Research On Speech Recognition Technology For Power Grid Dispatching

With the expansion of distribution network scale and the promotion of information construction,the information involved in distribution network command keeps increasing.Dispatchers need to perform a lot of repetitive work of issuing,receiving and checking orders every day,which is too much and may increase the risk of mistakes.Meanwhile,traditional grid dispatching usually uses telephone form for information transmission,which is easy to form information congestion and blockage when dealing with large faults,resulting in behavior mismatch and disposal failure.Therefore,today’s large-scale grid dispatching poses a huge challenge to dispatchers.With the rapid development of artificial intelligence technology,intelligent interaction is beginning to enter people’s lives.Speech recognition technology is also being gradually applied to grid dispatching.This also gives rise to the need to adopt intelligent virtual dispatchers to replace repetitive and tedious manual labor.Among them,the speech recognition link is related to the virtual dispatcher’s accurate understanding of the information reported by field personnel and is the basis for the correct processing and sending of dispatching instructions.Through speech interaction,it can quickly complete information transmission,effectively avoid information congestion,greatly improve the accuracy and timeliness of power grid dispatching commands,and effectively reduce accidents caused by wrong commands.However,there are some problems with speech recognition for power grid dispatching.First of all,grid dispatchers work in a noisy environment,and the noisy environment is an important factor affecting the speech recognition rate.Moreover,grid dispatching industry,like other industries,has its own domain-specific professional vocabulary,and it is difficult for public domain-oriented speech recognition technology to recognize the professional information vocabulary specific to grid dispatching.Therefore,it is necessary to study speech recognition with noise reduction capability and for grid dispatching.In this thesis,this project carries out a study on end-to-end grid dispatch speech recognition.It focuses on model noise immunity and recognition of specialized vocabulary for grid dispatching.The main research contents are as follows:(1)To enhance the noise immunity of the model,deep residual shrinkage network and gated convolutional network are introduced into the grid dispatching speech recognition in this thesis.The feature extraction ability of the convolutional neural network is improved by removing the redundant information in the threshold region through the shrinkage module in the deep residual shrinkage network,and the effective context is captured by the gated convolutional network.Based on this,and improving it,we propose a residual shrinkage convolutional network and a gated convolutional feedforward network to construct the end-to-end grid dispatching speech recognition model of RSCN-GCFN with joint CTC.(2)Since the end-to-end model based on CTC method does not introduce language model in decoding,which often involves domain-specific semantic understanding.Therefore,to solve the problem of recognition of specialized vocabulary for grid dispatch,the BERT network is applied and improved to adapt it to the grid dispatch language model task.Then it is connected with the end-to-end model based on CTC approach to make its model on the specialized vocabulary of grid dispatching with better recognition results.