TFN: An Interpretable Neural Network with Time-Frequency Transform Embedded for Intelligent Fault Diagnosis (2209.01992v2)

Abstract: Convolutional Neural Networks (CNNs) are widely used in fault diagnosis of mechanical systems due to their powerful feature extraction and classification capabilities. However, the CNN is a typical black-box model, and the mechanism of CNN's decision-making are not clear, which limits its application in high-reliability-required fault diagnosis scenarios. To tackle this issue, we propose a novel interpretable neural network termed as Time-Frequency Network (TFN), where the physically meaningful time-frequency transform (TFT) method is embedded into the traditional convolutional layer as an adaptive preprocessing layer. This preprocessing layer named as time-frequency convolutional (TFconv) layer, is constrained by a well-designed kernel function to extract fault-related time-frequency information. It not only improves the diagnostic performance but also reveals the logical foundation of the CNN prediction in the frequency domain. Different TFT methods correspond to different kernel functions of the TFconv layer. In this study, four typical TFT methods are considered to formulate the TFNs and their effectiveness and interpretability are proved through three mechanical fault diagnosis experiments. Experimental results also show that the proposed TFconv layer can be easily generalized to other CNNs with different depths. The code of TFN is available on https://github.com/ChenQian0618/TFN.

• The paper introduces the TFconv layer that integrates time-frequency transforms into CNNs for enhanced interpretability in fault diagnosis.
• It demonstrates superior diagnostic accuracy, nearing 100% on benchmarks like the CWRU dataset compared to baseline models.
• The method facilitates real-time attention visualization in the frequency domain, supporting rapid training and effective few-shot learning.

Analyzing TFN: An Interpretable Neural Network for Intelligent Fault Diagnosis

The paper "TFN: An Interpretable Neural Network with Time-Frequency Transform Embedded for Intelligent Fault Diagnosis" introduces an innovative approach to mechanical system fault diagnosis by addressing the opacity of traditional Convolutional Neural Networks (CNNs). While CNNs excel in feature extraction and classification, their black-box nature weakens their reliability in critical applications where understanding the decision process is essential. This paper introduces the Time-Frequency Network (TFN), embedding a time-frequency transform (TFT) into CNNs, thereby enhancing interpretability without compromising performance.

Key Methodological Advancements

The core innovation of this research is the TFconv layer, a novel preprocessing layer that integrates with existing CNN architectures. Unlike traditional CNNs, which employ real-valued kernels initialized randomly, TFconv layers incorporate complex-valued kernels derived from time-frequency transforms. These layers utilize kernel functions parameterized to simulate traditional signal processing methods like the Short-Time Fourier Transform (STFT), Chirplet Transform (CT), and Morlet Wavelet Transform (WT). This integration enables the network not only to extract robust fault-related features but also to visualize the CNN’s attention within the frequency domain.

Three different kernel functions correspond to three types of TFconv layers: TFN-STTF, TFN-Chirplet, and TFN-Morlet. Each kernel function is governed by tunable parameters, such as frequency factors, which the model learns during training. The amplitude-frequency response—a core component of interpretability—is derived from analyzing the TFconv layer's output, which provides insight into fault-relevant frequencies utilized by the CNN.

Empirical Validation and Observations

The authors validate the TFN using three datasets: CWRU public bearing, planetary gearbox, and aerospace bearing datasets. Across different experiments, TFN models consistently outperform both the baseline CNN and contrast models like SincNet and Wavelet Kernel Network (WKN) by achieving higher diagnostic accuracies. For instance, TFN models attain accuracy rates nearing 100% on the CWRU dataset, notably surpassing other models particularly as the number of TFconv layer channels increases. Moreover, TFN models manifest a clearer frequency domain interpretability that aligns with the dataset's information frequency bands, thus supporting the hypothesis regarding CNN attention and fault feature relevance.

The detailed analysis also encompasses the TFN's convergence speed and few-shot learning capability, both of which are superior compared to traditional CNNs. The implications are substantial for scenarios requiring rapid model training and deployment with limited dataset availability. Furthermore, the paper illustrates that the TFconv layer is transferable across different CNN architectures without sacrificing performance, suggesting broad applicability.

Implications and Future Directions

The significance of this research lies in its dual contribution to both enhancing CNN interpretability in fault diagnosis applications and maintaining or improving diagnostic performance. By introducing a physically grounded interpretative layer in CNNs, the TFNs bridge the gap between expert-driven feature extraction and modern data-driven techniques, which is crucial for adoption in industries wherein interpretability and accuracy are equally vital.

Future research directions may involve expanding the adaptability of kernel functions, possibly incorporating advanced signal processing techniques for improved fault feature extraction. Additionally, exploring the integration of TFNs within other neural architectural paradigms such as autoencoders could expand their applicability beyond classification problems, potentially enhancing unsupervised fault detection frameworks.

Overall, this research provides a compelling case for adopting TFNs in domains demanding interpretable diagnostic AI systems, thereby positioning them well in advancing machine intelligence in industrial applications.