JP6816783B2 - crane - Google Patents

Description

The present invention relates to a crane capable of controlling the behavior of a load suspended from a hook with high accuracy.

Conventionally, a crane, which is a typical work vehicle, has been known. The crane is mainly composed of a traveling body and a turning body. The traveling body is equipped with a plurality of wheels and is capable of self-propelling. In addition to the boom, the swivel body is equipped with wire ropes and hooks so that it can be carried while the luggage is suspended.

By the way, model prediction control is a control method in which the behavior of a controlled object is predicted by a prediction model that captures dynamic characteristics, and the control input is determined by solving an optimization problem (see Patent Document 1 and Patent Document 2). However, it is difficult to express the dynamic characteristics of a complicated hydraulic system or mechanical system by a mathematical formula, and there is a problem that the accuracy becomes coarse due to the imperfections of the prediction model. Therefore, there has been a demand for a crane that accurately predicts the behavior of a load suspended from a hook and determines the control input of the actuator based on this. In other words, there has been a demand for a crane that can control the actuator with high accuracy and, by extension, the behavior of the load suspended from the hook with high accuracy.

JP-A-2018-41150 JP-A-11-21077

We provide a crane that can control the behavior of luggage suspended from hooks with high accuracy.

The first invention is
With the boom
A wire rope hanging from the boom and
A hook that moves up and down by winding and unwinding the wire rope is provided.
In a crane that carries a load while the load is suspended from the hook,
Various actuators used to carry the luggage and
It includes a prediction model for predicting the behavior of the load and a controller composed of an optimizer that determines the control input of the actuator by solving an optimization problem.
The prediction model in the controller is artificial intelligence using a multi-layer neural network, and when the control data of the actuator is input to the input layer of the neural network, the behavior data of the luggage is obtained from the output layer of the neural network. It is the one to output.

The second invention is the crane according to the first invention.
The optimizer in the controller identifies the control input of the actuator that minimizes the error of the actual path with respect to the target path of the load.

The third invention is the crane according to the first or second invention.
The neural network imitates a brain neural circuit in which neurons are connected by synapses, and the weighting of the circuit corresponding to the synapse is adjusted by using the combination of the control data of the actuator and the behavior data of the luggage as learning data. It is a thing.

The fourth invention is the crane according to any one of the first to third inventions.
Equipped with a display
The controller displays the target route and / or the actual route of the luggage on the display.

The fifth invention is the crane according to any one of the first to fourth inventions.
The controller can acquire the learning data via a communication network and update the prediction model.

The sixth invention is the crane according to any one of the first to fourth inventions.
The controller is capable of acquiring constraints and updating the optimizer via a communication network.

The crane according to the first invention includes a controller composed of a prediction model for predicting the behavior of a load and an optimizer for determining a control input of an actuator by solving an optimization problem. Then, the prediction model in the controller is artificial intelligence using a multi-layer neural network, and when the control data of the actuator is input to the input layer of the neural network, the behavior data of the load is output from the output layer of the neural network. According to such a technical idea, it is possible to accurately predict the behavior of the load suspended from the hook and determine the control input of the actuator based on this. Therefore, the actuator can be controlled with high accuracy, and the behavior of the load suspended from the hook can be controlled with high accuracy.

In the crane according to the second invention, the optimizer in the controller identifies the control input of the actuator that minimizes the error of the actual path with respect to the target path of the load. According to such a technical idea, it is possible to improve the followability of the actual route with respect to the target route of the luggage suspended from the hook.

In the crane according to the third invention, the neural network imitates a brain neural circuit in which neurons are connected by synapses, and the combination of actuator control data and luggage behavior data corresponds to synapses as learning data. The weighting of the circuit is adjusted. According to such a technical idea, it is possible to acquire a function of accurately predicting the behavior of luggage.

In the crane according to the fourth invention, the controller displays the target route and / or the actual route of the luggage on the display. According to such a technical idea, the operator can visually recognize the target route and / or the actual route of the cargo.

