⚡ electrical engineering

FE-MCFormer: a novel time-frequency interpretable architecture for machinery fault diagnosis under strong noise environments

This paper proposes FE-MCFormer, a novel time-frequency interpretable architecture featuring a frequency adaptive learning layer and multiscale fusion mechanism to achieve robust and interpretable machinery fault diagnosis even under severe noise conditions down to -10 dB SNR.

Original authors: Yuhan Yuan, Xiaomo Jiang, Haibin Yang, Haixin Zhao, Shengbo Wang, Xueyu Cheng, Jigang Meng

Published 2026-07-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuhan Yuan, Xiaomo Jiang, Haibin Yang, Haixin Zhao, Shengbo Wang, Xueyu Cheng, Jigang Meng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but the crime scene is a chaotic, deafening rock concert. Your job is to find a specific, tiny sound—a single sneeze or a whisper—that tells you exactly what went wrong. In the world of heavy industry, this "crime scene" is a massive machine like a turbine or a compressor, and the "sneeze" is a subtle vibration that signals a broken part. The problem is that these machines are often so loud and the environment so messy that the helpful clues get drowned out by a wall of noise. For a long time, scientists have built computer programs (called deep learning models) to listen for these clues. However, most of these programs are like detectives who get confused by the noise; they either miss the clue entirely or, worse, they guess the wrong thing and can't explain why they guessed it. In the real world, if a computer says a machine is broken, engineers need to know how it knows, or they won't trust the advice. This paper dives into the messy, noisy corner of industrial science to build a smarter detective that can hear the whisper through the roar and show you exactly where it heard it.

The researchers behind this study, Yuhan Yuan and their team, have built a new digital detective named FE-MCFormer. Think of this new system as a super-powered pair of noise-canceling headphones combined with a magnifying glass. Its main superpower is that it doesn't just guess; it explains its reasoning in a way humans can understand, looking at both the rhythm of the sound (time) and the pitch of the sound (frequency).

The secret sauce of FE-MCFormer is a special first step called the Frequency Adaptive Learning Layer (FALL). Imagine you are trying to hear a friend's voice in a crowded room. Most people just shout louder, but this layer is like a smart filter that instantly learns which frequencies belong to your friend and which belong to the background chatter. It actively suppresses the "noise" (the chatter) while keeping the "signal" (your friend's voice) clear. The paper shows that this layer acts like a spectral shaper, explicitly cutting out the messy, noisy parts of the sound wave while preserving the specific harmonic structures that indicate a machine is sick.

Once the sound is cleaned up, the system uses a Multiscale Time–Frequency Fusion (MSTFF) module. If the first step was cleaning the audio, this step is like having a team of detectives with different-sized flashlights. Some flashlights are small and zoomed in to catch quick, sharp "impulses" (like a sudden clank), while others are wide to see the big picture of how the sound waves interact over time. By combining these views, the model can spot faults that are hidden in the noise, whether they are tiny, sharp cracks or larger, slower vibrations.

The team tested this new detective on two very different "crime scenes." The first was a standard dataset of rolling bearings (the wheels inside machines), and the second was a real-world dataset from a massive centrifugal compressor used in industry. They didn't just test it in a quiet room; they blasted the data with heavy noise, simulating environments as loud as -10 dB (which is extremely noisy, where the noise is much louder than the signal).

The results were impressive. In these noisy conditions, older methods like standard CNNs or Transformers started to stumble, with their accuracy dropping significantly as the noise got worse. For example, on the bearing dataset at -10 dB noise, a standard model might only get about 50% of the diagnoses right (basically guessing). In contrast, the new FE-MCFormer managed to get 72.99% right, and its lighter version, FE-MCFormer-s, got 70.44%. As the noise got slightly quieter (moving to -2 dB), FE-MCFormer soared to 99.11% accuracy, while other models struggled to keep up. On the real-world compressor dataset, the results were even more striking, with the model achieving nearly 100% accuracy (99.98%) at -2 dB and still holding strong at 93.48% even in the heaviest -10 dB noise.

Perhaps the most exciting part is that the paper proves this isn't just a "black box" that gives lucky guesses. The researchers showed that when FE-MCFormer identifies a fault, it is actually focusing on the right physical parts of the sound. In the visualizations, the model's "attention" (where it looks) lands exactly on the specific frequencies where a broken bearing or a cracked blade would sing. It successfully filters out the random noise and highlights the specific "sneezes" of the machine.

In short, this paper suggests that by teaching a computer to first clean up the noise in the frequency domain and then look at the sound through multiple time-scales, we can build fault diagnosis systems that are not only incredibly accurate in loud, messy factories but also trustworthy because they show us exactly what they are listening to. It's a step toward machines that can tell us they are sick, even when the factory floor is screaming.

Technical Summary: FE-MCFormer for Time–Frequency Interpretable Fault Diagnosis

1. Problem Statement

Rotating machinery is critical to industrial safety and efficiency, yet its operational reliability is frequently compromised by unexpected failures. While data-driven fault diagnosis (FD) methods have advanced significantly, they face two primary challenges in real-world industrial settings:

Strong Noise Interference: Harsh operating environments introduce severe background noise (e.g., electromagnetic interference, variable loads), which submerges fault-related impulsive features and degrades the performance of existing models.
Lack of Interpretability: Many deep learning models operate as "black boxes." Their internal decision-making mechanisms are rarely investigated, reducing trust and practical applicability in industrial operations. Furthermore, existing interpretable methods often focus on either the time domain or the frequency domain in isolation, failing to provide a unified time–frequency interpretability that is crucial for diagnosing complex faults where information may be distributed across both domains.

