JPH0733340A - Group supervisory operation control device for elevator - Google Patents

Abstract

PURPOSE:To set optimum control parameters for individual time zones by providing a partial model means modeling the relationship between control parameters of evaluation indexes and group supervisory operation control response results as the partial model assembly constituted of function models synthesized by a neural network. CONSTITUTION:Control parameters set for a fixed period at the different time are transmitted to a group supervisory operation controller 1 as the optimum evaluation index weighting values. The group supervisory operation controller 1 calculates the evaluation values of the evaluation indexes based on the information from unit controllers 2-1 through 2-N, applies the optimum weighting to the evaluation values based on the control parameters, and gives a control command to the optimum elevator. The group supervisory operation controller 1 models the relationship between the weighting values of the evaluation indexes and the group supervisory operation control response results as the partial system model assembly, calculates the weighting values for the partial system models based on the traffic demand, calculates the group supervisory operation control response inference result from the synthesis of the inference output and the partial model output, independently evaluates individual demands of a weighting values.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention operates a plurality of elevators on a plurality of floors, performs a predetermined evaluation calculation on a common hall call that has occurred, and evaluates each elevator to be optimal. TECHNICAL FIELD The present invention relates to an elevator group management control device for allocating various units to hall calls, and particularly to an elevator group management control device capable of performing allocation control satisfying an optimum evaluation standard for each building.

[0002]

2. Description of the Related Art In recent years, when a plurality of elevators are installed side by side, in order to improve the operation efficiency of the elevators and the service to the elevator users, elevators that respond to hall calls on each floor are Reasonable and prompt allocation is performed by using a small computer such as a microcomputer.

That is, when a hall call is generated, an optimum elevator serving the hall call is selected and assigned, and other elevators are prevented from responding to the hall call.

In the group management control device of such a system, recently, inter-floor traffic of each building such as measurement of car call registration data when responding to each hall call in real time, data measurement during getting on and off, etc. The demands peculiar to the building are grasped based on the above measurement data and used for the allocation control of the optimum number of units.

In such a situation, the optimal machine number allocation control of the elevator evaluates each evaluation index on the group management performance with respect to the occurrence of a hall call, weights and adds each evaluation value, and selects the optimum machine number as a comprehensive evaluation. I have decided.

However, in general, since each evaluation index on group management performance greatly changes depending on traffic demand, it is necessary to select an optimum weighting value (control parameter) depending on the situation of the elevator group system, and it continuously changes. It is difficult to optimize the control parameters by following the changing traffic demand.

The "optimum" evaluation criterion for group management performance is used for buildings such as hotels, tenant buildings, one-company-owned buildings, elevator users, building management companies, etc. Since it depends on each building, it is necessary to match the group management control performance with the evaluation criteria for each building, but in the conventional system, it is impossible to automatically set the optimum control parameters to the evaluation criteria unique to each building. .

[0008] Therefore, there is also a method in which a building-specific evaluation standard is shown in advance by the user, and a large number of simulations are performed to preset the control parameters that match the user's evaluation standard and input them into the system. This requires a large number of simulations to be performed, and the simulation results do not completely reflect the system's unique situation (the situation after operation).
Since the optimal control parameter changes depending on the congestion level,
There is a problem that the allocation control may not be optimal with respect to the evaluation standard intended by the user.

Further, there is a method of setting control parameters by storing and learning all data obtained online, but there are problems such as memory restrictions and deterioration of CPU performance.

[0010]

As described above, in the conventional group management control device for the reta, there is a problem that the allocation control that satisfies the optimum evaluation standard cannot be performed for each building. An object of the present invention is to form a plurality of relational models between each evaluation index weighting value in allocation control and a group management control response result represented by a hall call response cumulative distribution, and to further relate this relational model to traffic demand. In addition, a finite number of relational models are associated to realize a target model that responds to a large number of traffic demands, and the control parameters are optimized by following the continuously changing traffic demands. By extracting only valid ones of the situation, control parameters, and actual control response results and learning the relationship, it is possible to perform allocation control that satisfies the optimum evaluation criteria for each building. An object is to provide a high elevator group supervisory control device.

[0011]

In order to achieve the above object, according to the present invention, a plurality of elevators are activated for a plurality of floors, and a predetermined evaluation calculation is performed on a common hall call that has occurred. In the elevator group management control device that conducts an evaluation for each elevator and assigns the optimal number to the hall call, the relationship between the weighting value of each evaluation index and the group management control response result is ambiguous depending on the traffic demand. The partial model means modeled as a set of partial system models composed of functional models synthesized by the segmented neural network, and the relationship between the partial system model and traffic demand are vaguely expressed by a plurality of membership functions, and the membership concerned Memorize the relation of connection between the function and the subsystem model, and use the subsystem model based on the traffic demand. And inference means for calculating the weighting value,
A synthesizing means for calculating a group management control response inference result by synthesizing outputs from the inference means and the partial model means, and a group management control response inference result from the synthesizing means for each demand of the building in which the elevator system is installed. Inference result evaluation means that evaluates independently and sets the optimum value of the evaluation index weighting value according to actual building demand, and the evaluation index weighting value set by the inference result evaluation means according to the traffic demand during system operation A group adjusting control is performed with an optimum value, and a system adjusting unit for changing each component of the partial system model, the inference unit, and the inference result evaluating unit from the obtained group management control response result is configured. .

Here, in particular, as the system adjusting means, those in which the degree of activity with respect to the partial system model based on the traffic demand output from the inference means exceeds a predetermined value are regarded as valid data, and the partial system model is calculated by the data. I am trying to re-learn.

As the system adjusting means, the evaluation index weighting value which becomes an input of the partial system model or the group management control response result which becomes an output from the effective data of the partial system model extracted from the actual data when the system is operating. The partial system model is re-learned by selecting and collecting the data whose similarity exceeds a predetermined value.

