Source-linked AI summary
Benchmarking Model Compression Techniques Across Heterogeneous Edge Platforms: A Predictive Maintenance Use Case for Rotating Machinery
Matheus Calheira Guimarães de Oliveira, Marco Aurélio da Silva Cruz, Regis Cardoso, Lucas Cambuim, Rafael Macieira, Arthur Paim, Helaine Neves, Alex Santos, Lilian Guarieiro, Jean Schmith, Prashant Kumar, Erick G. Sperandio Nascimento
TL;DR
Predictive-maintenance deep-learning models face limited practical evidence on compression across heterogeneous, resource-constrained edge hardware. This study compares compression approaches and evaluates models across platforms; knowledge distillation was the most effective among those tested, with the smallest model and shortest average inference time while keeping performance close to baseline.
Problem
Evidence is limited on how compression affects predictive-maintenance fault-diagnosis models across heterogeneous edge hardware, including inference time and accuracy.
Source: §1 · p. 2, §2.5 · p. 8Method
The study compares compression strategies for a predictive-maintenance CNN and evaluates model performance across diverse computing platforms.
Source: §1 · p. 2, §2.5 · p. 8Results
Knowledge distillation was the most effective technique evaluated, with the smallest model and shortest average inference time while maintaining performance metrics close to the baseline.
Source: §4 · p. 11, §5 · p. 12Takeaways & Limitations
Deployment suitability should be assessed across execution performance, predictive accuracy, and reliability on diverse platforms.
Source: §2.5 · p. 8
Abstract
from arXiv · showhide
To enable Industry 5.0, where autonomous systems are widely adopted and integrated with people to promote sustainable development in industries, it is imperative to adapt complex deep learning models to operate efficiently on resource-constrained hardware for edge inference. However, adapting these models remains challenging due to the hardware heterogeneity of edge computing in industrial settings, performance degradation caused by model compression, and the lack of clear guidance on which compression techniques are practical and effective for applications such as predictive maintenance. To address these challenges, this paper presents a comprehensive comparative analysis of four state-of-the-art model compression techniques applied to a high-performance convolutional neural network for predictive maintenance using a real-world dataset of rotating machinery, equipment typically used in industry. The compressed models were evaluated across a heterogeneous set of nine computing platforms, ranging from highly constrained edge computing devices to general purpose and high-performance processors, reflecting the diversity and fragmentation typical of industrial edge environments. By jointly prioritizing classification performance, inference speed, and compression ratio, we found that compression using knowledge distillation emerged as the most effective approach. The model had more than 90% speedup across all devices while maintaining a 91.22% balanced F-score, with a very small drop in predictive performance relative to the uncompressed model, along with a 98.9% size reduction. Through successful deployment across multiple edge platforms and a systematic comparison of compression strategies, we demonstrated a clear path to achieving Industry 5.0 targets, drastically cutting computational costs and, consequently, energy consumption without sacrificing predictive performance.
1 Introduction
Deep-learning predictive-maintenance models face practical deployment gaps on resource-constrained edge hardware, despite the potential benefits of local processing. This study compares compression approaches across computing platforms to examine efficiency, accuracy, and hardware trade-offs.
1 Introduction
- Local edge processing can reduce reliance on continuous connectivity and latency, while embedded AI may support equipment monitoring and reduce downtime and maintenance costs.Source: §1 · p. 2
- Prior studies reported strong fault-diagnosis performance, but practical deployment behavior on resource-constrained edge devices remained underexplored.The stated gaps include limited evidence about compression effects on heterogeneous hardware, inference time, and accuracy.Source: §1 · p. 2
- D-PdMCNN was reported more than 90% faster across all devices and 98.9% smaller than the fault detection model.Source: §1 · p. 2
- The study benchmarks across microcontrollers, edge devices, and general-purpose and high-performance processors.Source: §1 · p. 3
- The work compares easy-to-implement deep-learning compression techniques for machinery fault diagnosis and examines accuracy, efficiency, and hardware constraints.Source: §1 · p. 3
2 Methods
The five-phase methodology moves from dataset definition and modeling through compression, hardware selection, and evaluation criteria. It compares model adaptation across diverse platforms using predictive performance, inference time, and model size.
