 |
Sticker Breakout Detection, First steps
| Before we begin |
Before we begin studying the implementation, let us see if we have to because as it's said in A guide to neural Computing applications about neural networks:
''It is easy to be carried away and begin to overestimate their capabilities. The usual consequence of this is, hopefully, no more serious than an embarrassing failure with concomitant mutterings about black boxes and excessive hype. Neural networks cannot solve every problem. Traditionnal methods may be better.''
|
| Who is concerned ? |
From: NN FAQ - Who is concerned with NN ?
''Neural Networks are interesting for quite a lot of very different people:
- Computer scientists want to find out about the properties of non-symbolic information processing with neural nets and about learning systems in general.
- Statisticians use neural nets as flexible, nonlinear regression and classification models.
- Cognitive scientists view neural networks as a possible apparatus to describe models of thinking and consciousness (High-level brain function).
- Neuro-physiologists use neural networks to describe and explore medium-level brain function (e.g. memory, sensory system, motorics).
- Physicists use neural networks to model phenomena in statistical mechanics and for a lot of other tasks.
- Biologists use Neural Networks to interpret nucleotide sequences.
- Philosophers and some other people may also be interested in Neural Networks for various reasons.
- Engineers of many kinds exploit the capabilities of neural networks in many areas, such as signal processing and automatic control.''
We obviously belong to the last categorie.
|
| The issues |
From Introduction to the Theory of Neural Computation, page 8
''Massive parallelism in computationnal networks is extremly attractive in principle. But in practice there are many issues to be decided before a successful implementation can be achieved for a given problem:
## What is the best architecture? Should the units be divided into layers, or not? How many connections should be made between units, and how should they be organized? What sort of activation functions should be used? What type of updating should be used: synchonous or asynchronous, deterministic or stochastic? How many units are needed for a given task?
## How can a network be programmed? Can it learn a task or must it be predesigned? If it can learn a task, how how many examples are needed for a good performance? How many times must it go through the examples? Does it need the right answers during training, or can it learn from correct/incorrect reinforcement? Can it learn in real-time while functioning, or must the training phase be separated from he performance phase?
## What can the various types of network do? How many different tasks can they learn? How well? How fast? How robust are they to missing information, incorrect data, and unit removal or malfunction? Can they generalize from known tasks or examples to unknown ones? What classes of input-to-output functions can they represent?''
|
| SENECA system |
From Application of Artificial Neural network Systems in the steel industry.
SENECA uses several Multi-Layer-Perceptrons for the 16 thermocouples of the upper sensor rows on the front and on the rear side of the mould. The network weights can be used by multiple MLPs (weight sharing). The networks observe and classify the temperature profiles that have been measured during the past one minute of the caster operation.
During off-line training the temperature profiles are presented to the net together with the respect classification as the target value (alarm/no alarm). The subjective classification has been done by experts on the basis of inspections of the slab surface as well as on the basis of evaluations of the temperature signals.
Comparison of classification performance of conventionnal and neural network based breakout prediction algorithm of SENECA:
+-------------------+---------------+--------+
| System | Conventionnal | SENECA |
+-------------------+---------------+--------+
| Sticker | 86 | 91 | Alarm was justified
| Misclassification | 171 | 25 | Alarm was not justified
| Breakout | 5 | 0 | No alarm has been generated
| Normal operation | 0 | 146 | No alarms
+-------------------+---------------+--------+
Off-line evaluation of the system performance based on the available data already showed that SENECA was able to correctly detect all real alarms. The classification ability was independant of the steel grades. Moreover, SENECA was able to identify five ruptures which have not been found by the conventionnal system and led to a real breakout. the rate of false alarms generated by the detection mechanism could be lowered to about 20% of the conventionnal algorithm.
Over a periodof several months the system has been tested in parallel to the conventionnal algorithm. The results of the off-line-evaluation have been fully confirmed. All real alarms have been detected by both systems. However, the alarms from the neural detection system came up to 14 seconds earlier then those of the conventionnal system. in online operation, SENECA's false alarm rate is about 25% compared to the conventionnal system. this can be improved by using the collected data for additionnal training.
|
| The answers |
From Application of Artificial Neural network Systems in the steel industry.
The ability of neural networks to detect complex patterns in noisy signals makes them a valuable tool in almost any fault detection system and helps in increasing the production's quality standards. SENECA is one example for a neural-network based system which has been trained using historical data and which can improve continuously by learning from experiences gathered during online operation.
Experience from the SENECA project shows, that alarm systems, well suited to specific pattern classification processes, can relatively easy be built by using the learning capabilities of neural network.
The concept of the trainable alarm system is flexible and not limited to breakout prediction. By training the networks accordingly, arbitrary patterns can be learned, classified, and suitable actions can be carried out.
Well it's fairly clear, let's go build our network...
|
|
|