Further, as the system adjusting means,
The effective data of the partial system model extracted from the actual data when the system is operating is discarded in order from the oldest, and the partial system model is relearned by the limited data.

On the other hand, as the system adjusting means, the number of valid data of the partial system model extracted from the actual data during the system operation is limited according to the configuration of the partial system model, and the partial data is converted by the limited data. 3. The elevator group supervisory control device according to claim 2, wherein the system model is relearned.

As the system adjusting means, the degree of connection in the inference means is changed so that the degree of activity with respect to the retrained partial system model is increased.

[0017]

Therefore, in the elevator group supervisory control apparatus of the present invention, the weight value for the partial system model is calculated by inputting the traffic demand for each certain period of each time of the building to the inference means.

A group management control for each control parameter combination is performed by inputting the control parameter combinations in which the control parameters of each evaluation index are changed within a predetermined range to the partial model means, and by the associative function of the combining means together with the weighting values. The response inference result is output.

Then, the inference result evaluation means evaluates the inference result according to an evaluation criterion that is judged to be "optimal" in accordance with the use, characteristics, or customer needs of each building where the elevator system is installed. The optimum inference result is extracted, and the evaluation index weighting value corresponding to this inference result is selected as the optimum control parameter.

Further, the group management control is performed by the system adjusting means at the optimum value of the evaluation index weighting value set according to the traffic demand when the system is operating, and from the obtained group management control response result, The constituent elements of the partial system model, the inference means, and the inference result evaluation means are changed to control the allocation of the optimum machine.

[0021]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A first embodiment of the present invention will be described in detail below with reference to the drawings. FIG. 1 is a block diagram showing an example of the overall system configuration of an elevator group management control device according to the present invention.

In FIG. 1, 1 is a group management control unit,
The group management control unit 1 includes a single control unit 2-1 to 2-N that controls each unit elevator and a learning control unit 1-1.
And a high-speed transmission line 6 which is a first transmission control means. The group management control unit 1, the learning control unit 1-1, and the single unit control units 2-1 to 2-N are configured by a single computer or a plurality of small computers such as microcomputers, and operate under the control of software. To do.

On the other hand, 3 is a hall call button provided on each floor, and 4 is a hall call input / output control section for inputting / outputting a hall call. Then, the group management control unit 1, the learning control unit 1-1, the single unit control units 2-1 to 2-N, and the hall call input / output control units 4 use the low-speed transmission line 7 as the second transmission control means. Connected through.

Here, the high-speed transmission line 6 has a unit controller 2-
1 to 2-N is a transmission control system that performs transmission between the group management control unit 1 and the learning control unit 1-1, that is, mainly between control computers in the machine room, and is connected by a high-speed and highly intelligent network. ing. Then, the control information necessary for the group management control is exchanged at high speed among the group management control unit 1, the learning control unit 1-1, and the individual unit control units 2-1 to 2-N.

The low-speed transmission line 7 is a control system for transmitting information mainly through the hoistway, such as the hall call button 3 in each hall and the monitoring board 5 in the monitoring room. Since it is low speed and long distance, it is configured by an optical fiber cable, etc., and is connected to the group management control unit 1, the learning control unit 1-1 and the single unit control units 2-1 to 2-N to exchange data. I am doing it.

When the group management control unit 1 is normal, the hall call button 3 is controlled by the group management control unit 1 via the low-speed transmission line 7, and when the hall call button 3 is pressed, the hall call gate is closed. The unit lamp control units 2-1 to 2-N sent via the high-speed transmission line 6 while setting the registration lamp
The optimum machine is determined based on the information of the
A control command is issued to the corresponding one of ~ 2-N. Then, the unit controller that has received the control command performs unit control using the control command as hall call information.

FIG. 2 shows a group management controller 1 and a unit controller 2-1 in the elevator group management controller according to the present invention.
It is a figure which shows an example of a 2-N software system. In FIG. 2, this software system uses a real-time OS 8, which is an operating system,
Individual control function task 9, group management control main function task 1
0, group management control sub-function task 11, transmission control task 1
2 is managed, and each task 9 to 12 is activated and held by the scheduler in the real-time OS 8.

The individual control function task 9 among these tasks 9 to 12 is a function task for operating the individual control units 2-1 to 2-N, and has a high priority.

The group management control main function task 10
Is a central functional task of the group management control unit 1, and collects information data for each machine from the group management control sub-function task 11 distributed to each individual control unit 2-1 to 2-N, and performs a comparison calculation. By doing so, the optimum number is determined, a control command is issued to the corresponding number, and hall call registration is controlled.

Further, the group management control sub-function task 11
Is a functional task that processes information of each unit of the group management control unit 1, and performs information processing under the control of the group management control main function task 10.

That is, the group management control main function task 1
The computer having 0 is configured to manage the start and end of the tax via the high-speed transmission line 6.
In accordance with a command from the group management control main function task 10 that is a master, distributed processing is performed in units of machine numbers, and data is conveyed to the group management control main function task 10 when the processing is completed.

Furthermore, the transmission control task 12 controls the transmission / reception of data on the high-speed transmission line 6 and the activation / termination of the group management control sub-function task 11. FIG. 3 is a block diagram showing a system configuration example of the high-speed transmission line 6 in FIG.

In FIG. 3, the transmission control is performed by using the microprocessor 13, for example, ISO.
Using the data link controller 14 and the media access controller 15 which are made up of hardware as a part for controlling the data link layer of the LAN network model layer proposed by (International Organization for Standardization), data transmission is performed with high intelligence. Thus, the ratio of the transmission control software managed by the microprocessor 13 to the high-speed transmission control is reduced.

