Using artificial intelligence in the foundry.
Knowledge is the key factor in running a successful foundry operation. Research into artificial intelligence Al) has resulted in new, effective methods for handling this knowledge. Since 1983, Al has been used to analyze foundry problems and build expert systems in Sweden. In the foundry industry, where much of the knowledge is vague and the casting process is complex, expert systems technology is ideal. With inductive type expert systems, the advantage is that the system can learn automatically from process data or from the experienced foundryman.
One essential ingredient in intelligent behavior is the ability to learn. Inductive learning is a method whereby general rules and search paths are automatically extracted from raw data. The method also allows knowledge to be elicited from experts. It has been implemented in a computer program, called FoundryExpert, and represents a user friendly and powerful way of solving problems and building expert systems.
This system has been developed in cooperation with Professor Donald Michie, an authority in Al and induction type systems and chief scientist at The Turing Institute/Glasgow.
This method is based on the principle that information or specific knowledge about conclusions can be represented as examples. The description is made by using attributes variables, features) relevant to the problem. Thus, an example is a vector consisting of attribute values that are typical for a specific outcome.
The information content in the examples can be quantified by means of a method where the entropy (information uncertainty) value is calculated. The inductive learning method automatically selects the attributes that reduce the original entropy the most. By doing this in a stepwise manner, optimized decision trees are created.
The entropy value is a measure of the information uncertainty in a message. Assume that the only knowledge we had about gas blows was in 10 examples where nitrogen was the primary cause. The information uncertainty would be zero. In order to learn what distinguishes nitrogen defects from other gas blows (H.sub.2, CO, air, etc.), we must also give examples of other defects. The entropy value will then increase.
To convert the information in the examples to useful knowledge, we have to find the symptoms (attributes) that most effectively reduce the original information uncertainty. This can be achieved by using an inductive learning algorithm that calculates the entropy value for each attribute in the examples. The attributes that reduce the original entropy the most are used to produce rules and search paths to cover the defined problem space.
Entropy is expressed as bits and the basic formula to calculate this value is based on the probability for each conclusion. The problem area is defined by entering relevant attributes (features, variables) and their values as well as the possible outcomes. All the attributes that the expert believes to be relevant are entered.
The next step is to enter typical cases using the defined attribute values. The examples can consist of typical cases from an experienced expert or raw data from the process. Once a sufficient number of examples has been entered, the inductive learning process can begin.
The learning process automatically examines the examples, combines them and produces general rules to be used for consultation in an expert system. Possible conflicts and areas lacking knowledge are automatically pointed out. The system reveals hidden patterns in the examples, removes redundant attributes and finds an optimal search path. The method is interactive because the developer can watch how the rules develop, in a graphical format, as decision trees. At any time, he can add or amend the attributes or examples and immediately study the effect.
Several methods have been developed to facilitate the knowledge acquisition" process. One way is to enter a main case followed by all its exceptions. Alternatively, examples can be entered unstructured and, when the system finds a conflict, it prompts the expert to provide an attribute to solve the conflict.
Another method useful in capturing an expert's silent knowledge is to let the system generate a truth table consisting of all possible combinations within the problem space. The expert then has specific cases to evaluate. His decision for each case is used when inducing the knowledge base.
Use in R&D
The induction method can be used in research and development as a tool to extract new knowledge from raw data. By entering cases with different outcomes, the most logical discriminating path can be found. The induction process traces dependency between variables, finds threshold values and removes redundant factors.
Both numeric and discrete attributes can be used in the same model. The result is presented as a graphical knowledge tree on the screen. By studying the knowledge tree, the expert will often gain insight into complex problems. The method has been used, for example, to optimize conditions in gating and risering systems by using examples of both successful and unsuccessful pattern schemes.
Induction excels over statistical methods, especially in areas where a firm hypothesis is difficult to formulate and in nonlinear variable-dependent domains.
Success in Foundries
The induction methodology can be used successfully on complex problems in foundries. Assume that you have a problem with poor surface finish on castings. The finish varies and you have been unable to find any correlation with sand properties using traditional charts.
Using an inductive system, you could enter all the attributes that might be important, such as sand properties, pouring temperature and mold hardness. As information, enter examples of cases when the surfaces were acceptable and unacceptable. More than two outcomes can be used (i.e., level of penetration as severe, moderate, some, none). Any pattern or rule that exists will be revealed by the inductive rule generator.
Often it is found that only a few of the attributes used in the examples are needed in order to classify the defects. By studying the knowledge tree, the expert can gain new insights into the problem. Thus, the system is used as a hypothesis generator, amplifying our logic reasoning.
During the first phase, redundant attributes are eliminated and a basic structure of the problem is displayed. The next step is to build a new model using only those attributes that proved to have classification values. Continuous attributes can now be treated as numeric in order to find important threshold values. The induction process should be invoked every so often, as examples are added, so that the development of the rules can be followed. About 30% of the data should be used as a test set to validate the rules.
A technique known as pruning allows inductive systems to handle noisy data and to calculate probabilities.
Building the System
Expert systems are basically a methodology to capture, refine, accumulate and distribute specialized knowledge. The system is built to solve a particular well-defined problem. Systems can be categorized in the following groups.
Analyzing Systems-analyzing data from processes, markets, etc. Such systems use machine learning methods in order to learn from raw data. inductive learning is the leading method. Typically the system answers questions such as: Why do we get pinholes? Which melting procedure maximizes the amount of eutectic graphite in ductile iron? Which pricing policy should we use?
