Toward Multimodal Industrial Fault Analysis: A Single-Speed Chain Conveyor Dataset with Audio and Vibration Signals

This paper introduces a comprehensive multimodal dataset comprising audio and vibration signals from a single-speed chain conveyor system, designed to benchmark robust industrial fault detection and classification under diverse operating conditions and noise levels through standardized evaluation protocols and baseline models.

The Gist

Imagine a busy factory floor where a giant, noisy conveyor belt is constantly moving boxes from one end to the other. This belt is the "heart" of the production line. If it starts to squeak, rattle, or jam, the whole factory could stop, costing money and time.

For a long time, engineers trying to predict these problems have been like doctors who only listen to a patient's heart with a single, old-fashioned stethoscope. They might listen to the vibration (how much the machine shakes) or the sound (what it noises like), but rarely both at the same time, and usually in a quiet, perfect laboratory rather than a messy, loud factory.

This paper introduces a new, super-detailed "health check" kit for these conveyor belts. Here is the story of what they built and why it matters, explained simply:

1. The "Super-Sensor" Setup

The researchers built a special conveyor belt system in a lab, but they didn't just set it up quietly. They wanted to make it feel like a real, chaotic factory.

• The Ears: Instead of one microphone, they used three different "ears": a professional recording device, an iPhone, and a Xiaomi phone. Why? Because different microphones hear things differently, just like how a dog hears a whistle that humans can't. This gives them a richer picture of the sound.
• The Feelers: They attached vibration sensors (like tiny accelerometers) to the motor and the other end of the belt. These feel the "shakes" and "tremors" of the machine.
• The Noise Machine: Real factories are loud. To mimic this, they played back recordings of actual factory noise through speakers while collecting data. This is like testing a car's engine while driving through a construction zone, not just in a quiet garage.

2. The "Sickness" List

They didn't just record the machine running perfectly. They intentionally broke it in four specific ways to see if their sensors could catch the trouble:

• Lean: The guide rails are crooked (like a train track that's slightly bent).
• Dry: The chain has no oil (like a rusty bike chain squeaking).
• Loose: The chains are too slack (like a loose belt on a fan).
• Screwdrop: A screw fell into the gears (like a pebble in a shoe).

They tested these "sick" machines at different speeds and with different weights (loads) on the belt.

3. The "Detective" Game

The researchers created a massive dataset (over 6,600 samples) and challenged computer programs to act as detectives. They set up two main games:

• Game A: The "Is Something Wrong?" Test (Fault Detection)

  • The Rule: The computer is only shown examples of the machine running perfectly (healthy). It never sees the broken ones during training.
  • The Challenge: When the computer hears a new sound, it has to say, "This sounds weird, something is broken!" without knowing what is broken.
  • The Result: Surprisingly, the sound (audio) was often better at spotting that something was wrong than the vibration sensors. It's like how you can hear a car engine misfire before you feel the car shake.

• Game B: The "What's Wrong?" Test (Fault Classification)

  • The Rule: Now the computer sees examples of all the broken types.
  • The Challenge: It has to guess exactly which problem it is (Is it dry? Is it loose?).
  • The Result: This is where the magic happened. When the computer combined both the sound and the vibration, it became a super-detective.
    • Analogy: If you drop a glass, the sound tells you it shattered (a sharp crash), but the vibration tells you how heavy the object was. Together, you know exactly what happened. The vibration was great at spotting the "screw in the gears" (a hard thud), while the sound was great at spotting the "loose chain" (a rattling noise).

4. Why This Matters

Before this paper, most "training data" for AI was like studying for a driving test in a quiet, empty parking lot. This new dataset is like a driving school that includes rain, traffic, and construction zones.

• Realism: It forces AI to learn how to ignore the background noise of a real factory.
• Teamwork: It proves that Audio + Vibration > Audio alone or Vibration alone. They are a dynamic duo.
• Open Source: The researchers made all the data and the "rules of the game" free for everyone. This means other scientists can use it to build better, smarter systems that stop factories from breaking down unexpectedly.

The Bottom Line

Think of this paper as the creation of the ultimate "Conveyor Belt Simulator." It gives AI a chance to practice diagnosing machine sickness in a noisy, realistic environment using both its ears and its hands. The lesson? To truly understand a machine, you need to listen to its voice and feel its heartbeat.

Technical Summary

Here is a detailed technical summary of the paper "Toward Multimodal Industrial Fault Analysis: A Single-Speed Chain Conveyor Dataset with Audio and Vibration Signals."

1. Problem Statement

Industrial condition monitoring is critical for preventing unexpected failures in automated production lines. However, existing data-driven fault analysis research faces three primary limitations:

• Component-Centric Focus: Most public datasets (e.g., CWRU, IIEE) focus on isolated components like bearings or motors rather than complete system-level operations.
• Single Modality: Many datasets rely solely on vibration or audio, failing to capture complementary fault information available across multiple sensing modalities.
• Unrealistic Noise Conditions: Existing benchmarks often use clean laboratory data or synthetic noise augmentation, which does not reflect the strong, structured environmental noise present in real factory settings.

