行车指挥调度是铁路运输的核心监控岗位，检测调度员的疲劳动作对保障铁路运营安全具有重要意义。为了能识别铁路行车调度员疲劳动作，降低安全生产风险，提出一种基于双向长短时记忆神经网络和支持向量机的自适应增强算法对调度员疲劳状态下的动作进行识别。首先，通过高分辨率网络(High-Resolution Net，HRNet)人体关键点检测模型，提取多个人体关键点及人体动作行为角度特征与长度比例特征。其次，搭建基于双向长短时记忆神经网络和支持向量机(Bi-directional Long Short-Term Memory-Support Vector Machine，BiLSTM-SVM)的动作识别模型，使用正交试验法对模型参数进行优化，并采用自适应增强算法(Adaboost，Adaptive Boosting)进一步提升疲劳动作识别。最后，基于调度仿真疲劳动作数据，对该模型的有效性进行验证。结果显示，该模型的精确率为0。97、准确率为0。96、召回率为0。96、F1分数为0。96。该模型提高了人体疲劳动作分类的准确率，为调度员疲劳检测提供了判断依据。
Research on railway dispatcher fatigue action recognition method based on BiLSTM-SVM adaptive enhancement algorithm
Train operation command and dispatch is the core monitoring position of railway transportation,and detecting the fatigue status of dispatchers is of great significance for ensuring railway operation safety.A railway dispatcher fatigue action recognition method based on bidirectional long and short-term memory neural network and support vector machine adaptive enhancement algorithm is proposed to reduce the human factor risks in safety production.Using a High-Resolution Network(HRNet)human key point detection model to extract multiple human key points,the angle and length ratio features of human action behavior were extracted.An action recognition model based on a Bi-directional Long Short-Time Memory-Support Vector Machine(BiLSTM-SVM)was constructed.The model parameters were optimized using orthogonal experimental methods,Finally,an adaptive boosting algorithm(Adaboost,Adaptive Boosting)was adopted to further enhance fatigue action recognition.Using accuracy,recall,precision,and F1 Score as the model evaluation metrics,we performed the ablation experiment on LSTM,BiLSTM,and BiLSTM-SVM.The experimental results show that BiLSTM-SVM-Adaboost yields the best performance.The accuracy of the prediction model is 0.96,which is respectively 0.12,0.04,and 0.02 higher than that of the comparative model.The Recall of the prediction model is 0.96,which is respectively 0.12,0.03,and 0.02 higher than that of the comparative model.The Precision of the prediction model is 0.97,which is respectively 0.08,0.03,and 0.02 higher than that of the comparative model.The F1 Score call of the prediction model is 0.96,which is respectively 0.14,0.04,and 0.03 higher than that of the comparative model.The experimental results show that the BiLSTM-SVM-Adaboost algorithm on the scheduling simulation fatigue action dataset achieved an accuracy and precision of 0.96 and 0.97,compared to the"Openpose+ANN"and"DBN+LSTM"network algorithms,there is a performance improvement.The model improves the accuracy of human fatigue movement classification,providing a basis for fatigue detection of dispatchers.
万方数据
维普
