WO2001096967A1

WO2001096967A1 - Loading calculation system for a roll set using a neural network

Info

Publication number: WO2001096967A1
Application number: PCT/FI2001/000545
Authority: WO
Inventors: Helena LEPPÄKOSKI
Original assignee: Metso Paper, Inc.
Priority date: 2000-06-13
Filing date: 2001-06-08
Publication date: 2001-12-20
Also published as: WO2001096967B1; AU2001266107A1; DE10196350T1; FI20001403A; FI20001403A0; FI115406B

Abstract

The invention concerns a load calculation system for the calender of e.g. a paper machine. The invention is characterised in that the calender load calculation routine is speeded up with the aid of a neural network (5) in order to control load control parameters (23) in on-line machine running in particular.

Description

LOADING CALCULATION SYSTEM FOR A ROLL SET USING A NEURAL NETWORK

The present invention relates to load calculation of roll systems, e.g. a calender. More specifically, the present invention concerns a load calculation system for a roll system, preferably a calender, in accordance with the introductory part of claim 1.

The present-day calculation of a roll system in the calender automation system is a calculation, which is heavy and slow to implement. Thus nowadays, when still using a decentralised automation system, optimising calculation of idle rolls working in a computer, such as a unix work station, is needed in addition to the optimising calculation of variable crown rolls. When changing the machine parameters of the calender, any change of the linear load profile will require that optimisation is carried out in the automation system, and the situation is the same when changing the linear load level. If the load distribution between nips is changed or if rolls are exchanged, then calculation of the entire roll system must be done at the unix work station.

Thus it is a basic feature of the present system that by using the mathematical model of both the loading and/or relieving mechanical system and the TK roll either one or both optimisation problems are solved for a new load situation every time when the load situation is to be changed. Thus, the optimisation problem is solved only when this is required for a machine run. On the other hand, it is a weakness of the system that old results are lost when the load situation changes, although results calculated from the same given values might be needed later.

At present the load calculation of a roll system, such as a calender, is a rather heavy and time consuming routine when done by an automation system. This results in a great need of calculation resources in on-line solutions, so that to have an economically sensible calculation capacity less strict demands must be made on the time used for calculation. A heavy and slow calculation routine makes it difficult to take quick on-line control steps concerning the whole roll system with the loading/relieving equipment of the roll system, such as the calender, e.g. in order to control the quality of the paper to be produced or in protection steps taken in trouble situations. For quality control, speediness is important in controlling situations of change, e.g. when using load to compensate for a change in temperature of chilled rolls when starting a machine run. It is especially difficult to carry out ramping or any advanced control or compensation methods, in case the time taken by calculating the loading/relieving pressure values corresponding with a new linear load situation is from half a minute to a few minutes, and if it also varies, that is, the calculation delay is a stochastic variable for the control algorithm.

The primary objective of the invention is to eliminate or at least to reduce the problems and weaknesses relating to the present load calculation and to bring about a new and inventive on-line application for load calculation.

This objective is achieved by an on-line application of load calculation of the kind mentioned in the beginning, the characteristic features of which are presented in the appended claims.

The invention is based on the new and inventive basic idea that the calculation routine for calculating the load of a roll system, such as a calender, is speeded up with the aid of a neural network, which is pre-instructed and which outputs pre- ins ructed optimisation of the given values of roll system optimisation and optimisation of the roll system. The relation between the results of the calculation is implemented by the given values of the neural network as control parameters for the roll system. Thus, in the system according to the invention, the neural network is pre- instructed in optimised target values. In this case the neural network thus operates as a data bank, from which such a target result or target values are retrieved when required for use, which have been calculated in advance for solving the problem. This saves both time and equipment costs. The input-output search performed by the instructed and/or constantly learning neural network is quick compared with iterative calculation. Savings are achieved in equipment costs, because in load calculation in an on-line machine run there is no longer any need for equipment required for searching an efficient iterative optimum solution (for each roll system/calender), but equipment performing iterative load calculation is only needed in the stage of designing the machine/roll system/calender to instruct the neural network intended to control the machine/roll system/calender. In addition, memory is nowadays cheap as an information technology equipment component, so the costs of recording solutions that were found earlier are not high.

