Patent Grant
7882049
1. Field of the Invention
The present invention relates to the measurement and control of manufacturing processes that produce web or sheet based products, and more specifically, to use of a neural network with spatially dependent data in the control of such processes.
2. Description of the Related Art
The quality of a manufactured product can often be more financially critical than the quantity that is produced. There are many standards worldwide that provide guidelines for quality assurance between suppliers and customers. Maintaining standards of quality for a product may require consideration of the specific properties of the product, as well as the product's final use. The quality of a product is the result of the physical integration of all the raw materials, equipment, and process and operator manipulations occurring during its manufacture.
Process control can be generalized as the collection of methods used to produce the best possible product properties and process economies during the manufacturing process. Many manufacturing processes fall into one of two categories based on the spatial or dimensional dependence of product properties—longitudinal or bulk manufacturing; and web or sheet based manufacturing. Longitudinal or bulk products can be considered dimensionally homogenous and can be measured or characterized with bulk properties. Examples include plastic dowels, polymer threads, fluids, and so forth. Web-based products can be measured or characterized with spatially dependant properties. Examples include rolls and sheets of plastic, paper, or other fibers, minerals and wood products, and even some food products. Note that as used herein, the term “sheet” may refer to both flat products and rolled products.
The challenges associated with web-based products require special consideration for the manufacturing process conditions and the product properties due to the dimensional nature of web-based products. Improper control of process conditions in web-based processes, in either the direction of manufacture or across the direction of manufacture, can result in products that are of little or no value to the final customer. In these situations the manufacturer will lose profit opportunity due to the need to recycle and remanufacture the product, or sell the product at a lower price. Many customers purchase web-based products for use as a raw material in their own processes, which then convert the web product into final end user consumer products. Less than first quality web-based products are not typically accepted by customers. The ability to effectively control web based processes and web-based product properties plays a significant role in determining the profitability of manufacturing operations.
Quality and Process Conditions for General Process Control
Controlling Process Conditions
As shown in
The automation of manufacturing process controls allows the production of products from complex manufacturing processes that cannot be controlled by manual operation. In addition to manufacturing products at higher rates that are more economically favorable, automatic process controls allow the products to achieve more desirable product properties, more consistently. These three factors: more production throughput, more desirable product properties, and a more economical operation, form the basis of process control, which can be summarized as utilizing scientific methods to gain economic leverage over the manufacturing process.
The process control tasks shown in
1) Setting of the initial process condition set points
2) Producing process condition measurements of the process conditions
3) Adjusting the controllable process states in response to process condition measurements
4) Producing product property measurements based on product properties of the manufactured product
5) Adjusting the process condition set points to in response to the product property measurements.
Steps 2 and 4 involve measurements of process conditions and measurements of product properties necessary for control and financial success of the manufacturing operation.
Thus, as
The example of plastic dowel extrusion shown in
In the general case, the actual product properties of a product produced in a process are determined by the combination of all the process conditions of the process and the raw materials that are used in the process. Process conditions can include, but are not limited to, the properties of the raw materials, the process speed, the mechanical manipulation of the process equipment, and the conditions within individual operations of the process, among many others. As mentioned above, the extrusion of a plastic dowel may be referred to as a longitudinal or bulk manufacturing process due to the relative insignificance of any latitudinal process or product considerations, i.e., due to the homogenous nature of the product in any direction other then the direction of manufacturing. Further examples of longitudinal or bulk products include liquids such as chemicals or petroleum products, solid particles of various sizes from polymeric raw materials to cement, or any other product where the properties have little or no cross manufacturing direction variability, and that can be considered homogeneous when measured over small increments of manufacturing time. The desired properties of the plastic dowel can be based on time or the relative product position in the manufacturing process.
