Received 3 August 2022, accepted 16 August 2022, date of publication 1 September 2022, date of current version 15 September 2022.

Digital Object Identifier 10.1109/ACCESS.2022.3203406

A Deep Learning Model for Remaining Useful Life

Prediction of Aircraft Turbofan Engine on
C-MAPSS Dataset
OWAIS ASIF1 , SAJJAD ALI HAIDER1 , SYED RAMEEZ NAQVI 1, JOHN F. W. ZAKI 2,

KYUNG-SUP KWAK 3 , (Life Senior Member, IEEE),

AND S. M. RIAZUL ISLAM 4 , (Member, IEEE)
1 Department of Electrical and Computer Engineering, COMSATS University Islamabad, Wah Campus, Wah Cantt 47040, Pakistan
2 Department of Computer and Systems, Faculty of Engineering, Mansoura University, Mansoura 35516, Egypt
3 Department of Information and Communication Engineering, Inha University, Incheon 22212, South Korea
4 Department of Computer Science, University of Huddersfield, Huddersfield HD1 3DH, U.K.

Corresponding authors: Kyung-Sup Kwak (kskwak@inha.ac.kr) and S. M. Riazul Islam (s.mr.islam@hud.ac.uk)

This work was supported by the National Research Foundation of Korea-Grant funded by the Korean Government (Ministry of Science
and ICT) under Grant NRF-2020R1A2B5B02002478.

1 ABSTRACT In the era of industry 4.0, safety, efficiency and reliability of industrial machinery is an
2 elementary concern in trade sectors. The accurate remaining useful life (RUL) prediction of an equipment
3 in due time allows us to effectively plan the maintenance operation and mitigate the downtime to raise the
4 revenue of business. In the past decade, data driven based RUL prognostic methods had gained a lot of
5 interest among the researchers. There exist various deep learning-based techniques which have been used
6 for accurate RUL estimation. One of the widely used technique in this regard is the long short-term memory
7 (LSTM) networks. To further improve the prediction accuracy of LSTM networks, this paper proposes a
8 model in which effective pre-processing steps are combined with LSTM network. C-MAPSS turbofan engine
9 degradation dataset released by NASA is used to validate the performance of the proposed model. One
10 important factor in RUL predictions is to determine the starting point of the engine degradation. This work
11 proposes an improved piecewise linear degradation model to determine the starting point of deterioration
12 and assign the RUL target labels. The sensors data is pre-processed using the correlation analysis to choose
13 only those sensors measurement which have a monotonous behavior with RUL, which is then filtered
14 through a moving median filter. The updated RUL labels from the degradation model together with the
15 pre-processed data are used to train a deep LSTM network. The deep neural network when combined with
16 dimensionality reduction and piece-wise linear RUL function algorithms achieves improved performance
17 on aircraft turbofan engine sensor dataset. We have tested our proposed model on all four sub-datasets in
18 C-MAPSS and the results are then compared with the existing methods which utilizes the same dataset in
19 their experimental work. It is concluded that our model yields improvement in RUL prediction and attains
20 minimum root mean squared error and score function values.

21 INDEX TERMS Deep learning, long short-term memory networks, remaining useful life, turbofan engine.

22 I. INTRODUCTION process industry, power systems to the feild of transporta- 26

23 We are living in an era of industrial automation where our tion [1]. Every system requires maintenance operation at 27

24 day-to-day activity depends heavily on a wide range of elec- some point in time [2]. There are three types of maintenance 28

25 trical and mechanical equipments varying from agriculture, techniques used in industries, reactive maintenance, preven- 29

tive maintenance and predictive maintenance. In reactive 30

The associate editor coordinating the review of this manuscript and maintenance repair operation is consider only when machine 31

approving it for publication was Fu-Kwun Wang . failure has occurred. In preventive maintenance the failure 32

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 10, 2022 95425
 O. Asif et al.: Deep Learning Model for RUL Prediction of Aircraft Turbofan Engine on C-MAPSS Dataset

33 can be prevented by performing the time-based maintenance challenging to gather run time-to-failure sensor data espe- 89

34 operation, while the predictive maintenance (PdM) lets you cially for new machines. One way is that we can intentionally 90

35 estimate the time-to-failure of an equipment for scheduling run a new system upto the failure mode but it is very pro- 91

36 the maintenance tasks [3]. PdM in contrast to the other two longed, highly undesirable and expensive approach. Due to 92

37 approaches needs to acquire various machine parameters for these limitations, researchers prefer some public datasets for 93

38 condition based monitoring (CBM) of industrial equipment. the evaluation. In this work we have used commercial mod- 94

39 This technique focuses on forecasting the error by model- ular aeropropulsion system simulation (C-MAPSS) dataset 95

40 ing the degradation trends between input sensors and time- which is basically a simulation of turbofan jet engine dataset 96

41 to-failure duration of the machine. So, the benefits of this provided by NASA prognostics center of excellence [19]. 97

42 maintenance strategy is that we can eliminate unplanned C-MAPSS dataset consist of four different multivariate time 98

43 downtime, reduced maintenance costs and maximize the series units with different number of engines and each engine 99

44 machine lifetime for safety critical circumstances. One such having different RUL. The dataset consists of twenty one 100

45 example is aircraft engine which requires continuous mon- sensors with three operating conditions with respect to time 101

46 itoring of the engine performance. The fault diagnostics cycle for each engine. 102

47 and prognostics of aircraft engine has gained great attention In recent times, various work has been done on estimating 103

48 over the last few decades [4], [5], [6], [7]. One important the RUL of turbofan engine using deep learning methods such 104

49 component in aircraft engine maintenance is to accurately as CNN, LSTM along with their combinations and variants. 105

50 determine its remaining useful life (RUL) for reducing the LSTM networks have shown better results as compared to 106

51 maintenance costs while attaining the reliability [8], [9]. RUL CNN based models [20], [21]. LSTM have shown excellent 107

52 prediction model is developed based upon the degradation results because they are suitable for time-series data, they 108

53 trends among the various condition monitoring sensors. This can learn the temporal features in multivariate system and 109

54 model helps in development of maintenance strategy in a minimize the root mean square error (RMSE) with respect 110

55 targeted manner to eliminate unplanned downtime and maxi- to target predictions. In this paper, a LSTM based model 111

56 mize machine lifetime for safety critical circumstances. Early has been proposed for RUL prediction of a turbofan engine. 112

57 anomaly detection and timely warning of a failure is vital for LSTM network can learn the association between target RUL 113

58 maximum utilization of the system. There are basically three values and sensor data but it alone cannot achieve state of the 114

59 types of prognostics techniques used for estimating RUL, art performance due to various limitations like outliers, noise 115

60 physical model-based approaches [10], [11], data-driven in the sensor values, un-normalized data and un-correlated 116

61 approaches [12], [13] and hybrid approaches [14]. sensor values. These shortcomings can reduce the perfor- 117

62 Model based approach initially required a comprehensive mance of a LSTM network [22]. In this paper, we are focusing 118

63 understanding of the physical architecture of the machine and on implementing some preprocessing steps on the sensor data 119

64 then applying the laws of physics to obtain the mathematical before it can be set as an input into the LSTM network. LSTM 120

65 model of the machine for RUL estimation [15]. Mathemat- network when combined with effective pre-processing steps 121

66 ical models often take some simplifying assumptions with have the power to estimate the RUL with highly accuracy. 122

67 uncertainty management for a complex industrial machinery, These added steps involve correlation analysis, data filtering, 123

68 which can impose serious limitations on these techniques and normalization, and a modified piece linear degradation model 124

69 hence degrade the RUL prediction accuracy [16]. for determining starting point of the degradation. It has been 125

70 Data-driven based prognosis approaches use various sta- shown that the starting point of degradation which is also 126

71 tistical and machine leaning (ML) algorithms to discover the called the initial RUL has a great impact in determining accu- 127

72 trends or patterns in the underlying sensor data to estimate rate RUL predictions [23]. Our proposed modified piecewise 128

73 RUL of the system. These techniques are suitable for com- linear degradation models help in efficiently calculating the 129

74 plex industrial machinery and further, it does not require a starting point of degradation which in combination with the 130

75 thorough understanding of a complete engine or the process. other pre-processing steps and LSTM network accurately 131

76 Hybrid method combines both the physics and data-driven predicts RUL for the given engines. 132

77 based model techniques [17]. The main contributions of our work are enumerated as 133

78 In the past decade, data-driven based prognostics meth- follows: 134

79 ods have been exploited by many researchers. These mod- 1) Novel piecewise linear degradation model for deter- 135

80 els estimate the RUL by analyzing the degradation trend mining the starting point of engine degradation is 136

81 and target trajectory of sensor data. Deep learning methods proposed. 137

82 like autoencoder, convolutional neural networks (CNN), long 2) An LSTM network with effective pre-processing steps, 138

83 short-term memory (LSTM) networks and their varianta and i.e. correlation analysis with data normalization and 139

84 combinations have achieved a massive success in the fields moving median filter is proposed, which when aug- 140