For example, as a data link controller 14 which is a controller for realizing the above-mentioned high intelligent transmission control, an i82586 which is an LSI (Intel Corporation) LSI, and a media access controller 15 which is also an i8250 of the same Intel.
Although 1 and the like have been put into practical use, by using them, a high-speed transmission function such as 10 Mbit / sec can be relatively easily performed while reducing the support ratio of the microprocessor 13.

In FIG. 3, 16 is a system bus, 17 is a control line, and 18 is a serial transmission system. On the other hand, FIG. 4 is a functional block diagram showing the flow of input / output signals of the learning control unit 1-1 according to the present invention, FIG. 5 is a diagram showing an example of the system configuration of the learning control unit 1-1 in FIG. 4, FIG. FIG. 7 is a diagram showing a detailed system configuration example of the inference unit and the partial model unit in FIG. 5, respectively.

First, in FIG. 4, the group management control unit 1
As described above, the group management control sub-function task 11 in the individual control units 2-1 to 2-N of the elevator group system 2
The hall call allocation control function is executed in cooperation with.

The evaluation algorithm used for this hall call allocation control evaluates each evaluation index on the group management performance,
The overall evaluation is performed by adding the respective evaluation values with optimal weighting.

Further, the learning control section 1-1 is arranged so that the elevator group system 2 transmitted via the group management control section 1 at regular intervals, that is, the individual control sections 2-1 to 2-N and the hall call input / output control. The traffic demand and group management control response result are generated based on the information from the unit 4, optimal control parameters are set from the traffic demand, and the group management control response result is used as base data for online learning.

Then, the learning control section 1-1 transmits the above-mentioned control parameters set for each fixed period for each time to the group management control section 1 as an optimum evaluation index weighting value.
The group management control unit 1 calculates the evaluation value of each evaluation index based on the information from the unit control units 2-1 to 2-N, and the learning control unit 1-
The evaluation parameter is optimally weighted by the control parameter transmitted from No. 1 and a control command is given to the optimal machine.

Next, the detailed operation of the learning control section 1-1 will be described with reference to the drawings starting from FIG. Explaining the functional configuration of the learning control unit 1-1, as shown in FIG.
The inference unit 21, the partial model unit 22, the combining unit 23, the inference result evaluation unit 24, and the system adjustment unit 25 are included.

[0041] The evaluation operation in group management control unit 1, in general, carried out for a plurality of evaluation indexes l, for i _{Unit, g 1 (i), g} 2 (i), ......, g l ( i), and the total evaluation value is obtained by weighting and adding the evaluation values assigned to each evaluation. The total evaluation value E _{i for} Unit _i is

[0042]

[Equation 1] Is expressed as

Here, α _j is a weighting value for each evaluation index j, and is a control parameter transmitted from the learning control section 1-1 to the group management control section 1. Therefore, the inference result evaluation unit 24 calculates the traffic demand for each fixed period for each time period, outputs the traffic demand to the inference unit 21, and generates a combination of the control parameters α _j within a predetermined range, The result is output to the model unit 22, and as a result, the group management control response inference result corresponding to the combination of each control parameter α _j is evaluated according to the use of the elevator installation building, and the optimum control parameter is set.

Assuming that the group management control response inference result is y and the input to the learning control section 1-1 is u, the calculation in the combining section 23 can be expressed as y = F (u) (1). In this case, y = (y ₁ , y ₂ , ..., Y _n ) ^T u = (u _s , u _e ) ^T = (C, α) ^T.

In this group management control response inference result y,
y ₁ , y ₂ , ..., Y _n represent inference results such as the occurrence rate of hall call response time, average multiplication rate, average service time, etc. in a fixed time, and evaluation reference data for determining group management performance. Becomes

Further, C in the input u represents traffic demand, and if C = (c ₁ , c ₂ , c ₃ ), then c ₁ , c ₂ , c ₃ are all average passenger occurrence intervals [s / person ], The average passenger occurrence interval on the standard floor,
It is data that represents the average interval between passengers heading to the standard floor, and represents the status of the system such as the congestion degree of the system and the flow of people.

Further, α is a weight value (control parameter) for each evaluation index, and as described above, α = (α ₁ , α ₂ , ..., α _l ) for a plurality of evaluation indices l. Represented.

Therefore, the inference unit 21, the partial model unit 22,
The target model composed of the synthesizing unit 23 is expressed by synthesizing m partial system models f _i (α), (i = 1, 2, ..., M), and the equation (1) can be rewritten as the following equation. It will be.

[0049]

[Equation 2]

Here, a _i (C) represents the activity of the partial system model f _i (α) in the traffic demand C, and the group management control response inference result y is the system situation and the portion obtained from the traffic demand C. It is determined by the connection relationship with the partial system model in the model unit 22.

Next, the system configurations of the inference unit 21, the partial model unit 22, the inference result evaluation unit 24, and the system adjustment unit 25 will be described. First, the inference unit 21
As shown in FIG. 6, the input unit 21-1 and the storage unit 21-2.
And an output unit 21-3 and a gate 21-4, and receives the traffic demand C from the inference result evaluation unit 24,
Depending on the system status obtained from these, a _i (C), (i = 1, 2, ...
To output.

Here, the input unit 21-1 has a k-dimensional state vector V consisting of k neurons, and the input traffic demand C is passed through a membership function φ _i so that each element has its member. The partial input vector c _i , (i = 1, 2, ..., M) composed of the ship grade is output. The M partial input vectors c _i are collectively input to the input state vector V as the input vector C.

The storage unit 21-2 comprises an r-dimensional state vector X consisting of r neurons, and the input unit 2
It corresponds to a storage unit that associates 1-1 with the output unit 21-3. Further, the output unit 21-3 is composed of an m-dimensional state vector Z composed of m neurons, and each element Z _i ,
(I = 1, 2, ..., M) corresponds to the partial system model f _i (α) of the partial model unit 22.

