Applying artificial intelligence to the modern foundry.
While the foundry industry has traditionally spent its available modernization dollars on its "Core" areas - melting, molding and coremaking - it has lagged behind other process-intensive industries in embracing emerging technologies in other areas. One such area that may offer revolutionary opportunities for metalcasters is that of neural network (NN) process control.
Computerized NNs - a form of artificial intelligence - have been successfully used in applications such as speech recognition, stock market modeling and prediction, and process control. This article reveals how NN-based process control can be employed in the foundry industry.
What are Neural Networks?
The computer algorithms that make up a NN are based on biological neurons and their interactions. Neurons - defined as grayish or reddish granular cells with specialized processes - are the fundamental units of nervous tissues, such as the brain. A biological neuron is shown in Fig. 1.
Neurons are densely interconnected in the brain, forming interactive networks of cells. The axon (output) path of a neuron splits up and connects to the dendrite (input) paths of other neurons through a junction referred to as a synapse. The transmission across this junction is chemical in nature and the amount of signal strength transferred depends on the synaptic strength.
Learning occurs at a neuron level when there is an increase in synoptic strength. In biological terms, learning takes place when the synaptic weight increases, which occurs when the input to a neuron is correlated with activity in the neuron or the presynaptic and postsynaptic activity are correlated.
Over time, synaptic bulbs decrease in size if activities aren't correlated. This corresponds to forgetting. The synaptic bulbs grow if reinforced via correlation. This corresponds to learning. It is in this manner that humans learn to accomplish complex tasks such as running, talking, or learning to associate tigers with danger and dogs with fun.
Computerized NNs are constructed and behave in a similar manner. A computerized NN is made up of individual processing elements that are modeled after biological neurons.
A schematic of a processing element (or node or neuron) is shown in Fig. 2. In computerized NNs made of processing elements, the weighting that occurs via growth in biological networks is accomplished via mathematical weighting. Applied with the use of various learning rules, this weighting corresponds to the learning that occurs in biological systems with synaptic growth.
NNs based on these concepts can recognize complex relationships between variables that would be difficult to recognize with traditional modeling techniques.
Arc Furnace Control
While the principles of electric arc melting are simple, the actual practice is challenging. This is because the electric circuit that reflects the condition of the furnace and determines the correct position of the electrodes is inherently unstable. It is constantly affected by factors such as line voltage, system impedance, and type and state of the charge.
As used today, arc furnace controllers monitor the phase voltage and current of the arc furnace during operation. These controllers seek to maintain a constant ratio of the phase voltage to phase amperes. This method of control is known as impedance regulation.
Impedance regulators use a set of equations relating phase voltage and current to model the behavior of the furnace. They adjust the electrode position so that a constant phase voltage and phase ampere ratio can be maintained.
Although effective, impedance regulation has one major deficiency - the impedance of the load (metal charge) of the furnace is assumed to be constant. This assumption is false. In fact, the impedance varies significantly throughout the course of a single heat as the charge scrap changes to a mixture of solid and liquid, and finally to liquid. An example of the variation in impedance during a heat is shown in Fig. 3.
A new arc furnace controller (known as the Intelligent Arc Furnace or IAF) was developed to address the problems associated with impedance regulators. It integrates high speed digital monitoring of the arc furnace electrical characteristics with NNs to predict the future state of the furnace and then take action to avoid that state, if undesirable. A block diagram of the IAF is shown in Fig. 4.
The IAF system consists of three NNs. These are the regulator emulator network (REN), furnace emulator network (FEN) and furnace controller network (FCN). The REN is designed to initially emulate the existing regulator, receiving the outputs from it while monitoring the furnace condition. Using this data, it quickly learns to mimic the existing regulator without using the traditional impedance regulator equations.
The FEN, meanwhile, is fed a time history of regulator outputs and furnace state conditions, which enables it to learn to predict the furnace's future state. This prediction ability allows the IAF system to "sense" impending degradation in the furnace state, such as a current overload situation, and position the electrodes to prevent the furnace from reaching that state.
