Training neural networks to filter noise in voice-driven energy systems

Moscow Aviation Institute MAI (National Research University), Volokolamsk highway, 4, 125993, Moscow, Russia

^* Corresponding author: vladivan25@yandex.ru

Abstract

The paper notes the labor-intensive and time-consuming process of compiling a training sample of a neural network, the influence of the compiled sample on the quality of network operation, and voice control tools for energy consumption systems, presents the main problems of training sample formation. It is noted that due to the simultaneous influence of a large number of factors on the quality of network operation, corrective changes in the training sample can lead to random results and disruption of energy consumption systems. In order to exclude randomness, it is suggested to use an approach based on comparison of the results of the trained neural network and the tested method. Recommendations on the volume and content of the sample depending on the nature of the filtered noise for homogeneous and dynamically changing noise are formulated. It is suggested to divide the training sample into groups with related properties in order to improve the efficiency of the training sample formation process. It is noted that due to the formal approach realized by neural networks, it is possible to incorrectly assess the impact of specific corrective changes on the quality of neural network performance. To reduce this factor and increase the efficiency of the process of forming the training sample, it is proposed to use an approach based on comparing the quality of noise filtering by a neural network and the results of filtering by a tested method based on physical principles of signal conversion. A process that improves the efficiency of neural network training sample formation and improving the quality of voice control of energy consumption systems is proposed. The requirements to the recommended method are formulated. It is proposed to use a transformed filtering method based on subtracting the noise spectrum from the signal spectrum containing noise. It is recommended to apply the method correctly depending on the frequency range of the noise and its characteristics.