The input unit 21-1 and the storage unit 21-
2. The storage unit 21-2 and the output unit 21-3 each have a mutual loop, and each unit 21-1 to 21-3 has a self loop.

This relationship has a discrete form and is expressed as follows. C (k) = φ (u (k)) (3.1) V (k + 1) = ψ (W _vc · C (k) + W _vv · V (k) + W _vx · X (k)) (3 .. 2) X (k + 1) = ψ (W _xv · V (k + 1) + W _xx · X (k) + W _xz · Z (k)) (3.3) Z (k + 1) = ψ (W _zx · X ( k + 1) + W _zz · Z (k)) ... (3.4) V (0) = V ₀ , X (0) = X ₀ , Z (0) = Z ₀ , k ≧ 0 where W _vc is a vector It is a matrix showing the load from C to the vector V, and is a synapse load with respect to the vector C of the neurons constituting the vector V. Also,
The same applies to W _vv , W _vx , W _xv , W _xx , W _zx , and W _zz .

Φ is a j-dimensional membership function,
ψ is a sigmoid function corresponding to each dimension, and the calculation of f (x) = 1 / {1 + exp (−x)} (3.5) is performed for each input element.

Further, k is a parameter representing time, and the unit time elapses each time it is incremented. The above (3.
By appropriately setting each W in the equations 1) to (3.4), the partial system model f _i (α), (i = 1, 2, ...,) for the input demand C (u (k)) The activity m) appears in the state vector Z of the output unit 21-3 with the lapse of time.

Furthermore, the gate 21-4 is opened after the set time T elapses, and outputs Z _i (T) as the activity a _i (C) of the partial system model f _i (α). Next, the partial model unit 22 has a system configuration as shown in FIG. 7, and by inputting the control parameter α, the group management control response result f _{i is obtained for} each partial system model.
It has a function of outputting (α).

Individual partial system models f _i (α), (i = 1, 2, ..., M) in the partial model unit 22.
Is composed of a multilayered neural network, as shown in FIG.
_i corresponds to, each of which stores an input output data i of the group management control response result of the real system and control parameter α for the demand. Then, the partial system model f _i (α) is learned by using the back-propagation method with the input / output data i as teacher data.

As shown in detail in FIG. 8, each subsystem model f _i performs the operation y (k) = f _i (u (k)) when given an input u. This calculation process is performed as follows: net (k) = W _hu (k) · u (k) (4.1) h (k) = ψ (net (k) + θ _h (k)) (4.2) nety (K) = W _yh (k) · h (k) + W _yu (k) · u (k) (4.3) y (k) = ψ (nety (k) + θ _y (k)) (4) .4) k ≧ 0 Here, W _hu , W _yh , and W _yu are matrices representing synaptic loads. Further, θ _h and θ _y are vectors representing bias values for the intermediate layer h and the output layer y, respectively.

Each partial system model f _i (u), (i =
1, 2, ..., M) have different synapse weights and biases and perform calculations. In the synthesis unit 23,
Outputs f ₁ (α), f ₂ (α), ..., F _m (α) of the partial system model input from the partial model unit 22
And the activity a ₁ (C), a ₂ (C), ..., A for each partial system model input from the inference unit 21.
_m (C) is synthesized according to the equation (2), and is output to the inference result evaluation unit 24 as the group management control response inference result y.

Next, in the inference result evaluation section 24, as will be described in detail later, as described above, for the target model including the inference section 21, the partial model section 22, and the synthesis section 23,
At the input u (= C, α) _T ), a combination of the control parameters α is generated within a predetermined range for the traffic demand of the actual system, and given as the input u, the combination of the control parameters α is obtained as a result. The group management control response result corresponding to is evaluated, the optimum control parameter α is set, and the result is transmitted to the group management control unit 1.

Next, based on the traffic demand from the group management control unit 1 obtained online, the group management control response result, and the control parameters from the inference result evaluation unit 24, the inference unit 21, which is performed by the system adjustment unit 25, Partial model section 22
Describe online learning.

As shown in FIG. 11, the system adjustment unit 25 includes a difference detection unit 25-1, an inference correction unit 25-2, an input / output data correction unit 25-3, and an input / output data group 25-4.
1 to 25-4m and partial system model correction unit 25-5
And the inference unit 21 based on the difference between the group management control inference result y and the response result from the group management control unit 1.
"Certainty" p _i , and the associated subsystem model f _i . The correction amount is the activity a _i
In direct proportion to the learning level R _i (k).

Here, when the difference between the inference result y and the response result y ₀ satisfies the expression (11), the difference detecting section 25-1 corrects | y−y ₀ | / y ₀ ≧ D _ef0 (11) The degree δN _i that is the basis of the quantity is calculated from the equation (6.4), and the inference unit correction unit 25-2 and the input / output correction unit 25 are calculated.
Output to -3. Here, D _ef0 is a constant and serves as a criterion for determining whether to correct.

In the inference unit correction unit 25-2, among the loops for correcting the certainty in the inference unit 21 shown in FIG. 6, the loop from the output unit 21-3 to the storage unit 21-2, that is, the matrix W. Limited to _xz . At this time, the inference unit correction unit 25-2 has the (i, j) element w of the matrix W _xz.
ij is modified as follows: w _ii = p _i , (i = 1, 2, ..., M) (5.1) w _ji = −p _i , (j ≠ i) (5.2).

Here, p _i ≧ 0 is a parameter representing the reliability of memory for the partial system model f _i (α), and is calculated by the following equation. p _i = ξ · R _i (k + 1) + ζ (6.1) R _i (k + 1) = 1-exp [−β (N _i (k + 1) + γ)] (6.2) N _i (k + 1) = N _i (k) + δN _i (6.3) δN _i = min {1, η · (a _i / R _i (k))} (6.4) where η, β, γ, ξ, ζ is a constant, and R _i and N _i are a proficiency level and a learning progress level of the partial system model f _i (α), respectively. Then, this learning progress N _i
(K) represents the progress of learning of the partial system model f _i (α) after learning k times, and is proportional to the activity level a _i and inversely proportional to the current proficiency level R _i (k). δN _i
It changes with (maximum 1). Then, with respect to the new learning progress N _i (k + 1) calculated by the equation (6.3), the new proficiency level R _i (k +) is determined by the equation (6.2) and finally corrected. The certainty p _i is obtained by the equation (6.1).

