STFT And LSTM Based Anomaly Detection Of Power Supply Data From Commercial Large Aircrafts

With the vigorous development of data mining and machine learning algorithms,a wave of big data has emerged in all walks of life.The aviation industry is no exception,major airlines have begun to attach great importance to big data,and have spent a lot of energy on research and analysis of aviation big data.China started late in the design and manufacture of the aviation industry,and the accumulation and analysis of aviation big data is even insufficient.In this research background,this thesis takes the AC power supply signal data of commercial large aircraft in aviation big data as the research object,focusing on the anomaly detection of key quality parameters in the frequency domain.At present,there are two shortcomings in the research methods for the key quality parameters of the power supply signal frequency domain.On the one hand,the timeliness of the anomaly detection results of the current method is very poor,and the detected abnormal information does not have much reference value.On the other hand,the accuracy of the anomaly detection results of the current method still needs to be improved.Based on the shortcomings of the above existing research methods,this thesis has carried out the improvement of the existing power supply signal frequency domain parameter anomaly detection methods,and the purpose is to design an anomaly detection method with high timeliness and high reliability.A new model based on frequency domain parameter value prediction was proposed and designed.At the same time,an airborne power source abnormality knowledge base was also established.As a result,the flight safety of commercial large aircraft is improved,and the experience accumulation of on-board power failure data is realized.There are two innovations in this research.One is to design and implement an anomaly detection model based on the Short Time Fourier Transform(STFT)and Long Short Term Memory(LSTM)recurrent neural network model.This model can realize the advanced prediction of the key quality parameter values in the frequency domain of the airborne power signal,effectively solves the shortcomings of the current method with poor timeliness,and has the characteristics of high timeliness.Another innovation is that the anomaly detection model has high detection accuracy and high reliability.Especially in the scenario where anomaly detection is performed a long time in advance,the accuracy of the model output results is much higher than other current methods.This research also did some in-depth expansion research work,including the optimization research of each module in the above model.Through the theoretical comparison and analysis between different implementation methods,we found the most suitable implementation method in this project scenario.At the end of this thesis is an experimental verification of all the contents of this research.On the one hand,a large number of experiments were performed on the anomaly detection model designed and implemented in this research.Compared with other current methods,this model has the advantages of high timeliness and high accuracy of prediction results.At the same time,the optimization effect of the abnormal knowledge base on the model is also verified.On the other hand,some comparative experiments for the purpose of optimization were carried out,and the experimental results were analyzed.