According to the invention, the neural network is instructed in the off-line state in solutions to the optimisation problem. Hereby it is suitable to use an efficient computer, such as e.g. a unix work station. Deterministic models of mechanical systems are needed in the instruction such as are in use also nowadays. This means that the models may be implemented in the manner of today. However, the invention does not require this. Any sufficiently accurate model a d a combined working optimisation method will do. The instruction algorithm is presented with a sufficient number of possible combinations of given data, which include normal variable data of the calender model, such as e.g. linear load levels, desired load distributions, properties of the rolls in the roll system, etc., and for which the instruction algorithm asks the optimisation algorithms to calculate a solution. The neural network learns the relation of "given optimisation data" "* "optimum solution of the problem" and a relation repeat mechanism is recorded as record data in the neural network's memory to define the operation of the neural network as parameters (= pressure coefficients), which can be used later in input-output search to control the loading of the roll system/calender optimally in changing machine run situations.

It is an advantage of this invention that the on-line implementation of the roll system calculation need no longer perform any iterative optimisation routines. In the on-line calculation only pre-calculated parameters of the neural network (= pressure coefficients) and a corresponding neural network structure are needed. All possible combinations of given data cannot be solved in the network instruction. However, the network can also cope with problems where it is asked for outputs corresponding with situations that do not occur in the instruction, when such a network structure is chosen for use, which is able by interpolating to produce new solutions on the basis of known solutions. This is taken into account in the choice of instruction cases - e.g. extreme states of all possible cases must be included in the instruction material.

Another advantage of the invention is that the solutions are calculated beforehand, so it is possible to study the rationality of results already at the design stage. In this way such surprises can be prevented, that when using the on-line application a certain combination of given data produces a solution which is feasible according to the model, but which had better not be implemented in production conditions by the calender's loading/relieving equipment judging by a criterion not included in the mathematical presentation of the problem, e.g. judging by common sense.

The invention also allows on-line learning of the neural network, but savings in equipment costs are not achieved hereby. Instead time is saved in that an optimisation problem once solved need not be solved again, but solutions earlier learnt by the network can be repeated over and over again. This function also makes demands on the structure of the network, instruction of the network must be quick and the network must not forget what it has learned earlier in the instruction. A solution of this kind does not either allow pre-checking of the results of learning before their on-line use as mentioned in the previous paragraph. The invention will be described in greater detail in the following with the aid of its one advantageous embodiment by referring to the appended patent drawing, which is a schematic view of neural network instruction indicated by the area delimited by a dotted line, and operation in an on-line machine run situation indicated by the area delimited by a double point dashed line.

In connection with the figure it should be underlined to begin with that it is simplified and that the system may of course include also other functions than those presented hereinafter. Likewise, data may be recorded in several different data stores and may come to the algorithm from several different sources.

The neural network illustrated in the figure may be e.g. a network of the back propagation or radial base type. The literature also presents other neural network solutions and combinations of fuzzy calculation, of which the DCA network (Dynamically Capacity Allocating network) may be mentioned as an example. According to the basic idea of the present invention, the neural network 5 is instructed in the off-line state in solutions to the optimisation problem. It is hereby suitable to use an efficient computer, and an advantageous one is the UNIX work station. The instruction uses deterministic models of mechanical systems, and suitable such are any sufficiently accurate model and a combined functioning optimisation method.