In the crane according to the fifth invention, the controller can acquire learning data via a communication network and update the prediction model. According to such a technical idea, it is possible to predict the behavior of luggage by the latest prediction model.

In the crane according to the sixth invention, the controller can acquire constraints and update the optimizer via the communication network. According to such a technical idea, it is possible to determine the control input of the actuator by the latest optimizer.

The figure which shows the crane. The figure which shows the inside of a cabin. The figure which shows the structure of the control system. The figure which shows the structure of a controller. The figure which shows the outline of the model predictive control. The figure which shows the structure of the prediction model. The figure which shows the learning phase of a neural network. The figure which shows the utilization phase of a neural network. The figure which shows the creation method of the learning data. The figure which shows the control mode at the time of carrying a load. The figure which shows the state which displayed the current position of a baggage. The figure which shows the state which displayed the target position of the baggage. The figure which shows the state which displayed the target route of a baggage. The figure which shows the state which displayed the actual route of a baggage.

The technical idea disclosed in the present application can be applied to other cranes in addition to the crane 1 described below.

First, the crane 1 will be described with reference to FIGS. 1 and 2.

The crane 1 is mainly composed of a traveling body 2 and a turning body 3.

The traveling body 2 includes a pair of left and right front wheels 4 and rear wheels 5. Further, the traveling body 2 is provided with an outrigger 6 for stabilizing the load L by grounding it when carrying the load L. In the traveling body 2, the swivel body 3 supported on the upper portion of the traveling body 2 is swiveled by an actuator.

The swivel body 3 is provided with a boom 7 so as to project forward from the rear portion. Therefore, the boom 7 can be swiveled by an actuator (see arrow A). Further, the boom 7 can be expanded and contracted by an actuator (see arrow B). Further, the boom 7 is undulating by an actuator (see arrow C).

In addition, a wire rope 8 is hung on the boom 7. A hook 9 is attached to the wire rope 8 hanging from the tip of the boom 7. A winch 10 is arranged in the vicinity of the base end side of the boom 7. The winch 10 is integrally configured with the actuator, and enables the wire rope 8 to be taken in and out. Therefore, the hook 9 can be raised and lowered by an actuator (see arrow D). The swivel body 3 is provided with a cabin 11 on the side of the boom 7. Inside the cabin 11, a swivel lever 31, a telescopic lever 32, an undulating lever 33, and a winding lever 34 are provided. In addition, a display 42, which will be described later, is provided.

Next, the maneuvering system 16 will be described with reference to FIG.

The control system 16 is mainly composed of a controller 20. Various levers 31 to 34 are connected to the controller 20. Further, various valves 35 to 38 are connected to the controller 20.

As described above, the boom 7 is rotatable by an actuator (see arrow A in FIG. 1). In the present application, such an actuator is defined as a swivel motor 12 (see FIG. 1). The swivel motor 12 is appropriately operated by the swivel valve 35, which is a directional control valve. That is, the swivel motor 12 is appropriately operated by the swivel valve 35 switching the flow direction of the hydraulic oil. The swivel valve 35 is operated based on the instruction of the controller 20. The turning angle and turning speed of the boom 7 are detected by a sensor (not shown).

Further, the boom 7 is expandable and contractible by an actuator (see arrow B in FIG. 1). In the present application, such an actuator is defined as a telescopic cylinder 13 (see FIG. 1). The expansion / contraction cylinder 13 is appropriately operated by the expansion / contraction valve 36 which is a directional control valve. That is, the expansion / contraction cylinder 13 is appropriately operated by the expansion / contraction valve 36 switching the flow direction of the hydraulic oil. The expansion / contraction valve 36 is operated based on the instruction of the controller 20. The expansion / contraction length and expansion / contraction speed of the boom 7 are detected by a sensor (not shown).

