JP3400062B2 - Plant control device and tunnel ventilation control device - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a plant control device, and more particularly to an image showing a state of a plant and an operation amount applied to the plant by an operator when the image is observed on the plant. A plant control device for accumulating a pair of with and further recording the time change of this image pattern and the pair of manipulated variables to obtain the control knowledge that the operator does not have a rule and performing control according to this knowledge. Is. The present invention also relates to a tunnel ventilation control device as a specific example.

[0002]

2. Description of the Related Art When controlling tunnel ventilation, blast furnace temperature, sewage sludge, steel plate rolling shape, traffic signals, etc., the data representing the conditions at many observation points of the plant is measured online, and the plant changes every moment. It is important to evaluate the condition of and to properly determine the operation amount. However, since the state change of these plants depends on a non-linear factor, it is difficult to obtain an appropriate manipulated variable by using a technique such as conventional sequence control.

As a solution to such a problem, a method based on knowledge engineering has been conventionally performed (for example, "intelligent technology of process control system", measurement and control, vol.3).
1, no.7, etc.). According to this method, the neural network learns the correspondence between the data and the state of the plant, describes the relation between the state of the plant and an appropriate manipulated variable in a rule format based on the empirical knowledge of the operator, and uses fuzzy logic. The amount of operation can be determined by inference. As a result, it is possible to realize a plant control device that can take in empirical control knowledge possessed by a skilled operator and perform advanced operation monitoring in consideration of nonlinear plant state changes.

[0004]

However, the above-mentioned conventional plant control device has the following problems. First, tunnel ventilation, blast furnace temperature,
In the control of sewage sludge, steel plate rolling shape, traffic signals, etc., the operator looks at the image showing the state at many observation points of the plant, evaluates the ever-changing plant state, and determines the operation amount. However, this image shows a large amount of information with a spatial spread, such as the distribution of soot concentration in the tunnel, the temperature distribution in the furnace of the blast furnace, the shape of sludge, the shape of the steel plate, and the distribution of vehicle volume. Since it includes the information, it is very difficult to clarify which information included in these images the operator pays attention to when determining the operation amount. Therefore, it has been difficult to apply the control technology by the conventional expert system or the like, in which the controlled variable must be clearly defined, to these plant control devices.

Secondly, when describing the empirical control knowledge possessed by the operator in a rule format, it is necessary to clarify what logical judgment the operator makes. In general, the knowledge possessed by the operator is often merely a piece of knowledge having no logical structure, and it has been difficult to extract the logical structure of knowledge from the operator, such as the order of application of rules. Furthermore, thirdly, the state transition of the plant often has dynamics, and in this case, the state of the plant cannot be identified only from the instantaneous image captured at a certain time point, and an appropriate manipulated variable is obtained. It was difficult. Fourthly, there was no function of organizing the logical structure of control knowledge obtained from the operator, finding an efficient control path, and performing optimum control.

The present invention has been proposed in order to solve the above-mentioned problems of the prior art. The first object is to determine the operation amount by the operator focusing on which information included in the image. It is an object of the present invention to provide a plant control device capable of automatically monitoring the operation of a plant based on the past operation record by a skilled operator in the plant control in which it is unclear whether or not the operation is being performed.

A second object of the present invention is to acquire the control knowledge of the operator without performing the complicated work of extracting the logical structure of the control knowledge of the operator such as determining the order of application of rules. It is to provide a plant control device that can perform.

A third object of the present invention is to obtain a temporal change in the correspondence relationship between the monitoring image and an appropriate operation amount,
Even if the state transition of the plant has dynamics,
An object is to provide a plant control device that can determine an appropriate manipulated variable.

A fourth object of the present invention is to provide a plant control device which generates the best control path from the control knowledge obtained from the operator and controls according to the best control path.

A fifth object of the present invention is to provide a tunnel ventilation control device which can acquire control knowledge by a skilled operator and can perform tunnel ventilation control automatically and with high accuracy. .

[0011]

A plant control apparatus according to claim 1 is for achieving the above first to fourth objects, and image input means for taking in a monitoring image showing the state of the plant, State database means for storing the correspondence between the monitoring image itself obtained through the image input means and the state of the plant as a pair, the monitoring image itself showing the state of the plant, and the state database means The correspondence between the state of the plant and the monitoring image stored in is created, a hypothesis of the current state of the plant is created, and state estimation means for calculating the certainty factor and state transition of the plant and its transition A control path database means for storing a pair with an operation amount to cause, a target state of the plant, and a hypothesis of the current state of the plant created by the state estimating means, Control path determination means for taking in a pair of the state transition of the plant stored in the control route database and the operation amount causing the transition, and calculating a route set for transition from the current state hypothesis of the plant to the target state, The set of route candidates from the current state hypothesis of the plant calculated by the control route determination means to the target state and the certainty factor of the hypothesis of the current state of the plant calculated by the state estimation means are taken in, And a control means for calculating an operation amount.

Further, a plant control apparatus according to a second aspect is for achieving the above first to fourth objects,
The plant control device according to claim 1, further presenting a monitoring image showing a state of the plant and a correspondence relationship between the state of the plant and the monitoring image stored in the state database means to an operator of the plant, Interface means for taking in the present state of the plant and an appropriate operation amount designated by a member, a monitoring image showing the state of the plant, the present state of the plant obtained through the interface means and the appropriate operation A pair with an amount is taken in, and a state transition represented by a pair of a state change sequence and an operation amount represented by a pair of a previous state and a current state is stored in the control path database means, and a monitoring image is also stored. And case collection means for storing a pair of current status in the status database means.

