Accompanying repository to the paper Using Expert Knowledge for Masking Irrelevant Data Streams in Siamese Networks for the Detection and Prediction of Faults which is currently under review.

Sketch of the proposed Siamese Encoder (CNN-2D GCN Masked, .c.f. Sect. 3):
In short: A multivariate time series is processed from left to right resulting in the output H_L and h_G. First, 2D convolutional layers (2D Convs) and finally fully connected layers (TDFC) are applied to each data stream (attribute-wise) for extracting temporal dependencies. Then for considering spatial dependencies between data streams, graph convolutions (GC) are applied. Repeating blocks of layers are shown as a trapezoid. The horizontal rectangles represent a time series and the square attached to it on the right represents the concatenation with its embedding. Lines connecting the rectangles represent connections modelled in the adjacency matrix. The crosses indicate the places where masking happens. The masking of data streams is class-dependent and, like the adjacency matrix, is based on expert knowledge.

• The sub directory supplementary_resources of this repository contains additional information about the data sets used and the architecture of the CNN-2D GCN Masked model(Sketch, Plotted Model Graph).
• The detailed logs for each of the experiments presented in Tab. 2 and 3 are provided.
• The raw data recorded with this simulation factory model was used to generate the training and evaluation data sets.
• The preprocessed data set we used for the evaluation.

1. Clone the repository
2. Download the preprocessed data set and move it to the data folder
3. Navigate to the neural_network folder and start the training and test procedure via python TrainSelectAndTest.py > Log.txt. This will reproduce the results for the proposed Model (CNN-2D GCN Masked, c.f. Sect. 3)

You can train and test another SNN architecture by changing the encoder configuration in line 130 in Configuration.py.

Used python version: 3.7.X
Used packages: See requirements.txt

• All settings can be adjusted in the script Configuration.py, whereby some rarely changed variables such as the relevant data streams per class are stored in the file config.json, which is read in during the initialization.
• The hyperparameters of the neural networks can be defined in the script Hyperparameter.py or a pre-defined configuration can be imported from a file in configuration/hyperparameter_combinations/.
• For training, selecting the best model on the validation set and testing it on the test set, use:TrainSelectAndTest.py
• The data/ directory contains all required data. Central are the pre-processed training data in [data/training_data/](https://github.com/PredM/SiameseNetwork-Masking/tree/master/data/training_data9 and the trained models in data/trained_models/. A detailed description of what each directory contains is given in corresponding parts of the configuration file.

The following section gives an overview of the most important packages, directories and included Python scripts in this repository. The implementation of some components is based on the one presented in NeuralWarp (GitHub).