The article is devoted to the modified multidimensional Kalman filter with Chebyshev points development to solve the task of diagnosing and parring off failures in the measurement channels of complex dynamic objects automatic control system, which will provide a more accurate and reliable assessment of system state in the presence of outliers in the data. An implementation of the proposed modified multidimensional Kalman filter with Chebyshev points is proposed in the form of a modified recurrent neural network containing a failure diagnostics layer, a failure parry layer, a filtering and smoothing layer, and a results aggregation layer. This structure of the modified recurrent neural network made it possible to solve the main problems of the method of diagnosing and parring off failures of the measuring channels of complex dynamic objects automatic control system, such as diagnosing failures with an accuracy of 0.99802, fending off failures with an accuracy of 0.99796, and assessing the state of the system with an accuracy of 0.99798. It is proposed to use a modified loss function of a recurrent neural network as a general loss function for diagnostics, fault restoring and system state assessment, which makes it possible to avoid retraining when there are a large number of parameters or insufficient data. It has been experimentally proven that the loss function remains stable on both the training and validation data sets for 1000 training epochs and does not go beyond –2.5 % to +2.5 %, which indicates a low-risk overtraining or undertraining of the model. It has been experimentally confirmed that the use of a modified recurrent neural network in solving the task of diagnosing and parring off failures of the measuring channels of complex dynamic objects automatic control system is appropriate in comparison with a radial basis functions neural network and a multidimensional Kalman filter without a neural network implementation, based on metrics such as the root mean square deviation, mean absolute error, mean absolute percentage error, coefficient of determination for the accuracy of reproducing previous data, and coefficient of determination for the accuracy of predicting future values. For example, the value of the standard deviation of the modified recurrent neural network is 0.00226, which is 1.65 times less than the radial basis function neural network and 2.20 times less than the multidimensional Kalman filter without a neural network implementation.

A new method of propaganda analysis is proposed to identify signs and change the dynamics of the behaviour of coordinated groups based on machine learning at the processing disinformation stages. In the course of the work, two models were implemented to recognise propaganda in textual data - at the message level and the phrase level. Within the framework of solving the problem of analysis and recognition of text data, in particular, fake news on the Internet, an important component of NLP technology (natural language processing) is the classification of words in text data. In this context, classification is the assignment or assignment of textual data to one or more predefined categories or classes. For this purpose, the task of binary text classification was solved. Both models are built based on logistic regression, and in the process of data preparation and feature extraction, such methods as vectorisation using TF-IDF vectorisation (Term Frequency – Inverse Document Frequency), the BOW model (Bag-of-Words), POS marking (Part-Of-Speech), word embedding using the Word2Vec two-layer neural network, as well as manual feature extraction methods aimed at identifying specific methods of political propaganda in texts are used. The analogues of the project under development are analysed the subject area (the propaganda used in the media and the basis of its production methods) is studied. The software implementation is carried out in Python, using the seaborn, matplotlib, genism, spacy, NLTK (Natural Language Toolkit), NumPy, pandas, scikit-learn libraries. The model's score for propaganda recognition at the phrase level was obtained: 0.74, and at the message level: 0.99. The implementation of the results will significantly reduce the time required to make the most appropriate decision on the implementation of counter-disinformation measures concerning the identified coordinated groups of disinformation generation, fake news and propaganda. Different classification algorithms for detecting fake news and non-fakes or fakes identification accuracy from Internet resources ana social mass media are used as the decision tree (for non-fakes identification accuracy 0.98 and fakes identification accuracy 0.9903), the k-nearest neighbours (0.83/0.999), the random forest (0.991/0.933), the multilayer perceptron (0.9979/0.9945), the logistic regression (0.9965/0.9988), and the Bayes classifier (0.998/0.913). The logistic regression (0.9965) the multilayer perceptron (0.9979) and the Bayesian classifier (0.998) are more optimal for non-fakes news identification. The logistic regression (0.9988), the multilayer perceptron (0.9945), and k-nearest neighbours (0.999) are more optimal for identifying fake news identification.

The work is devoted to the development of the radial basis functions (RBF networks) neural network new architecture – a polymorphic RBF network in which the one-dimensional radial basis functions (RBFs) in the hidden layer instead, multidimensional RBFs are used, which makes it possible to better approximate complex functions that depend on several independent variables. Moreover, in its second layer, the summing the RBF outputs one by one from each group instead, multiplication is used, which allows the polymorphic RBF network to better identify relations between independent variables. Based on the training classical RBF networks evolutionary algorithm, the polymorphic RBF network training algorithm was created, which, through the initializing weight coefficients methods use taking into account the tasks structure and preliminary values, using the mutations tournament selection, adding additional criteria to the fitness function to take into account stability and speed training a polymorphic RBF network, as well as using an evolutionary mutation strategy, allowed us to obtain the lowest errors in training and testing a polymorphic RBF network compared to known RBF network architectures. The created polymorphic RBF network practical application possibility is demonstrated experimentally using the helicopters turboshaft engines (using the example, the TV3-117 turboshaft engine) operating process parameters optimizing solving task using a multicriteria optimization algorithm. The optimal Pareto front was obtained, which made it possible to obtain the engine operation three additional modes: maximum reduction of specific fuel consumption at the total pressure in the compressor increase degree increased value by 5.0 %, specific fuel consumption minimization at the total pressure in the compressor increase degree reduced value by 1.0 %, the total pressure in the compressor increases degree optimal value with a slight increase in specific fuel consumption by 10.5 %. Future research prospects include adapting the developed methods and models into the general concept for monitoring and controlling helicopter turboshaft engines during flight operations. This concept is implemented in the neural network expert system and the on-board automatic control system.

This work focuses on developing a universal onboard neural network system for restoring information when helicopter turboshaft engine sensors fail. A mathematical task was formulated to determine the occurrence and location of these sensor failures using a multi-class Bayesian classification model that incorporates prior knowledge and updates probabilities with new data. The Bayesian approach was employed for identifying and localizing sensor failures, utilizing a Bayesian neural network with a 4–6–3 structure as the core of the developed system. A training algorithm for the Bayesian neural network was created, which estimates the prior distribution of network parameters through variational approximation, maximizes the evidence lower bound of direct likelihood instead, and updates parameters by calculating gradients of the log-likelihood and evidence lower bound, while adding regularization terms for warnings, distributions, and uncertainty estimates to interpret results. This approach ensures balanced data handling, effective training (achieving nearly 100% accuracy on both training and validation sets), and improved model understanding (with training losses not exceeding 2.5%). An example is provided that demonstrates solving the information restoration task in the event of a gas-generator rotor r.p.m. sensor failure in the TV3-117 helicopter turboshaft engine. The developed onboard neural network system implementing feasibility on a helicopter using the neuro-processor Intel Neural Compute Stick 2 has been analytically proven.