[0013]

[0014]

Further, a tunnel ventilation control device according to a third aspect is for achieving the fifth object, and the control device according to the second aspect is applied to the tunnel ventilation control. A monitoring camera that captures a monitoring image indicating the state of, a soot state example database that stores the correspondence between the monitoring image obtained by the monitoring camera and the state of the plant, a monitoring image that indicates the state of the plant, and the soot A neural net that takes in the correspondence between the plant state and the monitoring image stored in the state case database, creates a hypothesis of the current state of the plant, and calculates the certainty factor thereof, and the state transition of the plant and its A tunnel state transition record database that stores a pair with an operation amount that causes a transition, and a target state of the plant,
The hypothesis of the current state of the plant created by the neural network and the pair of the state transition of the plant stored in the tunnel state transition record database and the operation amount that causes the transition are taken in, and the current state of the plant is acquired. A control device for calculating a set of paths that transit from a state hypothesis to a target state, the control device including a set of route candidates from the current state hypothesis of the plant to the target state, and the plant calculated by the neural net. The operation amount of the plant is calculated on the basis of the certainty factor of the current state hypothesis.

[0016]

In the plant control apparatus according to the first aspect, the monitoring image captured by the image inputting unit is input to the state estimating unit, and the state estimating unit causes the case having an image close to the monitoring image to be in the state. Retrieved from database means. Next, the state estimating means calculates the certainty factor by using the state of the retrieved case as a hypothesis indicating the current state of the plant. Then, the control route determination means updates all the route sets from the state hypothesis obtained from the state estimation means to the target state of the plant. Next, the control unit takes in the certainty factors calculated by the state estimation unit for all state hypotheses, and determines the manipulated variable U such that the higher the certainty factor, the higher the probability of transition to the next state. , The plant is controlled by adding this manipulated variable to the plant. Therefore, according to the plant control device of the first aspect, the plant control can be automatically performed based on the past operation record.

Further, in the plant control apparatus according to the second aspect of the present invention, the interface means reads, from the state database means, a state list which is a candidate of the state of the plant when the monitoring image is observed, and together with the monitoring image. As a result of being presented to the operator and judged by the operator based on the monitoring image and the state list, the state input by the operator and an appropriate operation amount are read, and a predetermined operation amount is added to the plant. Furthermore, the state-manipulation amount pair input by the operator is generated and sent to the case collection means.
Next, the case collection means reads the monitoring image and the state-operation amount pair, generates a monitoring image-state pair, registers it in the state database means, and also pairs the past state and the current state. And a manipulated variable are generated and registered in the control path database means. Next, the variable indicating the past state is updated with the value of the current state. In addition to these operations, the same operation as the plant control device according to claim 1 is further performed.

Therefore, according to the plant control apparatus of the second aspect, the control knowledge of the skilled operator can be acquired as a new case, and based on this, automatically and more Highly accurate plant control can be performed.

Further, the tunnel ventilation control device according to claim 3 shows a case where the plant control device according to claim 2 is applied to the tunnel ventilation control, and similarly, the operator accumulated the experience. By acquiring control knowledge that is difficult to rule, it is possible to perform sophisticated automatic control. Further, it becomes possible to generate and execute an optimal control method that the operator did not have the awareness based on the acquired control knowledge.

[0020]

Embodiments of the present invention will be described below in detail with reference to the drawings. The present invention provides a plant control device for extracting and learning control knowledge of a trained operator who cannot be verbalized in plant operation control, and automatically monitoring and controlling the plant. Therefore, the present invention is, for example, an empirical technique that is difficult for a trained operator to verbalize and explain, such as tunnel ventilation control, blast furnace temperature control, sewage sludge control, steel plate rolling shape control, traffic signal area control, etc. It can be used for automatic control of various plants controlled by knowledge.

(First Embodiment) The plant control system of this embodiment is constructed as shown in FIG. That is, the image input means 1 for taking in the monitor image observed on the plant 9, the state database means 5 for storing the correspondence between the state of the plant 9 and the monitor image, the monitor image observed on the plant 9, and the state. The correspondence between the state of the plant 9 and the monitoring image stored in the database means 5 is fetched,
State estimation means 2 for creating a hypothesis of the current state of the plant 9 and calculating the certainty factor thereof, control path database means 4 for storing a pair of state transition of the plant 9 and an operation amount causing the transition. , A target state of the plant 9, a hypothesis of the current state of the plant 9 created by the state estimating means 2, a state transition of the plant 9 stored in the control path database means 4 and an operation amount causing the transition. , And the control path determining means 3 for calculating a set of paths for transition from the current state hypothesis of the plant 9 to the target state, and the current state of the plant 9 calculated by the control path determining means 3. A set of paths from the hypothesis to the target state and the plant 9 calculated by the state estimating means 2
The certainty factor of the current state hypothesis of
And a control means 8 for calculating the manipulated variable of.

The image input means 1 varies depending on the type of plant, and monitoring images for monitoring the image of the soot concentration in the tunnel from a monitoring camera and for measuring the temperature in the furnace of the blast furnace are used. A pattern image in which outputs from multiple temperature sensors (thermocouples) installed in various places inside the furnace are arranged in a two-dimensional plane, images of cameras installed in a sedimentation site that captures the shape of sewage sludge, and steel plate locations There is a pattern image in which the outputs of the sensors for measuring the thickness of the vehicle are arranged in a two-dimensional plane, or an image showing the traffic situation of the vehicle volume taken from surveillance cameras installed at various places on the road.

