Anomaly detection in piston engines plays a pivotal role in maintaining the continuous airworthiness and operational safety of general aviation aircraft. However, real-life data present several challenges, including diverse operating conditions, intermixed normal and abnormal samples and complex fault phenomena, all of which complicate model training and manual labeling tasks. To address the challenges, an improved deep autoencoder-based framework under multiple working conditions (IMDAEF) is proposed to enhance the accuracy of anomaly detection and the performance evaluation process. First, in the proposed method, a clustering-guided approach is introduced for working condition identification. Piston engine parameters are used for K-means clustering, and aircraft flight phase information is incorporated to fine-tune the results. Second, during the deep autoencoder training process, a sample selection mechanism and an equivalent threshold method are conducted, considering the presence of outliers in normal operation data. Finally, during the model evaluation stage, recognising the complexity of real-life fault phenomena that hinder manual labeling, an outlier region division and selection mechanism is proposed to re-annotate subtle fault occurrences in the fault datasets, thereby improving model evaluation accuracy. Results from four representative fault datasets confirm that the proposed framework maintains strong detection performance against complex faults. Moreover, the experiments demonstrate its ability to mitigate the negative impact of real-life data characteristics while meeting real-time performance requirements, offering valuable insights for future research.
