Abstract: The rapid advancement of seismic observation systems has ushered in an era of seismic big data, encompassing both artificial and natural events. This development presents new challenges that necessitates precise seismic event classification to enhance earthquake response strategies and seismological research.　　 This paper presents a novel approach for classification seismic events, integrating Gram’s angle field （GAF） with a multi-scale residual neural network, designated as QxceptionNet. The GAF method converts seismic wave time series into a graphic representation, preserving the interdependencies among different sampling points. The application of multi-scale convolutional analysis in this context reveals a more detailed feature set within the GAF images. Additionally, the residual structure of networks is designed to prevent performance degradation, even as network depth increases. The synergy of these elements forms the foundation of our proposed methodology.　　 The approach utilizes a dataset meticulously compiled and labeled by the Jiangsu Earthquake Agency, encompassing station records of natural earthquakes, artificial blasts, and collapses. Each record consists of three-channel waveforms. As an initial step, the raw seismic waveforms are normalized in amplitude, linearly detrended, and filtered to reduce effects from the recording instruments. To maintain consistency, all waveforms are truncated to a standardized length of 40 seconds. Subsequently, those waveforms are transformed into GAF images. These images, along with their respective category labels, serve as inputs for the QxceptionNet. The network’s training results in outputting probabilities, categorizing the waveform data into one of the three seismic event types. Our results indicate impressive classification accuracies across various seismic event categories, emphasizing the efficacy of this approach in seismic classification. The test results show that the accuracy for 1 078 natural seismic event records, 981 blast records, and 830 collapse records are 92.55% for classifications based on individual waveform records and 96.36% for those based on single station records. Notably, these accuracies surpass those achieved by other methods, such as support vector machine （SVM）, multi-layer perceptron （MLP） and short-time Fourier transform （STFT） combined with convolutional neural network （CNN）. These findings suggest that the combined use of the Gram’s angle field and multi-scale residual neural networks is highly effective in distinguishing different types of seismic events.　　In summary, this paper elucidates the process of transforming seismic waveform data into GAF images, the architectural design of the network model, and the comprehensive experimental setup and results. Looking ahead, broadening the scope of the classification system to encompass a more diverse array of seismic occurrences and integrating a wider variety of samples will be essential in creating an all-encompassing and intelligent seismic event detection system.