unit that acquires time-series sensor data indicating at least one of the following physical quantities, such as power, current, pressure, rotational speed, load, vibration, inclination, flow rate, speed, or construction management value, as a physical quantity indicating at least one of the following: the cutter, the pump communicating with the intake port, the winch driving the ladder, the hull, or the construction state, in response to the operation of a second operator; An input unit that inputs the sensor data acquired by the sensor data acquisition unit as explanatory variables to a trained model generated by the machine learning device described in claim 1, An instruction information output unit outputs instruction information corresponding to swing operation amount prediction data obtained as the target variable from the trained model in response to the input of the sensor data. An instruction information output device characterized by comprising:

5. The instruction information output unit outputs instruction information corresponding to the swing operation amount prediction data to the second operator as operation guidance for the ladder. The instruction information output device according to claim 4.

6. The instruction information output unit outputs instruction information corresponding to the swing operation amount prediction data as a ladder operation command to the control device that controls the ladder's movement. The instruction information output device according to claim 4.

7. The sensor data includes at least one of the following: cutter data indicating the operating state of the cutter, rudder pump data indicating the operating state of the rudder pump, rudder winch data indicating the operating state of the rudder winch, main pump data indicating the operating state of the main pump, swing winch data indicating the operating state of the swing winch, hull data indicating the state of the hull, or construction data indicating the state of the dredging work. The instruction information output device according to feature 4.

8. When there are multiple sensors that detect the aforementioned sensor data, and each of them has a different detection period, The system includes an interpolation processing unit that performs interpolation processing on the sensor data of multiple sensors so that the sensor data of each sensor has a common detection period. An instruction information output device according to any one of claims 4 to 7.

9. In a pump dredging device that sucks up soil excavated by a cutter installed on a ladder through an intake port, the acquisition step involves acquiring time-series sensor data indicating at least one of the following physical quantities representing the state of the cutter, the pump communicating with the intake port, the winch driving the ladder, the hull, or the construction state: power, current, pressure, rotational speed, load, vibration, inclination, flow rate, speed, or construction management value, and time-series swing operation amount data indicating the amount of swing operation of the ladder by a first operator. A machine learning step involves associating the acquired time-series sensor data and the swing operation amount data on a time axis, generating training data with the sensor data as explanatory variables and the swing operation amount data corresponding to the sensor data as the target variable, and performing machine learning using the training data. A storage step of storing the trained model generated by the aforementioned machine learning, A machine learning method characterized by comprising the following features.

10. Sensor data acquired in the first operator's new work is input as explanatory variables into the trained model generated by the machine learning method described in claim 9. The trained model is retrained using only data where the difference between the swing operation amount prediction data, which is the target variable output from the trained model, and the swing operation amount data corresponding to the sensor data acquired in the new operation is within a threshold, as training data for retraining. A machine learning method characterized by the following features.

11. A pump dredging device that sucks up soil excavated by a cutter provided on a ladder from an intake port, comprising a sensor data acquisition step of acquiring time-series sensor data indicating at least one of the following physical quantities, such as power, current, pressure, rotational speed, load, vibration, inclination, flow rate, speed, or construction management value, as a physical quantity indicating the state of at least one of the following: the cutter, the pump communicating with the intake port, the winch driving the ladder, the hull, or the construction state, in response to the operation of a second operator, An input step in which the sensor data acquired in the sensor data acquisition step is input as explanatory variables to a trained model generated by the machine learning method according to claim 9 or 10, An output step that outputs information corresponding to the swing operation amount prediction data obtained as the target variable from the trained model in response to the input of the sensor data. A method for outputting instruction information, characterized by comprising:

12. A program for causing a computer to perform the method described in claim 9 or 10.

13. A program for causing a computer to perform the method described in Claim 11.

Images