滚动轴承是旋转机械中广泛应用的重要零部件，其正常运行对于保障设备的稳定运行和延长使用寿命具有至关重要的作用。然而，滚动轴承在长期运行中可能会出现各种故障，如磨损、疲劳、润滑不良等，滚动轴承作为关键部件，一旦发生意外情况，将会造成严重的经济损失及安全隐患。传统故障诊断方法在处理高度复杂非线性信号时存在局限性，且过度依赖专家经验。本文通过深度学习方法进一步优化模型的诊断精度，提高不同工况和大批量轴承数据条件下的诊断精度。主要研究内容如下：

(1) 针对传统故障诊断方法通常仅采用单域信息输入，往往会出现部分信息丢失或信息的不完整使用，导致故障诊断精度低的问题，提出了一种基于多通道GM与改进支持向量机的轴承故障诊断识别方法。首先利用快速傅里叶变换和连续小波变换得到原始振动信号的频域和时频域信息；然后将时域、频域和时频域三种不同域信息作为样本输入构建的多通道GM网络进行多域特征融合；最后输入到采用改进的麻雀搜索算法优化支持向量机分类层进行故障诊断分类识别。实验结果证明，与其他算法相比，该算法在三个轴承数据集上的故障诊断识别准确率分别达到99.82%、98.86%和97.92%。证明该方法具有良好的诊断识别效果。

(2) 针对传统故障诊断网络模型对多变工况和大批量滚动轴承数据条件下，诊断识别准确率较低的问题，提出了一种基于自适应特征融合与多通道GM的轴承故障诊断识别方法。首先设计了多有效通道注意力机制通过赋予不同通道不同的权重，灵活地捕捉不同域信息全局和局部的关系；其次在时域、频域和时频域信息进行特征融合时引入自适应特征融合算法，使得模型能够对不同域融合重要特征进行动态加权，实现特征加权融合诊断；最后，与其他方法相比，该方法在三个轴承数据集上的准确率分别达到99.91%、99.84%和99.48%，召回率分别达到99.79%、99.83%和99.47%；并在不同实验台和工况条件下的组合轴承数据集原始振动数据中引入高斯噪声，验证了该方法具有良好的泛化和抗噪性能。

本文以滚动轴承故障诊断为背景，针对传统方法的局限性提出了基于深度学习的新方法。通过多通道GM和改进支持向量机，以及自适应特征融合与多通道GM相结合的方法，有效提高了故障诊断识别精度和泛化性能。为轴承故障诊断提供了新的思路和解决方案。

Rolling bearing is an important component widely used in rotating machinery, and its normal operation plays a vital role in ensuring the stable operation and prolonging the service life of the equipment. However, rolling bearings may have various failures in long-term operation, such as wear, fatigue, poor lubrication, etc. As a key component, rolling bearings will cause serious economic losses and safety hazards in case of accidents. Traditional fault diagnosis methods have limitations in dealing with highly complex nonlinear signals and are overly dependent on expert experience. In this paper, we further optimize the diagnostic accuracy of the model through deep learning methods to improve the diagnostic accuracy under different working conditions and large amount of bearing data. The main research content is as follows:

(1) Aiming at the problem that traditional fault diagnosis methods usually use only single-domain information input, which often results in partial loss of information or incomplete use of information, leading to low fault diagnosis accuracy, a bearing fault diagnosis and identification method based on multi-channel GM with improved support vector machine is proposed. Firstly, the frequency domain and time-frequency domain information of the original vibration signal is obtained by using the fast Fourier transform and continuous wavelet transform; then the three different domains of time domain, frequency domain and time-frequency domain information are input as samples into the constructed multi-channel GM network for multidomain feature fusion; finally, the inputs are optimized into the optimization of the classification layer of the support vector machine for the classification and identification of the fault diagnosis by using the improved sparrow searching algorithm, and the experimental results proved that the algorithm can achieve the best performance in the three bearing data compared with the other algorithms. Compared with other algorithms, this algorithm achieves 99.82%, 98.86% and 97.92% fault diagnosis and identification accuracy on the three bearing data sets, respectively. It proves that the method has good diagnosis and identification effect.

(2) Aiming at the problem that the traditional fault diagnosis network model has low diagnosis and recognition accuracy under the conditions of variable working conditions and large amount of rolling bearing data, a bearing fault diagnosis and recognition method based on adaptive feature fusion and multi-channel GM is proposed. Firstly, a multi-effective channel attention mechanism is designed to flexibly capture the global and local relationships of different domain information by giving different channels different weights; secondly, an adaptive feature fusion algorithm is introduced in the feature fusion of time domain, frequency domain and time-frequency domain information, which makes the model capable of dynamically weighting different domains to fuse the important features, and realizes the feature-weighted fusion diagnosis; lastly, the method is more accurate compared with other methods on the three bearing datasets reaches 99.91%, 99.84% and 99.48% accuracy and 99.79%, 99.83% and 99.47% recall, respectively; and Gaussian noise is introduced into the original vibration data of the combined bearing dataset under different experimental platforms and working conditions, which verifies that the method has a good generalization and anti-noise performance.

In this paper, a new method based on deep learning is proposed to address the limitations of traditional methods with the background of rolling bearing fault diagnosis. The fault diagnosis recognition accuracy and generalization performance are effectively improved by multi-channel GM and improved support vector machine, and adaptive feature fusion combined with multi-channel GM. New ideas and solutions are provided for bearing fault diagnosis.