In the plant control device of this embodiment having such a configuration, the plant control can be performed as described below. That is, as shown in FIGS. 1 and 2, the monitoring image X, which is the controlled variable referred to by the operator for controlling the plant to be controlled, is controlled by the image input means 1 from the plant 9 to the plant control unit. It is taken into the device and further inputted to the state estimating means 2.

The state estimating means 2 shows the correspondence between the monitored image X _i which is the controlled variable observed on the plant 9 in the past and the state S _i of the plant 9 when it is observed. A case (X _k , S _k ) including an image close to the monitored image X, which is the current controlled variable input from the state database means 5 that stores the case (X _i , S _i ). Search for 1 or 2 or more. next,
The state estimation unit 2 sets the retrieved state S _{k of} the case as a hypothesis (state hypothesis) indicating the current state of the plant 9, and the state hypotheses S _k of all the retrieved cases have the current state of the plant. A certainty factor b (S _k ) representing the degree of certainty that is the hypothesis S _k is calculated.

Subsequently, the control path determining means 3 sends all paths p from the state hypothesis S _k obtained from the state estimating means 2 to the target state T of the plant 9 given from the outside.
The route set P, which is the set of (S _k , T), is updated. That is, first, for all the state hypotheses S _k indicating the current state of the plant 9, on each of the routes that are the elements of the route set P, the next state of the state hypothesis one time before (that is, the present state). (State of) is not S _k
Excluded from. When the route set P becomes empty by this,
That is, when there is no route to the target state T of the plant 9, the control route database means 4 is searched for all the current state hypotheses S _k , and the route from the current state hypotheses S _k to the target state T is searched. Are obtained, and these routes are newly set as elements of the route set P. As a result, the route set P
Will be updated.

Here, regarding the update of the route set P,
This will be described in more detail. That is, the control path database means 4 is means for storing a pair of a state transition of the plant 9 and an operation amount causing the transition, and for example, the state S _i of the plant 9 and the plant 9 are in the state S _i . Sometimes, a pair {(S _i , S _j ), which is a pair of the next state S _j that transits when the operation amount U _ij is added and the operation amount U _ij at that time,
The state transition data represented by U _ij } is stored. Further, the path from the state S _i to the state S _j means that the operation amount U _{i (i + 1)} added when the plant 9 is in the state S _i transits to the state S _{i + 1} and Manipulated variable U
_{(i + 1) (i + 2)} causes a transition to the state S _{i + 2} , and a series of operations applied in the same manner causes a gradual transition of the states. Finally, when the plant 9 is in the state S _{j- 1.} An order sequence of transitions when the plant 9 is in the state S _j by the manipulated variable U _{(j-1) j} added to:

[Equation 1] Is represented.

Next, the control route determining means 3 determines the route set P.
The next state of p is updated for all the routes p included in.
That is, the route p, which is an element of the route set P, has an order sequence of transitions in a part thereof:

[Equation 2] , But the next state of p is then S _j . Note that S _k is the current state hypothesis.

Next, the control means 8 controls the state estimating means 2
Is the confidence factor b calculated for all state hypotheses S _k
(S _k ) is taken into consideration, and the operation amount U such that the higher the certainty factor is, the higher the probability of transition to the next state is, the certainty factor b is set.
(S _k ) and the manipulated _variable U _kj for making a transition from the current state hypothesis S _k to the next state S _j . Then, the control means 8 performs plant control by adding the manipulated variable U determined as described above to the plant 9. In this way, the monitoring image X, which is a new controlled variable, is sequentially read from the plant 9 by the image input means 1 and input to the state estimation means 2, and the above processing is repeated.

As described above, according to the present embodiment, it is possible to provide the plant control device capable of automatically controlling the operation based on the past operation record by the operator.

(Second Embodiment) The plant control apparatus of the present embodiment is the same as the plant control apparatus of the first embodiment,
An interface means having the functions described below and a case collection means are added. The same means as those in the first embodiment are designated by the same reference numerals and the description thereof will be omitted.

That is, as shown in FIG. 3, the monitoring image observed on the plant 9 is fetched by the image input means 1 and the state of the plant 9 stored in the state database means 5 is also stored. Interface means 7 is provided for presenting the corresponding relationship of the monitoring images to the operator, and for taking in the current state of the plant 9 designated by the operator and an appropriate operation amount.

Further, the monitoring image observed on the plant 9 and the pair of the present state of the plant 9 obtained through the interface means 7 and an appropriate manipulated variable are taken in, and the immediately preceding state and the present state are taken. The state transition represented by the pair of the state change order and the manipulated variable represented by the group is stored in the control path database means 4, and the pair of the monitoring image and the current state is stored in the state database means 5. A case collection means 6 for storing is provided. The other structure is the same as that of the first embodiment.

In the plant control apparatus of this embodiment having such a configuration, the operator's control knowledge can be acquired as described below, and the plant control is performed based on the acquired control knowledge. be able to. That is,
As shown in FIGS. 3 and 4, a monitoring image X, which is a controlled variable referred to by an operator in order to control the target plant 9, is changed from the plant 9 to the plant control device by the image input unit 1. And is input to the interface means 7 and the case collection means 6.

