PaperlayerView exact revision on arXiv

Source-linked AI summary

Lightweight Time-Frequency Vibration Sensing for Consumer Device Condition Monitoring

Po-Heng Chou, Jen-Yu Chiu, Chun-Yu Yeh, Wei-Lung Mao, Walid Saad

arXiv:2610.13040v1eess.SPcs.LG

TL;DR

Consumer-device vibration monitoring needs lightweight diagnosis of non-stationary signals, while correlated data splits can make benchmark evaluations overly optimistic. This paper evaluates a transform-and-classify framework with structured isolation, finding a favorable accuracy-latency tradeoff for CWT and distinct accuracy and efficiency strengths across models.

Abstract

from arXiv · show

Consumer rotating devices, including washing machines, air conditioners, electric fans, robotic vacuum cleaners, and home fitness equipment, increasingly rely on vibration sensing for intelligent condition monitoring and predictive maintenance. Ensuring reliable operation in these resource-constrained appliances requires efficient, lightweight artificial intelligence (AI) solutions capable of processing non-stationary signals under real-world operational interference. This paper proposes a lightweight AI-assisted vibration sensing framework that employs continuous wavelet transform (CWT) and Hilbert-Huang transform (HHT) to convert one-dimensional (1-D) vibration time-series signals into two-dimensional (2-D) time-frequency representations. These representations are subsequently analyzed using lightweight deep learning architectures, specifically MobileNetV3-Small, ConvNeXt-Tiny, and the You Only Look Once version 11 classification model (YOLOv11-cls). Public bearing benchmark datasets, including Case Western Reserve University (CWRU), Paderborn University (PU), and Intelligent Maintenance Systems (IMS), are utilized as representative rotating components to investigate model generalization. The proposed framework is designed with eventual embedded deployment in consumer electronics in mind. Experimental results demonstrate that CWT combined with YOLOv11-cls achieves high diagnostic accuracy, while MobileNetV3-Small provides an attractive low-latency alternative for computationally constrained edge platforms. Conventional random splitting is also compared with dataset-specific structured isolation. The results highlight the importance of avoiding data leakage and quantify tradeoffs in classification accuracy, parameter count, floating-point operations (FLOPs), model footprint, preprocessing cost, and central processing unit (CPU) inference latency.

I. INTRODUCTION

Consumer-device vibration monitoring must handle non-stationary signals while meeting strict edge hardware constraints. The paper proposes a lightweight framework and evaluates structured isolation, time-frequency transforms, and model tradeoffs for that setting.

I. INTRODUCTION

  • Vibration signals are non-stationary, transient, and noise-affected, while embedded monitoring must balance diagnostic accuracy with computational complexity, memory footprint, and latency.Source: §I · p. 1
  • CWT and HHT convert 1-D vibration signals into 2-D time-frequency images for CNN feature extraction, addressing the temporal-localization limits of conventional frequency analysis,,,.Source: §I · p. 1
  • Unlike YOLO-based localization of fault-related time-frequency regions, this study classifies spectrograms and emphasizes computational efficiency and generalization under dataset-specific structured isolation.Source: §I · pp. 1–2
  • The proposed lightweight AI-assisted framework targets consumer rotating devices, using CWRU, PU, and IMS bearing datasets as representative components for a methodology intended to transfer to appliances.Source: §I · p. 2
  • Dataset-specific structured isolation strategies are proposed to reduce leakage and provide more realistic generalization evaluation across benchmark data structures.Source: §I · p. 2
  • The study compares CWT and HHT preprocessing costs and transient preservation, and evaluates MobileNetV3-Small, YOLOv11-cls [9], and ConvNeXt-Tiny across accuracy and computational costs.Source: §I · p. 2

II. RELATED WORK

Prior time-frequency methods address limits of stationary frequency analysis, but their properties and computational costs differ.