To instruct the neural network 5, instruction algorithm 4 is presented with several different combinations 10 of given data 1, which include normal variable data of the calender and roll system, such as e.g. linear load levels, desired distributions of the load, properties of the rolls in the roll system, etc. With the aid of an optimisation request 10b and parameters 10c and an optimisation algorithm 2 to be entered in the mathematical representation, that is, a simulation 3, describing the mechanical properties or the mechanical process of the machine, which is preferably a roll system or a calender, an optimum solution 12b to the problem is calculated from the given data by simulating with the simulation model 3, which optimum solution 12b is input to the instruction algorithm 4. In this way the neural network 5 is instructed with the aid of the relation "given data of optimisation" ^→ " optimum solution of the problem", and the relations are recorded as record data 14b in the memory of the neural network's parameter table 6, which parameters 22 of the neural network can then be utilised in an on-line machine run situation in a search implementing "given data" 21 of roll system optimisation ^→ "control parameter 23 of roll system" to control the machine load values 23, e.g. hydraulic pressures 7, of the machine, such as a roll system/calender, to be optimum ones in varying machine run situations.

The figure is a schematic view of neural network instruction in roll system calculation. The given data 1 contains various given data alternatives, for which the optimum solution is sought. Given data 1 is e.g. characteristics of the rolls, the roll system and the hydraulic system and the desired linear load profiles in the different nips. This constitutes the input part 10 of the neural network's 5 instruction data, which is sent to the neural network's 5 instruction algorithm 4. In order to calculate the output part of the neural network's 5 instruction, that is, the target values, the given data and the optimisation request 10b are sent to the optimisation algorithm 2, and of the given data the part 10c of simulation model parameters is relayed to the simulation to be done by a mathematical model 3 of the machine, which is preferably a roll system or a calender. The optimisation algorithm 2 seeks iteratively by simulating with the mathematical model 3 of the system those optimisation results 11, e.g. pressure values for the hydraulic system, with which the desired values 12, such as the desired linear loads, are best realised. When sufficiently good optimisation results, that is, given values, are found, the optimisation algorithm 2 will relay the best found given values 12b, such as pressure values, to the neural network's instruction algorithm 4, whereupon the instruction algorithm 4 will relay the instruction data 13 to the neural network 5. The instruction data consists of the given data 10b used in optimisation and of the corresponding optimised given values 12b. Of these the latter thus function as target values for the instruction of the neural network. In order to calculate the neural network parameters or pressure coefficients 22 controlling the calculation of the neural network 5 and needed in rurming the machine, such as a roll system or a calender, the instruction algorithm 4 seeks from the memory 6 of the neural network parameters 22 by iterative calculation with the neural network 5 such parameter values 22, with which the distance of the given values 14 produced by the network 5 from the optimised given values 12b, such as given pressure values, is short enough. The neural network's parameter values 22 obtained by iterating in the neural network instruction situation are recorded as record data 14b in memory 6 of the control system's parameters 22, from which they can be sought by a simple keying of given value 21 -control parameter 23 as optimum load control values for the machine, such as a roll system or a calender, in varying machine run situations.

In order to use the neural network according to the invention for load calculation in an on-line machine run situation, the neural network 5 is supplied with network parameters 22 describing the mechanical and/or hydraulic system and with the desired load values 21. From these the neural network 5 by keying given value 21- roll system's control parameter 23 produces optimum control or regulating values 23, such as hydraulic pressures. Thus, the neural network 5 with the aid of network parameters 22 seeks or calculates the desired loading situations 1, 21, the best realisable control or regulating values 23 for the roll system, such as setting values for hydraulic pressures.

The neural network 5 in accordance with the invention is pre-instructed in the possible optimisation results. Thus, in an on-line machine run situation the neural network 5 functions as a data bank, from which the results are sought for use, which have been calculated beforehand to solve the problem. This saves both time and equipment costs. The input-output search performed by the instructed neural network is also essentially quicker than iterative calculation.

It should be noted that also in an on-line machine run situation the neural network 5 may be arranged to be learning. This is preferably done in such a way that the instruction algorithm 4 and also the optimisation algorithm 2 and the mathematical model 3 are kept in connection with the neural network 5, whereby, when the input of the neural network 5 is a new formerly unknown given value 21, a first simulation or iteration is first performed with the new desired given value 10b, 10c with the aid of the optimisation model 3 and the optimisation algorithm 2 and the optimised given value 12b is relayed as input to the instruction algorithm 4, which relays it as instruction data to the neural network 5. In order to get new parameters 22 needed when running the machine, the instruction algorithm 4 by interpolating with the neural network 5 calculates such new control parameter values 22, with which the distance of the output 14 produced by the network is sufficiently short from the given values 12b. The interpolated new parameter values 22 of the network are recorded as record data 14b in the memory 6 of the network parameters 22 of the control system according to the invention, from which they can again be found for use by a simple search done on the basis of the given data.

