深度学习网络是一种端到端的黑箱模型，对网络进行可解释分析可以更深入理解网络内部运行机理以便合理优化网络结构和调参。目前基于Transformer网络的轴承故障诊断模型必须借助时频分析等方法对时域信号进行升维，使其转换为时频图等二维图像以便于可解释分析，这种方法存在着升维过程中参数固定、网络参数量过大和网络可解释性较差等缺点。针对以上问题提出融合网络内自适应升维方法、卷积自注意力模块（Convolutional self-attention module）以及中间层类激活热力图（Mid Layer Class Activation Map，ML-CAM）的卷积自注意力自适应升维网络（Convolutional Self-attention Adaptive Dimension-Increasing Network，CSADI-Net）。卷积自注意力模块使用卷积层获取特征图的查询（Query,Q）、键（Key,K）和值（Value,V），极大的减少了可训练参数量；网络内自适应升维方法通过内部特征图拼接等操作将升维过程与网络训练融合，使其具有自适应参数调节的能力；中间层类激活热力图可以将网络内升维方法所得二维特征图上各个部位的关注度以热力图的形式展现，是一种直观的可视化可解释分析方法。此外，对CSADI-Net和ML-CAM进行了测试，CSADI-Net在凯斯西储大学轴承数据集上的准确率可以达到97.32±0.12%并且可以完全分类大连大学实测轴承故障数据集，同时使用ML-CAM对CSADI-Net在两数据集各样本上绘制了类激活热力图，解释了网络运行机理，证实CSADI-Net具有高准确率，高抗噪性能，可解释性强等优点。

Deep learning networks are considered as end-to-end black-box models. Performing interpretable analysis of these networks can facilitate a deeper understanding of their internal operations, enabling rational optimization of network architecture and parameter tuning. Currently, bearing fault diagnosis models based on Transformer networks require the use of methods such as time-frequency analysis to elevate the dimensionality of time-domain signals, transforming them into two-dimensional images like time-frequency graphs for interpretable analysis. However, this approach has drawbacks, including fixed parameters during the dimensionality expansion process, excessive network parameterization, and limited interpretability. To address these issues, we propose a Convolutional Self-Attention Adaptive Dimension-Increasing Network (CSADI-Net) that combines the integration of an in-network adaptive dimensionality expansion method, a Convolutional Self-Attention Module, and Mid Layer Class Activation Maps (ML-CAM). The Convolutional Self-Attention Module efficiently computes query (Q), key (K), and value (V) matrices from feature maps using convolutional layers, significantly reducing the number of trainable parameters. The in-network adaptive dimensionality expansion method seamlessly incorporates the dimensionality expansion process with network training, endowing it with adaptive parameter tuning capabilities. ML-CAM visualizes the network's focus on various regions of the two-dimensional feature maps obtained through the in-network dimensionality expansion method, offering an intuitive and visually interpretable analytical approach. Furthermore, we conducted tests on CSADI-Net and ML-CAM. CSADI-Net achieved an accuracy rate of 97.32±0.12% on the Case Western Reserve University bearing dataset and achieved complete classification on the experimentally measured bearing fault dataset from Dalian University. Additionally, ML-CAM was employed to generate class activation heatmaps for each sample in both datasets, elucidating the network's operational mechanisms. These results affirm that CSADI-Net possesses merits such as high accuracy, robust noise resistance, and strong interpretability.