The interface means 7 reads from the state database means 5 the state list S _i which is a candidate for the state of the plant 9 when the monitoring image X is observed, and the operator along with the monitored image X which is the controlled quantity. To present. Next, as a result of the operator's judgment based on the monitoring image X and the status list S _i , the status S _c input by the operator
And loads the appropriate operation amount U _c, adding the operation amount U _c to the plant 9. In addition, the state S input by the operator
A state-manipulation amount pair (S _c , U _c ) indicating the correspondence between _c and the manipulated variable U _c is generated and sent to the case collection means 6.

Next, the case collection means 6 displays the monitored image X which is the controlled quantity observed in the plant 9 and the state-manipulation quantity pair (S _c , U _c ) determined by the interface means 7. The monitoring image-state pair (X _c , S _c ) indicating the correspondence between the reading image and the state of the plant is generated and registered in the state database means 5. Further, one unit time versus the previous state S _b the current state S _c (S _b,
S _c ) and the manipulated variable U _bc one hour before {(S _b ,
S _c ), U _bc } and generates the control path database means 4
Register with. Next, the variable S _b indicating the state one time before is updated with the value of the variable S _c indicating the current state, and the variable U _bc indicating the operation amount one time before is changed to the value of the variable U indicating the current operation amount. Update with. The above processing is repeated until the monitored image, which is the controlled variable observed on the plant 9, runs out.

In this way, the monitoring image of the plant 9 and
The correspondence with the current state of the plant 9, which is determined by the operator by looking at this monitoring image, is collected in the state database means 5. At the same time, the control path database means 4 shows a correspondence relationship between the pair of the state of the plant 9 one hour before and the current state of the plant 9 determined by the operator, and the operation amount for the plant 9 one hour before. Collected.

In other words, according to this embodiment, the plant 9
The knowledge of the operator who identifies the state of the plant 9 when the specific monitoring image is obtained above and the specific state taken by the plant 9 one time ago are changed to another state closer to the target state. Therefore, the knowledge of the operator who determines the operation amount of the plant 9 to be added can be acquired. The operation at the time of plant control described in the first embodiment is similarly performed in this embodiment. Therefore, in plant control, plant control can be automatically performed based on the acquired control knowledge of the skilled operator.

(Third Embodiment) In this embodiment, the plant control device having the structure shown in the second embodiment is used for ventilation control in a tunnel of a highway. Generally, in tunnel ventilation, maintenance and management in pursuit of economic efficiency is a problem. In addition, the purpose of ventilation control is to provide a safe driving environment, and more specifically, to reduce the visibility due to soot in the tunnel and the air pollution concentration due to harmful substances such as carbon monoxide. The cost is within the allowable range.

In recent years, the mainstream of such a tunnel ventilation system is a vertical flow system for ventilating the inside of the tunnel in the longitudinal direction. Further, the wind speed in the tunnel fluctuates depending on natural conditions such as the amount of vehicles and seasonal winds, so the diffusion process of soot is affected by disturbance. Therefore, for the control of tunnel ventilation, a sequence control such as “when the carbon monoxide concentration exceeds a predetermined threshold value, increase the number of blowers operating and the number of revolutions” has been conventionally used. However, since such a control method is wasteful, it has been required to automate economical control by a machine, which an experienced operator uses empirical knowledge.

FIG. 5 is a longitudinal sectional view of a tunnel to which the plant control device of the present invention is applied. Inside this tunnel, as shown in FIG.
Group (HB1, HB2), 2 dust collectors (QC1, QC
2), two vertical fans (VB1, VB2) are installed. The horizontal blower is installed to increase the wind speed in the tunnel, and the dust collector is installed to purify soot and smoke in the tunnel. In addition, these cross blowers and dust collectors are installed on the ceiling of the tunnel. On the other hand, the vertical blower is installed in the road surface in order to diffuse pollutants that easily fall and settle to the upper part of the tunnel in order to secure the visibility of the lower part of the tunnel and keep the carbon monoxide concentration below a certain level. ing. In addition to these ventilation equipment, in order to monitor the soot and smoke condition in the tunnel, a total of 6 monitoring cameras (C1L to C3L, C1U to C) are provided at each of the three locations within the tunnel, one at the bottom and one at the top.
3U) is installed.

When performing ventilation control of a tunnel equipped with such ventilation equipment, the operator determines the soot concentration at each location, the direction of smoke flow, the flow velocity, etc. from the images captured by the surveillance camera. After grasping the current situation of the entire tunnel and predicting changes in the situation in the near future,
Ventilation in the tunnel is controlled by properly instructing the increase / decrease in the number of revolutions of the blower and the start / stop / revolution of the number of revolutions of the dust collector.
At this time, skilled operators use not only the local conditions inside the tunnel that can be monitored by surveillance cameras, but also the knowledge about the ventilation characteristics of the entire tunnel obtained from the operating experience of various ventilation equipment, for more economical control. It can be performed.
For example, “At a certain point, if the density of soot at the point visible from the camera is high at all 6 places where surveillance cameras are installed, but if the concentration immediately before that was relatively low, it is a transient phenomenon. Therefore, it is not necessary to increase the rotation speed of the blower. " Once such knowledge gained from experience can be verbalized as in the above example, if
-Then It can be expressed in the form of rules, and an expert system technology can be used to create an automatic control device.

However, such knowledge is rarely held in a logically organized form, and it is very difficult to extract knowledge from the operator (knowledge acquisition). Therefore, it is difficult to automate the tunnel ventilation control by applying the conventional technology such as the expert system.

