The developed library fulfills 3 main tasks:
- Preprocessing of raw and fuzzy data.
- Assessment of the current state of complex technical objects.
- Forecasting of the future state of complex technical objects.
Those tasks were fulfilled on two main technical objects:
- DSC: overhead crane when driving an L-shaped path with different loads (0kg, 120kg, 500kg, and 1000kg); each driving cycle was driven with an anti-sway system activated and deactivated; each driving cycle consisted of repeating five times the process of lifting the weight, driving from point A to point B along with the path, lowering the weight, lifting the weight, driving back to point A, and lowering the weight (based on the Driving Smart Crane with Various Loads dataset).
- HAI: testbed that comprises three physical control systems, namely a GE turbine, Emerson boiler, and FESTO water treatment systems, combined through a dSPACE hardware-in-the-loop; using the testbed, benign and malicious scenarios were run multiple times (based on the HIL-based Augmented ICS Security Dataset).
Let's consider each task and its modules in more detail.
This task is fulfilled with the help of the following modules:
- Preprocessor is a module for preprocessing of raw and fuzzy data. It can help with data types correction (
preprocessor/check_data_types.py), filling of the empty values of features (preprocessor/cluster_filling.py), and reduction of features in accordance with their informativity (preprocessor/informativity.py) and multicolinear analysis (preprocessor/multicolinear.py). - Extractor is a module for extracting knowledge from the data describing the behavior of complex objects in the form of class association rules, which is designed to extract fragments of knowledge from the available data about the layer in the form of association rules (in “If <premise>, then <consequence>" form) containing only class label in the right part (consequence). The algorithm implements the functions of a strong AI in terms of building a knowledge-based model.
Examples that are describing the work with Preprocessor module are presented in examples/preprocessor_examples.py.
- Function
preprocessor_example_basic()– basic example of the Preprocessor module work. In this example, Preprocessor module is applied to the generated data. All data reduction steps are printed in console. - Function
preprocessor_example_titanic()– example of the Preprocessor module work on the titanic dataset. This dataset is suitable, because it contains categorical and numerical features, while some values of features are empty. Moreover, the dataset is small, which helps to receive results fast.
Examples that are describing the work with Extractor module are presented in examples/extractor_examples.py.
- Function
extractor_basic_example()– basic example of Extractor module work. In this example, Extractor module is applied to generated balanced dataset with 2 classes. Example rule and information about transformed dataset are printed. - Function
extractor_ieee_data()- IEEE_smart_crane example of Extractor module. In this example, Extractor module is applied to IEEE_smart_crane dataset. RandomForestClassifier from sklearn is trained on original and transformed datasets. Information about original and transformed datasets are printed, as well as accuracy metrics for both classifiers. - Function
extractor_hai()- HAI example of Extractor module. In this example, Extractor module is applied to HAI dataset. RandomForestClassifier from sklearn is trained on original and transformed datasets. Information about original and transformed datasets are printed, as well as accuracy metrics for both classifiers.
This task is fulfilled with the help of the Assessor module. Its work consists in applying a pre-trained machine learning model that allows you to determine the state of an object from its descriptive attributes. Conventionally, the algorithm can be presented in the form of three stages:
- Extracting the attributes of the analyzed object.
- Normalizing the attributes.
- Assessing the state of the object.
Within the framework of the tasks considered in the project, the first stage involves reading attributes from a text file with parsing of the corresponding fields. At the second stage, the fields are preprocessed in order to bring them to a single interval. Finally, at the third stage, a pre-trained neural network is launched, which determines whether an object belongs to a particular state based on its attributes.
Examples that are describing the work with Assessor module are presented in examples/assessor_examples.py.
Examples of the application of this algorithm cover the task of ensuring the cybersecurity of critical resources and objects, as well as the task of determining the trajectory of a vehicle (crane). In the first case, the data obtained from the sensors of the system of steam turbines and pumped storage power plants are considered as input data. In the second case, the input data are parameters that describe the operation and movement of the overhead crane under various loads.
The essence of the experiment was to test the suitability of a pre-configured model as part of the task of assessing the state of a critically important object. During the experiment, two phases were distinguished: the training phase and the testing phase. At the first phase, the weights of the neural network were adjusted, and at the second phase, the calculation of performance indicators for estimating the state of the analyzed object was carried out.
This task is fulfilled with the help of the Forecaster module. It performs:
- training a model for forecasting the states of complex systems based on historical data in the form of a time series, which is a sequence of feature vectors of system states;
- autonomous forecasting of system states described by a certain feature vector for a given period of time.
To train the model, the length and sequence of features must be constant for the state of the system at any given time. The model predicts only the numerical parameters of the system states.
Examples that are describing the work with Forecaster module are presented in examples/forecaster_examples.py.
- Function
example_forecaster_model_training(dataset_name, suf='', mode=1)– example of training prediction and data normalization models (dataset_name: name of dataset (str),suf: suffix for naming the output (str),mode: boot mode, for developers (integer)). - Function
example_forecaster_forecasting(dataset_name, suf='', mode=1, independently=True, sample_type='test')- example of data forecasting based on an existing model, including predictive estimation (dataset_name: name of dataset (str),suf: suffix for naming the output (str),mode: boot mode, for developers (integer),independently: sequence is predicted depending on past values or not (boolean),sample_type: type of forecasting sample for estimation - train or test (str)).
If param independently is True then all feature value vectors are predicted independently of each other. At each forecasting stage, an element of the target sample is added to the batch. If param independently is False then each predicted vector becomes an element of a new package for subsequent forecasting.
If param sample_type is 'train' then forecasting time window is equal to the length of training sample from the second to the last batch. Batch for forecasting is the first batch of the training sample. True values for estimation is values of the training sample from the second to the last batch. If param sample_type is 'test' then forecasting time window is equal to the length of all test sample. Batch for forecasting is the last batch of the training sample. True values is values of the test sample.
An example of Extractor (preprocessing) and Forecaster (forecasting) modules integration is presented in examples/extractor_and_forecaster_examples.py.
For additional information, please, check the following documents:
- Programm_description.pdf
- Guide_for_programmers.pdf
Those documents are stored in the guides folder. Note that documents are in Russian.
Documentation in English was built with the help of Sphinx Autodoc and stored in the docs folder.
Levshun D., Kotenko I. A survey on artificial intelligence techniques for security event correlation: models, challenges, and opportunities. Artificial Intelligence Review. 2023. P. 1-44. DOI: 10.1007/s10462-022-10381-4. URL: https://link.springer.com/article/10.1007/s10462-022-10381-4. (Scopus, WoS, Q1)
The study is supported by the Research Center Strong Artificial Intelligence in Industry of ITMO University as part of the plan of the center's program: Development and testing of an experimental prototype of a library of strong AI algorithms in terms of generative design of physical and/or digital objects (components of cyberphysical/socio-cyberphysical systems) characterized by a set of interrelated parameters and satisfying the specified requirements.
This repository is also presented on GitLab.