2.1 Data acquisition and preprocessing
- MaFaulDa supplied 1,951 multivariate time series; three underhang accelerometers were FFT-transformed into five target classes and 8,580 windows split 5,550/1,220/1,810 for training, validation, and testing.The model used axial, radial, and tangential accelerometer signals.Source: §2.1 · pp. 4–5
2.2 Modeling
- The PdM-CNN reproduces a 300K-parameter architecture and adapts its original 10 fault classes to five for the study.Its components include Conv1D, MaxPool1D, global average pooling, dropout, and dense layers.Source: §2.2 · p. 5, Figure 3 · p. 5
2.3 Compression Techniques
- The baseline was evaluated with QI8, QDRI8, QF16, and knowledge distillation, selected as documented, reproducible techniques with different deployment trade-offs.The stated trade-offs include model size, inference latency, hardware compatibility, and accuracy.Source: §2.3 · p. 5
2.3.1 Post-Training Quantization Methods
- The three post-training quantization variants were full-int8 weights and activations, dynamic-range int8 weights, and float16 weights.Dynamic-range quantization keeps activations in floating point and quantizes them dynamically during inference; float16 can be most effective on platforms with native support.Source: §2.3.1 · p. 6
2.3.2 Knowledge Distillation
- Knowledge distillation retrained a 1,748-parameter student from teacher outputs, using a temperature of 40 and a joint loss balancing labels with teacher guidance.The loss assigned 90% of the training signal to teacher guidance and 10% to ground-truth labels.Source: §2.3.2 · pp. 6–7
2.4 Target Hardware Platforms
- The nine-device benchmark covered general-purpose processors, microcontrollers, and edge devices; TensorFlow Lite models could not use CUDA GPU acceleration on Jetson platforms.In that inference mode, only the baseline model in .h5 format was run on Jetson Nano and Jetson Orin Nano.Source: §2.4 · p. 8, Table 1 · p. 8
2.5 Evaluation
- Evaluation measured precision, recall, F1-score, specificity, and accuracy, alongside average inference time per device and flatbuffer model size.Inference was run on each of the 1,810 test samples.Source: §2.5 · pp. 8–9, §2.4 · p. 8
3 Results
The results assess compression through model size, inference time, and classification performance across hardware. D-PdMCNN is reported to balance these objectives favorably for heterogeneous edge deployment.
3 Results
- Figure 5 summarizes model-footprint reductions, but size savings alone do not establish practical value without efficient execution.The results emphasize balancing storage reduction against computational overhead.Source: §3 · p. 9, Figure 5 · p. 9
- Inference time is reported as an average for each device and compression technique, to be considered alongside classification reliability and model footprint.A method that accelerates execution but substantially degrades classification performance may not be suitable for practical deployment.Source: §3 · pp. 9–10, Table 2 · p. 9
- D-PdMCNN achieves a favorable balance of predictive performance, execution time, and memory consumption across heterogeneous edge hardware.Table 3 reports classification performance across devices; the accompanying comparison describes the model's trade-off across performance and compression rate.Source: §3 · p. 10, Table 3 · p. 10, Table 4 · p. 10
4 Discussion
Compression effects on inference time vary by technique and hardware, whereas knowledge distillation consistently reduces inference time across devices. The discussion also reports substantial size reduction and preserved predictive performance for the distilled model.
4 Discussion
- Inference-time effects vary across platforms: QDRI8 is 69.49% slower on HPC but 64.22% faster on Jetson Orin Nano (CPU).QI8 ranges from 87.41% slower on Jetson Orin Nano (MCPU) to a 49.92% speedup on STM32MP157F-DK2.Source: §4 · p. 11
- Knowledge distillation is the only evaluated technique reported to reduce average inference time across all devices, including a 99.38% reduction on Jetson Orin Nano (CPU).The authors attribute the speed advantage to D-PdMCNN's smaller architecture rather than changed numerical precision.Source: §4 · p. 11
- The KD model is 13.520 kilobytes, approximately 98.9% smaller than the 1,222.696-kilobyte baseline.QDRI8 is 318.616 kilobytes, while QF16 is 616.348 kilobytes.Source: §4 · p. 11
- KD accuracy drops by 1.37%, while KD precision decreases by 4.32%; performance metrics are reported as consistent across devices.The same-architecture D-PdMCNN trained from scratch differs by more than 19% in F1-score and recall, and the authors identify knowledge distillation as important to preserving performance.Source: §4 · pp. 11–12
- The reasons quantization maintains model size and strong metrics remain unclear; the authors suggest convolutional feature extractors may be robust to reduced precision.For KD, they argue that the student learned from the baseline and had its parameters effectively optimized.Source: §4 · p. 12
5 Conclusion
The conclusion identifies knowledge distillation as the most effective evaluated compression technique, balancing model size, inference time, and performance. It also identifies generalization and direct energy measurement as areas for future work.
5 Conclusion
- Compression choices involve trade-offs across CNN performance and devices, so model-size reduction and inference time both matter, especially on less robust processors.The conclusion calls for testing more diverse datasets and architectures to assess generalizability.Source: §5 · p. 12
- KD is reported as the most effective evaluated technique, producing the smallest model and shortest average inference time while keeping metrics close to baseline.The conclusion also reports F1 and recall losses greater than 20% for the same-architecture D-PdMCNN trained from scratch.Source: §5 · p. 12
- Direct energy consumption was not measured; inference latency and model size serve as proxies, and future studies should assess energy directly.The conclusion also calls for investigating pruning or hybrid methods to explain how compression can preserve performance.Source: §5 · p. 12