II. RELATED WORK

B. Lightweight Deep Learning for Condition Monitoring

The reviewed work motivates lightweight models and realistic evaluation, then outlines a consumer-device sensing pipeline that transforms vibration signals before classification.

B. Lightweight Deep Learning for Condition Monitoring

  • MobileNetV3 uses efficient depthwise convolutions and squeeze-and-excitation, ConvNeXt modernizes convolutional networks, and YOLOv11-cls offers a classification configuration; correlated random splits can inflate evaluation estimates.Source: B. Lightweight Deep Learning for… · p. 2
  • Randomly splitting correlated vibration segments can place samples from the same physical record in both training and testing, producing overly optimistic estimates,.Source: B. Lightweight Deep Learning for… · p. 2

III. PROPOSED FRAMEWORK AND METHODOLOGY

  • The proposed consumer-electronics workflow captures accelerometer vibration, converts it to a time-frequency representation, classifies it, and supplies the predicted condition to monitoring.Source: §III · p. 2
  • The classifier distinguishes Normal, Ball Fault, Inner-Race Fault, and Outer-Race Fault conditions.Source: §III · p. 2

A. Time-Frequency Representation Techniques

  • CWT and HHT are investigated to represent 1-D vibration signals as 2-D time-frequency images that capture non-stationary shock impulses and frequency-dependent characteristics.Source: A. Time-Frequency Representation… · p. 2

1) Continuous Wavelet Transform (CWT):

The section presents CWT and HHT as time-frequency approaches for vibration signals. CWT localizes analysis in time and scale, while HHT uses adaptive decomposition and Hilbert spectral analysis, with EMD processing carrying noise and mode-mixing sensitivities.

1) Continuous Wavelet Transform (CWT):

  • CWT provides localized analysis in time and scale, using the Morlet wavelet as its mother wavelet.The transform is defined in terms of scale, translation, and the mother wavelet.Source: 1) Continuous Wavelet Transform (CWT) · pp. 2–3
  • The proposed framework links vibration acquisition and preprocessing to time-frequency representation, lightweight AI classification, and condition monitoring.The diagram names these stages without specifying their implementation.Source: Figure 1 · p. 3
  • CWT preserves the temporal evolution of vibration energy across multiple scales, making its representation suitable for CNN-based feature extraction.The figure caption identifies a CWT spectrogram for a CWRU fault.Source: 1) Continuous Wavelet Transform (CWT) · p. 3, Figure 2 · p. 3

2) Hilbert-Huang Transform (HHT):

  • HHT combines empirical mode decomposition with Hilbert spectral analysis; EMD decomposes a vibration signal into intrinsic mode functions and a residual component.The section also identifies the Hilbert transform of each IMF as part of the HHT procedure.Source: 2) Hilbert-Huang Transform (HHT) · p. 3
  • HHT provides adaptive decomposition and instantaneous-frequency information that can benefit transient vibration events, but EMD processing is iterative and sensitive to noise and mode mixing.The figure caption identifies an HHT spectrogram for a CWRU fault.Source: 2) Hilbert-Huang Transform (HHT) · p. 3, Figure 3 · p. 3

B. Lightweight Deep Learning Architectures

The study evaluates three convolutional classifier families for potential embedded consumer-electronic constraints, ranging from low-complexity MobileNetV3-Small to the relatively higher-capacity YOLOv11-cls.

B. Lightweight Deep Learning Architectures

  • MobileNetV3-Small uses depthwise separable convolutions and lightweight squeeze-and-excitation mechanisms for low computational complexity while maintaining effective feature extraction.It is the lightweight architecture in the comparison.Source: B. Lightweight Deep Learning… · p. 3
  • ConvNeXt-Tiny combines large-kernel depthwise convolutions and other architectural improvements with a convolution-based structure for visual feature extraction.Source: B. Lightweight Deep Learning… · p. 3
  • YOLOv11-cls is the YOLO11 family’s classification configuration, incorporating C3k2 blocks and C2 position-sensitive attention-based feature processing.It is evaluated as relatively higher-capacity than MobileNetV3-Small.Source: B. Lightweight Deep Learning… · p. 3

IV. EXPERIMENTAL SETUP AND EVALUATION

The evaluation uses three public bearing benchmarks as representative rotating components and applies consistent signal preprocessing and training settings. The implementation also records the computing environment used for training and CPU latency measurement.

