Supporting method and system for process operation

• US 5,845,052 A
• Filed: 01/02/1996
• Issued: 12/01/1998
• Est. Priority Date: 03/13/1989
• Status: Expired due to Fees

• Current Assignee: Hitachi, Ltd.
• Original Assignee: Hitachi, Ltd.
• Inventors: Baba, Kenji, Matsumoto, Hiroshi, Yoda, Mikio, Watanabe, Shoji, Hara, Naoki, Enbutsu, Ichiro, Matsuzaki, Harumi, Nogita, Shunsuke, Yahagi, Hayao, Maruhashi, Fumio
• Primary Examiner(s): Davis, George B.
• Application Number:
  US08/582,018
• Time in Patent Office: 1,064 Days
• Field of Search: 395/23, 395/24, 395/22
• US Class Current: 706/33
• CPC Class Codes: C02F 1/008 Control or steering systems not provided for elsewhere in subclass C02F
• CPC Class Codes: C02F 1/5245 using basic salts, e.g. of aluminium and iron
• CPC Class Codes: C02F 1/66 by neutralisation; pH adjustment for degassing C02F1/20; using ion-exchange C02F1/42; for flocculation or precipitation of suspended impurities C02F1/52; for removing dissolved compounds C02F1/58
• CPC Class Codes: C02F 1/76 with halogens or compounds of halogens C02F1/4674 takes precedence
• CPC Class Codes: C02F 2001/007 Processes including a sedimentation step
• CPC Class Codes: C02F 2209/02 Temperature
• CPC Class Codes: C02F 2209/05 Conductivity or salinity
• CPC Class Codes: C02F 2209/06 pH
• CPC Class Codes: C02F 2209/07 Alkalinity
• CPC Class Codes: C02F 2209/11 Turbidity
• CPC Class Codes: C02F 2209/29 Chlorine compounds
• CPC Class Codes: C02F 2209/42 Liquid level
• CPC Class Codes: C02F 2303/04 Disinfection
• CPC Class Codes: G05B 13/027 using neural networks only
• CPC Class Codes: G05B 13/0285 using neural networks and fuzzy logic
• CPC Class Codes: Y02W 10/37 using solar energy
• CPC Class Codes: Y10S 706/903 Control
• CPC Class Codes: Y10S 706/906 Process plant

Abstract

A method for causing a neural circuit model to learn typical past control results of a process and using the neural circuit model for supporting an operation of the process. The neural circuit model is caused to learn by using, as input signals, a typical pattern of values of input variables at different points in time and, as a teacher signal, its corresponding values of the control variable. An unlearned pattern of input variables is inputted to the thus-learned neuron circuit model, whereby a corresponding value of the control variable is determined. Preferably, plural patterns at given time intervals can be simultaneously used as patterns to be learned.

Claims

1. A method for extracting as knowledge relationships between input variables and an output variable of a neural circuit model, said neural circuit model being of a hierarchical structure constructed of an input layer, at least one hidden layer and an output layer, said neural circuit model having performed learning a limited number of times by determining weight factors between mutually-connected neuron element models in different layers of the input layer, hidden layer and output layer, said method comprising the steps of:

determining a relationship between a particular input variable and a particular output variable with respect to plural routes extending from a neuron element model corresponding to the particular input variable of the input layer to a neuron element model corresponding to the particular output variable of the output layer by way of individual element models of the hidden layer,

wherein said determining step includes the steps of;

obtaining a product of the weight factors for each of the plural routes and a sum of the products of the weight factors for the plural routes, and

employing said sum of the products of the weight factors as a measure of the relationship between the particular input variable and the particular output variable.

2. An apparatus for extracting as knowledge relationships between input variables and an output variable of a neural circuit model, comprising:

a neural circuit model being of a hierarchical structure constructed of an input layer, at least one hidden layer and an output layer, said neural circuit model having performed learning a limited number of times by determining weight factors between mutually-connected neuron element models in different layers of the input layer, hidden layer and output layer; and

a control device for determining a relationship between a particular input variable and a particular output variable with respect to plural routes extending from a neuron element model corresponding to the particular input variable of the input layer to a neuron element model corresponding to the particular output variable of the output layer by way of the individual element models of the hidden layer, said control device obtaining a product of the weight factors for routes and a sum of the products of the weight factors for the plural routes, and employing said sum of the products of the weight factors as a measure of the relationship between the particular input variable and the particular output variable.

