GB2283834A - Adaptive process control system - Google Patents

Description

1 ADAPTIVE PROCESS CONTROLLER AND METHOD

BACKGROUND OF TEE INVENTION

Field of the Invention:

2283834 This invention generally relates to the field of process controllers and, more particularly, to process controllers which employ models of the process being controlled.

Description of the related art including information disclosed under 37C. F.R. 51.97 - 1.99

Recently, artificial neural networks (ANNfs) have been exploited for identification and control of dynamical systems as seen in Narendra, K.S., and Parthasarathy, K., March 1990, "Identification and Control of Dynamical Systems Using Neural Networks", IEEE Transactions on Neural Networks, Vol. 1, No. 1, pp. 4 - 27. Artificial neural networks have been applied to motion control, process control and aircraft control. Lan, M.S. , November 1990, "Learning Tracking Controllers for Unknown Dynamical Systems Via Neural Network Approach", Proceedings of the 1990 IEEE International Conference on Systems, Man, and Cybernetics, pp. 29 - 31. Neural Networks for Control, 1990, edited by T. Miller III, R.S. Sutton, and P. J. Werbos, MIT Press. Several attributes of ANN's make them attractive for process modeling and control. These attributes include the ability to approximate arbitrary nonlinear relations and adaption and learning. Moreover, when implementing ANN's in hardware such as in very large scale integration (VLSI), the parallel distributed processing architecture allows fast processing and provides a degree of robustness through fault tolerance and graceful degradation.

1.1 2 However, a problem with these controllers is that the neural networks in which the models are embodied respond to all detected disparities between a selected attribute of the actual output or result of the process and a desired attribute result even when those disparities or errors are caused by known causes or known types of causes with respect to which modification of the model is not desired. Consequently. in such systems, the models are caused to change in inappropriate ways that are not necessary to achieve the desired result, since such disparities are either intended or are from known or measured causes for which other aspects of the control system are designed to correct. Generally, the neural network which adapts according to a learning algorithm to produce the desired result is used to greatest advantage when dealing only with unknown variables that cannot or, for practical reasons, are not measured. When the variables are not measurable, they cannot provide the basis for control in conventional linear control systems.

SUMMARY OF THE INVENTION

It is therefore the principal object of the invention to provide a process controller and method in which models of the process are modified only if the cause is not of a type which is known in order to eliminate unneeded and disadvantageous modifications to the control model.

This object is achieved by provision of a process controller, comprising a model of the process being controlled, means for determining different types of causes of errors in the process and means for modifying the model of the process to correct one of the errors only if the determined cause for said one of the errors cannot be attributed to at least a preselected one of the different types of causes. In a preferred embodiment, the process controller includes means for modifying the model in response to a measured change in at least one of (a) a constrained a. 1 3 variable and (b) an uncontrollable variable.

f Preferably, the process control system includes a feed forward controller responsive to the model for producing corresponding changes in the process that includes an artificial neural network embodying another model corresponding to the model of the process and means responsive to the artificial neural network to control the process.

The object of the invention is also achieved in part by provision of a method of controlling a process, comprising the steps of (a) providing a model of the process, (b) determining different types of causes of an error in the process and (c) modifying the model of the process to correct said error only if the determined cause for said error cannot be attributed to at least a preselected one of the types of causes.

Also, the object of the invention is partially obtained by providing a process controller, comprising means for sensing changes in some conditions which influence the proper functioning of a process, a process model for establishing control parameters for operation of the process, means for detecting disparity in attributes between products actually produced by the process and preselected desired attributes and means for altering the process model in response to said disparity only if the disparity is not attributable to at least one of the sensed changes in conditions.

Thus, obtainment of the object is also achieved partly by provision of a method of controlling a process, comprising the steps of (a) sensing changes in some conditions which influence the proper functioning of a process, (b) establishing control parameters by a process model for operation of the process, (c) detecting disparity in attributes between products actually produced by the process and preselected desired attributes and (d) altering the process model in response to said disparity only if the disparity is not attributable to one of the sensed changes in conditions.

As will be seen, the primary object of the invention is also obtained by providing a printing press for producing 1 9 4 copies of a printed product with a controller, comprising means for sensing changes in some press conditions which influence the proper operation of the press, a process model for establishing control parameters for operation of the press and means for detecting disparity between printing attributes of copies produced by the press and preselected desired printing attributes and means for altering the process model in response to said disparity detecting means only if the disparity is not attributable to one of the sensed press conditions. Preferably, the sensed press conditions includes at least one of (a) dampener motor speed setting,-(b) ink key settings, (c) press speed, (d) press temperature.

- BRIEf-DESCRIPTION OF THE DRAWINGS

The foregoing objects and advantageous features of the invention will be explained in greater detail and others will be made apparent from the detailed description of the preferred embodiment of the present invention which is given with reference to the several figures of the drawing, in which:

Fig. 1A is a general functional blocked diagram of the preferred embodiment of the process controller of the present invention; Fig. 1B is a schematic illustration of a lithographic color printing press employing the process controller contructed in accordance with the present invention; Fig. 2A is a schematic illustration of the preferred embodiment of an artificial neural network employed in the printing press process controller of Fig. 1 which is preferably locally connected to form a forward process model used as the model of the printing press of the printer of Fig. 1; Fig. 2B is a schematic illustration of the preferred embodiment of an artificial neural network employed in the printing press process controller of Fig. 1 which is 1 3 preferably locally connected to form a feedforward controller as used in the printing press of Fig. 1; Figs. 3A, 3B and 3C are bar graphs of ranges of absolute color prediction errors for L (lightness), a (red/green value), b (yellow/blue value) space based on a comparison on the actual colors printed by a printing press employing the printing press controller of Figs. 1 and 2 and the colors predicted by a process model of the printing process controller in which the top of the bars represents the maximum error while the bottom of the bars represent the minimum error for the nine number ink keys of the press; Fig. 4 is a general functional block diagram of the preferred embodiment of the local mapping section, or on-line, printing press color controller section of the printing press process controller; and Fig. 5 is a functional block diagram of the preferred embodiment of the printing press process controller of Fig. 1 including both the local mapping section of Fig. 4 and a global mapping section for automatic establishment of initial, or preset, conditions.