85 of computer vision, speech recognition, video segmentation mented with the linear degradation model leads to an 141

86 and predictive maintenance [18]. The major drawback of deep improved RUL prediction. 142

87 learning algorithm is that it requires a large volume of data 3) Hyperparameter for the proposed prediction model 143

88 for offline training and in the field of prognostics, it is very has been selected through iterative grid search based 144

95426 VOLUME 10, 2022

 O. Asif et al.: Deep Learning Model for RUL Prediction of Aircraft Turbofan Engine on C-MAPSS Dataset

145 approach [24] to further improve the accuracy of our features in both direction and it can run training pass from 200

146 framework. forward to backward and backward to forward with back 201

147 The organization of remaining paper is given as fol- propagation algorithm. Wang et al. [40] proposed a hybrid 202

148 low. Section 2 gives the detailed literature review on network for turbofan engine in which trends and hidden 203

149 the existing methods on turbofan engine RUL estimation. pattern in long sequence sensor data is identified through 204

150 Section 3 explain the dataset taken for experimental testing. LSTM network and short duration sequence was analyzed 205

151 Section 4 discusses the proposed methodology with compre- through time window method with gradient boosting regres- 206

152 hensive block diagram that elucidate the entire work flow. sion (GBR). This method has two stages, offline stage to 207

153 Section 5 discusses the experimental results. Section 6 is the learn degradation pattern with LSTM network and TW-GBR 208

154 conclusion. used in online stage for extracting short sequence data. It also 209

implemented standardization and sensor selection criteria. 210

155 II. RELATED WORK Babu et al [41], proposed deep CNN regression network for 211

156 This section briefly reviews the existing literature on the tur- RUL estimation. The network consist of two dimensional 212

157 bofan engine RUL estimation. [25]. Traditional model-based convolutional layers for feature extraction followed by a fully 213

158 techniques usually employ algorithms like Kalman filter connected regression layer for prediction. For 2-D convolu- 214

159 (KF), extended Kalman filter (EKF) and particles filters to tion, first we have to convert our 1-D sensor data into 2-D with 215

160 come up with mathematical formulation of machine based on one dimension was taken as time and other taken as sensors 216

161 multi sensor time series sequence data [26], [27], [28]. Clas- amplitude. This model also extracts the spatial features very 217

162 sical degradation method such as Eyring model or Weibull efficiently. Li et al [42], proposed deep convolutional neural 218

163 distribution was implemented in [29]. Salahshoor et al. [30] network (CNN) for RUL estimation. The architecture of CNN 219

164 used a unified framework of EKF based design for sen- is modelled in such a way that feature can be extracted from 220

165 sor data fusion algorithm to further enhanced the detec- prepared 2-D sensor data by passing raw data into convolu- 221

166 tion and diagnosis of degradation trends and system faults. tion layers, then flattened layer is added to convert extracted 222

167 Ordonez et al. [31] implemented the auto-regressive inte- 2-D features into 1-D so that it can be given as an input to 223

168 grated moving average (ARIMA) model and support vector multilayer perceptron model with dropout layer for predicting 224

169 machine (SVR) methods collectively to estimate the RUL. RUL. Jayasinghe et al [43], proposed temporal convolution 225

170 The desired features can be created by analyzing prior learn- in which combination of CNN-LSTM network was used for 226

171 ing about the degradation models as presented in [32]. In [33], turbofan engine dataset. The layers of the model have stacked 227

172 it is suggested that failure thresholds or degradation state esti- by first applying data augmentation to create similar type of 228

173 mation is no longer required in learning-oriented approach. data for avoiding overfitting followed by data normalization 229

174 Khelif et al. [33] presented machine learning based support which was then followed by 1-D convolution for feature 230

175 vector regression (SVR) model to project the direct associa- extraction, lastly fully connected layer act as the bridge 231

176 tion between multivariate sensor data or health index and the between output of 1-D convolution layer and input of LSTM 232

177 aircraft turbofan engine RUL. layer. LSTM layer was then followed by fully connected layer 233

178 Across all these techniques for turbofan engine RUL pre- for output prognosis prediction. Hong et al. [44], proposed a 234

179 diction, deep neural network-based methods have gained vast similar kind of network by stacking a 1-D convolution layer, 235

180 popularity. Zhang et al. [34] introduced a multi-objective residual layer, LSTM layer and a Bi-LSTM layer. Correlation 236

181 evolutionary algorithm to expand and organized the deep analysis on sensor data for turbofan engine dataset was also 237

182 belief network into multiple parallel networks simultane- performed. Mo et al. [45], proposed multi-head neural net- 238

183 ously to accomplish the two convicting objectives i.e. diver- work for RUL prediction of turbofan engine. This network is 239

184 sity and accuracy. These networks attained a fine RUL different from the series network in such a way that they have 240

185 prediction accuracy especially in case of complicated oper- implemented the parallel branches of CNN layer in series 241

186 ations and in the presence of noise in input data [35], [36]. with LSTM network. Furthermore, fisher method in com- 242

187 Saeidi et al. [37], proposed a naive Bayesian classification bination with recursive least squares and single exponential 243

188 algorithm to measure the health index for turbofan engine. smoothing was also employed to find the prediction error 244

189 The pre-processing step takes the sensor data as input and and given it as an additional input into CNN-LSTM head for 245

190 apply moving average filter for removing the noise. It further optimum performance. Zhao et al. [46], proposed an adjacnet 246

191 categories the dataset into four different categories on the neural network model for leanring the degradation pattern in a 247

192 basis of time cycles i.e. time cycle values between 0 to 50 is sensor data. The degaradation pattern mapping learns through 248

193 labeled as urgent case which need immediate maintenance morkov property i.e. estimating the next state of sequence 249

194 and further categorization is also done in a similar manner. with the assist of only present states. 250

195 Zheng et al. [38] proposed LSTM network combined with Many researchers have used a piecewise linear degradation 251

196 piece wise linear function for RUL for estimating the degra- model in RUL prediction techniques. In this model, the start- 252

197 dation trends. It achieves good results by applying piece ing point of the degradation is estimated often referred to as 253

198 wise linear function and data normalization. Wei et al. [39] the initial RUL, many authors [47], [48], [49] have chosen its 254

199 proposed a Bi-LSTM network which can learn high level value on the basis of observations and no clear mechanism 255

VOLUME 10, 2022 95427

 O. Asif et al.: Deep Learning Model for RUL Prediction of Aircraft Turbofan Engine on C-MAPSS Dataset

256 of selecting it has been proposed. Lan et al. [23], proposed accurate estimation of RUL. This model have fourteen input 310

257 an LSTM algorithm for RUL prediction, it presented a piece parameters related to five rotating components of engine to 311

258 wise linear degradation model. The dataset is divided into simulate different degree of fault and deterioration of the 312

259 time windows and difference between geometric distance model. A total of twenty one variables out of fifty eight differ- 313

260 (Euclidian) of each window is used to determine the value ent sensor responses is considered from the model for predict- 314

261 of initial RUL or the starting point of degradation. They have ing the RUL. The three operating parameters of C-MAPSS 315

262 validated their model on only FD001 sub-unit of the dataset simulation model are given in table-1 and the details of 21 316

263 and selection of window size and thresholds are also not sensors are given in table-2. The legends of last column in 317

264 explained. Inspired by the results of this model, we have pro- table-2 Trends indicates the degradation pattern of sensor data 318

265 posed an improved version of the algorithm, which is applied with respect to time, where ‘‘∼ ’’ represents irregular sensor 319

266 to the complete dataset and results in overall increased in the behavior, I represent the parameter increasing with time and 320

267 accuracy of our prediction model. lastly, D is the variation of parameter that decreases with time. 321

268 The existing methods presented so far showed impressive

269 performance on RUL tracking; our proposed method further TABLE 1. Operating parameters of C-MAPSS.
270 improves the performance of these networks − summarized
271 in the results and discussion section. This is mainly due to the
272 fact that we have incorporated and implemented effective data
273 preprocessing steps into our prediction pipeline. In the previ-
274 ous works, pre-processing stages have not been completely
275 exploited in order to find the right combination for enhanc-
276 ing the learning capability of deep learning algorithms. So,
277 we propose one such combination along with the training of The main components of turbofan engine include nozzle, 322

278 LSTM with a hyperparameter selected through grid search low pressure turbine (LPT), high pressure turbine (HPT), 323

279 algorithm in order to achieve the optimum performance. fan, low pressure chamber (LPH) and high pressure chamber 324

(HPC). There are total of fourteen editable input parameters 325

280 A. PROBLEM STATEMENT such as fuel flow, HPC efficiency modifier, LPT efficiency 326

281 The predictive maintenance techniques is different from the modifier etc. that allows you to simulate various operating 327

282 reactive and preventive maintenance techniques in terms of behaviors. C-MAPSS sensor trajectories are further divided 328

283 safety, reliability, efficiency and unnecessary downtime for into four different units namely FD001, FD002, FD003, and 329

