Eciton: Very Low-Power LSTM Neural Network Accelerator for Predictive Maintenance at the Edge

Chen, Jeffrey; Hong, Sehwan; He, Warrick; Moon, Jinyeong; Jun, Sang-Woo

doi:10.1109/fpl53798.2021.00009

Cited by 16 publications

(8 citation statements)

References 42 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, the proposed model has far fewer parameters, its storage size is much lower than the LSTM’s one and it is also faster in the training stage. A trade-off between efficiency and effectiveness was thus achieved, which is of paramount importance in industrial contexts where the relationship between performance obtained and resources allocated is to be optimized (Chen et al ., 2021; Markiewicz et al ., 2019). In addition, the overall accuracy performances are comparable with the best techniques of the literature.…”

Section: Discussionmentioning

confidence: 99%

See 1 more Smart Citation

A deep attention based approach for predictive maintenance applications in IoT scenarios

Luca

Ferraro²,

Galli³

et al. 2023

JMTM

View full text Add to dashboard Cite

PurposeThe recent innovations of Industry 4.0 have made it possible to easily collect data related to a production environment. In this context, information about industrial equipment – gathered by proper sensors – can be profitably used for supporting predictive maintenance (PdM) through the application of data-driven analytics based on artificial intelligence (AI) techniques. Although deep learning (DL) approaches have proven to be a quite effective solutions to the problem, one of the open research challenges remains – the design of PdM methods that are computationally efficient, and most importantly, applicable in real-world internet of things (IoT) scenarios, where they are required to be executable directly on the limited devices’ hardware.Design/methodology/approachIn this paper, the authors propose a DL approach for PdM task, which is based on a particular and very efficient architecture. The major novelty behind the proposed framework is to leverage a multi-head attention (MHA) mechanism to obtain both high results in terms of remaining useful life (RUL) estimation and low memory model storage requirements, providing the basis for a possible implementation directly on the equipment hardware.FindingsThe achieved experimental results on the NASA dataset show how the authors’ approach outperforms in terms of effectiveness and efficiency the majority of the most diffused state-of-the-art techniques.Research limitations/implicationsA comparison of the spatial and temporal complexity with a typical long-short term memory (LSTM) model and the state-of-the-art approaches was also done on the NASA dataset. Despite the authors’ approach achieving similar effectiveness results with respect to other approaches, it has a significantly smaller number of parameters, a smaller storage volume and lower training time.Practical implicationsThe proposed approach aims to find a compromise between effectiveness and efficiency, which is crucial in the industrial domain in which it is important to maximize the link between performance attained and resources allocated. The overall accuracy performances are also on par with the finest methods described in the literature.Originality/valueThe proposed approach allows satisfying the requirements of modern embedded AI applications (reliability, low power consumption, etc.), finding a compromise between efficiency and effectiveness.

show abstract

Section: Discussionmentioning

confidence: 99%

“…(2022) introduced another LSTM-based attention mechanism to improve the ability of the model to analyze a sequence of signals in survival analysis. In turn, Chen et al . (2021) proposed a PdM system based on LSTM network adapted on FPGA, whose aim is to jointly reduce power consumption and management cost.…”

Section: Related Workmentioning

confidence: 99%

A deep attention based approach for predictive maintenance applications in IoT scenarios

Luca

Ferraro²,

Galli³

et al. 2023

JMTM

View full text Add to dashboard Cite

show abstract

“…Noticing this problem, Chen et al [4] implemented a similar accelerator on a much smaller FPGA iCE40 UP5K in 2021. Thanks to the ultra-low static power (at µW scale) of this FPGA, the overall power consumption during inference is approximately equal to the dynamic power of the FPGA, which is 17 mW.…”

Section: Related Workmentioning

confidence: 99%

“…In both works [4,6], researchers applied fixed-point logic to simplify the design and reduce the loss of precision compared to aggressive quantisation. The activation functions tanh() and sigmoid() were replaced with hard tanh() and hard sigmoid() respectively, which simplifies the computations but leads to a large reduction of precision [10].…”

Section: Related Workmentioning

confidence: 99%

Enhancing Energy-Efficiency by Solving the Throughput Bottleneck of LSTM Cells for Embedded FPGAs

Qian

Ling

Schiele

2023

Communications in Computer and Information Science

View full text Add to dashboard Cite

To process sensor data in the Internet of Things(IoTs), embedded deep learning for 1-dimensional data is an important technique. In the past, CNNs were frequently used because they are simple to optimise for special embedded hardware such as FPGAs. This work proposes a novel LSTM cell optimisation aimed at energy-efficient inference on end devices. Using the traffic speed prediction as a case study, a vanilla LSTM model with the optimised LSTM cell achieves 17534 inferences per second while consuming only 3.8 µJ per inference on the FPGA XC7S15 from Spartan-7 family. It achieves at least 5.4× faster throughput and 1.37× more energy efficient than existing approaches.

show abstract

“…Due to the computation-intensive nature of neural networks, the use of power-efficient hardware accelerators is one of the most prominent methods of bringing the benefits of deep neural networks to resource-constrained embedded systems. Embedded neural networks can reduce decision-making latency by removing the requirement for querying a remote server [22,23], as well as reducing power-hungry wireless network transmission requirements [24]. Some acceleration approaches include using mobile GPUs [1], customdesigned application-specific integrated circuits (ASICs) [25][26][27][28], as well as FPGAs [2].…”

Section: Embedded Neural Network Accelerationmentioning

confidence: 99%

MobileNets Can Be Lossily Compressed: Neural Network Compression for Embedded Accelerators

Lim

Jun

2022

Electronics

Self Cite

View full text Add to dashboard Cite

Although neural network quantization is an imperative technology for the computation and memory efficiency of embedded neural network accelerators, simple post-training quantization incurs unacceptable levels of accuracy degradation on some important models targeting embedded systems, such as MobileNets. While explicit quantization-aware training or re-training after quantization can often reclaim lost accuracy, this is not always possible or convenient. We present an alternative approach to compressing such difficult neural networks, using a novel variant of the ZFP lossy floating-point compression algorithm to compress both model weights and inter-layer activations and demonstrate that it can be efficiently implemented on an embedded FPGA platform. Our ZFP variant, which we call ZFPe, is designed for efficient implementation on embedded accelerators, such as FPGAs, requiring a fraction of chip resources per bandwidth compared to state-of-the-art lossy compression accelerators. ZFPe-compressing the MobileNet V2 model with an 8-bit budget per weight and activation results in significantly higher accuracy compared to 8-bit integer post-training quantization and shows no loss of accuracy, compared to an uncompressed model when given a 12-bit budget per floating-point value. To demonstrate the benefits of our approach, we implement an embedded neural network accelerator on a realistic embedded acceleration platform equipped with the low-power Lattice ECP5-85F FPGA and a 32 MB SDRAM chip. Each ZFPe module consumes less than 6% of LUTs while compressing or decompressing one value per cycle, requiring a fraction of the resources compared to state-of-the-art compression accelerators while completely removing the memory bottleneck of our accelerator.

show abstract

Eciton: Very Low-Power LSTM Neural Network Accelerator for Predictive Maintenance at the Edge

Cited by 16 publications

References 42 publications

A deep attention based approach for predictive maintenance applications in IoT scenarios

A deep attention based approach for predictive maintenance applications in IoT scenarios

Enhancing Energy-Efficiency by Solving the Throughput Bottleneck of LSTM Cells for Embedded FPGAs

MobileNets Can Be Lossily Compressed: Neural Network Compression for Embedded Accelerators

Contact Info

Product

Resources

About