Research On The Safety Warning Model Of Power Grid Dispatching Operation Based On Speech Emotion Recognition

In recent years,with the rapid expansion of the scale of the power grid,the structure of the power system has become more and more complex,and safety accidents in power production have occurred from time to time.Studies have concluded that human error is the cause of more than 70% of power production safety accidents,and personal emotional state is one of the main factors leading to human error.Smart grid development needs to minimize the mistakes of power production personnel.Emotional changes of power grid dispatchers are influenced by external environment,psychological and physiological factors,and are reliable monitoring indicators to reflect dispatchers’ working status.By detecting the fluctuation of their voice emotion change,when the unfavorable emotion value is persistently very high,the dispatching supervisors get the early warning prompt,which is conducive to timely intervention and effective measures to reduce the risk of affecting the safe and reliable operation of the power system.Therefore,the research of voice emotion recognition in this context has great practical significance and value.In order to better meet the requirements of the power grid for dispatching safety,the following research work is carried out in the thesis:Firstly,the paper analyzes the relationship between safe and reliable operation of power system and the emotional state of power practitioners,and shows the importance of monitoring the emotional fluctuations of power dispatchers through speech emotion recognition technology in this context.It also compares the research status of domestic and foreign research in the field of speech emotion recognition from three aspects: speech emotion recognition features,speech emotion recognition methods and speech emotion recognition applications.Next,the theory related to speech emotion recognition is outlined,including the basic process of speech emotion recognition,speech emotion description model,speech emotion database and the process of pre-processing speech signals,and the features related to speech emotion are summarized.The methods of feature fusion and the evaluation criteria of speech emotion recognition results are detailed,and the theoretical basis of relevant deep learning models used for speech emotion recognition techniques is introduced.Then,a one-dimensional convolutional neural network(1DCNN)model is constructed to address the problem of insufficient utilization of traditional features,and a hybrid neural network speech emotion recognition model based on 1DCNN and long short-term memory network(LSTM)and based on a combination of 1DCNN and gated recurrent unit(GRU),respectively,is established considering the relationship between speech emotion features and historical information.To address the current situation that the size of the emotional speech inventory of the experimental corpus cannot meet the amount of data required by the deep learning model,three better data enhancement methods are proposed: adding Gaussian white noise(AWGN),time stretching(TS)and waveform displacement(RS)to improve the recognition rate of a single classifier by about 29%.Finally,a weighted average integrated model is proposed to obtain comprehensive and stable speech emotion recognition results,with recognition rate improvement of more than 2% compared to all three single classifiers,and the improvement of the weighted average integrated model reaches about 30% using data augmentation.Finally,the idea of speech emotion recognition early warning for grid dispatchers is sorted out and the corresponding implementation plan is proposed in conjunction with the actual scenario of grid dispatching room.The weighted average integration model combined with attention mechanism is used as the speech emotion recognition module.For the problem of how to quantify emotion and reflect emotional fluctuations,the discrete emotion to dimensional emotion and combined with the central limit theorem are used to propose a security warning module,and the model combining the two modules provides a new solution for the research of reducing the risk of grid dispatching.