The FCN integrates the outputs of the other two networks to regulate electrode positioning to maintain the desired operation setpoint. This system of three networks allows the IAF to continuously adapt to operating changes in the furnace. This adaptation occurs for short term changes such as system impedance, line voltage fluctuations and electrode length; or longer time frame changes such as charge makeup variation or electrode positioning system degradation.
Plant Trial Results
A test of the IAF was conducted at John Deere Foundry Waterloo. Capable of producing 300,000 tons of iron/year, melting is handled by six 16-ton electric arc furnaces. The test of the IAF was conducted on Furnace 4. Previously, its operations were controlled by a microprocessor-based impedance controller.
The IAF unit was installed in mid-April 1994. The unit monitored the performance of the old regulator for about two weeks. During this time, the REN learned to mimic the impedance regulator on the furnace.
The IAF unit was put on-line as the primary furnace controller at the beginning of May '94. The unit initially controlled to the same power profile as the old impedance regulator, and immediately proved itself capable of controlling the furnace (the IAF didn't use any of the preprogrammed relationships found in the impedance regulator).
The performance of Furnace 4 controlled by the IAF was evaluated using two sets of data. The first set compared the performance of the furnace between heats controlled by the impedance regulator and heats controlled with the IAF.
The second set of data compared the performance of the IAF-controlled furnace vs. the performance of the other furnaces in the melt shop, which only differed in that they were controlled by the impedance regulator.
Current Stability - An examination of current vs. time profiles for impedance regulator-controlled and IAF-controlled heats showed that the latter's variation was significantly less than that of the former. This was particularly evident in the middle portion of the heat, where scrap cave-ins occur due to sudden changes in the furnace impedance.
While the impedance regulator showed great fluctuations in current (an average current range as large as 50 in the first 20 min of the heat), the IAF's ability to sense and respond to changes in the furnace, such as scrap cave-ins, resulted in relatively steady current levels (a difference of fewer than 25). The increased current stability can be quantified by calculating the standard deviation of the difference between the setpoint current and the actual current.
This analysis yielded a standard deviation value of the 3434 amperes for the impedance regulator-controlled heat and 1783 amperes for the IAF-controlled heat. This indicates that the IAF's ability to minimize current variation is about twice that of the impedance regulator. Since power is a function of current, less variation in current also leads to a more efficient transfer of energy to the charge.
Heat Times - The increased efficiency of thermal energy transferred to the charge with the IAF reduces the time needed to prepare a heat. On the previously used microprocessor-based impedance controller in April '94, 81 min were needed to complete a heat of iron. After the IAF was installed in May '94, heat times immediately decreased to 78 min. Interestingly, Furnace 4's improvement compared to the control furnaces was continued through several major changes in the charge makeup that occurred throughout the period.
Electrode Consumption - The improved control capability of the IAF results in a more stable arc, which - combined with such factors as reduced heat time - reduces electrode consumption. The electrode consumption for Furnace 4 and the control furnaces is shown in Fig. 5. Prior to the IAF installation, electrode consumption for all furnaces averaged 6.2 kg/ton (13.6 lb/ton).
When the IAF was put on-line in May '94, electrode consumption immediately dropped to 5.4 kg/ton (12.0 lb/ton). Further decreases in electrode consumption were achieved in the ensuing months to about 5.0 kg/ton (11.2 lb/ton).
The improvements seen with the NN-based IAF led to the installation of these controllers on all six arc furnaces.
Green Sand Control
Based on the success of NN-based control to arc furnaces, a project was initiated to investigate applications for green sand mulling. The first phase of the project, designed to verify the modeling ability and the feasibility of this approach, is complete.
Molding System - The green sand molding system selected for NN development is a high-pressure molding system that receives prepared molding sand from two continuous mullers. The system pump pressures are 1900 psi on the cope and 2200 psi on the drag, with a resultant applied pressure to the mold sand of about 144 psi in the mold.
Each muller is powered by a 350 hp motor that can deliver 230 tons of prepared mold sand per hour for 120 molds/hour. Actual molding practice is 210 tons of prepared molding sand for 110 molds/hour. The bond and water additions are controlled by an automatic compactibility tester located at the muller.