Prognostic or Forecasting Systems-based on earlier events and experience, the system can deduce the most likely outcome. Typical applications are in process control, market evaluation and staff assessments.
Advisory Systems-the system uses knowledge from experts to give advice, typically on production problems, alloys, gating and risering, etc.
Diagnostic Systems-basically faultfinding systems for equipment, these are also used for analyzing casting defects, etc.
Configuration Systems-a configuration system could be used to select production methods for a casting. In a configuration system, the conclusions for various parts are dependent on each other. Certain alloys can only be manufactured, for example, by certain molding methods. In order to select the optimal cleaning method it is necessary to know about the molding method.
An expert system often uses a combination of the above five categories. The first step in developing an expert system is to define the problem and to divide it into smaller problems. Problems are often solved in an hierarchical way. A system to diagnose casting defects could consist of one module that analyzes the symptoms and decide which type of defect it is-shrinkage, surface defect, gas blows, inclusions, dimensions and material problems, for example.
For each subclass, a smaller module is built to analyze the effect in more detail. Gas blows can be separated after the main cause-nitrogen, hydrogen, carbon monoxide, air, core gases and so on. The knowledge base in each module is based on typical examples for various defects using a language that the expert has personally defined. Inductive analysis is then used to generate the final rules to be used deductively in the expert system. The modules(each a small expert system) are linked to each other to make up the complete expert system.
By building the system in small homogeneous modules, it is easier to validate the knowledge base. It also is much easier to modify the system or extend it with further knowledge. New modules can also be added.
When the system is used in the consultation mode, it will ask the relevant questions on the screen and automatically select the optimal search path through the knowledge, depending on the answers from the user.
The technology can be used as an integrated part of conventional programs. in a solidification simulation program, an integrated expert system could be used to suggest optimal locations for risers, suggest orientation of the part or where to apply chill, etc.
Building expert systems using inductive analysis allows the domain expert to communicate with the system directly when developing the system.
Practical Foundry Usage
It has been shown that the inductive method is extremely powerful, especially as a hypothesis finder in process control and problem solving.
An adaptive process control system for green sand systems has been built using this methodology. The system learns by examples from laboratory data. The data is fed into the system daily and stored in a database containing the data from the past 20 days.
This data, in combination with information about the amount of additives used, as well as the casting program, are used by the system to adaptively produce rules for controlling the process. These synthetically produced rules are automatically created and used to produce sand mixtures based on the predicted influence of the patterns and the current quality of the additives, muller efficiency, etc. Thus, the control strategy used in the expert system is to predict what will happen when a certain pattern is used and to adjust the sand mixture in advance, before the fact. Most sand systems are controlled after the fact, which often times explains the variations in sand properties.
The result is a sand system with far fewer variations and lower consumption of additives. One Swedish foundry has reported a 5% production increase owing to fewer molds and castings being scrapped as a result of sand defects.
The inductive system has been used to solve several other difficult problems where traditional methods have failed. One example is to predict liquidus and solidus temperatures in steel alloys having up to 10 elements. Another, developed in a German foundry, was used to discover the reasons for severe surface defects in gray iron castings.
An expert system was developed for analyzing casting defects and has been in use since 1987. The system is structured in a hierarchical way and consists of many smaller systems, each containing knowledge about certain types of defects, linked to each other. The main types of defects covered are: inner and outer shrinkages; surface defects; gas blows; inclusions; and anomalies in the matrix or graphite shape.
Because the system uses inductive learning, it is very easy to modify and extend. Thus, the original system has been updated almost every month as new information about casting defects becomes available from colleagues, research reports and literature.
Expert systems are of prime interest in areas where explicit knowledge is difficult to ascertain or where it is imperative that the knowledge is applied correctly. Most problems in the foundry are rooted in missing or incomplete knowledge. The industrial applications are therefore numerous.
The new Al methods offer new and exiting possibilities to capture, refine and distribute knowledge. This methodology has been adapted for use in other areas such as medicine, organic chemistry and stock market evaluation. The foundry industry-where a complex process with a vast number of variables influences the results and where only a fraction of the knowledge is explicitly available-is ideally suited for this technology.
More Foundry Examples
It has been reported recently that a U.S. foundry had experienced a very high scrap rate, which had gradually ruined its financial viability. By using the Al system developed for foundries, the foundry was able to successfully reduce its scrap rate. The company went from a loss situation to a profit situation within a few months.
A Swedish foundry, Nya Perfecta AB has reported that it was able to take on 103 additional jobs by utilizing these same programs. Traditional methods for designing gating and risering would have allowed the processing of perhaps 30-40 orders.
A research project also has been concluded recently in conjunction with the Swedish Foundrymen's Assn. The aim of this project was to find the cause of a fracture type fault that occurs in steel castings. The fracture is referred to as 'rock candy," owing to the appearance of its microscopic structure. By using inductive analysis of data taken from test pour's at a steel foundry, researchers pinpointrf the two main elements that caused the problem (nitrogen and aluminum) and established critical threshold values (Xpertrule and Assistant Professional were used for this system).
|Printer friendly Cite/link Email Feedback|
|Title Annotation:||Special Report: International Metalcasting Trends|
|Date:||Dec 1, 1991|
|Previous Article:||Quality and productivity remain key issues for foundries.|
|Next Article:||Green sand & the environment: the European scene.|