Research On Detection Technology Of Abnormal Driving Behavior Of Train Drivers Based On Unsupervised Learning

In recent years,with the rapid development of the transportation industry,the issue of train operation safety has become increasingly prominent.More than 1/5 of railway traffic accidents are caused by abnormal driving by train drivers.Therefore,detecting and effectively controlling abnormal driving behavior of train drivers is crucial for ensuring train operation safety.Traditional abnormal behavior detection methods have strong invasiveness,poor robustness,and high detection costs.Based on the mainstream Unsupervised learning "non positive or negative" anomaly detection idea,this thesis uses the Mel Frequency Cepstrum Coefficient(MFCC)chart of the driver’s call response during duty to detect whether the driver’s physiological state is abnormal,so as to determine whether there is abnormal driving behavior,and uses on-board monitoring video to detect whether the driver has abnormal driving behavior caused by bad habits.The main research content of this article is as follows:1.In response to the abnormal behavior of train drivers in abnormal physiological states,this thesis constructs an anomaly detection model based on generative adversarial recoding network.The model is based on the anomaly detection idea of reconstruction,which converts the driver’s call response voice into Mel frequency cepstrum coefficient graph through preprocessing and voice feature parameter extraction,and reconstructs it through the improved convolutional Autoencoder.According to the reconstruction error,judge whether the input MFCC graph is abnormal to determine whether the corresponding time is abnormal.The anomaly detection model based on generative adversarial recoding network has made the following improvements: adding a recoding network to recode the reconstructed image and calculate the reconstruction encoding error,making the judgment of abnormal Mel frequency cepstrum coefficient map more comprehensive;The Generative adversarial network is introduced,the decoder is used as the generative network and the local discriminator for countermeasure training,the convolutional Autoencoder is used as the generalized generator,and the global discriminator is used for countermeasure training to improve the reconstruction ability of the model;Transfer learning is introduced into the model classifier to adaptively modify the model classifier to adapt to the Mel frequency cepstrum coefficient map of drivers’ accents in different regions,so as to improve the generalization ability of the model on large-scale data sets.By using the Mel frequency cepstrum of the call response voice,abnormal behavior detection of drivers can be achieved without specifying the specific type of abnormality.The detection accuracy on the self-made dataset reached 89.7%,and the recall rate reached 62.7%.2.In response to abnormal driving behavior caused by bad habits of train drivers,this article constructs an anomaly detection model based on step-by-step video prediction.The first prediction network predicts future frames based on existing video frame sequences,and then selectively fuses the predicted frame sequence with the corresponding frame difference map sequence through a feature fusion module.The second prediction network predicts future frames and obtains the final predicted frame.Input the predicted frame into the convolutional layer for convolution to obtain abnormal probability values,and use the GAN model for adversarial training to force the prediction performance of the secondary prediction network to be better.Determine abnormal frames by combining image space loss and predicted frame anomaly probability values.The experimental results demonstrate that the stepwise video prediction network has good prediction performance,with significant differences between abnormal frames and corresponding prediction frames,and can effectively detect abnormal behavior frames.On the self-made dataset,the Receiver operating characteristic of the anomaly detection model in this thesis is more smooth,which shows that the algorithm in this thesis is progressiveness.