Further, the boom 7 is undulating by an actuator (see arrow C in FIG. 1). In the present application, such an actuator is defined as an undulating cylinder 14 (see FIG. 1). The undulating cylinder 14 is appropriately operated by the undulating valve 37, which is a directional control valve. That is, the undulating cylinder 14 is appropriately operated by the undulating valve 37 switching the flow direction of the hydraulic oil. The undulating valve 37 is operated based on the instruction of the controller 20. The undulation angle and undulation speed of the boom 7 are detected by a sensor (not shown).

In addition, the hook 9 can be raised and lowered by an actuator (see arrow D in FIG. 1). In the present application, such an actuator is defined as a winding motor 15 (see FIG. 1). The winding motor 15 is appropriately operated by the winding valve 38, which is a directional control valve. That is, the winding motor 15 is appropriately operated by the winding valve 38 switching the flow direction of the hydraulic oil. The winding valve 38 is operated based on the instruction of the controller 20. The hanging length and lifting speed of the hook 9 are detected by a sensor (not shown).

In addition, the control system 16 has a laser sensor 41 and a display 42. The laser sensor 41 and the display 42 are each connected to the controller 20.

The laser sensor 41 scans the ground surface. The laser sensor 41 is attached to the boom 7 so as to irradiate the ground surface with laser light from directly above or diagonally above (see FIG. 1). Further, the laser sensor 41 receives the reflected light returned from the ground surface and transmits a signal corresponding to the reflected light to the controller 20. Specifically, the laser sensor 41 receives the reflected light returned from the ground surface and transmits a signal corresponding to the time and angle at which the reflected light returns to the controller 20. In this way, the controller 20 can create a three-dimensional map M that reproduces the terrain, buildings, and the like in the virtual space (see FIGS. 11 to 14).

The display 42 displays various images. The display 42 is attached to the front side inside the cabin 11 so that the operator can visually recognize it while operating the various levers 31 to 34 (see FIG. 2). The display 42 is connected to the controller 20 (see FIG. 3). Therefore, the controller 20 can provide information to the operator through the display 42. On the other hand, since the display 42 is a so-called touch panel, it can be said that it is an input device for the operator. Therefore, the operator can also provide information to the controller 20 through the display 42.

Next, the controller 20 will be described with reference to FIGS. 4 to 9.

The controller 20 constitutes an information storage unit 21 by a ROM. The information storage unit 21 stores a program for determining control inputs of actuators (swivel motor 12, telescopic cylinder 13, undulating cylinder 14, winding motor 15).

Further, the controller 20 constitutes the information receiving unit 22 by a CPU, a ROM, a RAM, and the like. The information receiving unit 22 can acquire various information stored in the remote server S via the communication network N. For example, information such as learning data Ds, which will be described later, can be acquired. The acquired information is stored in the information storage unit 21.

Further, the controller 20 constitutes the information processing unit 23 by a CPU, ROM, RAM, and the like. The information processing unit 23 enables so-called model predictive control (MPC: Model Predictive Control) by the predictive model 54 and the optimizer 55. The prediction model 54 predicts the behavior of the luggage L suspended from the hook 9. The optimizer 55 determines the control input of the actuators (swivel motor 12, expansion / contraction cylinder 13, undulation cylinder 14, winding motor 15) by solving the optimization problem.

Here, the model predictive control will be briefly described. It is assumed that the target route Rt of the luggage L has already been given (see FIG. 5). Further, u (t) represents a control input and y (t) represents a control output.

The controller 20 calculates the reference route Rr that gradually approaches the target route Rt based on the behavior of the luggage L at the current time k. Further, the controller 20 predicts the behavior from the current time k to the time k + Hp when the predetermined time Hp has elapsed (calculates the prediction path Rp). The predetermined time Hp at this time is called "predicted horizon", and the actuator (swivel motor 12, telescopic cylinder 13, undulating cylinder 14, winding) so that the reference path Rr and the predicted path Rp intersect at time k + Hp. The control input of the motor 15) is determined. Then, the controller 20 determines that the control input is performed between the current time k and the predetermined time Hu called "control horizon". After that, the controller 20 repeats the same routine with the time k + Hu as the current time k + Hu. The prediction interval T in the present application is synonymous with "prediction horizon".