284 aircraft turbofan engine [50]. These methods ensure reliable FD004 corresponding to different operating conditions and 330

285 solution managing the health of turbofan engine to reduce fault modes. This dataset contains 709 engines for the training 331

286 the downtime, which leads to significant loss in revenue. and 707 engines for testing which are of same type but with 332

287 Therefore, failures in a turbofan engines can cause catas- distinct manufacturing variation and initial wear, unknown to 333

288 trophic accidents due to its sensitive nature and it needs to be the researcher. The description of four sub dataset units with 334

289 estimated prior in time so that we can provide maintenance train and test trajectories and other details are given in table-3. 335

290 services in order prevent any fatal incident. In the start, all the engines in each sub-dataset are operating 336

291 In current scenario, due to the implementation of normally as seen from sensor behavior and originate the 337

292 cyber-physical system (CPS, that link the cyber world with fault sometime later in their life cycle. In training sequence, 338

293 a physical world, called smart manufacturing), the industrial complete run-to-failure data is available with a specified RUL 339

294 sector such as health care, nuclear power plant etc. gener- labels as faults grows in the system and in test time degrada- 340

295 ate enormous volume, velocity, veracity and variety of data. tion values are given up to some time prior to engine failure. 341

296 Therefore, with the rise of AI and availability of hardware Moreover, with different initial health conditions, there are 342

297 computing resources, data driven based artificial intelligence distinct number of time cycle even for the same engine in 343

298 (AI) predictive maintenance models have a capability to pro- dataset. The objective of this dataset is to predict remaining 344

299 cess big amount of real-world machines data with ease and useful life cycle of engines in each sub-unit. The actual RUL 345

300 predict heath index of aircraft turbofan engine in time before label are given in the test data, which is used to validate the 346

301 failure to prevent unwanted breakdown [51]. prediction results. 347

It can be observed from table 3 that different sub-units of 348

302 III. DESCRIPTION OF C-MAPSS TURBOFAN ENGINE the main dataset have different running life cycle time e.g. 349

303 SIMULATION DATASET in FD004 test data, the maximum life cycle time is 486 and 350

304 C-MAPSS dataset released by NASA is developed in minimum time 19. 351

305 MATLAB environment as a tool for simulation of turbofan

306 engines. C-MAPSS dataset was published in 2008 for 1st IV. PROPOSED METHODOLOGY 352

307 International conference on PHM [19]. This dataset was pub- This paper proposes LSTM based RUL prediction model for 353

308 lished some time ago but still it has been actively used in turbofan engines, which proves to be more robust than most of 354

309 research for evaluating the prognostics model with a focus on the existing models available in the literature. The increased 355

95428 VOLUME 10, 2022

 O. Asif et al.: Deep Learning Model for RUL Prediction of Aircraft Turbofan Engine on C-MAPSS Dataset

TABLE 2. Output parameters C-MAPSS turbofan engine datasets. which have a very little or zero correlation with RUL, this 378

includes some parameter in engines that are basically con- 379

trolled by a feedback controller and results into a near con- 380

stant values or having an oscillatory behavior. These kinds 381

of parameters do not play much part in RUL predictions. So, 382

the selected feature signals are then given as an input into the 383

data filtering stage. Statistical evaluation of turbofan engine 384

degradation dataset gives us certain insight into the multivari- 385

ate sensor data and furthermore reach towards the conclusion 386

that whether a considered sensor is adequate for training 387

the network or not. We can accomplish this abstraction by 388

computing the value of correlation coefficient ’r’ which is a 389

relationship between the sensors and RUL labels. 390

conv(x, y)
r = (1) 391
sx sy
Pn
(xi − x̄)(yi − ȳ)
r = pPn i=1 Pn (2) 392

i=1 (xi − x̄)( i=1 (yi − ȳ))

where conv (x,y) is covariance between input sensor data (x) 393

and output RUL label (y), x̄ and ȳ are the mean of input sensor 394
TABLE 3. Each sub-unit of C-MAPSS dataset. data and RUL label, n is the number of variables in a dataset 395

and Sx & Sy are standard deviation of the two signals x and y 396

respectively. 397

In [44] a correlation analysis is employed for dimension- 398

ality reduction to obtain the accurate results and to reduces 399

the complexity of sensor data. This technique is primarily 400

limited to FD001 sub-unit of C-MAPSS dataset, we have 401

extended this approach to entire degradation dataset and com- 402

prehensively investigate the trends and irregular behaviors by 403

analyzing the correlation matrix heat map [54] of C-MAPSS 404

turbofan engine dataset. The correlation matrix heat map 405

cells show the association of three operating settings and 406

356 robustness stems from the fact that we have integrated effi- 21 sensors with output RUL labels as shown in Fig. 2. The 407

357 cacious techniques from multiple alternatives and presented correlation matrices are converted into percentage with dark 408

358 two types of improvements to be made in the LSTM based green color representing higher correlation as opposed to 409

359 prediction model. The first type of improvement is the addi- light color which depict a low correlation value. The number 410

360 tion of more effective pre-processing steps, which in return of sensors selected from each sub-unit after the correlation 411

361 puts a great influence on LSTM training. Secondly, enhancing process are given in table 4. After analyzing the correlation 412

362 the LSTM training procedure by a grid search approach for coefficient of sensor degradation dataset, we have concluded 413

363 computing the effective hyper parameters. This framework that 14 out 24 parameters in FD001 and 16 out of 24 parame- 414

364 manages to achieve higher prediction accuracy, measured ters in FD003 reflect high strength of correlation with defined 415

365 in temrs of RMSE and score function values, than several or monotonous behavior and we have omitted those variables 416

366 existing works − see Section 5 for details. The framework that indicates less than 10% of correlation with the output 417

367 of our proposed model is shown in Fig. 1. The specific steps RUL label. For the case of FD002 and FD004, correlation 418

368 of the prediction model are explained below. coefficient values in a heat map illustrated that these variables 419

have little degree of association with output RUL labels and 420

369 A. CORRELATION ANALYSIS

370 First, we must prepare our health-to-failure data into an
TABLE 4. Sensors selection based on correlation matrix.
371 appropriate form for improving the accuracy of the LSTM
372 network for effective training operation. C-MAPSS dataset
373 consists of three operational settings and twenty one sen-
374 sor signals of engines with machine life span length. These
375 signals are then given as an input to correlation analysis
376 method [31], [44], [52], [53] to discover the relevance of
377 features with RUL. The algorithm excludes the sensor values

VOLUME 10, 2022 95429

 O. Asif et al.: Deep Learning Model for RUL Prediction of Aircraft Turbofan Engine on C-MAPSS Dataset

FIGURE 1. Proposed Framework for RUL predictions.

421 this correlation is more than 15% for best possible case. The choice of filter is made on its ability in removing the 430

422 After analyzes the correlation matrix, un-correlated sensors outlier while preserving the high and low frequency contents 431

423 are removed from the training process which are listed in table in sensor data and avoids any loss of data values. The time 432

424 5. So, only highly correlated sensor data are used for training window size of moving median filter is adaptive and vary with 433

425 the LSTM network which also increases the speed of learning respect to sensor values. The moving median filter belong to 434

426 for estimating RUL. a type of non-linear digital filter, which is used to remove 435

random unwanted noise especially when there is a high spike 436

427 B. FILTERING and short-term outlier present in the data points but preserving 437

428 The correlated data is passsed through a moving median filter the high frequency information contents [55]. Median filter is 438

429 for removing the outliers and noises in the sensors data. used to identify such sensor values in turbofan engine which 439

95430 VOLUME 10, 2022

 O. Asif et al.: Deep Learning Model for RUL Prediction of Aircraft Turbofan Engine on C-MAPSS Dataset

and then apply the following operation on each sensor output. 468

xic − µn
yni = (4) 469
σn
where, yni is the normalized value at ith time cycle for sensor 470

n, µn is the mean value of all output of sensor n, σ n is the 471

standard deviation of all n sensor output. 472

D. IMPROVED PIECEWISE LINEAR DEGRADATION MODEL 473

It is observed that RUL is linear decreasing function with 474

FIGURE 2. Correlation Matrix of C-MAPSS Dataset. respect to time as the efficiency of the system degrades. How- 475

ever, as the system starts their operation, there is no degrada- 476

tion present in the sensor readings. This pre-processing step is 477

TABLE 5. Excluded sensor list.
basically implemented on output labelled data that takes input 478

from previous correlation analysis stage and employs a piece 479

wise linear degradation function on sensor values for finding 480

the initial RUL or the starting point of degradation. All the 481

labels till this deterioration point are constant out to the initial 482

RUL value while the remaining RUL lables are represented 483

as a linear line from that degrdation point up to zero life cycle 484

time. 485

In this paper, we have presented an improved version of 486

the automatic piece wise linear function [23] for output RUL 487

labeling. This approach is self-governing that is sensitive on 488

440 are considerably different from the other data points and then variation of the degradation trends and will automatically 489

441 eliminate them. The algorithm of moving median filter is calculate the early point of sensor deterioration. The com- 490