Quality and Process Conditions for Web-based Process Control
For the case of a process specifically designed to produce a web or sheet based product there are both longitudinal and latitudinal considerations related to the raw materials, the manufacturing process, and the product properties. Web-based product properties are similarly determined by the combination of all the process conditions of the process and the raw materials that are used in the process. Web-based products can require that dimensional (i.e., 2 dimensional) considerations be given to the raw materials as part of the process being controlled. The previous example of a manufacturing process to produce plastic dowels can be compared to a corresponding manufacturing process for the production of a continuous plastic sheet or web 402, as illustrated in
A simplistic generalization can be made that the manufacturing processes for the production of a plastic dowel and for the production of a plastic sheet involve approximately similar process component functions affecting the raw materials with corresponding manipulations of temperature, pressure, flow, etc., over time. The resulting products (e.g., dowels 304 and sheets 402) differ with respect to their desired product properties and how the process conditions are controlled to achieve the desired properties. Note that the plastic sheet manufacturing process and its product properties differ from the plastic rod manufacturing process and its product properties due to the (two-) dimensional nature of the processes and properties. Like the extruded plastic rod, the desired properties of the plastic sheet can be measured based on its position in the manufacturing process and can be referenced by time; however, the web-based plastic sheet must also have measurements of its manufacturing process and its product properties in the latitudinal directions.
For typical web manufacturing processes producing web-based products, the latitudinal dimension for a process condition or a product property is referenced perpendicular to the direction of manufacturing. This position reference perpendicular to or across the manufacturing direction is typically referred to as the cross direction position or CD position, while the product property position referenced to the manufacturing direction is typically referred to as the manufacturing or machine direction position or MD position, each of which is illustrated in
Measuring Process Conditions and Product Properties.
As described above, there are specific steps in a generalized process control strategy that require measurements of process conditions and measurement of product properties, however, there are manufacturing process measurements and product property measurements that can be difficult to obtain due to the inherent nature of the physical measurement, the location at which the desired measurement must be taken, or the time needed to procure an accurate measurement. In other words, certain process condition measurements can be difficult to reliably acquire due to location, environment, accuracy or other considerations that limit the usefulness of the process condition data in a process control system or strategy, and various product property measurements data can be difficult to acquire do to similar considerations. Product property measurements have an additional constraint on their usefulness associated with the time required to produce an accurate and reliable measure of the specific product property. It is not uncommon for property measurements of certain products to require hours, even days or weeks before an accurate product property measurement is available, e.g., product properties involving physical performance or destructive testing such as strengths, shelf life, wear, color fastness, etc.
The economic viability of a manufacturing operation can be critically dependant on the timely availability of accurate process condition measurements and product property measurements. The inability to obtain accurate and timely measurements can affect the efficiency of the manufacturing process as well as the quality of the products produced.
Web-Based Measurements
It can thus be appreciated that the dimensional nature of web-based process conditions and web-based product properties that have the additional requirement of cross manufacturing direction measurements associated with any point in time, requires unique consideration.
Referring to the previous examples of the extruded plastic dowel and the extruded plastic film, a comparison of the two indicates that the web-based product may require additional measurements of the same desired property across the web-based product at a specific instant in time to characterize the desired product property, as compared to the characterization of the longitudinal or bulk product. In other words, as may be seen in
For many web-based products, there are important product properties that relate to the final end use and quality of the product and thus require additional, or subsequent, product property measurements in order to be acceptable. For example, the printability of paper may not be known (as it is being produced on a paper making machine), until it is shipped to a printer for testing. It is also common in the case of web-based products that some important web-based product property (or properties) cannot be measured directly. For example, some important properties may relate to the rate of variance of a property, i.e., may be based on minute differences that are spatially adjacent. In other words, the product's value or quality, and thus, the product's acceptability, may relate to the magnitudes of spatially adjacent product properties as they vary across the product. As an example, a thickness variance of 1 millimeter in sheet glass that occurs over a meter may not be noticeable, while that same variance over a centimeter may produce a noticeable distortion in the glass's transmissivity, e.g., a visible “ripple”, and so may result in an unacceptable product.
