A method for automatic recognition and anomaly detection of electrocardiogram signals based on deep neural network structure search has been proposed. Firstly, the raw ECG signals are converted into various image representations, including Gram angle field, recursive mapping, Markov transition field, etc., which enables the deep learning model to better handle these complex signal features. Meanwhile, this study utilizes convolutional neural networks for feature extraction and learns the complex relationships between features through fully connected layers. The results demonstrated that the improved method achieved a maximum accuracy of 98.5% and an average accuracy of 94.0% on the PhysioNet MIT-BIH dataset. Additionally, on the PTB dataset, the average recall rate of the improved method reached 98.4%, surpassing the performance of traditional neural networks and Canny algorithm. The experimental results indicate that the research method effectively optimizes the key patterns’ recognition ability in electrocardiogram signals and has excellent performance in detection results. This study offers a more reliable tool for early diagnosis and health management of human health diseases.