3. An apparatus for supporting an operation of a process, comprising:

a neural circuit model of a hierarchical structure constructed of an input layer, at least one hidden layer, an output layer and a teacher layer; and

a control device for determining a control value of a control variable for a target, to be controlled, in accordance with values of time-dependent input variables so as to bring the target closer to a desired state,

said control device causing the neural circuit model to learn, out of information on a past operation history of the process, a typical pattern of values of input variables at different points in time as input signals and a value of the control variable, the control value corresponding to the typical pattern, as teacher signal, said values of input variables at different points in time being simultaneously applied to said input layer while said value of the control variable being applied to said teacher layer so that the neural circuit model generates, after the learning, the value of the control variable from said output layer in response to said typical pattern of values of input variables given to said input layer, and

said control device inputting as the values of the input variables, an unlearned pattern to the thus-learned neural circuit model, to determine its corresponding value of the control variables.

4. An apparatus according to claim 3, wherein after plural input signals are learned by simultaneously using, as the input signals, a pattern of values of the input variables at a given time point and a pattern of values of the input variables at a time point a predetermined time interval before the given time point and using a value of the control variable at the given time point as a teacher signal a pattern of values of the input variables of a present time point and a pattern of values of the input variables at a time point a predetermined time interval before the present time point are simultaneously inputted to the neural circuit model, whereby a value of the control variable for the present time point is determined.

8. An apparatus according to claims 3, 4, 5, 6 or 7 wherein relationships between the values of the input variables and the corresponding values of the control variable are extracted on the basis of the results of the learning by the neural circuit model and the operation of the process is supported based on the relationships.

9. An apparatus according to of claims 3, 5, 6 or 7, wherein the neural circuit model has an input layer formed of plural neuron element models, at least one hidden layer formed of plural neuron element models for receiving outputs from the neuron element models of the input layer and an output layer formed of at least one neuron element model for receiving outputs from the neuron element models of said at least one hidden layer, the input variables are assigned to the respective neuron element models of the input layer and the control variable is assigned to the neuron element model of the output layer and the learning is performed by controlling weight factors applied to connections between the individual neuron element modes.

10. An apparatus according to claim 9, wherein the connection strengths between the individual input variables and the individual control valuables are determined on the basis of the results of the learning by the neural circuit model, and the network of the neural circuit model is modified based on the magnitudes of the connection strengths.

11. An apparatus according to claim 9, wherein the connection strength between a specific input variable and a specific control variable is defined by the sum of products of weight factors for individual routes from the neuron element model, corresponding to the specific input variable of the input layer to the neuron element model, corresponding to the specific control variable of the output layer via the neuron element model of the hidden layer.

12. An apparatus according to claim 10, wherein, the network of the neural circuit model is modified by eliminating the connection between particular neuron element models.

13. An apparatus according to claim 10, wherein the network of the neural circuit model is modified by using a varied number of hidden layers.

14. An apparatus according to claims 3, 5 or 7, wherein a neural circuit model which has learned information on a steady-state operation history and another neural circuit model which has learned information on a non-steady-state operation history are provided independently, and in supporting the operation the neural circuit models are switched over depending on whether the operation is to be in a steady-state or in a non-steady-state.

15. An apparatus according to claims 3, 5 or 7, wherein a plurality of neural circuit models which have learned information on different operation histories, respectively are provided for variations of the pattern of values of the input variables, respectively, and in supporting the operation, the plurality of neural circuit models are switched with one another depending on the variation of the pattern of values of the input variables.

5. An apparatus for supporting an operation of a process, comprising:

a neural circuit model of a hierarchical structure constructed of an input layer, at least one hidden layer, and an output layer and a teacher layer; and

a control device for determining a value of a control variable for at least one target to be controlled in accordance with values of time-dependent input variable such that the target can be brought into a desired state,

said control device causing the neural circuit model to successively learn a plurality of patterns of values of input variables, each by using, as input signals, patterns of values of the input variables at different times of attainment of control of the target into the desired state, and as teacher signals values of the control variables, said values corresponding to the patterns,

wherein in the learning of one pattern of the values of input variables said values of input variables at different times are simultaneously applied to said input layer while a corresponding value of the control variable are applied to said teacher layer so that the neural circuit model generates, after the successive learning, a desirable value of the control variable from said output layer in response to an unlearned input pattern of values of input variables given to said input layer, and

said control device inputting, as the values of the input variables, a given pattern of values of the input variables to the thus-learned neural circuit model to determine a value of the control variable for said given pattern.