Having these spatially dependent measurements across a web-based product would improve the overall control of the process conditions and/or product properties. Moreover, it may be beneficial for these measurement data to be spatially coherent, where, as used herein, the term “spatially coherent” refers to data that preserve their spatial relationships, e.g., that have associated position data whereby such spatial distribution may be preserved, or that are organized in such a way that preserves the spatial relationships or relative distribution of the data, e.g., the spatial relationships of the measurements (actual or synthesized). It should be noted that “spatially coherent” does not mean that the positions of the data are necessarily regularly spaced, or in any particular arrangement, but only that the spatial distribution of the data or spatial relationships among the data (which could be randomly distributed) are preserved, although such regular spacing is certainly not excluded.
Note that as used herein, the terms “array”, “array data”, “spatial array data” and “spatially coherent data” may be used to refer to data sets whose elements include positional information for the data contained therein, or to data arranged to preserve the positional information, not to the particular type of data structure used to store the data.
The same requirement for multiple measurements applies to web-based manufacturing process conditions. Referring to the previous examples of the extruded plastic dowel and the extruded plastic film, web-based manufacturing processes may require additional measurements of the same desired process condition across the web-based process at each specific instant in time to characterize the desired process measurement in a manner corresponding to the characterization of the longitudinal process, as illustrated in
There are many instances in web-based manufacturing processes where critical data from multiple measurements of process conditions or product properties occur at the same instant in time, or that are reported as having occurred at the same instant in time, e.g., stored with a single time stamp or other order denotation. These multiple process condition measurements and multiple product property measurements can be contained in a data array with an information structure that can establish the positions of the individual measurements within the array structure relative to the cross manufacturing direction position (CD) of the process condition or the product property. As noted above, within a data array, the spatial or positional relationships of the individual measurements can be maintained structurally, e.g., implicitly, via the data structure that contains the data, or explicitly, i.e., via additional information included or associated with the data.
It is also common in web based process industries that process condition and product property measurements are taken with devices that require some time to acquire or process the measurements, such as, for example, traversing measurement sensors that move across a field making a series of measurements in succession, in which case, the entire measurement data array may be reported as having occurred at the same instant in time. In other words, although the series of measurements occurred over a span of time, the resulting data array may be reported and/or stored as a measurement data array with a single order or time stamp. It should be noted that instead of including a time or order stamp, the data from successive measurements may simply be maintained (e.g., stored) in such a way as to preserve their order. In other words, the temporal ordering may be implicit (e.g., organizational), rather than explicit (i.e., including time or order stamps).
Thus, the data array can also contain information or be organized in a way that establishes the measurement position of the data array relative to the manufacturing direction (MD) or to time, which may be referred to as temporally coherent or ordered. In some web-based manufacturing processes these data arrays containing process condition measurements or product property measurements can be referred to as ‘profile arrays’ or simply as ‘profiles’.
Depending on the source of the product property measurement, directly from product property measurement device within the process or from a product property measurement device subsequent to the process, the data array can be order or time stamped or otherwise marked to indicate the measurement array data's relative position in the web-based manufacturing process, its relative position within the web-based product, or its occurrence in time.
The multiple measurement data of process condition or product properties arising from web-based products must be mathematically reduced to a single ‘average’ or otherwise representative order or time stamp value to accommodate the limitations of current neural network modeling technology and methods.
Neural Networks as Predictors of Process and Property Measurements
Current computer fundamental models, computer statistical models, and neural network models can address certain specific process condition measurement and product property measurement deficiencies related to these physical or time constraints in certain manufacturing processes. An exemplary current neural network based approach to process measurement and control is disclosed in U.S. Pat. No. 5,282,261 to Skeirik, which is incorporated by reference below.
