 |
|
|
|
|
|
| Home > Articles > Industrial Applications Of Neural Networks |
|
|
Industrial Applications Of Neural Networks |
| ...Continued from page1 |
|
|
Select a suitable NN model based on the nature of problem.
2. Construct a NN according to the characteristics of the application domain.
3. Train the NN with learning procedure of selected model.
4. Use the trained network for making inference or solving problems. If performance is not satisfactory, go to one of the processes.
The neural network learning algorithm: A general view
Given `n training instances.
1. Initialise the network weight. Set i = 1
2. Present the instance to the network on input layer.
3. Obtain the activation levels of output units using inference algorithm. If the network performance meets the pre-defined standard (or stopping criteria) then exit.
4. Update the weights by learning rule of the network.
5. If i = n then reset i = 1, otherwise increment i by 1. Go to step 2.
The neural network inference algorithm: A general view
1. Present the instance to the network on input layer.
2. Calculate the activation levels of nodes across the network
3. For a feed forward network, if the activation levels of all the output units are calculated, then exit. For a recurrent, if the activation levels of all the output units become (near) constant, then exit, else go to step 2. However, if the network found unstable then exit and fail.
Learning: Learning denotes the change in the system that are adaptive in the sense that they enable the system to do the same task or tasks drawn from the same population more efficiently and more effectively, the next time; Or learning can refer to either acquiring new knowledge or enhancing or refining skills.
Learning is of two types.
i. Supervised learning, e.g. Backpropagation (Rumelhart, Hinton)
ii. Unsupervised learning, e.g. ART (Carpenter And Grossberg 1985), Kohonen (1988).
Types of neural network: 3 types
1. Multilay feed forward
2. Hopfield Network
3. Kohonen self organising network.
Why neural network preferred?
Extensive computer study carried out during past years have revealed that neural network enjoys numerous practical advantages over conventional methods. In view of their architecture they are more fault tolerant and less sensitive to noise and they are more easily implanted in hardware because of parameterisation.
Neural network applications
Neural networks are well suited to solving the same type of problems as our brain. You would not enlist a NN to perform arithmetic operations, although in theory that could be done. Applications in new technologies such as robotics manufacturing, space technology and medical instrumentation as well as those in older technologies such as process control and aircraft control are creating a wide spectrum of control problems in which non-linearities, uncertainty and complexities play a major role. For the solution of many of these problems technique based on ANN are beginning to complement the conventional control technique. Rather a NN excel at recognition and classification type of problems. Here are just a few examples of problems to which NN have been applied.
Radar and Sonar signal classification: Neural network can distinguish among different types of radar returns (weather, birds, aircraft) with greater accuracy than conventional systems. A NN has been trained to distinguish between sonar returns from a rock and returns from a steel cylinder.
Speech to text conversion: Seznowskis famous NETTALK program is a NN that has trained to read text and convert it to speech.
Medical applications: NNs are finding many uses in medical classification and diagnosis applications. One example is a NN that use image data to classify reactions of blood cells to antibodies. (Donlin 1992)
Loan applications, credit card use and counterfeit checks: Banks are finding multiple use of NN. Loan applications are now being screened by neural networks, to determine credit risks. NNs are used to detect anomalous pattern in credit card use that may indicate that the card has been stolen. Some points of sale check scanners now incorporate a NN to determine whether a check may be counterfeit.
Investing and tracing: A fortune 500 company now uses a NN to select all of the stocks for its pension funds. A major brokerage firm runs one of its mutual funds using NN. Among security traders, NNs have become a well established tool.
Refinery control: Process control systems being usually complex and highly non-linear are difficult to control and derive a dynamic model. A NN refinery control scheme is described here.
Puget sound refinery of Tereaco with a capability of 120,000 barrels oil per day has been applied a NN model to control a debutaniser which separates and condenses hydrocarbons according to their molecular weights. The NN used has seven inputs of controls and disturbance variables and two outputs of manipulated variables. For training a network, 1440 data sets were taken. A feedback mechanism has been provided to eliminate unexpected errors. It has been reported that this NN model remains in control about 80% of time and even more during unstable processing.
Temperature controls: An inverse dynamic model of a laboratory water bath temperature control process is developed with three layered NN with eight hidden units and one output unit. The performance of this real time neural controller is compared with that of a PI controller. It is shown that the NN performs better under load disturbance for variable dead time process.
Quality Control: A chemical process plant of Cleveland Ohio uses a NN system for quality control. Usually trained experts are required for controlling the quality of a chemical product. In this plant, the extracted product samples during changeovers and other process stages are analysed on an infrared spectroscope to verify proper chemical ratio and absence of contaminants. A NN is trained with a training set of known contaminants and their associative outputs. It has the advantage of verification of product quality without programming knowledge of the technician, or the spectroscopy aspects.
Air traffic control: Netrologic of San Diego has developed a NN for air traffic control (ATC) using a stimulation program called TRACON. This uses a local associative NN for each mentioned craft. Each local network receives information such as conflicts in the next minutes and whether the next aircraft is a tower regarding other craft (one network for each aircraft) as input and conflicts are avoided by extrapolating current velocities and positions of aircraft on the screen on to the future. The outputs of the network are "change speed""turn light" and "increase altitude", etc. Netrologic has also developed a network based real time decision aid called an ATC conflict field interaction network.