442 accomplished by sliding the window of appropriate length putation of initial RUL starts by dividing entire sensor time 491

443 over the sensor data entries by entries and then replacing the cycle with non-overlapping pattern into equal sized window 492

444 corresponding values by computing the median of neighbor- length of w and extract the sensor data from a given windows. 493

445 ing points specified in the window. Mathematically it can be We then calculate the centroid of each considered window 494

446 expressed as: by determining their mean values and geometric distance 495

calculation is performed by subtracting the two subsequent 496

j j j windows to generate the trends in sensor data. As there are N 497

447 yi [n] = median(xi [n], . . . , xi [n + T ]) (3)
number of time cycles for given variable and window length 498

of w results into (g=N/w) geometric points for a given dataset. 499

448 where n = total number of sensor data points in each engine, These geometric distances are computed using Euclidean 500
j
449 i = sensor variable, j = engines, xi [n] denotes input data for distance method which is then squared and the degradation 501
j
450 sensor i in engine j and yi [n] returns the medians values across pattern from g values is evidently detectable from the result- 502
451 each sensor variables from FD001 to FD004 with a same ing plot as shown in Fig. 3. The centroid of window w1 is 503
452 dimension as of input data. first computed and subtracted from the other windows in a 504
453 The algorithm of filter is accomplished by sliding the sequence to compute the variation in sensor values to reach 505
454 window of length T over the neighboring elements in sensor on a point of deterioration based on the threshold value. The 506
455 data and computed the median for each considered window inflection point of the curve indicates the increase in sensor 507
456 and median operation is then performed on these array vector trends which is the initial RUL value. 508
457 to get the filtered output after applying ascending operation. The proposed algorithm is given as Algorithm-1 is imple- 509
458 So this processed data is then given to the next stages and mented for each engine. The minimum value of initial RUL 510
459 hence put a significant impact on the output. among all the engines in a sub-unit is taken as the initial RUL 511

for that sub-unit. Threshold level is dependent on the rate of 512

460 C. DATA NORMALIZATION rise in raw sensor data, its visual perception and how early 513

461 The range of sensor output after analyzing from the graphs we need to predict the faults in the machines for maintenance 514

462 is from tens to thousands and if we use these raw values purpose. 515

463 for training the network then accuracy will drop signifi- In this paper, we have set different ranges of threshold 516

464 cantly [56]. The filtered signals are normalized to have same (0.01 to 0.2) and window size (5,12) for calculating the initial 517

465 degree of range for efficient training of the network. Z-Score RUL and validating the performance of our model. We have 518

466 normalization [57] is used in this paper which first compute used different values of window sizes in order to compute 519

467 the mean (µ) and standard deviation (σ ) of each feature vector the knee point in sensor data effectively. This choice stems 520

VOLUME 10, 2022 95431

 O. Asif et al.: Deep Learning Model for RUL Prediction of Aircraft Turbofan Engine on C-MAPSS Dataset

FIGURE 4. Structure of a LSTM cell.

in a sequential manner with different number of hidden units 537

FIGURE 3. Degradation process represented in the form of windows of and a dropout layer is also added in between the LSTM 538
time cycle.
layers for enhancing the generalization of network to avoid 539

over fitting. It is then followed by two fully connected layer 540

Algorithm 1 Improved Piece-Wise Linear Function for Initial
with dropout layers and the final layer is the regression layer 541
RUL Calcultion for Each Engine
as shown in Fig. 1. Fig.4 shows a basic LSTM cell that is 542
1: Inputs: Time cycle values (tc) and Sensors data after
essentially consist of three control gates: input gate, forget 543
filtering and normalization gate and output gate. The output of the cell is denoted by ht , 544
2: Parameters: w = window length, g = total num-
which is a short-term memory sate in a network and Ct is 545
ber of windows, w1 ,w2 . . . wg = subsequent windows, considered as a long-term cell state. The first gate in an 546
Th=threshold LSTM cell is forgot gate ft , which is used to unlearn selective 547
3: Output: Initial RUL label (irul)
information stored in previous LSTM cell. The forget gate 548
4: e = Extract sensors values from the given windows
equation is given below. 549
(w1 ,w2 . . . wg )
5: m = Calculate the centroid of each window by ft = σ (Wf .[ht−1 , xt ] + bf ) (5) 550

computing the mean for each window

where σ ( ), is called sigmoid activation function, which can 551
6: for i = (2 to g − 1) do
control operation of forget gate. Wf is the weight matrix, ht−1 552
7: s = Subtract the mean of two windows (w1 ,wi )
short term state from previous cell, xt is the input of cell, and 553
8: sq = square (s)
bf is the bias vector of LSTM cell. The input gate controls the 554
9: if sq >= Th then
new information entering into the cell through following two 555
10: irul = tc - w * i
equations: 556
11: else
12: i = i+ 1 it = σ (Wi .[ht−1 , xt ] + bi ) (6) 557
End
C̄t = tanh(Wc .[ht−1 , xt ] + bc ) (7) 558

where tanh = (ex − e−x )/(ex + e−x ). This value is calculated 559
521 from the fact that by further increasing the window size, the by the same short term state vector ht−1 which is further used 560
522 algorithm bypasses the knee point, which is the starting point to update the new state of cell. Wi and bi is the weight matrix 561
523 of increasing/decreasing trend in sensor variables. Fig. 3 con- and bias vector of an input gate respectively. The C̄t computed 562
524 firms that the trend in sensor data changes within few cycles, from above equation is first filtered by it and then added to 563
525 and if we further increase the window size then we cannot the long term state of the cell. Wc and bc are the weight matrix 564
526 achieve true inflection point in these variations. As a result, and bias vector. After computing the value of forget gate (ft ), 565
527 initial RUL computation will not reflect the ideal knee point input gate (it ) and (C̄t ), long term state Ct of LSTM cell is 566
528 in the sensors’ behavior. updated after applying given below matrix operation 567

529 E. LSTM MODEL DEVELOPMENT Ct = ft ⊗ Ct−1 + it ⊗ C̄t (8) 568

530 The data from the normalization stage with updated RUL where, ⊗ is basically element wise matrix multiplication 569
531 labels from the degradation model is used to train a deep operation between a specified variable and Ct−1 is the pre- 570
532 LSTM network. Our proposed LSTM model consists of vious state of LSTM cell. Finally the output of LSTM cell is 571
533 LSTM layers, dropout layers, fully connected layers and the generated by the following two equation, 572
534 regression layer. The output node of fully connected layer is
535 a regression layer that gives the estimated RUL of turbofan ot = σ (Wo .[ht−1 , xt ] + bo ) (9) 573

536 engine. Our proposed model consists of four layers connected ht = ot ⊗ tanh(Ct ) (10) 574

95432 VOLUME 10, 2022

 O. Asif et al.: Deep Learning Model for RUL Prediction of Aircraft Turbofan Engine on C-MAPSS Dataset

575 The output state of LSTM cell ht , is obtained by filtering Algorithm 2 Proposed Methodology
576 the output gate ot equation with a matrix of Ct . Wo and bo 1: Inputs: Training Data,{δT = (x11 ,y11 ),(x21 ,y12 ). . . .(xNE ,yEN )}
577 are the weight matrix and bias vector of the output equation. 2: Parameters: N ← No. of sensor variables, E ← No.
578 The values of the weight matrix and biases are computed by of Engines, δN ← Filtered data, δC ← Correlated sen-
579 training the LSTM network. sors, M ← No of correlated sensors, δD ← Normal-
ized data, δP ← Piece-wise linear RUL function, LN ←
580 F. DROPOUT LAYER No. of layers in LSTM, LD ← Doprout layers, RL ←
581 The dropout layer is added to avoid the overfitting which Regression layer, LF ← No. of neuron in FC layer, W
582 inherently occur while training the deep neural network [58]. ← Window Length, α ← Learning rate, GD ← Gradient
583 This regularization layer is added in between the fully con- descent optimizer, Vs ← Validation set, Sarc t ← Network

584 nected layers and LSTM layers to increase the generalization architecture parameters
585 of the whole algorithm to better track the predicted RUL 3: Output: Performance Parameter,Opp ←[RMSE,Score]
586 with high accuracy. It drops some random portion of neuron 4: for i = (1 to E) do
587 according to the probability parameter defined in a network 5: for j = (1 to N ) do
588 while remaining weights are trained by the backpropagation 6: δC = Correlation Analysis (xji ,yij )
589 algorithm [59]. The expression for a dropout layer by choos- end for j
590 ing a linear activation function and considering the activity in end for i
591 unit i at layer h is given by the expression. 7: for k = (1 to M ) do
XX 8: δN = Median Filter(δCk ,W)
592 Sih (I ) = whl l 0
ij Sj with Sj = Ij (11) 9: δD = Data Normalization(δNk , mean(δN ), SD(δN ))
I <h j
10: δP = Piece-wise linear function(δCk )
593 where w is the weight and I is the input vector. end for k
11: t ← (L ,L ,L ,α,G ,L ,V ,R ,δ ,δ )
Sarc N H F D D s L D P
594 G. FULLY CONNECTED LAYER 12: for l = (1 to T ) do
13: FD =Forward Pass (Sarct )
595 The fully connected layers gets the data from the final com-
596 bination of LSTM and dropiut layer, so that the features 14: Et = Error(FD )
597 extracted from LSTM layer are used to generate the out- 15: Bp = Back Propagation(Et )
598 put [60]. Due to its fully connected nature between all the 16: Wnew = Wold + α*derivative(Bp )
599 neuron present in the network, it has a large amount of 17: bnew = bold + α*derivative(Bp )
600 weight parameters which needs to be computed by training end for l
601 the network. Fully connected layer along with a dropout 18: Oipp ← Calculate RMSE using equation 14.
602 layer is followed by single regression layer for predicting the 19: Oi+1
pp ←Calculate score function using equation 15.
603 RUL. The mathematical calculation of fully connected layer
604 is expressed below.
605 H0 = I (12) the model. These parameters include learning rate, batch 623