DESCRIPTION OF THE PREFERRED EMBODIMENT

Referring to Fig 1A, a general functional block diagram of the process controller 33 of the present invention is seen as controlling a process 60 which makes a product 30. The physical attributes of the product vary in accordance with changes in the control settings that determine such factors as speed, position and temperature. The control settings, in turn, respond to signals at inputs 45. The input signals to the control settings are automatically varied to eliminate disparity between desired product attributes 44 and actual product attributes 48 detected by a product sensor, or error.

In accordance with the object of the invention a model, preferably a neural network forward process model 52, is provided which contains a model of the process 60 which is 1. 3 6 modified to correct an error only if the cause of the error cannot be attributed to at least a preselected one of a plurality of known causes. These known or measured causes initiate changes in the control setting 46 themselves, which are set by a feedforward controller, preferably a neural network feedforward controller 42 with outputs 45. The different types of causes or different causes of errors in the process are determined by comparing at 67 the predicted product attributes 64 predicted by the neural network forward process model 52 with the actual product attributes 48. These predicted attributes 64 are determined, based on the value of the control setting 46 applied as inputs 51 to the model 52 with the actual product attributes 48. If the predicted product attributes 64 coincides with the actual product attributes 48 determined by product sensor 63, then the model is not modified to eliminate any error detected from a comparison at 65 between the actual product attributes 48 and the desired product attributes 44.

If the predicted product attributes determined by the neural network forward process model 52 based on the existing control settings 46 and the weights given to the inputs to the neural network forward process model 52 do not coincide, the weights of the forward process model are modified in accordance with a learning algorithm suitable for the process in question. In such case, the error is backpropagated through the process model 50 to update the weights at 54 and 55 the forward process model 50 and the feed forward controller. In the case of a color printing press described below with reference to Figs. 1B-5, a learning algorithm based on the sigmoid function has been found suitable as applied to a locally connected three layer, three input, three output, artificial neural network shown in Figs. 2A and 2B.

The updating of weights 55 of the feedforward controller 42 causes the process model of the feedforward controller 42 to conform to the forward process model. The control settings 46 are changed accordingly to correct the disparity between the actual product attributes 48 and the desired product 1 7 attribute. The new setting results in new predicted product attributes for comparison at 67 with the actual product attribute 6.

Referring now to Fig. 1B, the process control system 32 of Fig. 1A will be described with reference to controlling a full color lithographic web fed printing press. In a lithographic printing press system, color is measured by using a device such as a colorimeter or spectrophotometer, which gives a CIELAB (L, a, b) triplet based on an international standard. L defines lightness, a defines the red/green value and b defines the yellow/blue value of color. The color printing process involves multiple control inputs and generates multiple process outputs. The desired output of the process is the quality of printed colors at multiple locations across the width of the p"aper-. However, the complexity and inherent nonlinearities arising from various sources such as printing inks, nonlinearities of ink keys and dampening solution actuators, paper quality, press temperature, environmental effects and crosswise flow effects, greatly complicate developing an analytical process model for the color printing process using conventional methods.

Accordingly, in accordance with the present invention and in appreciation of the complexities of the printing process which have prevented prior attempts for on-line color control of a lithographic printing press, it has been determined that sufficient control can be obtained by use of artificial neural network or ANN described below. Unlike the use of analytical process models for printing press control, use of ANN for purposes of modeling and controlling the printing process enables an ability of the printing press process controller to approximate arbitrary nonlinear relations and to adapt and learn to enhance the control during on-line operation of the press. Reference should be made to "Modeling and Control of the Lithographic Offset Color Printing Process Using Artificial Neural Networks" authored by the co-inventors of this application appearing in PED-Vol. 57, Neural Networks in Manufacturing and Robotics, published by the American Society 8 of Mechanical Engineers in Book No. G00706-1992, on November 13, 1992, "Neural Networks as a Control Technology", Guha, A., Haggerty, A. and Jelinek, J. published in the winter 1989 edition of Scientific Honeyweller, published by the Sensor and System Development Center and the references cited therein for further details with respect to neural networks and their use in controlling dynamic systems. As noted above, artificial neural networks (ANN's) have been exploited for identification and control of dynamical systems as seen in Narendra, K.S.,, and Parthasarathy, K., March 1990, "Identification and Control of Dynamical Systems Using Neural Networks", IEEE Transactions on Neural Networks, Vol. 1, No. 1, pp. 4 - 27. Artificial neural networks have been applied to motion control, process control and aircraft control. Lan, M.S., November 1990, "Learning Tracking Controllers for Unknown Dynamical Systeiiii Via Neural Network Approach", Proceedings of the 1990 IEEE International Conference on Systems, Man, and Cybernetics, pp. 29 31. Neural Networks for Control, 1990, edited by T. Miller III, R.S. Sutton, and P. J. Werbos, MIT Press.

Referring to Fig. 1, a prefered envolvement of the printing press system jo is provided with a printing press process controller 32 in accordance with the present number inking train of rollers 12 and dampening liquid train of rollers 14 deliver ink and dampener liquid to a plate cylinder 22, respectively. The inking train of rollers 12 receive ink 15 from ink feeder 16 through a plurality of inking keys 18. Dampening solution 17 is carried from a reservoir 20 onto a plate having inked image areas on the surface of a printing plate carried by the plate cylinder 22. The inked image on the printing plate communicates with an offset blanket cylinder 24 that is in rolling contact with an impression cylinder 26. The impression cylinder 26 presses a paper web 28 against the image on the blanket cylinder 24 to produce the printed product 30. For further details of an exemplary printing press system with a plate cylinder, inking train of rollers, and dampening train of rollers, reference can be made to U.S. patent No. 5,107,762 of Fadner et al. entitled "Ink 0 9 Dampener For Lithographic Printing Press" issued April 28, 1992 and U.S. patent No. 4,684,925 of Van Kanegan et al. entitled "Simplified Lithography Using Ink and Water Admixtures" issued September 12, 1989. An on-line, printing press prodess controller 32 is employed in the printing press to control the printing operations of the press is described below.