On the other hand, the present invention can automatically acquire the empirical knowledge of a skilled operator, and can perform tunnel ventilation control with higher accuracy. That is, FIG. 6 is a block diagram when the plant control device according to the present invention is applied as a ventilation control device for a tunnel as shown in FIG. The control signals to the various ventilation facilities and the images from the surveillance camera are transmitted and received by the ventilation control device via the communication line.

In FIG. 6, the surveillance cameras (C1L to C3L, C1U to C3U) installed at the lower and upper portions at the three points in the tunnel correspond to the image input means 1 in FIG. The operation console 10 and the control operation computer 11 in FIG. 6 correspond to the interface means and the case collection means 6 in FIG.
The soot state example database 12 and the neural net 14 in FIG. 6 are the state database means 5 in FIG.
6 corresponds to the control path database means in FIG. 3, and the control device in FIG. 6 corresponds to the control path determining means in FIG. 3 and control means 8.

* Collecting Cases * First, the process of collecting operator's control knowledge will be described.
That is, as shown in FIG. 6, monitoring cameras C1L to C3L that are arranged in the lower part of the tunnel and monitor the state of the lower part of the tunnel.
And a monitoring image X (for example, an image showing a soot state at a key point in the tunnel) sent from a total of 6 cameras of the monitoring cameras C1U to C3U that are arranged at the top of the tunnel and monitor the state of the tunnel top It is input to the control operation computer 11. The monitoring image X corresponds to the monitoring image X in FIG.

The monitor image X is displayed on the monitor television connected to the operation console 10 operated by the operator by the control operation computer 11. The operator determines an appropriate operation command U by referring to the image showing the soot and smoke state displayed on the monitor TV, and inputs the operation command U to the keyboard connected to the console 10. The input operation command U is sent to and executed by each ventilation device. The operation command U corresponds to the operation amount _Uc in FIG.

Next, the control operation computer 11 searches the soot state case database 12 for the state name S _i added to the soot state of the tunnel observed in the past,
It is displayed on the monitor TV on the console 10. With reference to this display, the operator selects a state name considered appropriate for the current soot state X of the tunnel from the state names S _i displayed on the monitor TV or determines a new state name. And specify it through the keyboard.
Here, the state name designated by the operator is S _X.

Subsequently, the control operation computer 11
Monitoring image X showing the soot state of the tunnel and its state name S
_{The X} pair is registered in the soot state case database 12.
Furthermore, the status name S _Y one time ago and the current status name S _X
・ Operation command U _YX instructed by the operator one hour before
The cost C _YX required for the operation is paired and registered in the tunnel state transition record database 13. When the state name is S _Y , the cost C _YX is the time and power required for the soot state of the tunnel to reach S _X as a result of adding the operation command U _YX to various ventilation equipment. .

When this operation is repeated, the soot state case database 12 stores various cases including the monitoring image X indicating the soot state and the state name S _X. In addition, the state name in the tunnel state transition record database 13 is S
_{When Y} , when the operation command U _YX is added to various ventilation equipment, the soot state of the tunnel transits to S _X, and a record of the state transition that the operation costs C _YX is accumulated.

Here, FIG. 7 is a diagram schematically showing a part of the record accumulated in the in-tunnel state transition record database 13 for ease of explanation. For example, it is shown that the state (1) is changed to the state (2) by adding the manipulated variable U _1,2 to the ventilation equipment, and at that time, the cost of C _1,2 is required. Note that in FIG. 7, each state is indicated by a number in a circle.
A series of operation amounts and costs for making a transition from one state to another state as shown in FIG. 7 is a state transition record.

Further, the examples of the image X of the soot state and the state name S _X accumulated in the soot state case database 12 are used for learning the neural network 14 having the structure shown in FIG. In FIG. 8, circles represent neuron elements, and arrows connecting the circles represent connecting lines between neuron elements. This connection line connects the neuron elements of all the input layers and the neuron elements of all the middle layers, and the neuron elements of all the middle layers and the neuron elements of all the output layers. Only part of it is drawn for clarity. Further, in FIG. 8, the symbol U
_j is a neuron element of the input layer, U _i is a neuron element of the intermediate layer, U _k is a neuron element of the output layer, X _j is an input to the gray value U _j of the pixel of the synthesized image, W
_ij is a coupling weight value which is a signal transmission efficiency on the coupling line connecting U _j and U _i , W _ki is a coupling weight value which is a signal transmission efficiency on the coupling line connecting U _i and U _k , and a _k is a neural network. 1
4 represents the reaction intensity value of U _k with respect to the image input to the 4th input layer.

The neural network 14 has an invention previously proposed by the present inventors, for example, Japanese Patent Application No. 4-45.
614 and Japanese Patent Application No. 5-102980 may be used. Further, as for the learning procedure, those described in these inventions may be used, and therefore, only a brief explanation will be given here.

That is, the neural net 14 has a three-layer structure, and each element of the input layer shows the state of soot in the tunnel captured by the six monitoring cameras (C1L to C3L, C1U to C3U). The grayscale value of each pixel of the image obtained by arranging the six images horizontally in a row is input. That is, the input layer is composed of the same number of elements as the number of pixels of the combined image. The image information input to the input layer is propagated to the element of the intermediate layer through the coupling line and further to the element of the output layer. The elements of the output layer include the same number of elements as the number of state names, and each element has a one-to-one correspondence with each state name. As a result of the reaction propagating from the input layer to the intermediate layer to the output layer in this way, the state name corresponding to the element of the output layer that has reacted is the hypothesis of the soot state of the tunnel.

