Aims: Electrocardiogram (ECG) is an effective and non-invasive indicator for the detection and prevention of arrhythmia. This study aimed to identify the clinical diagnosis of a total of nine conditions including atrial fibrillation (AF) and ventricular premature beat complex (PVC) from 12-lead ECG records. Methods: In this paper, we propose a deep learning-based diagnosis approach, which uses 34-layer 1D deep residual network (ResNet34) to capture the correlation and signal characteristics between the 12-lead ECG records. Considering the long time series and multiple leads, we increase the size of the convolution kernel in each residual block from 3 to 7, thus increasing the receptive field of the network and achieving a significant performance improvement. Results: In the scoring system of PhysioNet / CinC Challenge, Our current score is F_2 Score-0.736、G_2 Score-0.518、Geometric Mean-0.617 and Run Time 17 minutes. However, our local test score(Divide the training set and validation set by a ratio of seven to three) is F_2 Score-0.804、G_2 Score-0.677. Conclusion: Based on the results, we believe that this method has good application prospects in clinical practice, especially in the auxiliary diagnosis of heart diseases of wearable devices.