Referring to Fig. 2A, the controller 32 preferably employs a multilayer feedforward network 34 for modeling the complex color printing process. This multilayer network 34 forms a complex nonlinear mapping between the printing process inputs 36 and outputs 38. As seen in Fig. 2A, a threelayer locally connected network is designed to address the spatial relationship of the ink keys 18, Fig. 1, and to capture the interference effect of ink flows between two adjacent inking zones. The term locally connected refers to the neurons 40, Fig. 2A, of one ink key 18 location being connected only to the neurons of the same ink key location as well as the adjacent locations. The cross-connections for the neurons 40 of the two adjacent ink key locations used as the forward process model 52, Figs. 4 and 5, occurs between the input layer 36, Fig. 2A, and a hidden layer 41. Neurons 40 of each ink key location are grouped into cells. The neurons 40 of each group are connected to the other neurons of the group and to the other neurons of a specified number of adjacent groups. The inputs 36 to the network 34 are the settings of multiple ink keys of each printing couple. The outputs 38 of the network 34 are predicted colors at the corresponding ink key 18, Fig. 1, locations. The artificial neural network 34, Fig. 2A, preferably has three inputs 36 respectively representative of the cyan, magenta and yellow subtractive primary colors. The three outputs 38 produce output signals representative of the color attributes of hue, saturation and lightness for the settings of each of the ink keys 18, Fig. 1.

Referring to Fig. 2B, a locally connected feedforward artificial neural network 34 is shown as used as a feedforward controller 42, Fig. 4 and 5. Preferably a cell in the input S layer 36 of the artificial neural network used as a controller is composed of three neurons 40, which accept L, a and b values respectively. Collectively the L, a and b values describe a color. The output layer 38 is composed of three neurons 40, which output ink key settings for the cyan, yellow and magenta printing couple, respectively. The number of neurons 40 in each cell of the hidden layer 41 is a selectable parameter. The cross-connections for the neurons 40 of two adjacent ink key locations occurs between the hidden layer 41 and the output layer 38.

The feedforward neural network 34 consists of multiple layers 36, 38 and 41 as show in Figs, 2A and 2B. Each layer consists of simple processing elements called neurons 40. Neurons 40 between two adjacent layers are connected with associated weights.-'There are no connections between neurons within the same layer. Each neuron 40 of a given layer receives an input, a weighted sum of outputs of the connected neurons in the preceding layer, and sends its output to each connected neuron 40 in the following layer. The relationship between the input and the output of each neuron is determined by an activation function. A range of activation functions is used. Preferably, the commonly used sigmoid function f(x) = I (1 + e') is used. The connection weights are adjusted each time the network 34 is presented with a range of input patterns and its output is compared to the desired output. The weights are adjusted by using an error backpropagation method which propagates backward through the network 34 the difference between the desired output and the network output. For further details on the error backpropogation method reference can be made to Runnelhart, D.E., Hinton, G.E., and Williams, R.J., 1986, 11 Learning Internal Representations," Parallel Distributed Processing: Explorations in the Microstructures of Cognition, Vol. 1: Foundations, MIT Press, pp, 318-362.

The most commonly used learning algorithm for adjusting connection weights is the generalized delta rule. In its 11 basic form, it is a simple gradient optimization procedure:

aE W ii (n+l) = wij (n) +,aw 4 (n), law 5 = - 11 aw ii where E is the cost function being minimized (usually the quadratic output error), wij is a weight in the network, and N is the learning rate. one drawback of this procedure is the slowness of convergence. Speed which can be increased through the use of a momentum term as described in (Rumelhart, Hinton and Williams, 1986 cited above). The modified algorithm is:

aE Awii (n) ii aw ii ' + cc AW Ii (n-1) The momentum term provides a damping effect that reduces the amount of oscillation during the learning process.

Referring to Figs. 3A, 3B and 3C, the ranges of the prediction errors for L, a and b are respectively illustrated. L denotes lightness, a denotes red/green value and b denotes yellow/blue value. The prediction of errors is based on a set of training data which is composed of a plurality of pairs of input and output patterns. The network 34, Fig. 2, is trained for ten thousand iterations. After the network 34 is trained, all seventeen input patterns are presented individually to the input layer 36 of the network 34 and the model predicted colors for each input pattern are generated at the output layer 38. The absolute values of color prediction errors are calculated by comparing the actual printed colors with the model predicted colors.

The results of Figs. 3A, 3B and 3C are shown that the artificial neural network was effectively trained to model the complex color printing process. Humans with normal (not color blind) vision cannot perceive color difference less than about 12 two or three vector units in L, a, b space, which places most of the prediction errors shown in Figs. 3A, 3B and 3C below the threshold of visual detection.

Referring to Fig. 4, a block diagram of the artificial neural network based on-line color control architecture is shown in which the artificial neural net-work controller 42 act as a feedforward controller which accepts as inputs desired colors 44 represented as (L, a, b) triplets as inputs, and provides outputs of the ink key, setting 46 control variables. With the transfer function of the controller 42 being a perfect inverse model of the actual printing process, the controller controls the process to produce desired colors. The feedforward controller 42 is implemented in a multilayer neural network 34, as illustrated in Fig. 2B, which is trained to learn the inverse model of the nonlinear printing process. The controller 42, Fig. 4, implemented in an artificial neural network is further adaptable to a new printing condition.