As a feature of the crane 1, the prediction model 54 in the controller 20 is artificial intelligence using a multi-layer neural network 60 (see FIG. 6). The "multi-layer neural network 60" refers to a neural network having three or more layers including an input layer 61, an output layer 62, and a plurality of hidden layers 63. In addition, "artificial intelligence" can be defined as "artificially created human-like intelligence." Such intelligence is realized by "a controller that enables expression learning", and more specifically, "a controller that can generate and model features from data".

The neural network 60 imitates a brain neural circuit in which neurons are connected at synapses. The neural network 60 is represented programmatically so as to connect a node 64 corresponding to a neuron with a circuit 65 corresponding to a synapse and transmit a signal from a node 64 in one layer to a node 64 in the next layer. The neural network 60 is weighted to represent the coupling strength for each circuit 65, and transmits by multiplying the given signal value by the weight W. Then, each time each node 64 exceeds the threshold value, a signal is transmitted to the next node 64.

The neural network 60 acquires a function of predicting the behavior of the load L suspended from the hook 9 in the learning phase (see FIG. 7). In the learning phase, the weighting for each circuit 65 is adjusted so as to derive the correct output data for the input data. The learning method for creating a neural network is generally called "machine learning". However, the learning method for creating this neural network 60 is "deep learning" established as an aspect of machine learning. Deep learning differs from machine learning in that features are generated and modeled by themselves without human intervention.

The neural network 60 created in this way predicts the behavior of the load L suspended from the hook 9 in the utilization phase (see FIG. 8). The “behavior of the cargo L” also includes a runout caused by the inertia of the load L and a runout caused by the deflection of the boom 7 (see arrow E in FIG. 1). Since the dynamic characteristics are not expressed by mathematical formulas, they are highly robust against disturbances. Further, since many of the materials of building structures have a shape determined by a standard, accurate behavior can be predicted if an appropriate learning phase using appropriate learning data Ds is performed. However, if a separate program is configured in the information processing unit 23 and it is determined that accurate behavior cannot be predicted, this may be canceled. In this case, it is conceivable to display the cancellation on the display 42.

By the way, in order to perform deep learning, learning data Ds is required. The learning data Ds in the present application is a combination of control data of actuators (swivel motor 12, expansion / contraction cylinder 13, undulation cylinder 14, winding motor 15) and behavior data of luggage L. In other words, it is a combination of data summarizing how the various levers 31 to 34 were operated and data summarizing how the luggage L moved accordingly (see FIG. 9). In the learning phase, thousands to tens of thousands of learning data Ds are used, and the value of the weight W is corrected until the behavior data of the luggage L can be predicted from the operation data of the various levers 31 to 34. In other words, the prediction and answer matching are repeated intently to find the value of the weight W that minimizes the restoration error. Such a method is called an error backpropagation method.

In addition, in the present crane 1, the learning data Ds can be created by the crane 1 lifting and transporting the luggage L. Therefore, the crane 1 can update the neural network 60 by using the learning data Ds obtained by itself. As a result, the crane 1 can always predict the behavior of the load L by the latest prediction model 54. In addition, it is possible to realize a forecast that incorporates deterioration over time. Further, the crane 1 can also acquire the learning data Ds via the communication network N and update the neural network 60. This also makes it possible for the crane 1 to always predict the behavior of the load L by the latest prediction model 54. In addition, it is possible to realize prediction based on abundant learning data Ds. However, it may not have such an update function.

On the other hand, the optimizer 55 in the controller 20 has an actuator (swivel motor 12, telescopic cylinder 13, undulation) that minimizes the error of the actual path Rr with respect to the target path Rt of the luggage L suspended from the hook 9. The control input of the cylinder 14 and the winding motor 15) is specified. At this time, the optimizer 55 expresses the degree of achievement by the cost function C so as to satisfy the constraint condition imposed on the behavior of the luggage L, and calculates so that the error is minimized. Then, the control input of the actuator (12, 13, 14, 15) is specified based on the calculation result. An example (Equation 1) of the cost function C related to the turning operation is shown below. "B" represents a barrier function. “Θ” is the turning angle, “Δpos” is the distance of load swing, and “u” is the amount of operation of the turning lever 31. Therefore, each term of the cost function C represents the distance to the target turning angle, the load swing distance, the maximum load swing limit, the maximum input limit by the swing lever 31, and the change in the input speed by the swing lever 31. ..

