时序数据多尺度缺失值填充方法OA

Due to the widely available missing data in data,missing value imputation has become an extremely important issue in data processing.Equipment health management,such as fault diagnosis,life prediction,and early detection of weak faults,relies on sensor data,which can seriously affect the analysis results of such methods when missing values are present in the sensor data.Most existing missing value imputation methods focus on overall trends while ignoring data details,but the details often reflect the true situation of the device,especially in weak fault detection.To address these issues,a wavelet transform-based missing value imputation method(WTMI)is proposed.The method first decomposes the data by wavelet to obtain multi-scale data,then reconstructs the multi-scale data separately using an auto-encoder,and finally obtains the final prediction value by using the inverse wavelet transformation.Considering the correlation between different scales and the problem of scale selection,we further proposed a multi-scale missing value imputation(MSMI).MSMI uses a stacked neural network to link data at different scales,so that the information at different scales can be fully utilized and more accurate predictions can be obtained.The experimental results of imputation and fault diagnosis on artificial and practical datasets show that the imputation accuracy of this method is similar to existing methods,and it improves the fault diagnosis accuracy by 2%compared with existing filling methods,especially by 5%for weak faults.