Training the controller 42 requires knowing the differences between the optimal settings of control variables for a given set of desired colors 44 and the ink key settings 46 generated by the controller for the same desired colors. Since the optimal settings of control variables for a given set of desired colors are not initially known, the differences between the desired colors 44 and the actual printed colors 48 are used to calculate the differences between the optimal and the controller 42 generated settings. This is done by means of error backpropagation the differences between the desired 44 and the actual printed colors 48 through a forward process model 52 of the process as illustrated along line 50. only a forward process model 52 in a form that can be differentiated is needed. Since no mathematical model of the printing process is available, an artificial neural network model or forward process model 52 is used to learn the forward model of the process, which is then used in the error backpropagation process along line 50 for adjusting the connection weights 55 in the neural network controller 42. This also provides the advantage that the controller 42 is 0 13 able to adapt to changes in process characteristics because the forward process model 52 is updated based on new printing data.

The multilayer network 34 of Fig. 2B, used as a controller 42, Fig. 4, is similar to the one for the process model as shown in Fig. 2A. The multilayer network 34 for the controller is a locally connected network and is composed of three layers 36, 38 and 41. The inputs to the network are the desired colors 44, described as (L, a, b) triplets, for multiple measuring locations, while the outputs are ink key settings 46 for the printing couples of three color inks. The cross-connections, however, occur only between the output 38, Fig. 2B, and hidden layers 41, instead of between the input layers 36 and hidden layers 41, as seen in Fig. 2A. Similar to the artificial neural network 34 of Fig. 2A, used as the forward process model 52, a cell in the input layer 36 of the neural controller network 42, Fig. 4, is composed of three neurons 40, Fig. 2B, which accept L, a, and b values, respectively. Since ink key settings 46, Fig. 4, are considered, a cell in the output layer 38, Fig. 2B, is also composed of three neurons, which output ink key settings for the cyan, yellow, and magenta printing couples, respectively. Since three color inks and nine ink keys on each printing couple are considered, a total of twenty-seven neurons 40 are used in the output layer 38 of the artificial neural network 34 of the controller. Likewise, a total of twenty-seven neurons 40, Fig. 2B, are used in the input layer 36, because a triplet is needed for describing a color, and colors preferably are measured at nine locations.

Referring again to Fig. 4, the artificial neural network controller or feed forward controller 42 receives generated reference signals representative of desired preselected color attributes 44 of the printed product. In response to receipt of the desired attributes of the printed product, the neural network feedforward controller 42 establishes the appropriate settings for the ink keys setting 46. The ink key settings 46 is an output of the neural network controller 42.

14 Based on the ink key settings 46, the printing press 10, Fig. 1, runs through the color printing process 60 by transferring ink and dampener solution through a series of rollers to establish a printed product 30. The printed product 30 is read by a color sensor 63 for ascertaining the color of the print 30. The output of the color sensor 63 is the actual color 48 representation of L, a and b of the printed product 30. For further details of the color sensor reference can be made to U.S. patent application serial No. 07/800,947, entitled "Control System for a Printing Press" of Wang, filed December 2, 1991.

The local mapper 33 controlling the on-line color printing process includes another artificial neural network 52 which functions as a model of the printing operation being controlled. The neural network model 62 responds to the output'__'_'___'_ signals of the neural network controller 42 to produce an output signal representative of the predicted color attributes 64 of the printed product 30. The signals representative of the actual color attributes 48 of the printed product 30 are compared with the output signals of the neural network model 52 representing the predicted colors 64. If there is a difference between the actual colors 48 and the predicted colors 64 from the neural network model 52, then the weighting factors 54 of the neural network model are correctly updated to account for the difference. Changing the weighting factors 54 as a result of a measured difference between the actual attributes 48 and the predicted attributes 64 results in a modification of the process model of the other neural network implemented as a forward process model 52. Furthermore, the changes in the weighting factors 54 of the neural network model 52 provides for corresponding changes in the operation of the neural network controller 42 which, in turn, changes the output control signals of the neural network controller which produce the printed product 30 with the desired color attributes 44 during on-line operation of the press 10, Fig. 1.

The desired color attributes 44 are also compared with 4 the actual color attributes 48 of the printed product 30. If a difference exists between the desired color attributes 44 and the actual color attributes 48 then the weights 55 of the neural network controller 42 are changed to produce a new control signal output of the neural network controller. This is accomplished by backpropagating the error represented along line 50, Fig.4, between the desired 44 and actual printed color 48 through the forward process model or neural network model 52.

The local mapper or on-line color controller 33 of Fig. 4, automatically responds to sensed attributes of the printed products 30 to establish a local control space of the color printing process 60. The local mapper 30 operates in accordance with a global mapper which defines a global control space establishing initial control settings as described in Fig. 5. The local mapper 33 includes the neural network model 52 of the press operation which is altered in accordance with changes in the color attributes of the printed product 30 to reduce disparity between the actual attributes 48 of the printed prodcuct and the original printed product.

Referring to Fig. 5, the architecture of the printing press process controller 32 is shown with the original image 70 for the desired printed product having the original color attributes sensed by the color separation scanner 61. The color separation scanner 61 separates the individual color components of the original image 70. The color separation scanner 61 ascertains the color of the original 70 product to be printed. The output of the color separation scanner 61 is used to develop separation negatives 72. Each film of the separation negatives 72 represents an individual color component of the original image 70. The output of the separation negatives 72 is utilized in the printing plates 74 in which each plate carries a color component of the original image. The printing plates 74 are an essential element in the printing process 60 to produce the final printed product 30 as described in Fig. 1. The printed area coverage 76, Fig. 5, derived from the separation negatives 72 and the printing 16 plates 74 in the unit coverage data that is provided by a plate scanner or a print area reader (not shown) from negative films.

The printed area coverage data 76 as well as with job data 78 is sent to global mapper 80. The job data 78 are the data for the press dynamic states and for the print materials. The job data 78 is required for the global mapper 80 to complete a mapping for each new print job. The global mapper 80 automatically responds to the printed area, job data 78, area coverage data and printing attributes of the original 70 to be printed for defining a corresponding global control space establishing initial control conditions of the press 10, Fig. 1. The global mapper 80, Fig. 5, performs global mapping based on the received data and directs the feedforward controller or neural:-network controller 42 to preset the press 10, Fig. 1, while the press is off line.