In addition, in the crane 1, the operator can set the constraint conditions. Therefore, the crane 1 can update the optimizer 55 so as to satisfy the constraint condition set by the operator. As a result, the crane 1 can always determine the control input of the actuators (swivel motor 12, telescopic cylinder 13, undulating cylinder 14, winding motor 15) by the latest optimizer 55. It is also possible to reflect the demands of the pilot. Further, the crane 1 can update the optimizer 55 by acquiring the constraint condition via the communication network N. This also makes it possible for the crane 1 to always determine the control input of the actuator (12, 13, 14, 15) by the latest optimizer 55. It is also possible to reflect the demands of others. However, it may not have such an update function.

Next, the control mode when the luggage L is carried will be described with reference to FIGS. 10 to 14. However, the control mode described below is an embodiment realized by the crane 1, and is not limited to this.

In step S1, the controller 20 calculates the current position Pc of the luggage L. That is, the controller 20 calculates the three-dimensional coordinates of the luggage L from the position of the traveling body 2 and the posture of the boom 7. At this time, the three-dimensional coordinates representing the current position Pc may be displayed on the three-dimensional map M of the display 42 (see FIG. 11: (◯, Δ, □)).

In step S2, the controller 20 grasps the target position Pt of the luggage L. That is, the controller 20 grasps the three-dimensional coordinates input as the target position Pt of the luggage L by the operator or another person. At this time, the three-dimensional coordinates representing the target position Pt may be displayed on the three-dimensional map M of the display 42 (see FIG. 12: (●, ▲, ■)).

In step S3, the controller 20 creates the target path Rt of the luggage L. That is, the controller 20 creates the target route Rt based on the current position Pc and the target position Pt of the luggage L, the shape of the luggage L, and the three-dimensional map M of the work site. At this time, the controller 20 may display the target route Rt on the three-dimensional map M so that the operator who sees the display 42 can also see the target route Rt (see FIG. 13). It is preferable to be able to select what is prioritized to create the target route Rt.

In step S4, the controller 20 moves the hook 9 to carry the luggage L. That is, the controller 20 operates the actuators (swivel motor 12, telescopic cylinder 13, undulating cylinder 14, winding motor 15) to move the hook 9, thereby transporting the luggage L. At this time, the controller 20 may display the actual path Rr on the three-dimensional map M so that the operator who sees the display 42 can also see the actual path Rr (see FIG. 14). Further, the controller 20 may display a comment or the like indicating that the transportation has been completed so that the operator who sees the display 42 can also know that the luggage L has reached the target position Pt.

Next, the technical ideas applied to this crane 1 and their effects will be summarized.

In this crane 1, the control input of the actuators (swivel motor 12, telescopic cylinder 13, undulating cylinder 14, winding motor 15) by solving the prediction model 54 that predicts the behavior of the load L and the optimization problem. The controller 20 is composed of an optimizer 55 for determining the above. The prediction model 54 in the controller 20 is artificial intelligence using a multi-layer neural network 60, and when the control data of the actuator (12, 13, 14, 15) is input to the input layer 61 of the neural network 60, the neural network 60 is used. The behavior data of the luggage L is output from the output layer 62 of the network 60. According to such a technical idea, it is possible to accurately predict the behavior of the luggage L suspended from the hook 9 and determine the control input of the actuator (12, 13, 14, 15) based on this. Therefore, the actuators (12, 13, 14, 15) can be controlled with high accuracy, and the behavior of the load L suspended from the hook 9 can be controlled with high accuracy.