606 Hl = φ(h(l−1) Wfc + bfc ) (13) size, number of layers, number of neurons in each layer, 624

optimizer and the training epochs. There is no definite rule 625

607 where I is the input vector, φ is activation function of a or process by which the best values of these parameters are 626

608 neuron which is primarily a ReLu fucntion, hl−1 is output selected, generally an iterative grid search method is used to 627

609 from the previous layer, Wfc is weight matrix of a specified get the best results. The model is run against different values 628

610 fully connected layer, bfc is the bias vector and Hl is output of these parameters and combination of values with best 629

611 at l th layer. result are recorded. Addition of more layers makes system 630

more complex and timing consuming for training but could 631

612 H. INTERPRETATION OF OUR APPROACH result in better predictions as compared to the single layer 632

613 We implement deep neural network combined with network which are more suited to the simple problems. The 633

614 per-processing steps for efficient RUL prediction of turbo- ranges/types of hyper parameters used for LSTM network 634

615 fan engine. Furthermore, we have separately discussed the training are given in table 6. 635

616 above-mentioned stages of prediction pipeline breifly in order

617 to delineate a comprehensive understanding of our work. J. EVALUATION METRICS 636

618 Therefore, to summarize our approach, we have presented After network development, training of LSTM network is 637

619 a pseudo code of our proposed pipeline in Algorithm 2. carried out and selection of hyper parameters is made on 638

the basis of root mean square error (RMSE) between actual 639

620 I. HYPER PARAMETERS SELECTION RUL and predicted RUL for each engine of the turbofan 640

621 LSTM training process involves many different parameters. dataset. Our proposed LSTM model is evaluated using two 641

622 These parameters have a great impact on the accuracy of widely used evaluation methods: RMSE and scoring function. 642

VOLUME 10, 2022 95433

 O. Asif et al.: Deep Learning Model for RUL Prediction of Aircraft Turbofan Engine on C-MAPSS Dataset

TABLE 6. Ranges/types of hyper parameters in LSTM model training. TABLE 7. Selected LSTM models with hyper-parameters values.

643 RMSE is the most used evaluation metric for predictions

644 used by the researcher for past decades. It is a symmetrical
645 scoring function which means it can assign equals weights or
646 penalties to both early and late prediction. The equation of
647 RMSE is given below.
s
PN 2
i=1 ei
648 RMSE = (14)
N
649 where ei , is the prediction error and N is the total number
650 of samples. In PHM08 data challenge competition, a score
651 function was employed to evaluate the performance of the
652 prediction model [61]. It is an asymmetrical score function
653 which means it can assign more weights to late prediction as
654 opposed to early prediction. It is described mathematically as:
N
X
655 score = sj (15)
j=1
 −h
e 13j − 1, if h < 0
j
656 sj = −h (16)
e 10j − 1, if h ≥ 0
j

657 The performance of the model will degrade at greater extent

658 in case of delayed prediction as compared to the early predic-
659 tion. The prediction error between the actual and predicted
660 RUL is close to zero then the value of score and RMSE is
661 also smaller with advancing towards zero.

662 V. RESULTS AND DISCUSSION

663 In this section, a detail analysis on RUL prediction results by FIGURE 5. RUL Prediction on Test Set, (a) FD001 (b) FD002, (c) FD003,
664 our proposed model is presented. There are various param- (d) FD004.

665 eters that affects the model’s accuracy. We can categorized

666 them into two sets, first are the parameters related to the initial
667 RUL value, these includes window size and threshold values. four LSTM models which gave us encouraging results were 678

668 Second ones are LSTM hyper parameters. Combining these further tested and tuned for best performance. The models 679

669 two sets of parameters gives a very large set of parameters along with the hyper parameters are given in table-7, LSTM 680

670 affecting the model’s accuracy. It is evident from the existing model structure remains same as described in the Fig.1. The 681

671 literature on deep learning techniques that there is no straight output of the pre-processing stages which processed sensor 682

672 forward way of determining a single best set of values for data and RUL output labels are given as an input into the 683

673 these parameters. This requires very extensive and detailed LSTM network. The network is trained using the different 684

674 simulations to analyze it further and an iterative grid search values of the hyper parameters given in the table-7 and the 685

675 approach is followed. The simulation model is developed in window size and threshold ranges defined in the last section. 686

676 MATLAB and hundreds of runs were carried out against the After following an iterative grid search approach, the best 687

677 parameters ranges defined in the last section. For comparison, hyper parameters are selected for the proposed model on the 688

95434 VOLUME 10, 2022

 O. Asif et al.: Deep Learning Model for RUL Prediction of Aircraft Turbofan Engine on C-MAPSS Dataset

TABLE 8. Model accuracy in term of RMSE and Score function values for the four LSTM models.

689 basis of lowest RMSE and score function values. The results
690 of the these four LSTM models against different thresholds
691 and window sizes values are given in table 8.
692 The results in table-8 show that there is no single set of
693 best values parameters for the whole dataset. The best values
694 of parameters for RMSE and score function are different for
695 each sub-unit. For FD001 case the best results are with using
696 LSTM Model-2 and window size and threshold values of
697 12 and 0.2 respectively. Similarly, for FD002 best results are
698 with using LSTM Model-2 and window size and threshold
699 values of 12 and 0.2 respectively. For FD003, best results are
700 with LSTM Model-3 and window size and threshold values
701 of 12 and 0.2 respectively and for FD004, best results are with
702 using LSTM Model-1 and window size and threshold values
703 of 12 and 0.2 respectively. It can be observed that although FIGURE 6. Histogram Distribution of Error, (a) FD001, (b) FD002,
704 LSTM model hyper parameters are different for different (c) FD003, (d) FD004.

705 sub-units, the best values of window size and threshold are
706 same. The minimum initial RUL value among all engines
707 in a sub dataset is selected as its initial RUL. The initial observed by histogram distribution of prediction error in 722

708 RUL values achieved from the training part of sub-units Fig. 6. 723

709 FD001 to FD004 are 78, 103, 79 and 87 respectively. These The histogram distribution of four sub-units indicates the 724

710 values are then used in the testing stage for each sub-unit. variation of RMSE across the dataset. The x-axis indicates 725

711 So, instead of using a single initial RUL value for the entire the error or difference between predicted observation and true 726

712 dataset which is used in most of the existing literature, we can RUL while the y-axis indicates the frequency of occurrence 727

713 achieve improve prediction accuracy by using separate initial for the given error. It can be seen from the figure that large 728

714 RUL value and LSTM hyper parameters values for each concentration for frequency of error lies in the range of 729

715 sub-unit. [-5 0] in FD001 & FD003 while for FD002 and FD004, 730

716 The prediction results using the best models are shown in it is concentrated in [-10 0]. The data description in section 731

717 Fig. 5, which shows highly accurate RUL predictions. The III shows that FD002 and FD004 has 6 operating conditions 732

718 x-axis denotes the number of engines in a specific sub-unit with more than 200 tracking trajectories. In correlation anal- 733

719 and y-axis is the result of RUL prediction for each engine. ysis, we have demonstrated that these two sub-units pos- 734

720 The red and blue legends on the graph indicate the predicted sess irregular behavior so RUL prediction for these complex 735

721 and true value of RUL. Moreover, these results are further sequences is a challenge for the prediction model. FD001 has 736

VOLUME 10, 2022 95435

 O. Asif et al.: Deep Learning Model for RUL Prediction of Aircraft Turbofan Engine on C-MAPSS Dataset

TABLE 9. Comparison of RMSE and score with other methods.