The global mapping is a process of searching the local control space for steady-state conditions. In the global mapping process the proper controllable variables (ink key and dampener settings) for the desired color and the constrained and uncontrollable variable, (press dynamic variables, print materials and image). During the process of global mapping the press control system 32 considers as many variables as possible. Global mapping is an adaptive presetting based not only on ink coverage but on print materials and press dynamic variables. The global mapper 80 defines a global control space establishing initial control conditions of a selectable area at which the press is located. The local mapper 33, of Fig. 4, operates in accordance with the global control space defined by the global mapper 80, Fig. 5, to establish a local control space 47 for the forward process model 52. The global mapper 80 contains a specification of preset values and a preset algorithm which is employed in defining the global control space.

The reference signals derived from the printed area coverage 76 by the color separation scanner 61 and the global mapper 80 representative of the desired color attributes of 17 the printed product are sent to the feedfoward artificial neural network controller 42. In response to receipt of the generated reference signals, the artificial neural network controller 42 produces associated output control signals to control the associated operation of the press and produce the desired attributes of the printed product. The color separation scanner 64 generates the reference signal representative of the preselected or desired color attributes of the original printed product. The artificial neural network controller 42 receives instruction signals from the global mapper So while the press is off-line and from the forward process model or other neural network 52 while the press is on-line.

The feedforward neural network controller 42 transmits output signals-to the dampener motor 78 and the ink key settings 46, to control the printing process 60 functionality. The printing process 60 also receives input from the printing plates 74 and the master ink control 80. The master ink control 80 determines the ink fountain roller speed based on the press speed 82. The settings of the dampener motor 78 determine the feedrate of the liquid dampening solution 17, Fig. 1, to the dampening train of the rollers 14. The ink feederates across the printing unit are controlled by adjustments to the ink key settings 46, Fig. 5. Upon completion of the printing process 60 controlled by the artificial neural network controller 42 a printed product having the actual color attributes of L, a and b 48 is produced. The actual color attributes of the printed product are compared with the desired original image 70 color attributes. If there is a difference between the desired color output and the actual color output then a control signal is sent to the feedforward neural network controller 42 to make the changes in the controllable process variables and adjust the appropriate press settings.

The forward process model also called the neural network model 52 is the core of the printing press process controller is 32 architecture. The forward process model 52 creates a process model as is needed while the press 10, Fig. 1, in online. The feedforward controller 42 establishes the control parameters for the operation of the press. The forward process model 52 further commands the feedforward neural network controller 42 to adapt to the ongoing process. The forward process model 52 works in the control space that is identified by the global mapper 80. The forward process model or other neural network 42 receives input signals representative of the dampener motor speed settings received from the dampener motor 78, the ink key settings 46, the press speed 82 and the press temperature 84. The forward process model produces output signals representative of the predicted color attributes 64 for the printed product 30, Fig. 1, to be produced in response to the input signals receives at the forward process model.

The feedforward neural network controller 42 controls the ink film optical density by making appropriate changes to the ink key settings 46. The color sensor 63, Fig. 4, senses the actual ink film optical density of the actual printed colors 48 of the finished printed product 30, Fig. 1. The predicted ink film optical density color attributes established as an output from the forward process model 52, Fig. 5, is compared with the actual ink film optical density sensed from the actual printed product 30, Fig. 1. The printing press process controller 32 determines if a difference exists between the actual color attributes of the printed product and the predicted color attributes established by the forward process model 52. If a difference does exist between the actual color printing attributes and the predicted color printing attributes then the weights are adjusted in the forward process model 52 in attempting to compensate for the difference.

As discussed above the forward process model or other artificial neural network 52 receives inputs representative of press conditions of the dampener motor speed setting 28, the ink key settings 46, the press speed 82 and the press 19 temperature 84. The forward process model 52 receives information regarding changes in these sensed press conditions which influence the proper operation of the press 10, Fig. 1. The printing press process controller 32 detects a disparity between the actual printing attributes 48 of the copies of the printed products and the desired printing attributes obtained from the original image 70. The forward process model 52 is altered in response to the detection of a difference between the actual printing attributes and the desired printing attributes only if this disparity is not attributed to a sensed press condition (i.e. dampener motor speed setting, ink key setting, press speed, press temperature).

The forward process model 52 works within the control space that is identified by the global mapper 80. When a output error (the difference between actual color attributes ------48 and predicted color attributes 64) occur and is monitored, the process control system 32 determines the causes of the monitored error, if the error is attributed to some of the measured constrained or uncontrollable variables that are changed, the feedforward neural network controller 42 adjusts the appropriate press settings.

The feedforward controller 42 presets the control variables to match the trends of the changing constrained and uncontrollable variables during press transients. However, if the monitored error cannot be attributed to a measured change in a constrained or uncontrollable variable, the step of modifying or updating the process model 86 is performed by changing the weights of the forward process model 52. The feedforward neural network controller 42, in turn, adapts to the updated process model of the forward process model neural network 52. The printing press process controller 32 determines if the disparity in a predicted color attribute 64 produced in accordance with the forward model 52 and the desired color attribute of the original 70 is attributed to a known measured cause, such as a change in the ink key settings 46. If the disparity between the attributes of the desired color and the predicted color 64 is not attributed to a known measured cause then the model is correspondingly changed or modified to remove the disparity in the color attributes. Thus, the relationship between the forward process model 52 and the feedforward controller 42 is that of master and slave. The forward process model 52 acting as master, continuously oversees the control process and updates the process model as needed. The forward process model 52 in turn, directs the feedforward controller 42 to adapt to the updated process. The feedforward controller 42 acts as a servo that is under the control of the forward process model 52.