Further, in the crane 1, the optimizer 55 in the controller 20 has an actuator (swivel motor 12, telescopic cylinder 13, undulating cylinder 14) that minimizes the error of the actual path Rr with respect to the target path Rt of the luggage L. -Specify the control input of the winding motor 15). According to such a technical idea, it is possible to improve the followability of the actual path Rr with respect to the target path Rt of the luggage L suspended from the hook 9.

Further, in the present crane 1, the neural network 60 imitates a brain neural circuit in which neurons are connected by synapses, and actuators (swivel motor 12, telescopic cylinder 13, undulating cylinder 14, winding motor). The weighting of the circuit 65 corresponding to the synapse is adjusted by using the combination of the control data of 15) and the behavior data of the luggage L as learning data. According to such a technical idea, it is possible to acquire a function of accurately predicting the behavior of the luggage L.

Further, in the present crane 1, the controller 20 displays the target path Rt and / or the actual path Rr of the luggage L on the display 42. According to such a technical idea, the operator can visually recognize the target route Rt and / or the actual route Rr of the luggage L.

Further, in the crane 1, the controller 20 can acquire the learning data Ds via the communication network N and update the prediction model 54. According to such a technical idea, the behavior of the luggage L can be predicted by the latest prediction model 54.

Further, in the present crane 1, the controller 20 can acquire the constraint condition via the communication network N and update the optimizer 55. According to this technical idea, the latest optimizer 55 makes it possible to determine the control input of the actuator (swivel motor 12, telescopic cylinder 13, undulating cylinder 14, winding motor 15).

Finally, in the present application, the subject of the invention is "a crane capable of controlling the behavior of a load suspended on a hook with high accuracy", but "a crane capable of controlling the behavior of a load suspended on a hook with high accuracy". It can also be regarded as a "control system for cranes". Therefore, the technical idea extends to this as well.

1 Crane 2 Traveling body 3 Swivel body 7 Boom 8 Wire rope 9 Hook 10 winch 12 Swivel motor (actuator)
13 Telescopic cylinder (actuator)
14 Cylinder for undulation (actuator)
15 Winding motor (actuator)
16 Steering system 20 Controller 21 Information storage 22 Information receiving 23 Information processing 41 Laser sensor 42 Display 54 Prediction model 55 Optimizer 60 Neural network 61 Input layer 62 Output layer 63 Hidden layer 64 Node 65 Circuit C Cost function Ds Learning Data for L Luggage N Communication network Rt Target route Rr Actual route W Weight

Claims (5)

With the boom
A wire rope hanging from the boom and
A hook that moves up and down by winding and unwinding the wire rope is provided.
In a crane that carries a load while the load is suspended from the hook,
Various actuators used to carry the luggage and
It includes a prediction model for predicting the behavior of the load and a controller composed of an optimizer that determines the control input of the actuator by solving an optimization problem.
The prediction model in the controller is artificial intelligence using a multi-layer neural network, and when the control data of the actuator is input to the input layer of the neural network, the behavior data of the luggage is output from the output layer of the neural network. To be done,
The neural network imitates a brain neural circuit in which neurons are connected at synapses, and a combination of control data of the actuator obtained by lifting and carrying a load and behavior data of the load is used as learning data. Adjust the weighting of the circuit corresponding to the synapse ,
When performing the swing operation of the boom, the optimizer calculates a cost function including the swing angle of the boom, the distance of the load swing of the load, and the operation amount of the swing lever of the boom as variables. The crane is characterized in that the control input of the actuator is specified based on the result of the above . The crane according to claim 1, wherein the optimizer in the controller identifies a control input of the actuator that minimizes an error in the actual path with respect to the target path of the load. Equipped with a display
The crane according to claim 1 or 2, wherein the controller displays a target route and / or an actual route of the luggage on the display. The crane according to any one of claims 1 to 3, wherein the controller can acquire the learning data via a communication network and update the prediction model. The crane according to any one of claims 1 to 3, wherein the controller can acquire a constraint condition via a communication network and update the optimizer.

Images