A neural network which operates in this way will be described in this embodiment. That is, the neural network 14
When the monitor image X is input to the input layer of, the state name S _{X of X} appears in the output layer. At the same time, the certainty factor b (S _i ) in which the state name of X is S _i is obtained as the reaction strength of the element representing each state name. On the other hand, in the process of learning, the element in the output layer is provided with information that specifies a correct reaction of the output layer element, which is called a teacher signal. This teacher signal is information indicating the state name for the surveillance image input to the input layer, and is given to correct an error in the estimated value of the state name for the surveillance image that appears in the output layer of the neural network 14. is there. Then, the learning of the neural network 14 is repeated until the signals corresponding to the teacher signals appear as the reaction of the elements of the output layer of the neural network 14 for the images of all the cases accumulated in the soot state case database 12. . Therefore, the learning here means to correct the weights on the coupling lines of all the elements so that an output that matches the teacher signal appears.

* At the time of control * Next, the process of performing ventilation control of the tunnel using the tunnel ventilation control device of the present embodiment, which has learned the examples as described above, will be described. That is, the monitoring image X, which is a controlled variable and is composed of six images showing the soot state in the tunnel obtained by the six monitoring cameras, is presented to the input layer of the neural network 14 through the communication line. Then, the neural network 14 calculates and outputs the state name S _X , which is a hypothesis of the state corresponding to the monitoring image X, and its certainty factor b (S _X ). Here, in general, it is extremely rare that an image that exactly matches the monitoring image X has been experienced in the past, so at this time, the neural network 14 selects the monitoring image X from the images learned in the past. Close, multiple state hypotheses S _i
Is output. That is, a plurality of pairs of state hypotheses and certainty factors that form a pair are output.

Further, the control device 15 reads the state name S _X and its certainty factor b (S _X ) from the neural network 14, and the operator through the console 10 and the control operation computer 11. The target state T designated by is read. Next, the control device 15 searches the in-tunnel state transition record database 13 and reads all the state transition records that are the routes from the hypothesis group S _X of the current state to the target state T. For example, it is assumed that the record registered in the in-tunnel state transition record database 13 is as shown in FIG. That is, when the current state hypotheses output by the neural network 14 are (9) and (12), and the respective confidence factors are 0.6 and 0.2 and the target state is (14), the current state hypotheses are There are three routes from (9) → (10) → (14) (12) → (13) → (14) (12) → (15) → (16) → (14) There is.

Next, the control unit 15 calculates which of the above three routes has the lower cost, the two routes whose current state hypothesis is (12), and determines the route with the higher cost as the route. Remove from candidates. To find a route with low cost, for example, a dynamic programming method (Dynamic Programming, commonly known as DP method) widely used as an optimization means for multistage decision
Should be applied. For DP method, see R. Bellma
n: Dynamic Programming, Princeton Univ. Press. Here, in this embodiment, (12) → (13)
Assuming that the route of (14) has a lower cost, the route of (12) → (15) → (16) → (14) having a higher cost is excluded from the candidates. Therefore, at this stage, as a candidate for the route from the current state hypothesis to the target state, the current state hypothesis (9) is used as a starting point (9) → (10) → (14).
And the current state hypothesis (12) as the starting point (1
Two routes of 2) → (13) → (14) remain.

Next, the controller 15 determines an operation command U for various ventilation equipment. This operation command U collectively represents the increment direction of the operation amount for each type of ventilation equipment. That is, in any of the horizontal fan, dust collector, and vertical fan that are the ventilation equipment of the tunnel to which the application is applied, the operation command U _{a, b} for making the number of revolutions of the motor an operation amount and transitioning from the state a to the state _b is 6 motors VB
1, VB2, HB1, HB2, QC1, QC2 rotation speed increment direction

[Equation 3] Is. Here, the values taken by the respective components of the operation command U _{a, b} are NB, ZZ and PB, which correspond to “decrease”, “maintain” and “increase” the rotation speed of the motor, respectively. Further, the relationship between the direction of the rotation speed increment of the motor and the actual value of the rotation speed increment is determined by the membership function shown in FIG.

That is, the current state hypotheses are (9) and (12).
Therefore, the control device 15 _changes the rotation increment of each motor based on the operation commands U _9,10 , U _12,13 for transitioning to the next states (10), (13) on the two paths. decide.

Here, these two operation commands U _9,10 and U
₁₂ , ₁₃ will be described by taking a command to the vertical blower VB1 as an example. The command regarding VB1 of U _9,10 is "N
B ", that is," reduce "the number of rotations of the motor, and U
It is assumed that the command relating to VB1 of ₁₂ , ₁₃ is "PB", that is, "increases" the rotation speed. At this time, the vertical blower VB1
The rotation speed increment of the motor is calculated by the center of gravity method in fuzzy reasoning, as shown in FIG. As the grade of each membership function, the certainty factor 0.6 of the state hypothesis (9) and the certainty factor 0.2 of the state hypothesis (12) are used. In the same manner, after determining the rotational speed increments of the remaining five motors, the rotational speeds of the motors of the respective ventilation devices are changed.

Next, the image from the surveillance camera is presented to the input layer of the neural network 14 again. As a result, the current state hypotheses are (13) and (16), and the certainty factors are 0.
Suppose that it is 8, 0.2. In this case, the control device 15
Is the state hypothesis expected as the next state among the candidates (9) → (10) → (14) (12) → (13) → (14) for the route to the target state set one hour ago ( The first route starting from the state (9) that did not reach 10) is excluded from the route candidates. Therefore, the increment of the rotation speed of the motor is
Operation command U from state hypothesis (13) to state hypothesis (14)
Determined from _13,14 only. Note that this decision is also made by fuzzy inference as in the case of one time ago.