Preferably, the feedforward controller 42, forward process model 52 and global mapper 80 as well as the comparison steps of the desired, actual and predicted color attributes are employed in a digital computer such as a Sun worksation or an - Intel - 80486 micro- 5proic --- ess"o-r- based IBM personal--'--'computer. However, equivalent digital computer or central processing with employing an associated memory may also be used in the printing press process controller. It is appreciated that separate components may alternatively be used for implementation of the artificial neural networks without changing the scope of the present invention. With reference to the printing press process controller, the software used for implementation of the artificial neural networks and successfully employed to obtain the results of Figs. 3A, 3B and 3C is seen in the preferred computer program listing attached hereto as Appendix A.

The development of a true real-time feedback controller 42 requires that the dynamics of the controlled process be explicitly accounted for in the combined feedback and feedforward elements of the on-line printing process control system 32. The time dependent parameters designed into the feedforward controller 42 ensures that the process control system 32 is responsive enough to be effective, yet not so responsive as to result in process instability. The most basic implementation of feedforward control is that embodied in ink presetting. The algorithms used in the presetting feature on the press 10, Fig. 1, are empirically determined in i 1 21 a constrained environment.

one set of process materials and operating conditions are assumed to be representative of all process materials and operating conditions which a given press encounters. A warm start on i newspaper press, for example, is a satisfactory mode of operation. This is not the case for a cold start on a heat set press. The reason for this is that, in a cold start on a heat set press, the variety of factors that determine the outcome of the printing process, specifically color on the page, vary more than is accommodated by a single set of values in a preset algorithm.

Adaptive feedforward control extends the idea of ink presetting by utilizing a valid algorithm for transforming the outcomes desired from the printing process (which are ink film optical densities in the case of conventional presetting) into adjustment signals for electomechanical devices under one set of circumstances. The same is done under many sets of circumstances. on-line adaptive feedforward control evaluates the current state of operation of a press 10, Fig. 1, to determine whether or not the algorithm which relates the desired outcome to the electromechanical adjustments is still valid, updates the algorithm if it is not currently valid, and resets the press if the algorithm has been changed.

The idea of retaining end-of-run ink key settings and comparing them to the preset values for the purpose of incrementing those presets which repeatedly differ from the operator settings has been considered in several instances. The printing press process control system 32 of the present invention compares the desired print attributes to the actual print attributes rather than comparing before and after settings of devices and that the modification of the feedforward algorithm takes place on-line during the press run rather than off-line after the run is over. The use of ink and dampener curves to track ink and water feed rates with press speed is also a basic form of adaptive control. The foregoing examples are illustrative of one-to-one mapping between a controllable parameter and a controlled output. The feedforward controller 42, Fig. 5, must have access to as many of the variables that affect the outcome of the process as possible. Thus, the present invention is based on a many-to-one mapping between the color attributes of the print and the controllable, uncontrollable, and constrained variables that drive the printing process. The controllable, constrained and uncontrollable process variables affect the color quality of printed products in lithographic offset printing. The controllable process variables include the ink feedrate and the dampening solution feedrate. The constrained process variables include the press speed, and the conductivity and pH value of the dampening solution. The uncontrollable variables include the press temperature, ambient humidity, paper color and paper texture.

The printing press process control system 32, Fig. 5, provides the press 10, Fig. 1, intelligence needed for on-line adaptation of the feedforward algorithms. Neural network technology provides for the mapping of the controllable and constrained adjustments available within a printing process in relation to the attributes which determine the quality of the printing. color on a unit area of the finished printed product 30, Fig. 1, varies as a consequence of the way the printing process takes place, only three parameters are needed to determine conformance or the lack thereof. There are a number of reference frames within which these three parameters are specified, some are standardized, such as the CIELAB color spaces, and some are not standardized, such as the red/green/blue values produced by color video. There are many controllable variables independent of the reference frame used to quantify the output of the process. These controllable variables include ink key setting, ink fountain roller speed, and dampener roller speed. constrained variables, such as press speed, fountain solution conductivity, and fountain solution temperature and some uncontrollable variables, such as press temperature and relative humidity, affect the process outcome. Neural network technology permits the process control system 32, Fig. 5, to learn the relationship between 23 these variables and the attributes of the printed product and thereby to adapt the feedforward controller 42, Fig. 4, by modifying the weighing functions it uses to do the mapping between the printing attributes and the electromechanical adjustments. The process of learning is driven by sensing the three attributes of color at specified areas on the printed web 28, Fig. 1, comparing these values to the desired values, and using the error between the two to teach the feedforward controller 42 the new relationships.

For further details of the learning and teaching skills of neural network technology, reference can be made to Lan, M.S., November 1990, "Learning Tracking Controllers For Unknown Dynamical Systems Using Neural Networks" Proceedings on the 1990 IEEE International Conference on systems, Man and Cybernetics, pp. 29-31; Lan,-M.S., Bain, L.J., and Lin, P., "Modeling and Control of Lithographic Offset Color Printing Process Using Artificial Neural Networks:, American Society of Mechanical Engineers, reprinted from PED - Vol. 57, Neural Networks in Manufacturing and Robotics, Book No. G00706, 1992; Widrow, B. and Lehr, M.A., 1130 Years of Adaptive Neural Networks: Perception, Madaline and Backpropagation", Proceedings of the IEEE, Volume 78, No. 9, September 1990, pp. 1415 - 1442; and Guha, A., Haggerty, A. and Jelinek, J., "Neural Networks as a Control Technology", Scientific Honeyweller, Winter 1989.

The feasibility of using an adaptive feedforward printing process controller 32, Fig. 4, based on neural network technology for on-line press control is not dependent upon any specific set of color attributes or any specific sensor technology or upon a particular group of press adjustments. The experiments used to establish feasibility are based on the use of the (L, a, b) triplet to define the process input and output and on the selection of ink key settings 46, dampener roller speed, and press temperature as the process variables. The color space coordinates are selected since they represent a standardized frame of reference that is device independent. Additionally, certain combinations of 24 process variables which yield identical results in (L, a, b) space yield varying results in cyan/magenta/yellow space. Knowledge of whether or not this is true is important in the determination of how a press is optimally controlled.