Currently available neural network technology can provide predicted values of process condition measurement data and product property measurement data that may not be readily measured. For example, the prior art technique referenced above requires that the input data be specifically time stamped for training and that the predicted data be specifically time stamped for further use in a controller or in a control strategy. Time stamped process condition data and product property data are considered discrete data points, in that each individual measurement data point is detached or independent from any other and clearly identified by a time stamp that can establish its position relative to the manufacturing process.
Prior art neural network applications in process control (see, e.g., U.S. Pat. No. 5,282,261) have evolved through various computer-based modeling strategies including, for example, first principles modeling, statistical and empirical modeling; and non-conventional neural networks. The current limitations for utilizing model-based neural network applications in web-based process control require statistical averaging or manipulations of the available spatial array data that significantly reduces the usefulness of those very process condition and product property spatial array measurements.
Averaging spatial array data to produce a single data point in order to accommodate modeling limitations effectively defeats the basic purpose of using computer modeling and neural network modeling to produce measurements that are difficult to obtain. This type of data handling renders the neural network technology relatively ineffective for treating web-based measurement for process conditions and/or product properties.
Currently there are considerable deficiencies in conventional approaches to obtaining desired measurements for web-based manufacturing process conditions and web-based product properties. Specifically, there are no available methods for directly utilizing spatial array based data that are typically derived from web-based manufacturing processes, nor are there well-defined web-based product properties for modeling and predicting desired measurements of process conditions and product properties. Web-based manufacturing processes and products thus present a unique challenge to existing neural network technologies due to the need to accommodate array based measurements that are referenced in both the manufacturing direction and the cross manufacturing directions, as well as in time.
Various embodiments of a computer-based neural network process control system and method capable of utilizing spatial array based input data (i.e., spatially coherent data) for the prediction of spatial array based process conditions and/or product properties (i.e., spatially coherent predicted process conditions and/or product properties) are described. In this approach, a trained neural network utilizing spatially coherent input data produces a spatially coherent array of predicted values of process conditions and/or product properties that cannot be readily measured. The predicted values, including their spatial relationships or distribution, may be stored in a data network, e.g., in a historical database, and supplied to a controller used to control a web based process for producing a product, or presented to a process operator for use to control a process for producing a web or sheet based product.
The computer neural network process control system and method may incorporate data in array forms, and optionally in discrete forms as well, where the data may be from multiple databases and/or from other data sources that may include the manufacturing process operation, the product property testing operation, and/or product customers operation, among others.
In one embodiment, the method for controlling a process with spatially dependent conditions for producing a product with spatially dependent properties may include providing input data to a neural network, where the input data include a plurality of input data sets, each including values for a set of one or more input parameters, where each input parameter in the set includes or represents a respective process condition or product property, and where the input data preserve spatial relationships of the input data. The neural network may generate output data in accordance with the input data, where the output data include a plurality of output data sets, each including values for a set of one or more output parameters, each output parameter including a predicted process condition or product property, where the output data preserve spatial relationships of the output data, and where the spatial relationships of the output data correspond to the spatial relationships of the input data. The output data are useable by a controller or operator to control the process. In preferred embodiments, the output data are provided to the controller or operator, and the controller or operator may then control the process in accordance with the output data.
It should be noted that the spatial coherence of the input data and/or the output data may be maintained in various ways. For example, in some embodiments, the data may include position information, e.g., included in, or associated with, each data set. In some embodiments, the data may be stored in such a way that the spatial relationships among the data are maintained implicitly, e.g., via the data structure itself.
In one embodiment, the neural network may be configured by a developer who supplies neural network configuration information, e.g., the number of layers in the network, the number of nodes in each layer, the rules governing the neural network's mechanisms and evolution, and so forth. In one embodiment, the developer may easily configure the neural network using a template approach. For example, a graphical user interface may be provided whereby the developer may interactively specify the neural network architecture, as well as other details of the network's operation.
