基于改进LSTM的桥梁传感器异常数据的检测方法

Detection Method of Bridge Sensor Abnormal Data Based on Improved LSTM

桥梁正常与否通常通过传感器来检测,但是庞大的数据量对于传统检测方来说存在很大挑战,因此提出基于长短期记忆模型循环神经网络(Long Short Term Memory、LSTM)的方法进行异常检测。首先利用小波变换与奇异谱分析(Singular Spectrum Analysis、SSA)对传感器数据进行预处理,之后利用两层LSTM对序列进行向量表示、逆序重构,利用贝叶斯优化算法对LSTM网络进行参数优化,最终通过极大似然估计(Maximum Likelihood Estimate、MLE)对该段序列进行异常得分估计,最终通过学习异常报警阈值实现时间序列异常检测并发现潜在异常。采用桥梁某部位的应变数据、风速数据与振动传感器数据进行仿真实验,验证了所提方法相比其他传统方法具有更高的精确性。

Whether the bridge is normal or not is usually detected by sensors,but the huge amount of data is a great challenge for the traditional detection parties.Therefore,a long short term memory(LSTM)based method for anomaly detection is proposed.Firstly,wavelet transform and singular spectrum analysis(SSA)are used to preprocess the sensor data,then two-layer LSTM is used to represent the sequence and reconstruct the sequence in reverse order,Bayesian optimization algorithm is used to optimize the parameters of LSTM network,and finally maximum likelihood estimation(MLE)to estimate the abnormal score of the sequence,and finally through learning the abnormal alarm threshold to detect the time series abnormal and find the potential abnormal.The strain data,wind speed data and vibration sensor data of a certain part of the bridge are used for simulation experiments,which verify that the proposed method has higher accuracy than other traditional methods.