The paper addresses these gaps by introducing a new dataset and evaluation framework designed for system-level, multimodal fault analysis under realistic industrial noise conditions.

2. Methodology and Dataset Construction

The SSCC Dataset

The authors constructed a dataset based on a Single-Speed Chain Conveyor (SSCC) system, a common mechanism in industrial transmission.

• System Setup: The experimental platform emulates a linear, unidirectional production line (preventing workpiece return) with variable speeds and loads.
• Sensing Modalities:
  • Audio: Captured via three heterogeneous devices (Zoom H5 Recorder, iPhone 11, Xiaomi Mi 9 SE) to introduce natural channel diversity.
  • Vibration: Captured via a single-axis sensor (motor) and a tri-axis sensor (system end) to measure localized and system-level responses.
  • Total Channels: 7 synchronized channels (3 audio + 4 vibration).
• Operating Conditions:
  • Speeds: 5 levels (20, 40, 60, 80, 100 nominal units).
  • Loads: Heavy, Medium, Light.
  • Noise: Data collected under both clean and noisy conditions. Real factory noise was recorded on-site and reproduced via loudspeakers during data acquisition to ensure realistic acoustic interference.
• Fault Types: The dataset covers normal operation and four representative fault types:
  1. Lean: Guide rail misalignment.
  2. Dry: Insufficient chain lubrication.
  3. Loose: Excessive chain looseness.
  4. Screwdrop: Foreign object intrusion causing obstruction.
• Scale: The dataset contains 6,669 samples (5-second clips), annotated with sample-level fault labels.

Evaluation Protocols

The authors established standardized protocols to ensure fair comparison:

• Unsupervised Fault Detection: Trained only on normal data (normal-only training). The test set includes normal data at a held-out speed (velocity 100) and all fault types across all regimes.
• Supervised Fault Classification: A multi-class classification task with strict data partitioning. Samples at velocity 80 and specific fault categories at velocity 100 are held out from training to test generalization.
• Baseline Method: A unified channel-wise k-Nearest Neighbor (kNN) framework is used. Features are extracted using pre-trained foundation models (BEATs, CED, DaSheng, EAT, ECHO, FISHER) without task-specific fine-tuning. Anomaly scores are derived from Euclidean distances to the nearest neighbors in a memory bank of normal samples.

3. Key Contributions

1. New Multimodal Benchmark: The first dataset specifically targeting system-level fault analysis for chain conveyors, featuring synchronized multi-channel audio and vibration.
2. Realistic Noise Modeling: Unlike synthetic augmentation, the dataset incorporates reproduced real-world factory noise, enabling robust evaluation of models under practical industrial acoustic conditions.
3. Unified Evaluation Framework: Introduction of standardized protocols for both unsupervised detection (normal-only training) and supervised classification, utilizing a task-agnostic kNN baseline to isolate representation quality from model bias.
4. Open Availability: The dataset, code, and demo are publicly released to facilitate reproducible research.

4. Experimental Results and Analysis

The authors evaluated six pre-trained audio foundation models across three views: Audio-only, Vibration-only, and Fused (Audio + Vibration).

• Fault Detection (Unsupervised):

  • Audio Superiority: Audio-based methods generally outperformed vibration-based methods across most encoders. This suggests that subtle sound variations are more discriminative for detecting anomalies in this system than vibration signals, which are often dominated by steady-state periodic components.
  • Fusion Benefits: Decision-level fusion (averaging anomaly scores) generally improved performance over vibration-only views and often matched or exceeded audio-only results.
  • Top Performance: The FISHER encoder achieved the highest AUROC (0.954) on audio, while the fused view reached 0.882.

• Fault Classification (Supervised):

  • High Accuracy: All encoders achieved high accuracy (>90% for most fused models).
  • Complementary Strengths:
    • Vibration excelled at detecting "screwdrop" faults (impulsive mechanical responses).
    • Audio excelled at detecting "loose" faults (frictional/rattling sounds).
  • Fusion Impact: The fused multimodal system achieved the best overall performance (e.g., DaSheng Fused: 96.7% Accuracy, 96.5% Macro-F1), demonstrating that integrating modalities captures complementary physical cues.

5. Significance and Conclusion

This work provides a critical step forward in industrial AI by shifting the focus from isolated component analysis to system-level, multimodal monitoring under realistic noise.

• Practical Impact: The dataset bridges the gap between laboratory benchmarks and real-world deployment, offering a robust testbed for noise-robust algorithms.
• Research Potential: The results indicate that current encoders have not yet fully exploited the rich information in the dataset, leaving significant room for improvement in advanced representation learning and fusion strategies.
• Future Direction: The dataset serves as a foundation for developing more resilient industrial fault diagnosis systems capable of operating in noisy, complex production environments.