In some embodiments, the training of the neural network may accomplished using training input data having predetermined dimensional array configurations or formats for both space and time. Training may be based on discrete time-based data as well as spatial array data with or without a time base. Training array data may be compared to predicted output array data values produced by the neural network.
In preferred embodiments, a modular approach may be utilized for the neural network. For example, certain specific neural network modules may be provided based on the specific manufacturing process and desired product properties.
In preferred embodiments, the neural network may be used to control a process with spatially dependent conditions for producing a product with spatially dependent properties, e.g., to control a web-based process for producing a web-based product. The process may be operated, and process conditions and/or product properties measured at a plurality of positions to generate measurement data including measured process conditions and/or product properties, where the measurement data preserve spatial relationships among the measurement data, i.e., the measurement data are spatially coherent. Additional data may optionally be synthesized based on the measurement data. The measurement data may be provided to a neural network as an input data array, and the neural network may produce output data in response to the measurement data, where the output data include predicted process conditions and/or product properties, and where the output data have spatial relationships that correspond to the spatial relationships of the measurement data.
Controller output data may be computed using the neural network output data as controller input data in place of measurement input data, and an actuator array controlled based on the controller output data to change a controllable process state using the actuators in accordance with the controller output data. This process of operating, providing, producing, computing, and controlling may be performed in an iterative manner to produce the product with desired properties.
Alternatively, the output data from the neural network may be provided to a human operator, who may then manually control the process, e.g., by controlling an actuator array based on the output data to change a controllable process state using the actuators in accordance with the output data. As with the automatic control embodiment, this process of operating, providing, producing, and manually controlling may be performed in an iterative manner to produce the product with desired properties.
A better understanding of the present invention can be obtained when the following detailed description of the preferred embodiment is considered in conjunction with the following drawings, in which:
While the invention is susceptible to various modifications and alternative forms, specific embodiments thereof are shown by way of example in the drawings and will herein be described in detail. It should be understood, however, that the drawings and detailed description thereto are not intended to limit the invention to the particular form disclosed, but on the contrary, the intention is to cover all modifications, equivalents, and alternatives falling within the spirit and scope of the present invention as defined by the appended claims.
Incorporation by Reference
The following references are hereby incorporated by reference in their entirety as though fully and completely set forth herein:
U.S. Pat. No. 5,282,261, titled “Neural network process measurement and control”, filed Aug. 3, 1990, issued Jan. 25, 1994, and whose inventor was Richard D. Skeirik.
FIGS. 11A and 11B—Embodiments of a System for Process Control Using An Array-Based Neural Network
As is well known in the art of process control, control of processes, e.g., manufacturing processes, generally involves a feedback loop between the process, and a controller or an operator performing manual control, where measurements of various properties of the process and product are fed back to the controller, which may adjust aspects of the process accordingly to maintain or modify product qualities. In predictive process control, the controller may include or be coupled to a model of the process, and may use the model to make predictions or estimates regarding the process behavior or product qualities or properties given certain process conditions. Based on these predictions or estimates, the controller (or operator) may make adjustments to the process to affect a desired outcome, e.g., to produce products with desired properties, or to manage the process in a desired manner. An overview of a typical manufacturing process control system is provided above with reference to
As
As is well known in the art of automated process control, the supervisory controller 1108 generally manages the regulatory controller 1110, i.e., determines the overall strategy, while the regulatory controller 1110 actually implements the strategy specified by the supervisory controller 1108, i.e., controls the process itself, e.g., via actuators. Note that the system of
As shown in
Conversely, information directed to actuator control 1124, labeled “A”, may be provided by the regulatory controller 1110 to the process 1116, thereby driving or specifying actuator behavior for the process 1116. Note that the term “actuator” may refer not only to mechanical effectors, but also to any means used to control the manufacturing process, e.g., mechanical, electrical, hydraulic, optical, logical, etc. As
In a preferred embodiment, the system uses spatially coherent data, e.g., in the form of a spatial data array, from the suitably designed and configured neural network 1104 to replace process spatial data measurements or laboratory spatial data measurements as input to a controller (1108 of
Thus, the neural network 1104 may receive input data in the form of a spatially coherent data set, e.g., an array of records that preserves the spatial distribution of the input data, and may produce corresponding output data, e.g., in the form of a spatially coherent set of predicted process conditions and/or product property values. These output data may then be provided as input to a controller or control strategy, as set points to a controller of control strategy, or as the set points to a manually implemented operator control strategy, as mentioned above. Note that the use of ordered or order/time stamped (spatially coherent) array data as inputs to a neural network that generates similarly spatially coherent output data rather than a neural network single point output facilitates or provides the spatially coherent predictive and control data needed to effectively control web-based manufacturing processes.