2. Methodology: FE-MCFormer

The authors propose FE-MCFormer (Frequency-Enhanced Multiscale Convformer), a novel architecture designed to achieve robust and interpretable fault diagnosis under strong noise conditions. The framework integrates three core components:

2.1. Frequency Adaptive Learning Layer (FALL)

Positioned at the input stage, FALL performs first-layer interpretable diagnosis through learnable spectral reconstruction.

Mechanism: It employs a discrete Fourier transform (DFT) to convert input signals into the frequency domain. A learnable spectral weight vector ( $W$ ) is applied via element-wise multiplication to attenuate noise-dominated frequency components while preserving fault-sensitive harmonic structures.
Function: Unlike conventional convolutional filters that implicitly learn frequency responses, FALL explicitly shapes the spectrum. It acts as a global encoder that suppresses broadband industrial noise before deep feature extraction begins.

2.2. Multiscale Time–Frequency Fusion (MSTFF) Module

This module jointly captures localized impulsive characteristics and structured global spectral interactions without the quadratic computational cost of standard Multi-Head Self-Attention (MHSA). It consists of two sub-components:

Multiscale Convolutional Attention Layer (MSCAL): Utilizes parallel branches with $1\times3$ and $1\times5$ convolutional kernels to extract multiscale temporal features. A softmax-based attention mechanism adaptively scales these features based on diagnostic relevance, allowing the model to focus on localized fault impulses that might be obscured in global sequences.
Time–Frequency Fusion Network (TFFN): Combines local time-domain modeling with global frequency-domain dependencies. It uses depthwise convolutions for local patterns and integrates the Frequency Adaptive Learning (FAL) operation within the frequency domain to model long-range dependencies and enhance non-linear representation.

2.3. Distillation Layers

Hierarchical distilling layers are placed between MSTFF blocks. These layers employ 1-D convolution and max-pooling to halve the temporal dimension at each stage. This reduces memory usage and computational burden while creating focused global feature maps, ensuring the model remains computationally efficient for long vibration sequences.

3. Key Contributions

The paper outlines four main contributions:

Novel Architecture: The development of FE-MCFormer, a framework that integrates time–frequency interpretable features to diagnose turbomachinery faults under strong noise.
Frequency Adaptive Learning (FALL): The introduction of a layer that adaptively encodes global spectral information and suppresses noise-dominated components in the first layer, enhancing both robustness and interpretability.
Multiscale Time–Frequency Fusion: The design of the MSTFF module, which combines multiscale attention and time–frequency fusion networks to capture both local impulsive behaviors and global spectral interactions.
Validation: Extensive experimental validation on two datasets (a benchmark rolling bearing dataset and a real-world centrifugal compressor dataset) demonstrating stable performance down to -10 dB SNR and providing interpretability analysis.

4. Experimental Results

The authors evaluated FE-MCFormer against seven state-of-the-art methods (including MSCNN-LSTM, ResNet50, DenseNet, Vanilla Transformer, and Li-convformer) on two datasets:

Case 1: Paderborn University (PU) Rolling Bearing Dataset

Conditions: 18 fault states (normal, outer-ring, inner-ring) with Gaussian white noise added at SNRs ranging from -10 dB to -2 dB.
Performance: FE-MCFormer achieved 99.11% accuracy at -2 dB SNR and 72.99% at -10 dB SNR.
Comparison: It consistently outperformed CNN-based and Transformer-based baselines. For instance, at -10 dB, FE-MCFormer (72.99%) significantly surpassed ResNet50 (58.85%) and Vanilla Transformer (43.43%).
Efficiency: The standard model required 338.95M FLOPs, and the lightweight version (FE-MCFormer-s) required 54.60M FLOPs, both lower than ResNet50 (1610.34M FLOPs).

Case 2: Real-World Centrifugal Compressor Dataset

Conditions: Five operating states (normal, blade rupture, block friction, oil whirl, rotor imbalance) with SNRs from -10 dB to -2 dB.
Performance: FE-MCFormer achieved 99.98% accuracy at -2 dB SNR and 93.48% at -10 dB SNR.
Comparison: It outperformed Li-convformer (a strong lightweight baseline) by margins ranging from 1.65% to 3.74% across all SNR levels.

Interpretability Analysis

Frequency Domain: Visualizations of the FALL reconstruction showed that the model successfully suppressed broadband noise while retaining fault-sensitive harmonics (e.g., BPFO/BPFI for bearings, rotational harmonics for compressors).
Time Domain: Attention heatmaps demonstrated that the model focused on physically meaningful fault impulse regions even under high noise (-8 dB), whereas models without FAL exhibited dispersed attention.

5. Significance and Claims

The paper claims that FE-MCFormer provides an effective framework for turbomachinery fault diagnosis in complex noisy environments. Its significance lies in:

Robustness: The ability to maintain high diagnostic accuracy under severe noise conditions (-10 dB) where other deep learning models fail or degrade significantly.
Interpretability: The architecture offers time–frequency interpretability from both the first layer (via FALL) and deep feature representations. This allows engineers to understand why a diagnosis was made by observing which frequency components were preserved and which temporal segments were attended to.
Industrial Applicability: The successful validation on a real-world centrifugal compressor dataset suggests the method is suitable for deployment in actual industrial condition monitoring systems, balancing high performance with computational efficiency.

The authors conclude that the proposed method addresses the limitations of current interpretable FD methods by explicitly handling strong noise and unifying time–frequency analysis, thereby improving the reliability of intelligent operation and maintenance systems.