Fault diagnosis method of telecom cloud platform based on deep CNN model

In response to the problem of fault location caused by massive alarm logs in the complex architecture of telecom cloud platforms, this study proposes a time-frequency image recognition model (WCNN) based on depthwise separable small convolution kernels, which replaces traditional pooling layers to achieve efficient feature extraction. We propose a time-frequency image recognition model based on depthwise separable small convolution kernels to address the issue of information loss caused by improper handling of fuzzy features in traditional pooling methods. The experimental results show that in extreme noise environments with a signal-to-noise ratio of -4dB, the WCNN model achieves a recognition accuracy of 90%, significantly better than FFT-SVM (<60%), FFT-KNN (<60%), FFT-BP (80%), and FFT-DNN (80%). In addition, under low noise conditions (signal-to-noise ratio>6), the accuracy of the WCNN model is further improved to 99.3%, and the model complexity is reduced by 42% compared to traditional convolutional neural networks, resulting in a 30% increase in computational efficiency. The research has verified the anti-interference ability and feature preservation advantages of the WCNN model in strong noise environments, providing an efficient solution for fault diagnosis in telecommunications cloud platforms.