As indicated above, in some embodiments, the data network 1106, e.g., historical database(s), may be used to provide spatially dependent process conditions and/or product properties, e.g., historical spatially distributed (and spatially dependent) process condition measurement data and/or product property measurement data, to the neural network, although it should be noted that data from other sources may also be used as desired, including live feeds from the process, laboratory, etc., synthesized data, and so forth.
In some embodiments, when new data are added to the data network, e.g., historical database(s), additional training of the neural network 1104 may be performed to maintain currency of the neural network 1104, i.e., to keep the operational behavior of the neural network 1104 up to date. The historical database(s) and processes used to manage the databases(s) may be referred to as “historians”. In one embodiment, new data provided to the historians may automatically initiate retraining, which may occur on-line and in real time, or off-line, as desired.
As mentioned above, in preferred embodiments, the output of the system and methodology may be incorporated into automatic control system structures, e.g., supervisory or/or regulatory, or as part of a manually initiated control procedure. In some embodiments, a modular architecture may allow the system to build multiple neural networks from multiple databases associated with a process. The system may provide control functions such as supervisory, expert, and/or statistical and analytical functionality, to support automatic and/or manual control. It should be noted that the controller or controllers used in various embodiments of the present invention may be any type of controller, i.e., may utilize any type of technology suitable for controlling the process, including neural networks, expert systems, fuzzy logic, support vector machines, and so forth, as desired.
It should also be noted that in some embodiments, other types of nonlinear models may be used in addition to, or in place of, the neural network 1104, including, for example, support vector machines, expert systems (e.g., rule-based systems), physical models, fuzzy logic systems, partial least squares, statistical models, etc.
FIGS. 12 and 13—Method for Utilizing an Array-Based Neural Network
Turning now to
As
The process conditions may include any attribute or aspect of the process related to the product manufacture. Example process conditions may include, but are not limited to, temperature, position (e.g., gap), pressure, humidity, voltage, and current, flow, speed, rate, feed properties (e.g., of materials), and raw material properties, among others. Example product properties may include, but are not limited to, weight, moisture content, color, strength, stiffness, composition, flatness, texture, thickness, gloss, runnability, printability, and hardness, among others. For example, the above parameters or properties may be important quality metrics for paper production, laminar products, and so forth.
As will be described below in more detail, in some embodiments, the input data may be provided in the form of multiple data sets, where each data set may include respective values for a set of input parameters, where each input parameter in the set comprises a respective process condition or product property. In preferred embodiments, each data set may include, or may be associated with, information indicating the spatial distribution of the input data. In other embodiments, the input data sets may simply be organized or arranged to preserve this spatial distribution, e.g., in a spatially coherent data structure.
Note that the input data may be provided from any of a variety of sources. For example, the input data may be obtained by measuring and/or synthesizing one or more process conditions at each of one or more spatial positions in a production line of the process, and by measuring and/or synthesizing one or more product properties at each of one or more spatial positions on the product.
As mentioned above, measurements may be made using one or more sensors operable to detect physical attributes of the product and/or process conditions, such as, for example, traversing sensors and/or static sensor arrays, although any other types of sensor or sensor configuration may be used as desired.
