Brakes in mechanical equipment are crucial for operational safety, and their effects are directly affected by friction performance. The fault signal induced by friction interface presents the phenomenon of multi-source, and the fault samples are difficult to obtain in practical engineering. Both aspects yield unsatisfactory recognition performance of diagnosis models. To address the issues, in this article, we proposed an algorithm based on a modified convolutional neural network (CNN) and support vector machine (SVM). First, dynamic features were extracted from the friction factor and friction surface temperature as samples, which could effectively present the state of brake friction. Next, CNN was used to learn feature knowledge from dynamic feature set, the Mish activation function, batch normalisation and dropout were employed to complete the training of modified CNN (MCNN). Then, the dynamic feature set was input into the trained MCNN again to learn the feature representations of friction state. Finally, the feature representations were migrated to SVM to establish the mapping between feature space and label space, and the final fault recognition was completed. The proposed algorithm fully combined the powerful feature learning ability of MCNN and the excellent classification performance of SVM on small samples. Experiment results showed that MCNN-SVM had faster convergence speed, and the accuracy of the proposed algorithm reached 100%. Its diagnosis effect was better than counterpart algorithms.

机械设备中的制动器对于操作安全至关重要，其效果直接受摩擦性能影响。摩擦界面引起的故障信号呈现多源现象，实际工程中难以获取故障样本。这两个方面都产生了不令人满意的诊断模型识别性能。为了解决这些问题，在本文中，我们提出了一种基于改进的卷积神经网络 (CNN) 和支持向量机 (SVM) 的算法。首先，从摩擦系数和摩擦表面温度作为样本提取动态特征，可以有效地呈现制动摩擦状态。接下来使用CNN从动态特征集中学习特征知识，Mish激活函数，采用批量归一化和 dropout 来完成修改后的 CNN (MCNN) 的训练。然后，将动态特征集再次输入到训练好的 MCNN 中以学习摩擦状态的特征表示。最后，将特征表示迁移到SVM，建立特征空间和标签空间的映射关系，完成最终的故障识别。该算法充分结合了MCNN强大的特征学习能力和SVM在小样本上的优异分类性能。实验结果表明，MCNN-SVM具有更快的收敛速度，所提算法的准确率达到100%。其诊断效果优于对应算法。将动态特征集再次输入到训练好的 MCNN 中以学习摩擦状态的特征表示。最后，将特征表示迁移到SVM，建立特征空间和标签空间的映射关系，完成最终的故障识别。该算法充分结合了MCNN强大的特征学习能力和SVM在小样本上的优异分类性能。实验结果表明，MCNN-SVM具有更快的收敛速度，所提算法的准确率达到100%。其诊断效果优于对应算法。将动态特征集再次输入到训练好的 MCNN 中以学习摩擦状态的特征表示。最后，将特征表示迁移到SVM，建立特征空间和标签空间的映射关系，完成最终的故障识别。该算法充分结合了MCNN强大的特征学习能力和SVM在小样本上的优异分类性能。实验结果表明，MCNN-SVM具有更快的收敛速度，所提算法的准确率达到100%。其诊断效果优于对应算法。该算法充分结合了MCNN强大的特征学习能力和SVM在小样本上的优异分类性能。实验结果表明，MCNN-SVM具有更快的收敛速度，所提算法的准确率达到100%。其诊断效果优于对应算法。该算法充分结合了MCNN强大的特征学习能力和SVM在小样本上的优异分类性能。实验结果表明，MCNN-SVM具有更快的收敛速度，所提算法的准确率达到100%。其诊断效果优于对应算法。