Model Development - First, the relevant measurements were identified that would be used for the model's inputs and outputs. Inputs consisted of incoming sand temperature, conductivity, measured water flow, muller motor amps and bond rate. Outputs included green strength and compactibility.
The time constant of the green sand mulling system was also considered. This represents the time it takes a change in an input variable to affect the output variables. In a mulling system, this is the time between a change in the bond or water addition rate and changes in discharged sand compactibility, or green strength. Based on knowledge of the existing mulling system, a time constant of about 3 min was estimated.
Neural models were created to determine the best architecture and learning algorithm for the process. All the networks consisted of an input layer, a hidden layer and an output layer with varying numbers of neurons based on the system's requirements. One model was created for the green strength prediction, while a second was created for the compactibility prediction.
Each model was then individually modified for the best performance by varying the number of neurons in the hidden layer and using different learning algorithms. The data was gathered into two files representing operation in April '95 and May '95.
For the models created using this data, about 25% of the samples were chosen randomly to train the network. The networks were then tested over the entire month's data. The performance of the networks was evaluated by calculating the error between the predicted values of the compactibility and green strength and the actual measured values.
The results of this evaluation indicate that NN-based models can predict the green strength properties of the discharged sand with an average error of only -0.072 psi, and predict the compactibility of the discharged sand with an average error of only -0.060 points.
Muller Control Systems - The ability of the neural models to predict the desired green strength and compactibility led to the current phase of applying NN to green sand control. A NN-based control system is being designed and is scheduled for testing in the first quarter of '96.
The diagram in Fig. 6 represents the proposed control system. This controller will be installed parallel to the existing controller. Production and maintenance staff will be able to switch between either controller for testing or maintenance.
The new sand control system will consist of these main components:
* multiple muller neural models for green strength and compactibility prediction;
* feedback validation algorithms for testing the automated sampler feedback and neural model drift;
* muller process variable validation for testing process inputs;
* neural muller control algorithm for optimized process control.
The neural muller model depicted in Fig. 6 is divided into three main components. The "dynamic" neural model provides the real-time green strength and compactibility feedback to the controller and to the operator display. This model is trained several times per hour using the lab-corrected automatic sampler results. This allows the model to quickly learn any new process abnormalities from unusual sand conditions or from sensor drift.
The second component is the "static" model. This model is trained strictly from the lab results. It acts as a safety net to verify that the dynamic model's output is within an acceptable range.
The component that ties these two models together is the feedback validation block. It uses simple rules to verify that the automated sampler results are valid, preventing the dynamic model from using failed samples in its training algorithm. Should the automatic sampler spoil more than a defined number of samples in a row or more than a defined number of samples in the last hour, an output will be sent to an appropriate alarm system.
The new control strategy is expected to offer these process improvements:
* the NN-based green sand control system will use the real-time neural muller model prediction as the compactibility feedback to the controller. This instantaneous feedback would reduce the controller response time from 90 sec to as little as 1 sec. This reduced response time will allow the controller to quickly adjust for process variations.
* the presence of an adaptive neural muller model will allow the system to adapt to the ever-changing conditions present in a green sand system. The model will automatically adjust its prediction for seasonal and other variations.
* the neural controller will adapt to optimally respond to changing muller conditions. This will provide a quicker control system response to correct any deviation from the compactibility.
RELATED ARTICLE: Candidates for Neural Network-Based Process Control
NNs excel in controlling difficult processes. They are best suited for consideration in systems/processes that possess some of these characteristics.
* Noisy data
* Multivariate (multiple inputs and outputs)
* Boundary conditions are well understood
* Automated data logging is available
* System state is dynamic
* Effect of all inputs isn't completely understood
* System can't be modeled well with traditional methods
|Printer friendly Cite/link Email Feedback|
|Date:||Feb 1, 1996|
|Previous Article:||Computer simulation aids V-process steel casting.|
|Next Article:||Videos offers foundrymen a new marketing venue.|