The invention was described above only with the aid of its one embodiment, which is considered advantageous. Of course, this is not intended to limit the invention, and as is obvious to professionals in the art, many alternative solutions and modifications are possible within the scope of the new and inventive idea defined in the appended claims.

Claims

1. Load calculation system for a roll system such as a calender in a paper machine or board machine or other such fibrous web machine, characterised in that the load calculation routine is speeded up with the aid of a neural network (5), which is pre-instructed and which gives roll system control parameters (23) corresponding with pre-instructed and desired given data or values (1, 21) and/or which in a machine running situation interpolates new roll system control parameters (23) from the neural network's control parameters (22) corresponding with the given data or values pre-instructed for the desired new given data or values (1, 21).

2. Load calculation system according to claim 1, characterised in that the given data or values (1, 21) are chosen from a set including: linear load levels and/or profiles, load distributions, properties of the roll system, the rolls and/or the hydraulic system, and that the control or regulating values (23) produced by the neural network (5) are preferably hydraulic pressure values.

3. Load calculation system according to claim 1 and 2, characterised in that the load calculation system is a computer system including: a mathematical model (3) of the mechanical process of the calender's roll system, which of the given data (1) receives the part (10c) of the simulation model's parameters, and an optimisation algorithm (2), which receives an optimisation request (10b) and which based on the given data (1) with the aid of a mathematical model seeks iteratively such optimised given values (12b), with which the desired given data or values (1) are realised; the neural network's (5) instruction algorithm (4), which receives the optimised given values (12b) and the given data part (10) of the neural network's (5) instruction; and a neural network (5), which from the instruction algorithm (4) receives instruction data (13) and to which a memory (6) of the neural network's parameters (22) is connected.

4. Load calculation system according to any one of the claims 1-3, characterised in that in order to get new roll system control parameters (23) needed in the machine rimning situation, the instruction algorithm (4) calculates by interpolating with the neural network (5) with the aid of recorded parameters of the neural network such new roll system control parameter values, whose distance from the corresponding values (12b) obtained by optimisation is sufficiently short.

5. Load calculation system according to claim 4, characterised in that the instruction algorithm (4) seeks through iterative calculation with the neural network (5) and records in the memory (6) of the network parameters (22) such values for the network parameters (22), with which the neural network produces network given values (12b), which as well as possible correspond with the given values (12b) of the optimisation calculation, and that the neural network (5) produces roll system control or regulation values (23), such as hydraulic pressures (7), corresponding with the desired given values (1, 21).

6. Load calculation system according to any one of the preceding claims 1-5, characterised in that through optimisation (2) by simulating with a mathematical model (3) target values (12b) are produced for the neural network (5) to instruct the neural network in the connection between the optimisation calculation of given data or values (1,21).

7. Load calculation system according to claim 6, characterised in that instruction of the neural network includes a given data part (10) of the neural network's instruction, which is relayed to the instruction algorithm (4), and a calculation part of target values for the neural network's instruction, wherein the given data (1) and the optimisation request (10b) are relayed to the optimisation algorithm (2) and wherein of the given data (1) the part (lie) of the mathematical model's parameters is relayed to simulation to be done with the mathematical model (3), that the optimisation algorithm (2) by iterative simulation with the mathematical model (3) seeks those values (11), with which the desired given data (1) is best realised, whereupon the optimised values (12b) are relayed to the instruction algorithm (4) of the neural network, which algorithm relays to the neural network (5) instruction data (13) including optimised given values (12b) corresponding with the given data (10b) relayed to the optimisation algorithm (2), which given values thus function as target values for the neural network's (5) instruction.