737 lowest RMSE and score because of one fault mode. Finally, linear degradation function which decides how early you 749

738 we have compared our proposed LSTM model with other need to provide the maintenance operation by determining 750

739 method published in the last few years on C-MAPSS sensor the starting point of the degradation. 751

740 degradation dataset. The validation of model is investigated The robustness of our proposed model is showed by pre- 752

741 by RMSE and score function values which is presented in dicting the full life cycle time of few randomly selected test 753

742 table 9. Our proposed model achieves state of the art per- engines in Figures 7-10. Fig. 7 shows the actual and predicted 754

743 formance on FD001 and FD003 sub-dataset with minimum degradation results of RUL of four different types of engines 755

744 RMSE, score values and it shows second best result on FD002 from a total set of 100 engines for validating the performance 756

745 and FD004 sub-dataset. The minimum values of FD001 and of model for FD001. In a similar manner, Fig. 8-10 shows 757

746 FD003 are lower than FD002 and FD004 because of the the actual and predicted results of full cycle for other three 758

747 irregular behavior of sensor degradation data. The accuracy sub-units. From these prediction graphs, we analyze that the 759

748 significantly depends on the improved automatic piece wise predicted and actual degradation of test engine in FD001 and 760

95436 VOLUME 10, 2022

 O. Asif et al.: Deep Learning Model for RUL Prediction of Aircraft Turbofan Engine on C-MAPSS Dataset

FIGURE 7. RUL Prediction of 4 engines in FD001 sub-dataset (a) engine FIGURE 9. RUL Prediction of 4 engines in FD003 sub-dataset (a) engine
# 6 (b) engine # 20 (c) engine # 42 (d) engine # 90. # 1 (b) engine # 18 (c) engine # 48 (d) engine # 75.

FIGURE 10. RUL Prediction of 4 engines in FD004 sub-dataset (a) engine

FIGURE 8. RUL Prediction of 4 engines in FD002 sub-dataset (a) engine # 15 (b) engine # 73 (c) engine # 155 (d) engine # 202.
# 1 (b) engine # 40 (c) engine # 84 (d) engine # 150.

To conclude, in what follows we reason why the pro- 771

761 FD003 sub-dataset are very efficient and accurate as shown in posed prediction framework is able to yield better results 772

762 Fig. 7 and Fig. 9 but for FD002 and FD004, we have seen that than the existing equivalent models. The proposed frame- 773

763 there was some irregular and abrupt changes in the predicted work consists of multiple stages from correlation analysis 774

764 results while tracking the actual target RUL of test engine due to LSTM network. The significant stages in this framework 775

765 to the complex nature of theses sub-datasets, this is shown that greatly enhance the overall accuracy of model are cor- 776

766 in Fig. 8 and Fig. 10. In case of prediction on actual RUL relation function to filter out irrelevant sensor variables as 777

767 having a plot similar to a straight line as shown in Fig. 9(b), given in Fig. 2, and the estimation of initial RUL value 778

768 our model performs quite good on tracking the degradation with piecewise linear function as given in algorithm 1. This 779

769 trends obtained by applying piecewise linear function on a function gives the starting point of degradation for sensor 780

770 target labels. data, and hence, we have used those values for training 781

VOLUME 10, 2022 95437

 O. Asif et al.: Deep Learning Model for RUL Prediction of Aircraft Turbofan Engine on C-MAPSS Dataset

782 our deep LSTM network. Moreover, the hyper-parameters [2] O. Merkt, ‘‘On the use of predictive models for improving the quality 837

783 selection for the neural network plays a significant role on of industrial maintenance: An analytical literature review of maintenance 838
strategies,’’ in Proc. Federated Conf. Comput. Sci. Inf. Syst., Sep. 2019, 839
784 the effectiveness of model. Therefore, in our research work, pp. 693–704. 840
785 the hyper-parameters have been picked through an iterative [3] Z. M. Çınar, A. Abdussalam Nuhu, Q. Zeeshan, O. Korhan, M. Asmael, and 841

786 grid search based approach as shown in Table-6. Thus, these B. Safaei, ‘‘Machine learning in predictive maintenance towards sustain- 842
able smart manufacturing in industry 4.0,’’ Sustainability, vol. 12, no. 19, 843
787 steps when combined in a single prediction pipeline achieved p. 8211, Oct. 2020. 844
788 optimum results compared to the existing methods on all the [4] B. Lu, D. Durocher, and P. Stemper, ‘‘Predictive maintenance techniques,’’ 845

789 four sub-units, and its tracking performance on the complete IEEE Ind. Appl. Mag., vol. 15, no. 6, pp. 52–60, Nov. 2009. 846
[5] J. J. M. Jimenez, S. Schwartz, R. Vingerhoeds, B. Grabot, and M. Salaün, 847
790 life cycle time prediction will be beneficial in preventing the ‘‘Towards multi-model approaches to predictive maintenance: A system- 848
791 real life failures of machines. atic literature survey on diagnostics and prognostics,’’ J. Manuf. Syst., 849
vol. 56, pp. 539–557, Jul. 2020. 850

792 VI. CONCLUSION & FUTURE DIRECTION [6] S. Behera, A. Choubey, C. S. Kanani, Y. S. Patel, R. Misra, and A. Sillitti, 851
‘‘Ensemble trees learning based improved predictive maintenance using 852
793 This work presented a LSTM based model for predicting IIoT for turbofan engines,’’ in Proc. 34th ACM/SIGAPP Symp. Appl. 853
794 the RUL of turbofan engines. The proposed work shows that Comput., Apr. 2019, pp. 842–850. 854

795 LSTM model when combined with effective pre-processing [7] K. T. P. Nguyen and K. Medjaher, ‘‘A new dynamic predictive maintenance 855
framework using deep learning for failure prognostics,’’ Rel. Eng. Syst. 856
796 steps results in highly accurate RUL predictions. In addition Saf., vol. 188, pp. 251–262, Aug. 2019. 857
797 to widely used pre-processing steps like feature selection, [8] C. Zheng, W. Liu, B. Chen, D. Gao, Y. Cheng, Y. Yang, X. Zhang, S. Li, 858

798 filtering and normalization, this work proposed an improved Z. Huang, and J. Peng, ‘‘A data-driven approach for remaining useful life 859
prediction of aircraft engines,’’ in Proc. 21st Int. Conf. Intell. Transp. Syst. 860
799 piecewise linear degradation model which estimated the ini- (ITSC), Nov. 2018, pp. 184–189. 861
800 tial RUL valuewhich is the starting point of the degradation. [9] V. Mathew, T. Toby, V. Singh, B. M. Rao, and M. G. Kumar, ‘‘Prediction 862

801 This value had a great impact on the RUL prediction accuracy. of remaining useful lifetime (RUL) of turbofan engine using machine 863
learning,’’ in Proc. IEEE Int. Conf. Circuits Syst. (ICCS), Dec. 2017, 864
802 LSTM network of the proposed model which consisted of a pp. 306–311. 865
803 combination of multiple LSTM layers, dropout layers, fully [10] N. Bolander, H. Qiu, N. Eklund, E. Hindle, and T. Rosenfeld, ‘‘Physics- 866
804 connected layers and a regression layer, was tested on multi- based remaining useful life prediction for aircraft engine bearing progno- 867

805 ple hyper parameter combinations to achieve the best result. sis,’’ in Proc. Annu. Conf. PHM Soc., vol. 1, 2009, pp. 1–10. 868
[11] C. Xiongzi, Y. Jinsong, T. Diyin, and W. Yingxun, ‘‘Remaining useful life 869
806 The proposed model was tested on the C-MAPSS turbofan prognostic estimation for aircraft subsystems or components: A review,’’ in 870
807 engine simulation dataset. the prediction accuracy depends Proc. IEEE 10th Int. Conf. Electron. Meas. Instrum., Aug. 2011, pp. 94–98. 871

808 upon various number of parameters which include parameters [12] P. Adhikari, H. Rao, and M. Buderath, ‘‘Machine learning based data driven 872
diagnostics & prognostics framework for aircraft predictive maintenance,’’ 873
809 for initial RUL calculations and hyper parameters for LSTM in Proc. 10th Int. Symp. NDT Aerosp. Dresden, Germany, 2018, pp. 1–15. 874
810 model. It was observed that due to different dynamics of each [13] L. Liu, Q. Guo, D. Liu, and Y. Peng, ‘‘Data-driven remaining useful life 875

811 sub dataset, it is better to use a separate prediction model for prediction considering sensor anomaly detection and data recovery,’’ IEEE 876
Access, vol. 7, pp. 58336–58345, 2019. 877
812 each sub-units instead of a single model for the whole dataset. [14] M. Aykol, C. B. Gopal, A. Anapolsky, P. K. Herring, B. van Vlijmen, 878
813 RUL predictions of the proposed model were compared with M. D. Berliner, M. Z. Bazant, R. D. Braatz, W. C. Chueh, and B. D. Storey, 879

814 recent benchmark publications and showed higher prediction ‘‘Perspective—Combining physics and machine learning to predict battery 880
lifetime,’’ J. Electrochem. Soc., vol. 168, no. 3, Mar. 2021, Art. no. 030525. 881
815 accuracy. The threshold value and the window size are two
[15] H. Oh, M. H. Azarian, M. Pecht, C. H. White, R. C. Sohaney, and E. Rhem, 882
816 important parameters in the degradation model which had ‘‘Physics-of-failure approach for fan PHM in electronics applications,’’ in 883
817 great impact on the model accuracy. In future, this approach Proc. Prognostics Syst. Health Manage. Conf., Jan. 2010, pp. 1–6. 884