Robotics: Neural networks have been extensively used in robotics and the areas where neurocontrollers have found applications can be conveniently classified as:
- Manipulator control
· Contact control
· Locomotion
· Co-ordination, grasping and manipulation
- Planning and design
Autonomous navigation: A truly interesting and naval application of a neural network adaptive control is described by Pomerleav and is in the area of vision based driving (autonomous driving). Based on images formed on board based video camera, a robot van equipped with motors for a steering wheel, braking and acceleration paddles determines its own trajectory. The most noteworthy feature of the system is the supervisory control of method used to train it. The NN is taught to imitate the driving reactions of a person. As the person drives, the NN is trained using backpropagation. The input to the NN is a 30 x 32 unit video image and output unit consists of 30 nodes each of which corresponds to a steering direction. A van trained in the above fashion was able to travel at 55 mi/h under different lighting and weather conditions, which according to the author is about 5 times as fast as any non-connectionist system has done using comparative hardware.
Neural network in aeronautics: Another area where NNs are finding applications is in aeronautics. In rotor craft, high intermittent blood torque causes rotor fatigue requiring replacement of critical components. Interaction between rotor blades and blade tip vortex can cause vibrations of the retreating rotor blades. Under such operating conditions, to maintain safety and operational readiness, new type of control systems are needed, i.e. NN.
Neural control of steel rolling mill: The problem of controlling the strip thickness of a steel rolling mill in simulation studies was reported by Sbarbara. The complex non-linear nature of problem as well as varying delays, make the use of NN.
Arc furnace: Neural networks for closed loop welding process control: Welding in a manufacturing context is much more complex than the familiar concept of the fusion process for joining materials. Generally, the planning and design begins with a specification of the weld joint geometry and environment and final required characteristics from selection of a suitable power source can be made. In order to make good welds consistently, consideration must be given to mentioning and controlling machine performance continuously.
One of the most impressive industrial applications of neuro-control is the intelligent arc furnace designed by Neural Application Corporation in Coralville. Electric arc furnace uses electric power to melt steel. A typical furnace is 15-30 ft in diameter and melts 55-150 tonnes of steel in about an hour. The amount of scrap power delivered to scrap metal is controlled by positioning three large electrodes each having 12-24 ft in diameter.
Neural networks for closed loop welding process control: Welding in a manufacturing context is much more complex than the familiar concept of the fusion process for joining materials. Generally, the planning and design begins with a specification of the weld joint geometry and environment and final required characteristics from selection of a suitable power source can be made. In order to make good welds consistently, consideration must be given to mentioning and controlling machine performance continuously.
The gas metal arc welding process, (GMAW) formerly called metal inter gas (MIG) welding, is extensively used in metal works industry to weld a variety of ferrous and non-ferrous metals. The process uses a continuous consumable wire electrode. The molten weld paddle is completely covered with a sealed gas, which is also fed through a welding gun. The GMAW process has been widely used as a metal joining technique because of two facts,
a. Its potential for increasing the productivity and quality of the weld,
b. Its use by most arc welding robots.
High performance automated welding offers the potential for excellent consistency, increased level of productivity and reduced cost. To achieve a high level of performance and quality at high gun travel speed, the weld path must be well defined and appropriate welding process variables must be correctly adjusted along the weld path. The multivariate environment suggests the need for an intelligent control system that can evaluate the process and determine the best adjustments and do so in real time. The full potential for an intelligent control system to evaluate and control the welding process goes beyond the shop-floor. Such technology makes it possible to integrate the weldment design and process planning activities more completely.
Arata and colleagues did a series of studies to investigate the characteristics of the welding arc sound. They concluded that welding arc sound is a good information signal to express the welding condition. Yu conducted a set of experiments to study the feasibility of using the welding arc sound as a means of a feedback to control the welding process. The results clearly indicate the excellent repeatability of arc sound signal and characteristic response of various process conditions - the potential for using the sound as the basis of the feedback control. In order to measure the weld condition and control the weld process in real time, a special pattern classification technique is required to classify welding arc sound in order to assess the condition of weld. In past few years, neural networks have been recognised and accepted as a powerful tool for correlating data without making strong assumptions about the problem. One of the most widely used neural network configuration is a multi-layer feed forward network structure, which is normally trained through an optimisation technique such as back-propagation.
Methodology
Architecture: Figure shows the neural network structure adopted in investigation for closed loop welding process control. The first neural network, the signal classification network (SCN) is used to classify the welding arc sound signal in order to assess the quality of weld. The second neural network, feedback control signal (FCN) provide the feedback control signal to adjust and control the welding process.
The closed loop control system starts setting the controllable variables of the welding system. The sound signal corresponding to this setting is then collected. The collected sound signal is fed into SCN (signal classification network). The output of the network indicates the present quality of weld, which is determined by measures: penetration, spatter, weld bead width and weld bead height. After comparing the present weld quality with base line weld quality, a quality difference is obtained and fed into FCN (feedback control network) signal. The FCN, then in turn, provides the feedback control signal to adjust the controllable variables and forms a closed loop control system.
|
|
|
|
|
| Posted : 10/27/2005 |
|
|
|
| |
|
|
|