6. An apparatus according to claim 5, wherein after plural input signals are learned by simultaneously using, as the input signals, a pattern of values of the input variables at a given time point and a pattern of values of the input variables at a time point a predetermined time interval before the given time point and using a value of the control variable at the given time point as a teacher signal a pattern of values of the input variables of a present time point and a pattern of values of the input variables at a time point a predetermined time interval before the present time point are simultaneously inputted to the neural circuit model, whereby a value of the control variable for the present time point is determined.

7. An apparatus for supporting an operation of a process, comprising:

a neural circuit model, having a hierarchical structure constructed of an input layer, at least one hidden layer, an output layer and a teacher layer, for supporting an operation of a process; and

a control device for determining a value of a control variable for a target to be controlled in accordance with values of time-dependent input variables so as to bring the target closer to a desired state,

said control device causing the neural circuit model to learn plural input signals by simultaneously using, as the input signals, at least two patterns of a pattern of values of the input variables at a given time point, a pattern values of the input variables at a time point a predetermined time interval before the given time point and a pattern of differences between the values of the input variables at the given time point and those at the time point a predetermined time interval before the given time point and using a corresponding value for the control variable at the given time point as a teacher signal,

wherein in the learning of the plural input signals said at least two patterns are simultaneously applied to said input layer while the corresponding value of the control variable are applied to said teacher layer so that the neural circuit model generates, after the learning, a desirable value of the control variable from the output layer in response to unlearned input signals of said at least two patterns given to said input layer, and

said control device inputting patterns at a present time point simultaneously to the thus-learned neural circuit model to determine a value of the control variable for the present time said patterns corresponding to said at least two patterns.

16. An apparatus for supporting an operation of a process, comprising:

a neural circuit model of a hierarchical structure constructed of an input layer, at least one hidden layer, an output layer and a teacher layer;

a transcendental model adapted to output a target value of a control quantity on the basis on an observation quantity of the process; and

a control device for calculating the difference as an error between an output value from the transcendental model and an actually measured value of the control quantity,

said control device inputting observation quantities at plural time points having a certain time interval in the past simultaneously to the input layer of the neural circuit model and inputting a corresponding error to the teacher layer of the neural circuit model, learning by the neural circuit model a relationship between the observed quantities and the error so that the neural circuit model generates, after the learning, the corresponding error from said output layer in response to the learned observation quantities given to said input layer, and

said control device inputting observation quantities at desired time points having said time interval to the thus-learned neural circuit model to have the neural circuit model predict an error at the corresponding time point, and outputting a target value of the control quantity on the basis of the output value from the transcendental model and the error predicted by the learned neural circuit model.

17. An apparatus for supporting an operation of a process, comprising:

a neural circuit model of a hierarchical structure constructed of an input layer, at least one hidden layer, an output layer and a teacher layer;

a transcendental model having a parameter adjustable to output a target value of a control quantity on the basis of an observed quantity of the process; and

a control device for inputting observation quantities at plural time points having certain time interval in the past simultaneously to the input layer of the neural circuit model and inputting a value for the adjustable parameter of the transcendental model at a corresponding time point to the teacher layer of the neural circuit model,

said control device causing the neural circuit model to learn a relationship between the observed quantities and the value of the parameter so that the neural circuit model generates, after the learning, the value of the parameter from said output layer in response to the learned observation quantities given to said input layer; and

said control device inputting observation quantities at desired time points having said time interval to the thus-learned neural circuit model to have the neural circuit model predict a value of the parameter at the corresponding time point, and obtaining an output value of the transcendental model from the learned neural circuit model by using the thus-predicted value of the parameter and then outputting a target value of the control quantity on the basis of the thus-obtained output value of the transcendental model.