随着智能化程度的提高，钻井行业对于实时识别随钻岩性的需求越来越迫切。基于钻孔过程中的钻进碎岩声音信号和振动信号（声振信号），提出了一种随钻岩性智能反演方法。首先，通过室内微钻实验钻进7种不同类型的岩石以获取原始信号样本，在采集过程中改变钻进参数（钻速、转速、钻头尺寸）并获得相应信号数据。针对采集的声振信号特征，通过短时傅里叶变换获得具有信号特征的时频图像。在此基础上，构建改进VGG16卷积神经网络模型以实现岩性智能识别，并通过超参数优选实现对模型的训练、评估和调优。之后引入迁移学习训练策略，分别以不同钻进参数作为数据标签，依据参数取值划分源域和目标域，以实现对小样本目标领域的快速识别。实验结果表明：随着钻进参数的改变，模型的迁移学习结果存在差异；基于声振信号训练得到的岩性反演模型预测精度高、泛化能力强，声音信号测试集准确率可达99%，振动信号测试集准确率可达100%；在机械钻速变化下，声振信号受到的影响最小，作为数据标签进行岩性反演时能取得更为卓越的效果，且小钻速作为目标域时岩性反演准确率最高。在岩性反演的过程中，不同的信号类型适用于不同的岩石。其中声音信号对粗黄砂岩的适用性最高、振动信号更适用于花岗岩。研究成果对于提升工作面钻进智能化程度具有一定借鉴价值。

With the improvement of intelligence, the drilling industry's demand for real-time identification of lithology while drilling was becoming more and more urgent. An intelligent inversion method of lithology while drilling is proposed based on the acoustic signal and vibration signal (acoustic vibration signal) of broken rock during drilling. Firstly, the original signal samples were obtained by drilling seven different types of rocks through indoor micro-drilling experiments. During the acquisition process, the drilling parameters(drilling speed, rotation speed, bit size) were changed and the corresponding signal data were obtained. According to the characteristics of the collected acoustic vibration signal, the time-frequency image with signal characteristics was obtained by short-time Fourier transform. On this basis, an improved VGG16 convolutional neural network model was constructed to realize the intelligent identification of lithology, and the training, evaluation and tuning of the model are realized by hyperparameter optimization. Then, the transfer learning training strategy is introduced, and different drilling parameters were used as data labels. According to the parameter values, the source domain and the target domain were divided to realize the rapid identification of the small sample target domain. The experimental results show that the transfer learning results of the model are different with the change of drilling parameters. The lithology inversion model based on acoustic-vibration signal training has high prediction accuracy and strong generalization ability. The accuracy of the acoustic signal test set is up to 99%, and the accuracy of the vibration signal test set is up to 100%. Under the change of penetration rate, the acoustic and vibration signals are least affected, which can achieve more excellent results when used as data labels for lithology inversion, and the accuracy of lithology inversion is the highest when the penetration rate is small as the target domain. In the process of lithology inversion, different signal types are suitable for different rocks. Among them, the sound signal has the highest applicability to coarse yellow sandstone, and the vibration signal is more suitable for granite. The research results have certain reference value for improving the intelligent degree of working face drilling.