GPU-Accelerated Multivariate Empirical Mode Decomposition for Massive Neural Data Processing

Mujahid, Taha; Rahman, Anis Ur; Khan, Muhammad Murtaza

doi:10.1109/access.2017.2705136

Cited by 10 publications

(7 citation statements)

References 28 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We first present an initial version. The sliding window 𝑤 𝑢𝑖 is applied sequentially from left to right to perform the CSI of 𝑢(𝑡) by (24) as shown in Fig. 1(b).…”

Section: Optimization Of the Input/working Memories And The Lmemd Alg...mentioning

confidence: 99%

“…Let us now explain why 𝑥(𝑡) and 𝑥 ̃(𝑡) can share the same memory in the AESW procedure. The 𝑢(𝑡) and 𝑙(𝑡) in (24) depend on the local extrema of 𝑥(𝑡) and the coordinates of all local extrema {𝛾 ℓ𝑗 , 𝛾 𝑢𝑖 } are already sorted by the AESW procedure. Then, (𝛾 𝑢 , 𝑥(𝛾 𝑢 )) or (𝛾 𝑙 , 𝑥(𝛾 𝑙 )) is computed sequentially from left to right.…”

Section: Optimization Of the Input/working Memories And The Lmemd Alg...mentioning

confidence: 99%

See 1 more Smart Citation

On the Memory Cost of EMD Algorithm

2022

View full text Add to dashboard Cite

Empirical mode decomposition (EMD) and its variants are adaptive algorithms that decompose a time series into a few oscillation components called intrinsic mode functions (IMFs). They are powerful signal processing tools and have been successfully applied in many applications. Previous research shows that EMD is an efficient algorithm with computational complexity 𝑂(𝑛) for a given number of IMFs, where 𝑛 is the signal length, but its memory is as large as (13 + 𝑚 𝑖𝑚𝑓 )𝑛, where 𝑚 𝑖𝑚𝑓 is the number of IMFs. This huge memory requirement hinders many applications of EMD. A physical or physiological oscillation (PO) mode often consists of a single IMF or the sum of several adjacent IMFs. Let 𝑚 𝑜𝑢𝑡 denote the number of PO modes and, by definition, 𝑚 𝑂𝑢𝑡 ≤ 𝑚 𝑖𝑚𝑓 . In this paper, we will propose a low memory cost implementation of EMD and prove that the memory can be optimized to (2 + 𝑚 𝑜𝑢𝑡 )𝑛 without aggravating the computational complexity, while gives the same results. Finally, we discuss the optimized memory requirements for different noise-assisted EMD algorithms.INDEX TERMS EMD, EEMD, CEEMD, memory cost.

show abstract

“…We first present an initial version. The sliding window 𝑤 𝑢𝑖 is applied sequentially from left to right to perform the CSI of 𝑢(𝑡) by (24) as shown in Fig. 1(b).…”

Section: Optimization Of the Input/working Memories And The Lmemd Alg...mentioning

confidence: 99%

Section: Optimization Of the Input/working Memories And The Lmemd Alg...mentioning

confidence: 99%

On the Memory Cost of EMD Algorithm

2022

View full text Add to dashboard Cite

show abstract

“…As compared to the graphical processing unit (GPU) based implementation of MEMD [24], the proposed FPGA-based hardware architecture offers the following advantages: i) delivery of high computational density per watt; ii) less power dissipation; iii) on-line and real time processing of multivariate data since the design in [24] can only perform batch processing. GPU based system has one major disadvantage, which is challenging for online real-time applications.…”

Section: Hardware Resources Utilization and Timing Analysismentioning

confidence: 99%

“…For general multivariate extension of EMD (MEMD), however, only software implementations; i.e., based on MATLAB and C programming language; or GPU based designs are available that limits its usage exclusively for offline applications. Specifically, the GPU based implementation for MEMD, using the compute unified device architecture (CUDA) architecture, was presented in [24] that exploits inherent parallelism within the MEMD algorithm. The parallel design improved the speed by up to seven times when compared to C programming based serial implementation of MEMD.…”

Section: Introductionmentioning

confidence: 99%

FPGA based design for online computation of Multivariate EMD (MEMD)

Gul¹,

Siddiqui²,

Rehman³

2020

Preprint

View full text Add to dashboard Cite

Multivariate or multichannel data have become ubiquitous in many modern scientific and engineering applications, e.g., biomedical engineering, owing to recent advances in sensor and computing technology. Processing these data sets is challenging owing to: i) their large size and multidimensional nature, thus requiring specialized algorithms and efficient hardware designs for on-line and real-time processing; ii) the nonstationary nature of data arising in many real life applications demanding new extensions of standard multiscale non-stationary signal processing tools. In this paper, we address the former issue by proposing a fully FPGA based hardware architecture of a popular multi-scale and multivariate signal processing algorithm, termed as multivariate empirical mode decomposition (MEMD). MEMD is a data-driven method that extends the functionality of standard empirical mode decomposition (EMD) algorithm to multichannel or multivariate data sets. Since its inception in 2010, the algorithm has found wide spread applications spanning different engineering related fields. Yet, no parallel FPGA based hardware design of the algorithm is available for its on-line and real-time processing. Our proposed architecture for MEMD uses fixed point operations and employs cubic spline interpolation within the sifting process. Finally, examples of decomposition of multivariate synthetic and real world biological signals are provided.

show abstract

“…Commonly, the typical methods can be summarized as time-frequency distributions [5], the empirical mode decomposition [6] and its multivariate extensions [7], the local mean decomposition [8], the reassignment method [9,10] such as the synchrosqueezing transform (SST) [9], the sparsification methods [11,12], and so on.…”

Section: Related Workmentioning

confidence: 99%

Intrinsic Mode Chirp Multicomponent Decomposition with Kernel Sparse Learning for Overlapped Nonstationary Signals Involving Big Data

Sun

Miao

2018

Complexity

View full text Add to dashboard Cite

We focus on the decomposition problem for nonstationary multicomponent signals involving Big Data. We propose the kernel sparse learning (KSL), developed for the T-F reassignment algorithm by the path penalty function, to decompose the instantaneous frequencies (IFs) ridges of the overlapped multicomponent from a time-frequency representation (TFR). The main objective of KSL is to minimize the error of the prediction process while minimizing the amount of training samples used and thus to cut the costs interrelated with the training sample collection. The IFs first extraction is decided using the framework of the intrinsic mode polynomial chirp transform (IMPCT), which obtains a brief local orthogonal TFR of signals. Then, the IFs curves of the multicomponent signal can be easily reconstructed by the T-F reassignment. After the IFs are extracted, component decomposition is performed through KSL. Finally, the performance of the method is compared when applied to several simulated micro-Doppler signals, which shows its effectiveness in various applications.

show abstract

GPU-Accelerated Multivariate Empirical Mode Decomposition for Massive Neural Data Processing

Cited by 10 publications

References 28 publications

On the Memory Cost of EMD Algorithm

On the Memory Cost of EMD Algorithm

FPGA based design for online computation of Multivariate EMD (MEMD)

Intrinsic Mode Chirp Multicomponent Decomposition with Kernel Sparse Learning for Overlapped Nonstationary Signals Involving Big Data

Contact Info

Product

Resources

About