818 can be considered for online problems and other machines. [16] C. S. Byington, M. Watson, D. Edwards, and P. Stoelting, ‘‘A model- 885
based approach to prognostics and health management for flight control 886
819 In future, we will further integrate novel deep learning actuators,’’ in Proc. IEEE Aerosp. Conf., vol. 6, Mar. 2004, pp. 3551–3562. 887
820 algorithms into our prediction pipeline for processing the [17] M. A. Chao, C. Kulkarni, K. Goebel, and O. Fink, ‘‘Fusing physics-based 888

821 complex and real-world sensor data as opposed to traditional and deep learning models for prognostics,’’ 2020, arXiv:2003.00732. 889
[18] W. Zhang, D. Yang, and H. Wang, ‘‘Data-driven methods for predictive 890
822 techniques. In the field of prognostics, normally we don’t maintenance of industrial equipment: A survey,’’ IEEE Syst. J., vol. 13, 891
823 get large amount of real-world data for industrial machines no. 3, pp. 2213–2227, Sep. 2019. 892

824 specially in-case of faulty conditions. Therfore, we need to [19] A. Saxena, K. Goebel, D. Simon, and N. Eklund, ‘‘Damage propagation 893
modeling for aircraft engine run-to-failure simulation,’’ in Proc. Int. Conf. 894
825 develop such novel methods that can be able to train on small Prognostics Health Manage., Oct. 2008, pp. 1–9. 895
826 dataset, but, it predict the output effectively on large scale [20] Y. Liao, L. Zhang, and C. Liu, ‘‘Uncertainty prediction of remaining useful 896
827 practical data. Basically, we requires such types of models, life using long short-term memory network based on bootstrap method,’’ 897

828 which have strong generalization capability with respect to in Proc. IEEE Int. Conf. Prognostics Health Manage. (ICPHM), Jun. 2018, 898
pp. 1–8. 899
829 output. Morever, for further improvements, we have to eval- [21] S. Zhao, Y. Zhang, S. Wang, B. Zhou, and C. Cheng, ‘‘A recurrent neural 900
830 uate the performance of models on other publicly available network approach for remaining useful life prediction utilizing a novel 901

831 dataset. trend features construction method,’’ Measurement, vol. 146, pp. 279–288, 902
Nov. 2019. 903
[22] S. Agarwal, V. Saxena, V. Singal, and S. Aggarwal, ‘‘LSTM based music 904
832 REFERENCES generation with dataset preprocessing and reconstruction techniques,’’ in 905
833 [1] A. L. Ellefsen, E. Bjørlykhaug, V. Æsøy, S. Ushakov, and H. Zhang, Proc. IEEE Symp. Ser. Comput. Intell. (SSCI), Nov. 2018, pp. 455–462. 906
834 ‘‘Remaining useful life predictions for turbofan engine degradation [23] G. Lan, Q. Li, and N. Cheng, ‘‘Remaining useful life estimation of tur- 907
835 using semi-supervised deep architecture,’’ Rel. Eng. Syst. Saf., vol. 183, bofan engine using LSTM neural networks,’’ in Proc. IEEE CSAA Guid., 908
836 pp. 240–251, Mar. 2019. Navigat. Control Conf. (CGNCC), Aug. 2018, pp. 1–5. 909

95438 VOLUME 10, 2022

 O. Asif et al.: Deep Learning Model for RUL Prediction of Aircraft Turbofan Engine on C-MAPSS Dataset

910 [24] K. T. Chui, R. W. Liu, M. Zhao, and P. O. D. Pablos, ‘‘Predicting students’ [45] H. Mo, F. Lucca, J. Malacarne, and G. Iacca, ‘‘Multi-head CNN- 985
911 performance with school and family tutoring using generative adversar- LSTM with prediction error analysis for remaining useful life predic- 986
912 ial network-based deep support vector machine,’’ IEEE Access, vol. 8, tion,’’ in Proc. 27th Conf. Open Innov. Assoc. (FRUCT), Sep. 2020, 987
913 pp. 86745–86752, 2020. pp. 164–171. 988
914 [25] A. P. Hermawan, D.-S. Kim, and J.-M. Lee, ‘‘Predictive maintenance of [46] Z. Zhao, B. Liang, X. Wang, and W. Lu, ‘‘Remaining useful life prediction 989
915 aircraft engine using deep learning technique,’’ in Proc. Int. Conf. Inf. of aircraft engine based on degradation pattern learning,’’ Reliab. Eng. Syst. 990
916 Commun. Technol. Converg. (ICTC), Oct. 2020, pp. 1296–1298. Safe., vol. 164, pp. 74–83, Aug. 2017. 991
917 [26] D. Simon and D. L. Simon, ‘‘Aircraft turbofan engine health estimation [47] A. Zhang, H. Wang, S. Li, Y. Cui, Z. Liu, G. Yang, and J. Hu, ‘‘Transfer 992
918 using constrained Kalman filtering,’’ J. Eng. Gas Turbines Power, vol. 127, learning with deep recurrent neural networks for remaining useful life 993
919 no. 2, pp. 323–328, Apr. 2005. estimation,’’ Appl. Sci., vol. 8, no. 12, p. 2416, Nov. 2018. 994
920 [27] T. Kobayashi and D. L. Simon, ‘‘Hybrid Kalman filter approach for aircraft [48] K. T. Chui, B. B. Gupta, and P. Vasant, ‘‘A genetic algorithm opti- 995
921 engine in-flight diagnostics: Sensor fault detection case,’’ in Turbo Expo: mized RNN-LSTM model for remaining useful life prediction of turbofan 996
922 Power for Land, Sea, Air, vol. 42371. Barcelona, Spain: ASME, 2006, engine,’’ Electronics, vol. 10, no. 3, p. 285, Jan. 2021. 997
923 pp. 745–755. [49] Y. T. Wu, M. Yuan, S. Dong, L. Li, and Y. Liu, ‘‘Remaining useful life 998
924 [28] S. Chen, M. Wang, D. Huang, P. Wen, S. Wang, and S. Zhao, ‘‘Remain- estimation of engineered systems using vanilla LSTM neural networks,’’ 999
925 ing useful life prediction for complex systems with multiple indicators Neurocomputing, vol. 275, pp. 167–179, Jan. 2018. 1000
926 based on particle filter and parameter correlation,’’ IEEE Access, vol. 8, [50] K. Liu, N. Z. Gebraeel, and J. Shi, ‘‘A data-level fusion model for devel- 1001
927 pp. 215145–215156, 2020. oping composite health indices for degradation modeling and prognostic 1002
928 [29] J. B. Ali, B. Chebel-Morello, L. Saidi, S. Malinowski, and F. Fnaiech, analysis,’’ IEEE Trans. Autom. Sci. Eng., vol. 10, no. 3, pp. 652–664, 1003
929 ‘‘Accurate bearing remaining useful life prediction based on Weibull Jul. 2013. 1004
930 distribution and artificial neural network,’’ Mech. Syst. Signal Process., [51] A. Kanawaday and A. Sane, ‘‘Machine learning for predictive maintenance 1005
931 vols. 56–57, pp. 150–172, May 2015. of industrial machines using IoT sensor data,’’ in Proc. 8th IEEE Int. Conf. 1006
932 [30] K. Salahshoor, M. Mosallaei, and M. Bayat, ‘‘Centralized and decen- Softw. Eng. Service Sci. (ICSESS), Nov. 2017, pp. 87–90. 1007
933 tralized process and sensor fault monitoring using data fusion based on [52] C. W. Hong, K. Lee, M.-S. Ko, J.-K. Kim, K. Oh, and K. Hur, ‘‘Multivariate 1008
934 adaptive extended Kalman filter algorithm,’’ Measurement, vol. 41, no. 10, time series forecasting for remaining useful life of turbofan engine using 1009
935 pp. 1059–1076, 2008. deep-stacked neural network and correlation analysis,’’ in Proc. IEEE Int. 1010
936 [31] C. Ordón̄ez, F. S. Lasheras, J. Roca-Pardiñas, and F. J. de Cos Juez, Conf. Big Data Smart Comput. (BigComp), Feb. 2020, pp. 63–70. 1011
937 ‘‘A hybrid ARIMA–SVM model for the study of the remaining useful [53] Y. Huang, ‘‘Pearson correlation coefficient,’’ in Noise Reduction in Speech 1012
938 life of aircraft engines,’’ J. Comput. Appl. Math., vol. 346, pp. 184–191, Processing (Springer Topics in Signal Processing), vol. 2. Cham, Switzer- 1013
939 Jan. 2019. land: Springer, 2009. 1014
940 [32] N. Gebraeel, A. Elwany, and J. Pan, ‘‘Residual life predictions in the [54] M. Friendly, ‘‘Corrgrams: Exploratory displays for correlation matrices,’’ 1015
941 absence of prior degradation knowledge,’’ IEEE Trans. Rel., vol. 58, no. 1, Amer. Statistician, vol. 56, no. 4, pp. 316–324, Nov. 2002. 1016
942 pp. 106–117, Mar. 2009. [55] M. Ohki, M. E. Zervakis, and A. N. Venetsanopoulos, ‘‘3-D digital filters,’’ 1017
943 [33] R. Khelif, B. Chebel-Morello, S. Malinowski, E. Laajili, F. Fnaiech, in Multidimensional Systems: Signal Processing and Modeling Techniques 1018
944 and N. Zerhouni, ‘‘Direct remaining useful life estimation based on (Control and Dynamic Systems), vol. 69, C. Leondes, Ed. New York, NY, 1019
945 support vector regression,’’ IEEE Trans. Ind. Electron., vol. 64, no. 3, USA: Academic, 199, pp. 49–88. 1020
946 pp. 2276–2285, Mar. 2017. [56] H. M. Elattar, H. K. Elminir, and A. M. Riad, ‘‘Towards online data-driven 1021
947 [34] C. Zhang, P. Lim, A. K. Qin, and K. C. Tan, ‘‘Multiobjective deep belief
prognostics system,’’ Complex Intell. Syst., vol. 4, no. 4, pp. 271–282, 1022
948 networks ensemble for remaining useful life estimation in prognostics,’’
Dec. 2018. 1023
949 IEEE Trans. Neural Netw. Learn. Syst., vol. 28, no. 10, pp. 2306–2318,
[57] C.-G. Huang, H.-Z. Huang, W. Peng, and T. Huang, ‘‘Improved trajectory 1024
950 Oct. 2017.
similarity-based approach for turbofan engine prognostics,’’ J. Mech. Sci. 1025
951 [35] J. Zhang, P. Wang, R. Yan, and R. X. Gao, ‘‘Long short-term memory for
Technol., vol. 33, no. 10, pp. 4877–4890, Oct. 2019. 1026
952 machine remaining life prediction,’’ J. Manuf. Syst., vol. 48, pp. 78–86,
[58] P. Baldi and P. Sadowski, ‘‘Understanding dropout,’’ in Proc. Adv. Neural 1027
953 Jul. 2018.
Inf. Process. Syst., vol. 26, 2013, pp. 2814–2822. 1028
954 [36] A. Elsheikh, S. Yacout, and M.-S. Ouali, ‘‘Bidirectional handshaking
[59] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdi- 1029
955 LSTM for remaining useful life prediction,’’ Neurocomputing, vol. 323,
nov, ‘‘Dropout: A simple way to prevent neural networks from overfitting,’’ 1030
956 pp. 148–156, Jan. 2019.
957 [37] M. Saeidi, M. Soufian, A. Elkurdi, and S. Nefti-Meziani, ‘‘A jet engine J. Mach. Learn. Res., vol. 15, no. 1, pp. 1929–1958, Jan. 2014. 1031