On the other hand, the correction of the partial system model f _i (α) is performed by the input / output data correction unit 25-3 and the input / output data group 2
5-4 to 4 m, and partial system model correction unit 25-
Step 5.

First, the input / output data correction unit 25-3 has the correction degree δN _i , the control parameter α ₀ , and the group management response result y _0.
Thus, the input / output data groups 25-41 to 4m included in each partial system model f ₁ (α) are rewritten. Then, the rewritten input / output data 25-4i is used as teacher data,
The partial system model correction unit 25-5 performs addition or re-learning by the back propagation method, and replaces the old partial system model f _i with the new partial system model f _i after learning.

Next, a method of correcting input / output data will be described. The input / output data correction unit 25-3 calculates the group management control response result based on the response result of the above-mentioned time zone when the fixed period of each predetermined time zone ends, and inputs the control result in the above-mentioned time zone together with the control parameter. Output data D ₀ = (u ₀ ) and u ₀ = (C ₀ * α ₀ ) ^T are generated.

[0071] At this time, input and output data with the partial system model _{f i (α) (D 1} , D 2, ......, D L) rewritten. All the input data are scanned from the input / output data groups 25-41 to 4m, the square of the distance between α and α ₀ is calculated, and dα = | α−α ₀ | ² (7.1) is obtained, and is close (dα). The two data D ^(1st) and D ^(2nd) are extracted in order from the smallest data.

Here, in the case of C ^(1st) = C ₀ and d ^(1st) α> 0 (7.1 '), D ₀ is exceptionally additionally registered as new input data D _{L + 1.} To do.

If the equation (7.1 ') does not hold, the square of the distance between C and C ₀ for the above two data D ^(1st) and D ^(2nd) d _C = | C -C ₀ | ² (7.2) is obtained, and then D ^(1st) and D ^(2nd) are corrected by the following equation. However, y ' ^(1st) and y' ^(2nd) are the corrected output data.

Y ′ ^(1st) = ρ ₁ · y ₀ + (1-ρ ₁ ) · y ^(1st) (7.3) y ′ ^(2nd) = ρ ² · y ₀ + (1-ρ ₂ )・ Y ^(2nd) … (7.4) κ = 1 / (γ ・ D ^(1st) α + 1)… (7.5) ρ ₁ = δN _i・ {1 / (λD ^(1st) C + 1)} ・ ・ ・ ・ ( 7.6) ρ ₂ = (1−κ) · δN _i × {1 / (λD ^(2nd) C + 1)} × {1 / (γD ^(2nd) α + 1)} (7.7) The above (7. 3) and (7.4), the input / output data in the partial system model is rewritten, and the rewritten input / output data is used as teacher data in the partial system model correction unit 25-5 by the back propagation method. Modify the model load matrix
Follow the procedure below.

Δ _y (k) = ψ ′ {nety (k) + θ _y (k)} * {y ^* (k) −y (k)} (8.1) δ _h (k) = ψ ′ { net (k) + θ _h (k)} * W ′ _yh · δ _y (k) (8.2) ΔW _yh (k + 1) = η _yh (k) · δ _y (k) · h ′ (k) + α _yh · ΔW _yh (k) (8.3) ΔW _hu (k + 1) = η _hu (k) · δ _h (k) · u ′ (k) + α _hu · ΔW _hu (k)… (8.4) ΔW _yu (k + 1) = η _yu (k) · δ _y (k) · u ′ (k) + α _yu · ΔW _yu (k) (8.5) W _yh (k + 1) = W _yh (k) + ΔW _yh (K + 1) (8.6) W _hu (k + 1) = W _hu (k) + ΔW _hu (k + 1) (8.7) W _yu (k + 1) = W _yu (k) + ΔW _yu (k + 1) (8) .8) ΔW _yh (0) = 0, ΔW _hu (0) = 0, ΔW _yu (0) = 0, Δθ _y ( 0) = 0, Δθ _h (0) = 0 where y ^* is teacher data, and k is incremented until (1/2) | y ^* −y | ² <ε holds for all y ^*. And repeat.

Here, A * B represents the product of each of the matrices A and B, and η and α are learning parameters. By the above calculation, the partial system model correction unit 25
In -5, the load matrix of the partial system model is corrected by the input / output data correction unit 25-3 every time the input / output data group 25-41 to 4m is rewritten to relearn.

Next, the detailed configuration of the inference result evaluation unit 24, the optimization setting of the control parameter α, and the certainty correction operation will be described with reference to FIGS. 9 and 10. As shown in FIG. 9, the inference result evaluation unit 24 includes a control parameter combination generation unit 24-1 and a traffic demand detection unit 24-2.
An inference result evaluation parameter setting unit 24-3, an inference result evaluation operation unit 24-4, and a control parameter setting unit 24-
5 and a certainty correction unit 24-6.

As described above, the inference unit 21, the partial model unit 22, and the synthesizing unit 23 of the learning control unit 1-1 respond to the arbitrary traffic demand of each building with the control parameters and the group management control response inference result. And the group management control response result y can be estimated by association with Equation (1).

Therefore, the inference result evaluation unit 24 determines the group management control response result y according to each building application and customer needs.
Optimal control parameter values are set based on the "optimal" criteria that suit each building.

Therefore, the inference result evaluation unit 24 detects the traffic demand in a predetermined time period and outputs the traffic demand to the inference unit 21, and the inference result evaluation parameter set in advance according to the purpose of each building and the needs of the customer. , The inference result evaluation parameter in the current time zone is selected.