A. Datasets and Preprocessing

  • CWRU, PU, and IMS provide representative bearing fault conditions for evaluating the framework.CWRU covers motor loads and fault types, PU covers operating conditions and damage categories, and IMS provides run-to-failure measurements.Source: A. Datasets and Preprocessing · pp. 3–4

B. Implementation Details and Hyperparameters

  • Signals were segmented into 1024-point windows with 50% overlap and Z-score normalized before time-frequency conversion.CWT used Morlet scales 1–128; HHT used the first three empirically selected IMFs, Gaussian smoothing, and 224 × 224 image resizing.Source: B. Implementation Details and… · p. 4
  • Training used AdamW for 50 epochs with batch size 32 and initial learning rate 1 × 10−3, with ImageNet-pretrained weights and 20 Optuna search trials.Optuna targeted learning rate and weight decay.Source: B. Implementation Details and… · p. 4
  • Experiments ran in PyTorch 2.1 and Python 3.9 on an RTX 4090, while CPU latency was benchmarked on an Intel Core i7-13700K.Source: B. Implementation Details and… · p. 4

C. Evaluation Protocols and Dataset-Specific Isolation Strategies

The evaluation isolates data according to each benchmark’s structure before segmentation, addressing leakage risks from correlated overlapping windows. Under the tested configuration, CWT preprocessing is substantially faster than HHT.

C. Evaluation Protocols and Dataset-Specific Isolation Strategies

  • Structured isolation assigns CWRU by load condition, PU by physical bearing ID, and IMS by chronological blocks before segmentation.Segmentation is performed independently within each subset, preventing overlapping windows from the same isolation unit from crossing subsets.Source: C. Evaluation Protocols and… · p. 4
  • For IMS, each class is split into chronological blocks 0–84 for training, 85–101 for validation, and 102–118 for testing, yielding 340, 68, and 68 segments.The four classes contribute equally to each set; the CWRU categories map to Normal, Ball Fault, Inner-Race Fault, and Outer-Race Fault.Source: C. Evaluation Protocols and… · p. 4

A. Preprocessing Latency: CWT vs. HHT

  • 15.2 ± 1.8 ms per sample for CWT versus 88.5 ± 7.4 ms for HHT; HHT therefore incurs approximately six-fold preprocessing latency.Measurements were mean ± standard deviation over at least 100 runs on an Intel Core i7-13700K CPU, batch size one.Source: A. Preprocessing Latency: CWT vs. HHT · p. 4

B. Classification Performance Comparison: CWT vs. HHT

Under dataset-specific structured isolation, YOLOv11-cls performs best among the evaluated architectures, while random splitting yields higher accuracy. CWT also outperforms HHT across the three datasets and has a more favorable accuracy-latency tradeoff in the tested comparison.

B. Classification Performance Comparison: CWT vs. HHT

  • CWT consistently outperforms HHT across CWRU, PU, and IMS with MobileNetV3-Small held constant in the comparison.Table II reports accuracy, precision, recall, and F1-score for both representations.Source: B. Classification Performance… · p. 4, Table II · p. 5
  • HHT’s adaptive instantaneous-frequency analysis is susceptible to mode mixing in noisy environments and requires approximately six times CWT’s preprocessing latency.Under the tested configuration, the authors report CWT as the more favorable accuracy-latency tradeoff for resource-constrained consumer devices.Source: B. Classification Performance… · p. 5

