Adaptive Wavelet Threshold Attention and Explainable Lightweight CNN for Strongly Time-Varying Fault Diagnosis and Feature Extraction Mechanism of Rotating Machinery

In the industrial world, the hum of a factory is often the sound of health, but it is also the language of failure. Inside massive gearboxes and spinning motors, tiny cracks and worn teeth begin to whisper long before they scream. For engineers, the challenge has always been to hear these whispers over the roar of the machine itself. When a machine runs at a steady speed, the sound of a fault is predictable, like a steady drumbeat. But in the real world, machines rarely stay still. They speed up, slow down, and change loads constantly. This constant shifting scrambles the sound of a fault, making it drift and fade into the background noise. Traditional tools for listening to these machines often struggle here; they are built for steady conditions and get confused when the rhythm changes, or they are so heavy and complex that they cannot be installed on the machines themselves to listen in real time.

Researchers Tengji Xia and Pengfei Pang have developed a new way to listen that is both sharp and simple. They created a digital system designed to act like a highly tuned ear that can adjust itself to the changing speed of a machine while filtering out the static of the factory floor. Their approach, described in their recent work, focuses on a specific type of artificial intelligence that is light enough to run on small, embedded computers but powerful enough to find hidden faults in bearings and gears. By combining a method that understands how sound waves stretch and compress with a smart filtering system that knows exactly when to ignore noise, their system can spot problems that other methods miss, even when the machine is running under difficult, changing conditions.

The core of their work addresses a problem that has long plagued fault diagnosis: the way faults change as a machine speeds up or slows down. When a bearing or gear is damaged, it creates a specific vibration pattern. However, if the machine's speed is not constant, the frequency of that vibration shifts, much like the pitch of a siren changing as an ambulance drives past. Older computer models often use fixed settings to look for these sounds, meaning they might miss a fault simply because the machine's speed has drifted slightly from what the model expected. To solve this, the researchers built a system that does not rely on fixed settings. Instead, it uses a mathematical tool called a wavelet, which is excellent at breaking down complex signals into their parts, but they made it "adaptive." This means the system can learn and adjust its own internal settings in real time to track the shifting pitch of a fault, ensuring it never loses the signal even as the machine's speed fluctuates wildly.

To make this system practical for the real world, the team had to overcome another hurdle: complexity. Many advanced diagnostic tools are like massive supercomputers; they are powerful but require too much energy and space to be placed directly on a factory machine. The researchers designed their solution to be "lightweight," meaning it uses very few computational resources. They achieved this by stripping the system down to a single layer of processing. Instead of stacking many layers of complex calculations on top of each other, which can sometimes blur the delicate details of a fault, their model uses a direct, streamlined path. This design allows the system to process data quickly and efficiently, making it suitable for installation on the edge devices that control industrial machinery, rather than requiring a distant server to do the work.

A critical part of their success lies in how the system handles noise. In a noisy factory, the signal of a small crack can be drowned out by the vibration of the floor or the hum of nearby motors. The researchers added a special module to their system that acts like a dynamic gatekeeper. This module constantly analyzes the incoming sound, calculating the average noise level and the variations in the signal. Based on this calculation, it sets a moving threshold. If a sound spike is just random noise, the gatekeeper ignores it. But if a spike matches the pattern of a fault, the system boosts it, making the weak signal stand out clearly against the background. This process happens automatically and continuously, allowing the system to enhance the true fault signals while suppressing the interference, a capability that proved vital in their testing.

The team tested their new system on two challenging datasets that mimic real industrial environments. The first involved a bearing platform where the speed was constantly changing, and the second involved a gearbox that was subjected to varying loads and speeds. In these tests, the system was pitted against several other advanced methods, including deep learning models that are much larger and more complex. The results were striking. Under conditions where the noise was so high that it would typically hide a fault, the new system maintained a high level of accuracy, correctly identifying the type of fault in 93.75 percent of cases on the gearbox data and even higher on the bearing data. In contrast, many of the other methods saw their performance drop significantly as the noise increased. The new system also proved to be incredibly fast, requiring far less time to analyze the data than its competitors, which is essential for catching problems before they cause a breakdown.

What makes this finding particularly robust is not just the high accuracy numbers, but the statistical proof that the improvement is real and not a fluke. The researchers ran rigorous statistical tests to compare their system against the others, and the results showed that the difference in performance was significant, especially in the noisiest conditions. They also looked inside the "black box" of their model to understand how it worked. They found that the adaptive wavelet layer successfully learned to track the shifting frequencies of the faults, and the noise-filtering module effectively cleaned up the signal, allowing the system to focus on the true problem. This internal clarity gives engineers confidence that the system is not just guessing, but is actually following the physical laws of how these machines vibrate.

The researchers also demonstrated that their system could be deployed on small, low-power devices, which is a major step forward for industrial applications. By keeping the model simple and efficient, they ensured that it does not require expensive hardware to run. This means that factories could potentially install these diagnostic tools directly on their machines, providing a continuous, real-time health check without the need for heavy infrastructure. The system's ability to handle strongly time-varying conditions—where speed and load are constantly in flux—fills a gap that many existing technologies leave open, offering a reliable way to monitor equipment in the most dynamic and difficult operating environments.

Ultimately, this work represents a shift in how we approach machine health. It moves away from the idea that we need massive, complex models to solve complex problems. Instead, it shows that a carefully designed, streamlined system that understands the physics of the signal can outperform heavier, more rigid models. By combining the ability to track shifting frequencies with a smart, adaptive noise filter, the researchers have created a tool that is both sensitive and resilient. For the industries that rely on the continuous operation of rotating machinery, this offers a new level of assurance: a way to hear the faintest warning signs of failure, even when the machine is running at its most chaotic, ensuring that maintenance can happen before a small issue becomes a catastrophic failure.