Controllable, constrained, and uncontrollable process variables as a group, consist of those which are expected to vary during one press run or one job. In many instances there are aspects of the process that do not vary within a single run or job but do vary from run-to-run or job-tojob. For instance, one job on a press is heat set, but another job on the same press is nonheat set. one run is process color on both sides and another run is process color on one side with black and a spot color on the other side. These variations in the process are termed global variations. The three types of variations that are expected to occur within a run or job are termed local variations. The preferred control strategy is partitioned into a global component and a local component.

The process of commissioning the press 10, Fig. 1, includes the execution of a series of test runs with the process materials and operating conditions that are expected to make up the production routine of the press. The test runs provide the basic training of the controller 42, Fig. 4, in relation to the global parameters that are varied and the essential relationships between the local variables within each global set. The global variables are of two kinds, those which determine the state of the press 10, Fig. 1, and those which identify the process materials.

The state of the press 10, Fig. 1, is determined by the installed rollers 12 and 14 and blankets, how these are set to their respective cylinders and drums, iron-to-iron settings, bearer loads, ink key 18 zeros, and other adjustments which are made in a maintenance activity but not as an operating adjustment. The control system 32 must be aware of whether the press is in a normal operating state or is being operated in an other than normal state. This is logically a default decision, the control system 32, Fig. 5, assuming the press 10, Fig. 5, to be in its normal operating state unless told otherwise.

The process materials are paper 28, ink 15, and fountain dampener solution 17. The paper 28 is alternatively coated or uncoated. The ink is alternatively heat set or non-heat set. The ink may also be an unusual color. The fountain dampening solution is alternatively acid, alkaline or neutral and may include an alcohol substitute. The process material variations occur in sets, so the combinations of paper 28, ink 15, and fountain solution 17 are limited. The control system 32 is aware of the combinations of process materials on the press 10.

When the printing press process controller 32 is informed as to the state of the press and the process material set, it executes a global map search to identify the control space within which it operates"for the run or job given. once the control space is selected, the printing press process controller 32 expects to receive information that specifies the printed area coverage for each control zone. With this information, the process controller 32 invokes the preset algorithm currently active in the selected control space and calls for readings of the constrained and uncontrollable variables. The on-line control system 32 issues instructions to preset the controllable variables and switches from the preset mode to the run mode.

When the press 10 is started, the on line color controller 33, Fig. 4, operates within the selected control space and modifies the parametric weighing functions which describe that space to achieve a minimum error between the desired color outcome and the actual color outcome. Through this process the printing press process controller 32, Fig. 5, adapts the shape of the local control space as newly acquired information dictates. The specific manner in which it does this is governed by the local strategy.

The printing press process controller 32 is programmed to sample the values of selected constrained and uncontrollable variables and to sample the attributes of the printed product. The constrained variables, press speed and fountain solution 26 properties, are ones which have a quasi-static effect on the outcome of printing. For example, having an inker track the press speed does not require that derivatives, or the rate of speed change, be invoked. All that is necessary is that when the chosen speed is reached, the inker be at the proper speed. Temperature, the primary uncontrollable variable, does call for a rate-of-change computation. The thermal inertia in the press 10 and in the ink 15 dictates that adjustments made only on the basis of the current temperature will always be behind the curve until the temperature comes to equilibrium. Ink adjustments which are made as the consequence of temperature changes lead rather than lag the temperature curve.

on-line measurement of the color attributes of the printed product are dependent upon the particular color space which has been chosen to be the control sjace. The color __ sensor 63, Fig. 5, used must has the capability to capture and process light reflected from the image in a manner that is sufficiently accurate and dimensionally consistent with the selected color space. For example, sensing the image with a spectrophotometer is consistent with the selection of an optical density based control space, but the use of a densitometer is not consistent with the selection of (L, a, b) dimensions for the control space. The attributes of printed color are also statistically distributed over multiple copies, so the measurement of these attributes for purposes of process control is accurate only within a statistical frame of reference. Multiple copies are sampled to determine what the true values of the attributes are. Comparison of these true values with the desired values provides error values the process control system 32 uses to determine what action to take.

Changes in any of the constrained or uncontrollable variables initiates a transition from one state condition in the press to another. The color attributes on the printed product do not change in exact phase with the press variable that is changing, so the error between the actual and desired values is zero, even though the state-to-state transition has begun. The printing press process controller or process control system 32, includes the time-related functions needed to initiate appropriate changes in the controllable variables in this circumstance, so the transient condition in the press 10 has thi minimum effect on the printed product 30. No adaptation of the printing press process controller 32 takes place, because the process of adaptation is driven by back propagation of errors between the desired and actual attributes, which are zero in this case.

Changes in the color attributes on the printed product occurs even when there are no indications of change in the constrained and uncontrollable variables. These are the result of influences that are external to the set being measured. For example, if the color of the paper 28, Fig. 1, is not one of the measured variables, splicing in a new roll can change the outcome of the process; and the only indication of this given to the controller is a change in the color attributes being measured. An error signal is generated when the desired and actual attributes are compared, and the weighing functions 55, Fig. 4, in the neural network feedforward controller 42 are modified so the control system 32 is adapted to the new paper. Control actions are initiated to bring the actual color attributes back into coincidence with the desired values. The new weighing functions remains in effect until another error signal, which cannot be attributed to a measured change in a constrained or uncontrollable variable, is received. If the press 10, Fig. 1, is turned off for a replate after having encountered and acted upon the new roll of paper 28. A new ink preset is called for. The feedback control systems of the prior art would have nothing to offer at this point and, thus, the ink would be preset as if the paper was always the same.

Once the press 10, Fig. 1, is restarted, the color attributes on the printed copy would be in error, and the printing press process controller 32, Fig. 5, goes though the same set of actions it had just completed before the replate to bring the printing into conformance. The adaptive.

28 feedforward control system or process controller 32 of the present invention executes a new ink preset with the weighing functions which are adapted to the new paper. The color attributes are not in error, and no running adjustments are needed to be made upon the restart of the press 10.

