Research On Bimodal Emotion Recognition Of Power Grid Dispatcher Based On Deep Learning

Human error is one of the main causes of power system accidents.The requirement of the national smart grid development is to reduce the human error in the grid dispatching operation system as much as possible.The mood change of the power grid dispatcher is an important manifestation of the physiological,behavioral and cognitive components.By monitoring the emotional change curve and connecting with the grid events,it is beneficial for the grid general dispatch to effectively intervene to deal with certain emergencies.According to the actual dispatching scenario of the power grid,the experimental scheme is designed by communicating with the power grid dispatching expert and the common research ideas of the “human-machineloop” system.The pre-experiment effect is used to modify the scheme,and finally the non-contact grid scheduling video recording and voice collection scheme is determined to obtain experimental data.For the grid dispatching scenario,most expressions of the dispatcher are not obvious when doing the normal work.The PPDN network input is two pictures with the same type but different intensity expressions.The PPDN architecture is beneficial to enhance the contrast learning of specific expressions.The convolution module uses the fine-tuned Inception-V3 network and uses the PGS back propagation algorithm to optimize the training process.After pre-training in the public CK+ dataset and achieving satisfactory results,the collected grid dataset is used for cascading training.Finally,the experiment achieves the accuracy rate of 80.4%.In the actual context,because the emotion is affected by the near phonemes,the paper designs to use the Bi-LSTM network architecture.For the speech sequence input,there is a problem that the speech segment has pauses and some silence,and there is invalid character.The paper designs a network with a connectionist temporal classification(CTC)method to solve the problem.The Bi-LSTM network without CTC module and the Bi-LSTM network with CTC module were separately trained.The experimental results show that the accuracy of using CTC is improved by 4.3%.After the expression and the voice emotion recognition are separately trained,a good recognition effect is achieved.According to the theory of emotional research,the paper designs the rules and fuses the data of two channels based on the decision level.The obtained emotion sequence data is divided into sliding window,and the research calculates each window to effectively reduce the dimensionality of the emotional sequence.The paper classifies a large number of complicated tasks of power grid dispatching,and combines the dispatch logs to organize the task categories according to the timeline.By correlating the emotional change trend with the scheduling task,the rationality of the proposed algorithm can be verified.