To improve the accuracy and objectivity of explosives’ impact sensitivity testing， machine learning methods were applied in the intelligent recognition of explosives’ impact response acoustic signals. Experiments on mixed explosives were conducted using a drop-weight impact sensitivity test device， with an audio acquisition system synchronously capturing acoustic signals during the impact process. One-dimensional time domain and frequency domain features， such as maximum value and bandwidth， were extracted. Short-Time Fourier Transform （STFT） was employed to convert audio data into frequency spectrograms. Data augmentation for one-dimensional data was performed using a Conditional Generative Adversarial Network（cGAN）， while a Deep Convolutional Generative Adversarial Network（DCGAN） was applied to enhance spectrogram data. Multiple machine learning models were employed for explosion classification， including Random Forest （RF）， eXtreme Gradient Boosting （XGBoost）， Back-propagation Neural Network （BPNN）， Support Vector Machine （SVM）， k-means clustering， Convolutional Neural Network （CNN）， and Vision Transformer （ViT）. Results demonstrate that RF， XGBoost， BPNN， and SVM achieve accuracy rates exceeding 99.5% on the real dataset and achieve 100% on the cGAN-augmented dataset. In contrast， k-means clustering initially reaches an accuracy of 98.5% on the real dataset， but accuracy shows a trend of increase followed by decline on augmented data. CNN and ViT achieve accuracies of 98.1% and 98.4% on the real dataset， respectively， and improved to 98.4% and 98.9% on augmented data. However， their performance remaine slightly lower than traditional models due to the constraints of small sample sizes and minor overfitting issues. The proposed deep learning-based intelligent recognition method for explosives’ impact sensitivity in this study achieved a high level of accuracy， demonstrating its reliability and practicality in the task of detecting explosive sound signals. At the same time， it effectively mitigates the subjectivity and inefficiency associated with traditional manual recognition methods， providing a reliable technical solution for the safety use of explosives.