When the measurements of the constrained or the uncontrollable variables indicate that one or more of the variables is changed by an amount sufficient to change the state of the press 10, Fig. 1, and there is a concurrent error indication from a color attribute comparator, the printing press process controller 32 the controller determines whether the change in the state of the press and the error indication are related through cause-and-effect. If they are related, appropriate control actions are taken. In certain instances, it is computational. ly inefficient to poll every incoming signal at every sampling interval. Some variables, fountain solution chemistry for example, do not change fast enough to justify short interval interrogation. The printing press process controller 32 uses color attribute error indications to initiateinterrogations of variables in order to determine whether updating their values would result in a decision to take control action but forego updating of the weighing values. Failure to identify any such cause-and-effect relationships results in action on the appropriate controllable variables and adaptation of the control space to the new set of conditions.

The statistical nature of the color attributes of printing dictate that the measurement of these color attributes be treated accordingly. A significant dimension of the local strategy component of the process controller 32 is the invoking of statistical process control concepts as they apply to continuous processes.

The outcome of the printing process is selectively described by a normal distribution of measurable entities. The measurable entities are the vectors in the selected color space. If the process is invariant, a statistically significant number of printed copies (meaning at least five) 29 is sufficient to quantify the set to which every copy belongs. A probability is computed that tells how likely it is that any single copy or any group of copies belong to the set. In discrete processes this is sufficient for the foundation of a statistical process control scheme. However, in a continuous process this is not adequate. An element of prediction is needed, because products continue to be produced while the assessment of quality is being made. The printing press process controller 32 not only determine the likelihood of a product sample being within or outside the set of conforming products, but also determine the likelihood that a product sample to be produced at a specified future time will be within or outside the set of conforming products. Direct control actions or adaptation of the weighing functions in the printing press process controller will be executed only when there is a sufficiently high probability that non-conforming products will be printed if no action is taken and/or the weighing functions are not updated. Thus, the printing press process controller 32 alters the printing process only in response to the probability of a copy being in a range of specifications for conforming products being below a preselected minimum level. Clearly, the numerical values that determine what is sufficiently high are dependent on the specific application being performed. However, the statistics of the process are well enough known to make this determination a straightforward task.

While a detailed description of the preferred embodiment of the invention has been given, it should be appreciated that many variations can be made thereto without departing from the scope of the invention as set farth in the appended claims.

Claims (25)

1. A process controller, comprising:

a model of the process being controlled; means for determining different types of causes of errors in the process; and means for modifying the model of the process to correct one of the errors only if the determined cause for said one of the errors cannot be attributed to at least a preselected one of the different types of causes.

2. The process controller of claim 1 including means for modifying the model in response to a measured change in at least one of (a) a constrained variable and (b) an uncontrollable variable.

3. The process control system of claims 2 including a feed forward controller responsive to the model for producing corresponding changes in the process.

4. The process control system of claim 3 in which the feed forward controller includes an artificial neural network embodying another model corresponding to the model of the process, and means responsive to the artificial neural network to control the process.

5. The process control system of claim 1 including means for defining a control space parameter of the model of the process being controlled.

6. The process control system of claim including means for comparing a predicted output for a change in condition of the process with the actual change in output.

7. The process control system of claim 1 in which the determining means includes means for detecting the error in the process control system by measuring a difference between an actual output of the process and a desired output of the process to determine if there is an error.

8. The process control system of claim 1 including a feed forward controller, and 31 means for adapting the feed forward controller to conform to the modified model.

9. The process control system of claim 1 in which the modifying means includes means for updating the model while the process is on-line.

10. A process controllerf comprising:

means for sensing changes in some types of conditions which influence the proper functioning of a process; a process model for establishing control parameters for operation of the process; means for detecting a disparity in attributes between products actually produced by the process and preselected desired attributes; and means for altering the process model in response to said disparity only if the disparity is not attributable to one of the sensed changes in conditions.

11. The process controller of claim 10 in which the model is contained within an artificial neural network having an output for each of the attributes and an input for each of the types of conditions being sensed.

12. The process controller of claim 10 including means responsive to changes in the process model for affecting changes in the process during on-line operation of the process.

13. The process controller of claim 12 in which the change affecting means includes an artificial neural network controller, and means for inserting a control model into the neural network which conforms to the process model.

14. The process controller of claim 10 in combination with a printing press for controlling the printing press during on-line operation of the press.

15. The process controller of claim 14 in which the process model is selectively not changed in response to an error caused by controllable variables including at least one of ink feed rate and dampening solution feed rate.

16. The process controller of claim 14 in which the . 1h 32 model is modified in response to disparities caused by at least one of (a) conductivity of dampening solution, (b) pH value of dampening solution, (c) press temperature, (d) ambient humidity, (e) paper color and (f) paper texture.

17. The process controller of claim 14 in which the process model includes an artificial neural network with a forward process model of the printing press

18. A method of controlling a process, comprising the steps of:

providing a model of the process; determining different types of causes of an-error in the process; and modifying the model of the process to correct said error only if the determined cause for said error cannot be attributed to a-E... least a - presilec-Eed "o'n'e- -of the - types of causes.

19. The method of claim 18 including the step of modifying the model in response to a measured change in at least one (a) a constrained variable and (b) an uncontrollable variable.

20. The method of claim 18 in which the step of determining includes the step of detecting the error in the process control system by measuring a difference between an actual output of the process and a desired output of the process.

21. The method of claim 18 in which the step in which the process model is embodied in an artificial neural network controller; and altering the model of the process is achieved by altering the weighting of the neural network in accordance with a neural network learning algorithm.

22. The method of claim 20 in which the disparity detecting means includes means for comparing predicted changes in output resulting from measured changes in conditions with the actual changes in the output product.

33

23. A process controller substantially as hereinbefor described with reference to and as shown in the accompanying drawings.

24. A method of controlling a process substantially as hereinbefore described with reference to the accompanying drawings.

25. A printing press including a process controller as in any one of Claims 1 to 17 or Claim 23; or operated by the method of any one of Claims 18 to 20 or Claim 24.