Abstract: Epilepsy is a common neurological disorder that is usually detected by electroencephalographic (EEG) signals. Visual inspection and interpretation of EEG is a slow, time-consuming process that is prone to errors and subjective variations. A discrete wavelet transform (DWT) and autoencoder (AE) coupled signal abnormality detection method is proposed to distinguish seizure signals from normal (seizure-free) signals. First, the wavelet transform is used to decompose the EEG signal into approximation and detail coefficients, and the number of wavelet coefficients is limited by rejecting insignificant coefficients according to the threshold criterion. Secondly, the DWT coefficients were encoded using a self-encoder. Finally, the EEG signal was analyzed to detect outliers, data reconstruction was performed by compressing the feature set, and a neural network classifier was used to detect epileptiform activity from epileptic-free signals. The results of this method were compared and validated with those of previous methods using the Bonn University database. The method achieved a good classification performance (99.93% accuracy, 100% specificity) using linear and nonlinear machine learning classifiers to detect seizure activity from EEG data. Therefore, this method can be considered as robust with good detection ability to distinguish seizure activity, seizure-free activity, and normal EEG activity with a simple linear classifier. The method is suitable for time series signal analysis with further detection and determination of abnormalities. The method of detecting abnormal EEG signals in epilepsy can provide an objective reference for diagnosing, treating, and evaluating epilepsy. Thus, it can reduce physicians' workload and improve treatment efficiency, which has important theoretical significance and practical application value.