Data synthesis may also be performed using any of a variety of methods. For example, in the case that some measurements (of process conditions and/or product properties) are available, but not at the spatial frequency or resolution desired, interpolation and/or extrapolation (e.g., linear or nonlinear) may be used to generate the “missing” data. In other embodiments, models may be used to generate the additional data. For example, nonlinear models, such as trained neural networks or support vector machines which may operate to model data sets under specified constraints, may receive the available measurement data as input, and may generate the additional data as output.
In some embodiments, the additional data may include values of the same parameters measured but at positions where such measurements were not made, e.g., may be “interstitial” data (e.g., interpolated), or extended data (e.g., extrapolated), such as temperature values for regions not covered by sensors. In some embodiments, the additional data may include values of parameters that may not be directly measurable, but that may be derivable from measurements. For example, assume that temperature and pressure are measured process conditions, e.g., in an enclosed portion of the product line, but that moisture content is (for some reason) not measurable in this region. Assume that the composition of the product is known. A first principles model based on the relationship between temperature, pressure, and moisture content, may then be used to generate the moisture data.
In 1204, the neural network may generate output data, where the output data include or describe predicted or estimated product properties, and where the output data preserve spatial distribution of the output data, e.g., preserve spatial relationships among the data, e.g., via included or associated position information, or via a spatially coherent data structure. For example, in a paper manufacturing process, data describing process conditions such as temperature, position (e.g., gap), pressure, material feed rates and mixes, etc., may be provided to the neural network, which may then predict a resulting product property, such as for example, the paper's moisture content, thickness, etc. In preferred embodiments, the predicted product properties are useable to control the manufacturing process.
One embodiment of a neural network suitable for implementing embodiments of the present method is illustrated in
As
As indicated in
As
It should also be noted that in some embodiments, in addition to the spatially coherent input data, a single or average data value may also be provided as input to the neural network. For example, one or more single or average data values may be provided corresponding to or representing one or more general process conditions or product properties, e.g., properties or conditions that are not specific to any particular position but are representative of all positions as a statistical or otherwise representative average.
As noted above, the neural network 1320 may have any of various architectures. For example, in some embodiments, the neural network 1320 may be a single network, e.g., may have a monolithic architecture. Alternatively, as indicated in
Each sub-network may operate to receive these parameter values, and as indicated, may generate respective one or more predicted product property and/or process condition values 1332. As
For example, following the same paper production example, each sub-network may provide a predicted value for the paper's moisture content corresponding to the input parameter values (temperature, mass, thickness, production rate, etc.), as well as (in this particular example embodiment) a respective predicted process condition, such as, for example, humidity (e.g., due to evaporation from the product). As mentioned above, the output data may be arranged or included in a spatially coherent output array 1330, and/or may be output with position information preserving the spatial distribution of the data. Note that in this particular embodiment, due to the correlation between the input data, the sub-networks, and the output data, the neural network 1320 may itself be considered to be spatially coherent, e.g., in a logical sense. For example, each respective input data set, sub-network, and output data set, may correspond to a respective CD zone, profile zone, or data box, with respect to the product.
Thus, the neural network may generate spatially coherent output data in accordance with the input data, where the output data comprises spatially coherent values for an output parameter comprising a predicted process condition and/or product property. As noted above, in preferred embodiments, the output data are useable by a controller or operator to control the process.
As shown in 1206, the predicted product properties, i.e., the spatially coherent output data, may optionally be provided to a controller or operator for use in controlling the manufacturing process. For example, referring back to
FIGS. 14A and 14B—Methods for Controlling a Process Using a Neural Network
As noted above, in various embodiments, some of the method elements shown may be performed concurrently, in a different order than shown, or may be omitted. Additional method elements may be performed as desired. Note that where method elements have been described earlier, the descriptions below may be abbreviated.