Here, the inference result evaluation parameter is
It is a parameter table for evaluating the group management control response result y, which is set for each building for each traffic demand. The group management control response result y is an index indicating the group management control performance, and is the evaluation reference data such as the hall call response time occurrence rate, the average multiplication rate, and the average service time as described above.

When the group management control performance is evaluated, the evaluation is performed based on the above evaluation reference data.
The weighting for the above evaluation standard data differs depending on the building application, customer needs, and traffic demand.For example, in general office buildings, the items such as hall call response time and average service time are highly weighted, and in hotels, on the contrary, the average The weight of items such as lowering the rate is high. Similarly, even in buildings with the same purpose, the weight of items varies depending on the time of day and customer specifications. Therefore, the inference result evaluation parameter is configured as a weighting value for evaluating each index of the group management control response result y with the traffic demand and the time zone as array elements according to the use of each building.

The inference result evaluation parameter β selected by the predetermined traffic demand C in this manner is sent to the inference result evaluation operation unit 24-2, and the group management control response output from the combining unit 23 here. The result y is evaluated, and the performance evaluation value PE which is the evaluation result is the control parameter setting unit 24-
5, the optimum control parameter α ₀ is set.

Regarding the operation of the inference result evaluation unit 24,
A detailed description will be given based on the flowchart shown in FIG. First, in the control parameter combination generation unit 24-1, the image control parameter α is changed by a minute change amount Δα within its possible range, and a finite number of combinations Pmax is determined for each α (steps S1 and S2). .

[0085] Next, the control parameter P for the combination _{P (α 1P, α 2P ......} , α lP) with respect to, the current time zone by traffic demand detection unit 24-2 traffic demand parameters C with the input u ( = (C, α) _T ), and the inference unit 2
1. Input to the partial model unit 22, and as a result, the combining unit 23
The group management control response inference result y _P is obtained via (step S3).

Based on the inference result y _P , the performance evaluation value PE is modeled as a mathematical model as an objective function indicating the group management control performance evaluation, and the performance evaluation is digitized (step S4). The evaluation result for the control parameter combination P is PE _P , and the inference result is evaluated as follows.

[0087]

[Equation 3]

Here, β (= (β ₁ , β ₂ , ..., β
_n )) is an inference result evaluation parameter,
It is a data array that is set in advance for each traffic demand and time zone according to customer specifications.

When the performance evaluation value PE _P ; (P = 0 to Pmax) is calculated for all the combinations P (step S5), the combination P having the smallest PE _P is set to P _0, and this combination P the control parameter P ₀ corresponding to _{P 0 (α 1P0, α 2P0} , ......, α lP0) for the best control parameter and transmits to the group management control unit 1 (step S6).

As described above in detail, in the elevator group management control device of this embodiment, the learning including the inference unit 21, the partial model 22, the synthesis unit 23, the inference result evaluation unit 24, and the system adjustment unit 25. By configuring the control unit 1-1, it is possible to optimize the control parameter, which is a weighting value for each evaluation index in the group management performance, according to the traffic demand, in the allocation control of the optimum number machine for the hall call. , It is possible to automatically set the optimum control parameters even for the evaluation criteria unique to each building.

The inference unit 21 and the partial model unit 22 are
Since the system adjustment unit 25 performs online learning based on the result of the group management control response, it is possible to form an autonomous system with high application capability.

The present invention is not limited to the above embodiment, but can be implemented in the same manner as described below. (A) In the present invention, it is obvious that the detailed configuration of the group management control response result, the control parameter, the traffic demand, etc. can be treated in the same manner even if they are changed within their intended ranges.

(B) In the first embodiment, the case where the elements are weighted by the inference result evaluation parameter and linearized for evaluation of the group management control response inference result has also been described. Not limited to this, it is also possible to set a reference value of an ideal response result in advance, determine the response result with the minimum deviation as the optimum, and set the control parameter. The evaluation of the control response inference result can be appropriately changed within the intended range.

(C) In the first embodiment, among the input / output data acquired online during the predetermined time period, data that exceeds the predetermined value D _{ef0 at} which there is a difference between the inference result and the response result is valid data. And the inference unit 2 based on this difference.
Although the certainty of 1 and the case of modifying the related partial system model f _i have been described, the present invention is not limited to this, and a method of extracting effective data from the input / output data in the time zone is described in the first embodiment. In addition to the above method, there is a method based on the activity for the partial system model.

The second embodiment in this case will be described below with reference to FIGS. 12 and 13. The method of evaluating the group management control response inference result corresponding to the combination of each control parameter from the traffic demand calculated by the inference result evaluation unit 24 for each constant period for each time zone and setting the optimum control parameter is As described in the first embodiment, in the process, the response inference result is calculated by the linear sum of the activities of the partial model 22 and the partial model 22 from the inference unit 21 as shown in the equation (2). To be done.

Each time the difference detecting unit 25-1 executes the calculation of the optimum control parameter for a certain traffic demand, as shown in FIG. 12, whether or not the activity level a _i exceeds a predetermined value A _1im . Check for all models. As a result, when the activity a _i exceeds the predetermined value A _1im , the set control parameter and response result are regarded as valid data, and the number is input / output data correction unit 25-3.
Output to. At this time, in the input / output data correction unit 25-3, the control parameter and the response are added to the table having the data structure as shown in FIG. 13 together with the partial model number input from the difference detection unit 25-1. Store the result.

Further, the input / output data correction section 25-3 sorts the data group stored based on the above number into any of the input / output data 25-41 to 4m, and updates the data when the data is updated. The relevant partial model is trained using the data as teacher data.

