Research On Anomaly Detection And Its Application Based On Machine Learning Algorithm

As the power core of modern ships,marine diesel engine has a long history.With the development of technology,the internal structure of marine diesel engine is becoming more and more complex,and the corresponding failure rate is also increasing.Abnormal detection of marine diesel engine running state can find the fault of marine diesel engine in time or even in advance,improve the safety of ship navigation and reduce the workload of maintenance personnel.The main contents of this paper are as follows:First of all,the operation data of marine diesel engine is cleaned.The method of one class support vector machine,deep belief network and autoencoder are used for anomaly detection.After comparing the effect of anomaly detection,the autoencoder is selected for improvement.Firstly,the convolution autoencoder based on one-dimensional convolution is used.The reason why the convolution autoencoder based on one-dimensional convolution is used is that the operation data of marine diesel engine is one-dimensional time series data,and one-dimensional convolution will convolute in time dimension,so that the model can consider the correlation of input data in time dimension.Then,the sliding window is added,and the convolution autoencoder based on one-dimensional convolution is changed into the convolution autoencoder based on two-dimensional convolution.The sliding window is used to add a dimension to the input data,so that the time series data can be connected in a larger span.Because a dimension is added,the convolution core also needs to add a dimension.Finally,VMD is added to decompose the data,and the convolution autoencoder based on two-dimensional convolution is changed into convolution autoencoder based on three-dimensional convolution.Using VMD to decompose the data is to make the complex time series data relatively simple,so that each mode only contains relatively concentrated frequency components,so that the signal changes from non-stationary signal to stationary signal,and stationary signal is easier to learn.Because the data is expanded by another dimension after VMD decomposition,the convolution core also needs to add another dimension.After adding the sliding window,the AUC score of model anomaly detection is improved by more than 0.05.After decomposing with VMD,the AUC score of anomaly detection is improved by about 0.02.Secondly,for the time series prediction of marine diesel engine running state data,four basic methods,support vector regression(SVR),xgboost,LSTM and ESN,are selected for comparison.Then,based on LSTM,an improvement is made by adding variational mode decomposition(VMD)to decompose the original signal according to frequency,and the signal changes from non-stationary signal to stationary signal,which is easier to learn.The relative error of the improved prediction model is reduced by more than 25%.Finally,the specific application process of anomaly detection and time series prediction in marine diesel engine is described.In the modeling process of time series prediction,the data are classified firstly,and different algorithms are used to train the prediction model for different kinds of data.VMD-WD-AE is directly used to train the model of anomaly detection.Then,an off-line learning and on-line recognition marine diesel engine anomaly prediction system is designed.The existing historical data are used to train and update the anomaly detection and prediction model,and then the trained model and related parameters are sent to the on-line recognition part,so that it can predict and detect the real-time data.The test results show that the AUC score of each abnormal condition can reach above 0.95.