As
As
As
As
Alternatively, as indicated in
As
As
In 1414, additional data may optionally be synthesized to augment and/or replace at least a portion of the measurement data, as also described above with reference to
Finally, as also shown in
Thus, in embodiments of the methods of
Note that in some embodiments, the neural network may be occasionally be trained, e.g., may periodically be updated, e.g., using data that have been accumulated regarding the operation and performance of the process. For example, each cycle of the process may include storing measured data regarding the process conditions and product properties in historical databases, as described above. These data may be used to provide further training of the neural network. In some embodiments, various analyses may be performed on the data, where the results of such analyses may also be used to train the neural network.
In various embodiments, the neural network may be trained offline or online. For example, in offline embodiments, the neural network may be updated or retrained without being in direct control of or connected to the operating manufacturing process, or, in preferred embodiments, a clone of the neural network may be trained and effectively swapped out with the production neural network, e.g., two or more copies of data defining the neural network's configuration may be maintained, where an offline version is trained using the historical (or other) data, and where the production neural network may be reconfigured to reflect the training, possibly without stopping the process at all. In one embodiment, a developer may configure the neural network using a template approach. For example, a graphical user interface may be provided whereby the developer may interactively specify the neural network architecture, as well as other details of the neural network's operation.
Alternatively, or in addition to, the neural network may be trained online, e.g., may be trained while the process is in operation, e.g., via incremental training, where the neural network is modified by small amounts over the course of operation of the process.
Various embodiments of the system and method described herein may substantially ameliorate problems related to process dead time, measurement dead time, data time frequency, data spatial frequency, and measurement variability in both process condition measurement array data and product property measurement array data, and so may improve control in both the time domain, i.e., in the manufacturing direction, and in the spatial domain, i.e., the cross manufacturing direction.
Various embodiments further include receiving or storing instructions and/or data implemented in accordance with the foregoing description upon a computer-readable medium. Suitable media include a memory medium as described above, as well as any other medium accessible by a computer, and operable to store computer-executable program instructions.
Although the system and method of the present invention has been described in connection with the preferred embodiment, it is not intended to be limited to the specific form set forth herein, but on the contrary, it is intended to cover such alternatives, modifications, and equivalents, as can be reasonably included within the spirit and scope of the invention as defined by the appended claims.
This application is a continuation of application Ser. No. 11/129,062, filed May 13, 2005, entitled “Neural Network Using Spatially Dependent Data For Controlling A Web-Based Process” in the name of L. Paul Collette, III et al.
| Number | Name | Date | Kind |
|---|---|---|---|
| 5121467 | Skeirik | Jun 1992 | A |
| 5805453 | Sasaki | Sep 1998 | A |
| 5812404 | Hamalainen et al. | Sep 1998 | A |
| 5930136 | Nakajima et al. | Jul 1999 | A |
| 6421575 | Shakespeare | Jul 2002 | B1 |
| 6718234 | Demoro et al. | Apr 2004 | B1 |
| 6743337 | Ischdonat | Jun 2004 | B1 |
| 6753965 | Kumar et al. | Jun 2004 | B2 |
| 6915180 | Sasaki | Jul 2005 | B2 |
| 6985779 | Hsiung et al. | Jan 2006 | B2 |
| 20020011319 | Shakespeare et al. | Jan 2002 | A1 |
| 20030046130 | Golightly et al. | Mar 2003 | A1 |
| 20030050728 | Sarabi et al. | Mar 2003 | A1 |
| 20040267395 | Discenzo et al. | Dec 2004 | A1 |
| 20050038538 | McDonald et al. | Feb 2005 | A1 |
| Number | Date | Country |
|---|---|---|
| 2283940 | May 1995 | GB |
| Number | Date | Country | |
|---|---|---|---|
| 20080300709 A1 | Dec 2008 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 11129062 | May 2005 | US |
| Child | 12194440 | US |