The input / output data correction unit 25-3 selects and aggregates similar data including the input / output data i25-4i from the data extracted by the method of the second embodiment. Then, there is a method in which the input / output data 25-4i is updated and the partial system model correction unit 25-5 learns the partial system model f _i using the above data as teacher data. As a result, the input / output space can be expanded and the response result (output data) can be inferred for various input data.

Here, an example of a method of selecting (cluster analysis) and aggregating similar data will be described with reference to FIGS. 14 and 15. Cluster analysis is a numerical method that collects "similar things" into clusters (clusters) based on multiple measured values for an object (individual). The representative longest distance method is used.

First, consider n individuals (data) as n clusters. Next, dαβ is defined as the distance between α cluster and β cluster. Now, let d _{h1 be the} minimum dαβ for all α and β. Then, two h clusters and l clusters having this minimum distance are fused. If such an operation is repeated (n-2) times, two clusters remain at the end. When the h cluster and the l cluster are combined to form the g cluster, the distance d _fg between the f cluster not belonging to the g cluster and the g cluster is
If it is obtained from only the distances d _fh , d _f1 , and d _h1 before being combined, the relationship of the formula (12.1) is considered as the calculation formula of d _fg .

D _fg = α _hd f _h + α _1d f ₁ + βd _h1 + γ | d _fh −d _f1 | (12.1) Here, α _h , α ₁ , β, and γ are all parameters (constants). is there. The above relationship diagram is shown in FIG.

The longest distance method adopts the distance between the furthest points in two clusters as the similarity, and (1
2.1), α _h = α ₁ = 1/2, β = 0, γ = 1/2
Then, the expression (12.2) is obtained.

D _fg = (d _fh + d _f1 + | d _fh −d _f1 |) / 2 (12.2) = d _fh (when d _fh ≧ d _h1 ) = d _h1 (d _fh <d _h1 14) As shown in FIG. 14, when h and l form one object each, and when they form a g cluster, d _fh ≧ d _h1
Then, d _{fg of the} longest distance method is equal to d _fh . When the center of the g cluster is m, and d _fh is taken as d _fg , the distance between f and m is d _fh, and the point f moves to the point f ′ in FIG.

As described above, when the clusters fuse with each other, the non-fused portions around the clusters are separated from the fused portions. That is, there is a tendency that objects that do not belong to the cluster move away from the cluster, in other words, spatial diffusion occurs. By the above method,
FIG. 15 shows the process of repeating the fusion of the clusters to make the last n clusters into one cluster (hereinafter referred to as a tournament diagram), and FIG. 16 and FIG. 17 show the processing procedure. .

It is shown that the distance increases as the tournament progresses upward, and the number at the bottom corresponds to the number of n data. The process of fusion will be explained using the clusters of Nos. 1, 2, and 5. First, the clusters of No. 2 and No. 5 are fused to become the clusters of (n + 1), and then the clusters of No. 1 and (n + 1). No. cluster is fused to become (n + i) cluster. This fusion will be repeated for all clusters.

Now, when clustering is performed at the distance d ₁ , 2nd and 5th, 3rd and 4th can be regarded as similar data, and 2nd and 5th, 3rd and 4th respectively. Center
By using the fusion data or something of each as the representative data, n pieces of data can be aggregated into (n−2) pieces.

Next, the fusion process will be described with reference to FIGS. 16 and 17. First, the distance DSM (i, j) between the points i and j is generated for all n data. Next, the minimum point combination in the distance DSN (i, j) is obtained. The combination of the points is taken as h, 1 point, and the g point is calculated by fusing the h, 1 point by the longest distance method. And g
The distance DSM (k, cnt) between a point and another point is generated.
When the above process is repeated (n-1) times, a tournament diagram as shown in FIG. 15 is completed.

In the third embodiment, the input / output data i2
5-4i selects similar data including past data,
Although the case of performing aggregation has been described, the invention is not limited to this.
Furthermore, in the fourth embodiment, among the input / output data, as shown in FIG.
There is a method of re-learning the partial model with new data only, by discarding the date data of the data structure shown in (1) after a certain period of time. As a result, it becomes possible to appropriately respond to the demand of the building which changes in time series, and to always perform optimum group management control.

In the above second to fourth embodiments, the case where all the extracted effective data are made to be learned by the partial model has been described, but the present invention is not limited to this, and there is a limit to the data that the partial model can learn. In the fifth embodiment, the number of valid data is further limited from the extracted data to relearn the partial model. As a method of limiting the number in this case, a combination of the third and fourth embodiments will be described here with reference to FIG.

The input / output data has a table structure, and newly extracted data is sequentially saved in the table. That is, the smaller the table index, the older the data. Now, a tournament diagram as shown in FIG. 15 is created by the longest distance method for n pieces of input data in which old and new data are mixed. In FIG. 15, the smaller the cluster number, the older the data.

A procedure for limiting the number of n data to (n-4) will be described below. First, clustering is performed at the distance d ₁ to aggregate n pieces of data into (n−2) pieces.
At this time, the 2nd and 5th data are fused to the (n + 1) th data, and the 3rd and 4th data are fused to the (n + 2) th data, respectively.

Next, in order from the left, the two data (n + 1) fused with the old data No. 1 are discarded, and (n-
4) Limit to only one. Finally, the sixth embodiment will be described.

In the first embodiment described above, the correction of the reliability of the inference unit is executed every time the actual data is input / output online, but in the sixth embodiment, it is extracted. When the partial model f _{i is} relearned by the input / output data i, the reliability of the memory for the partial model f _i is corrected by the equations (5.1) and (5.2).
This makes it possible to improve the efficiency of processing and correct the accuracy of memory.

