Application of ANN for Fault Detection in Overhead Transport Systems for Semiconductor Fab Note
Zhakov, A., Zhu, H., Siegel, A., Rank, S., Schmidt, T., Fienhold, L., & Hummel, S. (2020). Application of ANN for fault detection in overhead transport systems for semiconductor fab. IEEE Transactions on Semiconductor Manufacturing, 33(3), 337–345.
Introduction
Why fault detection is important in OHT systems?
(1) Proper operation of OHT process is necessary for highly utilized production tools and high output of the factory.
(2) Individual fault related failures in the rail network cause considerable damage and a loss of production.
This paper’s mainly contribution:
(1) apply artificial neural network(ANN) for fault detection and classification(FDC) in OHT rail system.
(2) the adaption of the ANN-based algorithm to the real condition.
Concept for automatic condition monitoring
The measuring platform(referred to as monitoring system). It allows carrying a wide range of different sensors suitable for various monitoring tasks.
As we know, during the transport operation, power supply cables can considerably expand longitudinally and bulge due to heating. In combination with vibrations they can detach from their holders and another common fault is the incorrect position of the cable holders.
The laser sensors capture all objects in measuring area as 2D profile in the XZ-plane. Each captured profile consists of 360 pixels with height information, so 3D data of the cables is recorded when the measuring platform moves.
To ensure real time fault detection at full vehicle speed, computation time is not allowed to exceed 1/(3.3 m/s) approximately 0.3 seconds per meter. Thus, the powerful mobile computer and applicable analysis methods are necessary. In this connection, this paper will discuss the potential ANN-based algorithm.
Algorithm for automatic fault detection
Problem defined
Focus on 3 main faults which causes more than 80% system failure: (1) upper power supply cable out of position; (2) lower power supply cable out of position; (3) twisted cable holder
Spec:
(1) the cable is out of position if its deflection exceeds 4mm
(2) the twisted of cable holder is defined a threshold of 2 mm
Input: the captured profile data consists of 360 pixels with height information
Output: y1, y2 and y3 which stands for the state of upper cable, lower cable and holder respectively. Once the fault is detected from the profile, the state will be set from 0 to 1.
Artificial Neural Network(ANN)
ANN consists of individual neurons that are connected each other. The activation of neuron is given by an activation function(here, they used ReLu as their transfer function).
Finding a right model provides sufficiently good results for the assigned task requires expertise and proper data, especially determined the hyperparameters like learning rate, amount and the size of hidden layer, mini-batch size, window size and the stride of average pooling.
In order to perform the prediction task with sufficient accuracy, this paper used supervised-learning. The training and test data needs to be labeled manually, however, it is a time-consuming process. Thus, they used a semi-automatic approach.
In some reason, it is impossible to collect the training data form fab’s OHT system, so they collected the data from their simulation laboratory. As the result there are differences between the source(lab) and target domain(fab) to expect and hence and ANN trained by the source domain. In this case, the domain generalization must be applied
In case of minor domain differences, the generalization by synthetically generated data via data argumentation leads to a significant improvement of ANN performance.
Experiment and results
Laboratory experiments and results
Experiments and results from a semiconductor fab
This paper used the training data captured in the laboratory. The ANN model trained with the dataset only correctly interprets blocks with profile data from able holders if their Xs- and Zs-positions match the positions in the laboratory. In this case, the ANN model reacts very sensitively to this kind of position vibration, so called overfitting.
Adaption of the ANN-based algorithm to fab conditions
As we know, the initial dataset is unbalanced and missing vibration of the cable holder position in the XZ-plane, the can be eliminated by synthetic methods. Common methods for balancing a training dataset are random minority over-sampling and random majority under sampling.
To tackle both unbalanced training dataset and overfitting problem, this paper extended the classes ‘twisted cable holders’ and ‘no faults’ with synthetically generated samples using the data argumentation. This process involves the following step:
(1) select all of samples that contains cable holder profiles
(2) generate new samples by randomly shifting the profile data along the Xs- and Zs-axis in the range of +-1.5mm
(3) add the generated data to the initial dataset
Results of adapted ANN-based algorithm
Conclusion and outlook
(1) demonstrates that an ANN-based approach is suitable for automatic fault detection and classification in OHT rail system
(2) the ANN model can correctly identify the main faults in the rails