By repeating the above operation until the target state (14) is reached, various ventilation devices can be controlled appropriately.

* Operation when the state transition of the plant has dynamics * Next, according to the plant control device of the present invention, it will be explained that appropriate control can be performed even when the state transition of the plant has dynamics. . That is, FIG. 11 shows a part of the routes accumulated in the tunnel state transition record database 13 when there is dynamics in the state transition of soot in the tunnel.

In FIG. 11, the circles indicate the states in the tunnel at the relevant time, the symbols in the circles indicate the state names, and the arrows indicate the transitions between the states. In addition, the character string enclosed in parentheses attached to the arrow means an operation command that causes a transition between the states connected by the arrow. This operation command is expressed by combining the name of the ventilation equipment that changes the rotation speed of the motor and the symbols (+, −) indicating the direction of the increase in the rotation speed of the motor. For example, the operation command (VB1 +, HB1 +) is the vertical blower VB.
1 and the number of rotations of the horizontal blower HB1 are increased respectively. Further, the operation command (0) indicates that the current rotation speed is maintained for all ventilation equipment.

Further, the number 2 described above the circle indicating the state
The letter string in the third row is a character string representing the observed controlled variable, that is, the situation of the surveillance image corresponding to the state, which is captured by the six surveillance cameras. In addition, the alphabetical string in the upper row is the three surveillance cameras C1U to C installed at the top of the tunnel.
The situation of the surveillance image captured by 3U is shown. The letters in the lower row are the three surveillance cameras C1 installed at the bottom of the tunnel.
The situation of the monitoring image captured by L to C3L is shown. In addition, the letter "L" means that the soot concentration is low, "m" is medium, and "h" is high. In addition, in fact,
The set of six monitoring images is the controlled amount, but here, for simplification of explanation, instead of showing the six images, the monitoring image is graded into three levels according to the soot concentration and the degree thereof is shown. The controlled variable will be represented by a set of letters. Further, the information actually read by the operator from the monitoring image includes not only the soot concentration but also the smoke flow direction and the wind speed, but here, only the soot concentration is considered.

As shown in FIG. 11, at time k, 6
It is assumed that the images obtained from the two monitoring cameras C1U, C2U, ..., C3L are mmmhh. As mentioned above,
The neural network 14 reads this image, and as a state hypothesis when the monitoring image is mmmhh, "a"
"1" and "A" are output. In addition, the control device 15 starts from these three state hypotheses “a”, “1”, and “A” from the in-tunnel state transition record database 13 to three routes (a) → (ro ) → (C) → (1) → (2) → (3) → (A) → (B) → (C) → The three routes are indicated by solid lines in FIG.

These three routes are control routes in order from the top of FIG. 11 when the vehicle amount decreases, when the vehicle amount is constant, and when the vehicle amount increases. The state hypothesis when the monitored image is mmmhhhh is , "A", "1", and "A", it is not possible to determine which of the above three situations the vehicle volume in the tunnel is from only the instantaneous image at time k. That is, it is impossible to determine which one of the three routes to reach the target state. next,
The control device 15 determines the operation amount by the fuzzy inference described above, and controls the ventilation device based on the operation amount. In addition, in the example shown in FIG. 11, since the operation command of any path is (0), all the increments of the rotation speed with respect to all the ventilation devices are 0.

Next, it is assumed that the image obtained from the surveillance camera is hhhhhh at time k + 1, which is one time later. The neural network 14 reads this image and outputs "2" and "B" as the state hypothesis. In addition, the control device 1
5 is a route in which the next state (that is, the state at time k + 1) of the state immediately before one of the three currently considered routes is not “2” or “B”, (a) →
Remove (b) → (c) from the route candidates. Further, the control device 15 then operates the certainty factor of the state hypothesis “2” and “B”, the operation command (HB1 +, HB2 +) of the path (2) → (3), and the operation of the path (B) → (C). Command (VB1 +, VB2 +, Q
From C1 +, QC2 +), the operation amount is determined by the fuzzy inference described above, and each ventilation device is controlled based on this operation amount.

Furthermore, it is assumed that the image obtained from the surveillance camera at time k + 2 is hhhmhm. The neural network 14 reads this image and outputs "C" as a state hypothesis. In addition, the control device 15 selects a route (2) → (3) in which the state next to the state one time before (that is, the state at the time k + 2) is not “C” among the currently considered two route candidates. ) Is excluded from the route candidates.

As shown in the above example, when the plant state transition has dynamics, the plant state cannot be identified only by observing the instantaneous controlled variable at a certain time point. In the above example, the controlled variable m at time k
There are three possible states for mmhh,
According to the plant control device of the present invention, since it is possible to gradually narrow down the state hypothesis, even when the state of the plant cannot be identified from the instantaneous value of the controlled variable observed at a certain time. Appropriate plant control can be performed by handling the time series of states, that is, the state transitions.

As described above, according to the present embodiment, the means for directly handling the data referred to by the operator at the time of control such as the monitoring image as the controlled variable, the means for acquiring the control knowledge of the operator, and the acquisition By providing a means for generating the best control path from the acquired knowledge, it is possible to incorporate control knowledge by a skilled operator and automatically perform plant control according to the optimum control path.