[0114]

As described above, according to the present invention, the relationship between the control parameter of each evaluation index and the group management control response result is composed of a function model synthesized by a neural network in which it is vaguely divided according to the traffic demand. Partial model means modeled as a partial model set, inference means relating the relationship between each partial model and traffic demand by a plurality of membership functions, and group management control response results are calculated by synthesis from the inference means and partial model means By having a learning control unit comprising a synthesizing means for, because it is possible to quantitatively estimate the relationship between the control parameter of each evaluation index and the group management control response result for any traffic demand for each building, For conventional systems where it was difficult to set optimal control parameters that changed due to traffic demand, for continuously changing traffic demand, It is possible to set the optimum control parameter for each time period.

Further, even when the "optimal" evaluation criterion differs depending on the building use such as a hotel, a tenant building, a single company building, etc., optimum control parameters can be set according to each building use, Individualization can be realized for each building. In addition, even if the “optimal” evaluation standard is changed in the middle due to the change in the customer's specifications, it is possible to quickly respond by changing the inference result evaluation parameter.

Furthermore, according to the present invention, even if there is no data that completely matches the traffic demand that has occurred after the building starts, each control parameter can be controlled by the associative function of the system composed of the inference means, the partial model means, and the synthesis means. It outputs the group management control response estimation result to the system and can respond to any traffic demand. Only valid data is extracted online from the actual system corresponding to a certain period, and the target part is collected. The internal structure of the model means and the inference means can be automatically changed to form a highly capable autonomous system, and the performance of the CPU is improved and the memory is saved. Due to these characteristics, it is possible to realize the optimal allocation control of the optimal number of hall calls for the various hall configurations and customer needs.

[Brief description of drawings]

FIG. 1 is a block diagram showing a first embodiment of an elevator group management control device according to the present invention.

FIG. 2 is a block diagram showing a configuration example of a software system of a group management control unit and a unit control unit in the embodiment.

FIG. 3 is a block diagram showing a system configuration example of a high-speed transmission line in the embodiment.

FIG. 4 is a functional block diagram showing a flow of input / output signals of a learning control unit in the embodiment.

FIG. 5 is a block diagram showing a system configuration example of a learning control unit in the embodiment.

FIG. 6 is a block diagram showing a system configuration example of an inference unit in the embodiment.

FIG. 7 is a block diagram showing a system configuration example of a partial model unit in the embodiment.

FIG. 8 is a schematic diagram showing a configuration example of each partial system model of the partial model section.

FIG. 9 is a block diagram showing a system configuration example of an inference result evaluation unit in the embodiment.

FIG. 10 is a flowchart for explaining the operation of the inference result evaluation unit.

FIG. 11 is a block diagram showing a system configuration example of a system adjusting unit in the embodiment.

FIG. 12 is a flowchart for explaining data extraction based on activity.

FIG. 13 is a diagram showing an example of a storage structure of learning data.

FIG. 14 is a conceptual diagram of the longest distance method.

FIG. 15 is a diagram showing an example of a tournament based on the longest distance method.

FIG. 16 is a flowchart for explaining a procedure for creating a tournament diagram by the longest distance method.

FIG. 17 is a flowchart for explaining a procedure for creating a tournament diagram by the longest distance method.

[Explanation of symbols]

1 ... Elevator group system, 1-1 ... Learning control unit, 2 ...
Elevator group system, 2-1 to 2-N ... Single control unit,
3 ... Hall call button, 4 ... Hall call input / output control unit,
5 ... Monitoring board, 6 ... High-speed transmission line, 7 ... Low-speed transmission line, 21 ...
Inference section, 22 ... Partial model section, 23 ... Synthesis section, 24 ... Inference result evaluation section, 25 ... System adjustment section.

Claims (6)

[Claims] 1. A plurality of elevators are put into service for a plurality of floors, a predetermined evaluation calculation is performed for a common hall call that has occurred, an evaluation is made for each elevator, and an optimum number of elevators is assigned to the hall. In an elevator group supervisory control device assigned to a call, a partial system model consisting of a functional model composed by a neural network in which the relationship between the weighting value of each evaluation index and the group supervisory control response result is ambiguously divided according to traffic demand. A partial model means that is modeled as a set of, and the relationship between the partial system model and the traffic demand is vaguely expressed by a plurality of membership functions, and the connection relationship between the membership function and the partial system model is stored, Inference means for computing a weighted value for the subsystem model based on demand; A combination means for calculating a group management control response inference result by combining outputs from the inference means and the partial model means, and a demand for the building in which the elevator system is installed, the group management control response inference result from the combination means. Inference result evaluation means that evaluates each independently and sets the optimum value of the evaluation index weighting value according to actual building demand, and the evaluation index set according to the traffic demand during system operation by the inference result evaluation means System management means for performing group management control with an optimum value of the weighting value, and changing each component of the partial system model, inference means, and inference result evaluation means from the obtained group management control response result. An elevator group supervisory control device comprising: 2. The system adjusting means considers that the activity of the partial system model based on the traffic demand output from the inference means exceeds a predetermined value as valid data, and the partial system is determined by the data. The elevator group supervisory control device according to claim 1, wherein the model is retrained. 3. The system adjustment means is a group management control response result which is an evaluation index weighting value or an output which is an input of the partial system model from effective data of the partial system model extracted from actual data during system operation. 3. The elevator group supervisory control apparatus according to claim 2, wherein the partial system model is re-learned by the data obtained by selecting and summing those whose similarity exceeds a predetermined value. 4. The system adjusting means discards the valid data of the partial system model extracted from the actual data when the system is operating, in order from the older one,
The elevator group supervisory control apparatus according to claim 2, wherein the partial system model is re-learned by the limited data. 5. The system adjusting means limits the number of valid data of the partial system model extracted from actual data during system operation according to the configuration of the partial system model, and uses the limited data according to the limited data. The elevator group supervisory control apparatus according to claim 2, wherein the partial system model is relearned. 6. The system adjusting means changes the degree of connection in the inference means so that the degree of activity with respect to the retrained partial system model is increased. Elevator group management control device described.