C. Classification Performance and Generalization

  • 3.5 to 11.1 percentage points: structured isolation lowers accuracy relative to random splitting across all evaluated architectures and datasets.The paper attributes potentially optimistic random-split results to highly correlated vibration segments being distributed across training and testing subsets.Source: C. Classification Performance and… · p. 5
  • 96.5% accuracy for YOLOv11-cls on CWRU versus 92.1% for MobileNetV3-Small under dataset-specific structured isolation.MobileNetV3-Small’s lower computational complexity makes it attractive for highly constrained edge platforms; ConvNeXt-Tiny is intermediate on CWRU but has a substantially larger footprint.Source: C. Classification Performance and… · p. 5

D. Complexity and Accuracy-Efficiency Tradeoffs

Under structured isolation, YOLOv11-cls has the strongest accuracy, while MobileNetV3-Small offers lower computational cost. End-to-end timing and potential appliance applications frame these tradeoffs for consumer-device use.

D. Complexity and Accuracy-Efficiency Tradeoffs

  • Complexity comparisons used 224 × 224 inputs; CPU latency measured batch-1 network execution on an Intel Core i7-13700K, excluding preprocessing.The measured latency excludes data loading and CWT/HHT preprocessing.Source: D. Complexity and Accuracy-Efficiency… · p. 5
  • Approximately 23.5 ms is required per sample with MobileNetV3-Small and CWT, combining 15.2 ms preprocessing with 8.3 ms inference.The paper favors MobileNetV3-Small for severely constrained platforms and YOLOv11-cls when higher diagnostic accuracy is prioritized.Source: D. Complexity and Accuracy-Efficiency… · p. 5

VI. POTENTIAL APPLICATIONS IN CONSUMER ELECTRONICS

  • The framework is proposed as transferable to consumer devices, with examples including washing-machine imbalance, fan bearing wear, and vacuum drive-wheel or gearbox problems.The application examples also include washing-machine bearing degradation and suspension wear, and vacuum mechanical interference from accumulated debris.Source: §VI · pp. 5–6
  • Table III compares CWT classification performance and generalization under random and dataset-specific structured isolation.The table caption identifies the comparison but does not provide its cell values.Source: Table III · p. 6

D. Complexity and Accuracy-Efficiency Tradeoffs

  • MobileNetV3-Small uses 2.54 M parameters and 0.06 GFLOPs; YOLOv11-cls uses 2.43 GFLOPs and achieves the highest structured-isolation accuracy across evaluated datasets.ConvNeXt-Tiny has 27.80 M parameters and a 109.1 MB footprint, the largest among the three models.Source: D. Complexity and Accuracy-Efficiency… · p. 5, Table IV · p. 6

VII. LIMITATIONS AND FUTURE WORK

The conclusion reports a favorable CWT accuracy-latency tradeoff and different accuracy-efficiency strengths for YOLOv11-cls and MobileNetV3-Small. It also bounds the evidence by noting the gap between standardized benchmarks and domestic appliances.

VII. LIMITATIONS AND FUTURE WORK

  • Standardized benchmark datasets do not fully represent domestic appliances’ acoustic, mechanical, and environmental complexity; device structure, sensor position, and load may affect cross-device generalization.The CPU latency results are a hardware-specific baseline and do not directly represent execution time on resource-constrained embedded microcontrollers.Source: §VII · p. 6
  • Future work will collect consumer-appliance vibration data and investigate cross-device generalization and hardware-aware optimization, including INT8 quantization, pruning, and knowledge distillation.Source: §VII · p. 6

VIII. CONCLUSION

  • CWT provides a more favorable accuracy-latency tradeoff than HHT; under structured isolation, YOLOv11-cls has the highest accuracy while MobileNetV3-Small is more efficient.The reported performance gap between random and structured splitting demonstrates the importance of preventing correlated segments from entering both training and testing sets.Source: §VIII · p. 6

Related papers

Loading 2610.13040v1…