(Other Embodiments) The present invention is not limited to the above-described embodiments, but various modifications can be made without departing from the scope of the invention. For example, in the above embodiment, the neural net is used as the state estimating means 2, but any means can be used as long as it is a means for classifying the controlled variables. For example, a statistical pattern based on the principal component analysis. Recognizing means (eg compound similarity method)
Alternatively, any clustering method may be adopted. Further, the DP method is used as the means for determining the route with low cost in the control route determining means 3, but any means may be used as long as it is a means for finding the optimum route. For example, the Floyd method is adopted. It doesn't matter (Reference: R.
W.Floyd: Algorithm 97, Shortest Path, C.ACM, Vol.5, No.
6).

In this embodiment, the tunnel ventilation control is targeted, but in addition to this, the temperature inside the blast furnace, the sewage sludge,
The same effect can be obtained by using the plant control device according to the present invention even in a rolled steel plate shape, a traffic signal, and the like.

[0074]

As described above, according to the present invention, the time change between the image given to the operator and the operation amount determined by the operator by looking at the image is recorded and inputted from this record. By estimating the plant state with respect to the image, determining the control path to the target state, and determining the manipulated variable for transitioning the plant state to the next state on the control path, the operator's control knowledge It is possible to provide a plant control device that can perform automatic operation control based on past operation results without performing the complicated work of acquisition.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a first embodiment of a plant control device of the present invention.

FIG. 2 is a diagram showing a flow of processing during control in the plant control apparatus shown in FIG.

FIG. 3 is a block diagram showing the configuration of a second embodiment of the plant control device of the present invention.

FIG. 4 is a diagram showing a flow of processing when collecting a case in the plant control apparatus shown in FIG.

FIG. 5 is a vertical cross-sectional view of a tunnel to which the plant control device of the present invention is applied.

FIG. 6 is a block diagram showing the configuration of a third embodiment (tunnel ventilation control) of the plant control device of the present invention.

FIG. 7 is a diagram showing an example of a control route recorded in a tunnel state transition record database.

FIG. 8 is a conceptual diagram showing an example of the structure of a neural network.

FIG. 9 is a diagram showing a membership function used when the control unit determines an operation amount.

FIG. 10 is a diagram showing a method of determining an operation amount in the third embodiment.

FIG. 11 is a diagram showing a control path when the state transition of the plant has dynamics.

[Explanation of symbols]

1 ... Image input means 2 ... State estimation means 3 ... Control path determining means 4 ... Control path database means 5 ... State database means 6 ... Case collection means 7 ... Interface means 8 ... Control means 9 ... Plant 10 ... Operation console 11 ... Computer for control operation 12 ... Soot state case database 13 ... Tunnel state transition record database 14 ... Neural network 15 ... Control device

─────────────────────────────────────────────────── ─── Continuation of the front page (58) Fields surveyed (Int.Cl. ⁷ , DB name) G05B 23/00-23/02 G05B 11/00-13/04

Claims (3)

(57) [Claims] 1. A pair of image input means for capturing a monitoring image showing a state of a plant, and a correspondence relationship between the monitoring image itself obtained via the image input means and a state of the plant .
And stores the state database means, the monitoring image itself showing the state of the plant, and the correspondence between the plant state and the monitoring image stored in the state database means, and creates a hypothesis of the current state of the plant. Then
Also, a state estimating means for calculating the certainty factor, a control path database means for storing a pair of a state transition of the plant and an operation amount causing the transition, a target state of the plant, and the state estimating means are created. The current state hypothesis of the plant to be taken in and the pair of the state transition of the plant stored in the control path database and the operation amount causing the transition are taken in, and a transition is made from the current state hypothesis of the plant to the target state. Control route determining means for calculating a route set, a set of route candidates from the current state hypothesis of the plant calculated by the control route determining means to a target state, and a current state of the plant calculated by the state estimating means And a control means for calculating the manipulated variable of the plant by taking the certainty factor of the above hypothesis. 2. A monitoring image showing the state of the plant and the correspondence between the state of the plant and the monitoring image stored in the state database means are presented to the operator of the plant, and the current state of the plant designated by the operator. Interface means for capturing the state and the appropriate manipulated variable, a monitoring image showing the state of the plant, and a pair of the current state of the plant and the appropriate manipulated variable obtained through the interface means, A state transition represented by a pair of a state change order and an operation amount represented by a set of a state and a current state is stored in the control path database means, and a pair of a monitoring image and a current state is stored in the state. The plant control device according to claim 1, further comprising a case collection unit that stores the data in a database unit. 3. A monitoring camera for capturing a monitoring image showing the state of the plant, a soot state case database storing the correspondence between the monitoring image obtained by the monitoring camera and the state of the plant, and the state of the plant are shown. A monitoring image, a neural network that captures the correspondence relationship between the plant state and the monitoring image stored in the soot state case database, creates a hypothesis of the current state of the plant, and calculates the certainty factor thereof, A tunnel state transition record database that stores a pair of a state transition of a plant and an operation amount that causes the transition, a target state of the plant, a hypothesis of the current state of the plant created by the neural network, and the tunnel. A pair of the state transition of the plant stored in the internal state transition record database and the operation amount causing the transition And a controller for calculating a route set that transitions from the current state hypothesis of the plant to the target state, wherein the controller is a set of route candidates from the current state hypothesis of the plant to the target state, A tunnel ventilation control device, wherein an operation amount of the plant is calculated based on a certainty factor of a hypothesis of a current state of the plant calculated by the neural network.