958 prognostic and diagnostic system based on Bayesian classifier,’’ in Proc. [60] S. Lee, H. Yu, H. Yang, I. Song, J. Choi, J. Yang, G. Lim, K.-S. Kim, 1032

959 12th Int. Conf. Develop. eSyst. Eng. (DeSE), Oct. 2019, pp. 975–977. B. Choi, and J. Kwon, ‘‘A study on deep learning application of vibration 1033

960 [38] S. Zheng, K. Ristovski, A. Farahat, and C. Gupta, ‘‘Long short-term data and visualization of defects for predictive maintenance of gravity 1034

961 memory network for remaining useful life estimation,’’ in Proc. IEEE Int. acceleration equipment,’’ Appl. Sci., vol. 11, no. 4, p. 1564, 2021. 1035

962 Conf. Prognostics Health Manage. (ICPHM), Jun. 2017, pp. 88–95. [61] E. Ramasso and A. Saxena, ‘‘Review and analysis of algorithmic 1036

963 [39] R. Wu and J. Ma, ‘‘An improved LSTM neural network with uncertainty to approaches developed for prognostics on CMAPSS dataset,’’ in Proc. 1037

964 predict remaining useful life,’’ in Proc. CAA Symp. Fault Detection, Super- Annu. Conf. Prognostics Health Manage. Soc., 2014, pp. 1–12. 1038

965 vision Saf. Tech. Processes (SAFEPROCESS), Jul. 2019, pp. 274–279. [62] J. Wang, G. Wen, S. Yang, and Y. Liu, ‘‘Remaining useful life estimation 1039

966 [40] S. Wang, X. Zhang, D. Gao, B. Chen, Y. Cheng, Y. Yang, W. Yu, Z. Huang, in prognostics using deep bidirectional LSTM neural network,’’ in Proc. 1040

967 and J. Peng, ‘‘A remaining useful life prediction model based on hybrid Prognostics Syst. Health Manage. Conf. (PHM-Chongqing), Oct. 2018, 1041

968 long-short sequences for engines,’’ in Proc. 21st Int. Conf. Intell. Transp. pp. 1037–1042. 1042

969 Syst. (ITSC), Nov. 2018, pp. 1757–1762. [63] J. Li, X. Li, and D. He, ‘‘A directed acyclic graph network combined with 1043

970 [41] G. S. Babu, P. Zhao, and X.-L. Li, ‘‘Deep convolutional neural network CNN and LSTM for remaining useful life prediction,’’ IEEE Access, vol. 7, 1044

971 based regression approach for estimation of remaining useful life,’’ in Proc. pp. 75464–75475, 2019. 1045

972 Int. Conf. Database Syst. Adv. Appl. Cham, Switzerland: Springer, 2016, [64] Z. Kong, Y. Cui, Z. Xia, and H. Lv, ‘‘Convolution and long short-term 1046

973 pp. 214–228. memory hybrid deep neural networks for remaining useful life prognos- 1047

974 [42] X. Li, Q. Ding, and J.-Q. Sun, ‘‘Remaining useful life estimation in tics,’’ Appl. Sci., vol. 9, no. 19, p. 4156, Oct. 2019. 1048

975 prognostics using deep convolution neural networks,’’ Rel. Eng. Syst. Saf., [65] H. Liu, Z. Liu, W. Jia, and X. Lin, ‘‘Remaining useful life prediction using 1049

976 vol. 172, pp. 1–11, Apr. 2018. a novel feature-attention-based end-to-end approach,’’ IEEE Trans. Ind. 1050

977 [43] L. Jayasinghe, T. Samarasinghe, C. Yuenv, J. C. N. Low, and S. S. Ge, Informat., vol. 17, no. 2, pp. 1197–1207, Feb. 2021. 1051

978 ‘‘Temporal convolutional memory networks for remaining useful life esti- [66] C. Peng, Y. Chen, Q. Chen, Z. Tang, L. Li, and W. Gui, ‘‘A remaining 1052

979 mation of industrial machinery,’’ in Proc. IEEE Int. Conf. Ind. Technol. useful life prognosis of turbofan engine using temporal and spatial feature 1053

980 (ICIT), Feb. 2019, pp. 915–920. fusion,’’ Sensors, vol. 21, no. 2, p. 418, Jan. 2021. 1054

981 [44] C. W. Hong, C. Lee, K. Lee, M.-S. Ko, D. E. Kim, and K. Hur, ‘‘Remaining [67] U. Amin and K. D. Kumar, ‘‘Remaining useful life prediction of aircraft 1055

982 useful life prognosis for turbofan engine using explainable deep neural engines using hybrid model based on artificial intelligence techniques,’’ in 1056

983 networks with dimensionality reduction,’’ Sensors, vol. 20, no. 22, p. 6626, Proc. IEEE Int. Conf. Prognostics Health Manage. (ICPHM), Jun. 2021, 1057

984 Nov. 2020. pp. 1–10. 1058

VOLUME 10, 2022 95439

 O. Asif et al.: Deep Learning Model for RUL Prediction of Aircraft Turbofan Engine on C-MAPSS Dataset

1059 [68] K. Yu, D. Wang, and H. Li, ‘‘A prediction model for remaining useful JOHN F. W. ZAKI received the master’s degree 1111
1060 life of turbofan engines by fusing broad learning system and temporal from Cambridge University, the M.B.A. degree 1112
1061 convolutional network,’’ in Proc. 8th Int. Conf. Inf., Cybern., Comput. from Glasgow University, and the Ph.D. 1113
1062 Social Syst. (ICCSS), Dec. 2021, pp. 137–142. degree from Mansoura University. He worked in 1114
1063 [69] Y. Liu, X. Zhang, W. Guo, H. Bian, Y. He, and Z. Liu, ‘‘Prediction of both technical and managerial positions in various 1115
1064 remaining useful life of turbofan engine based on optimized model,’’ multinational companies. He is experienced in 1116
1065 in Proc. IEEE 20th Int. Conf. Trust, Secur. Privacy Comput. Commun. establishing tech start-ups, project management, 1117
1066 (TrustCom), 2021, pp. 1473–1477.
and business processes. He is currently the Vice 1118
Director of the Communication and Information 1119
Technology Centre and the Director of e-learning 1120
at Mansoura University. He is also an Assistant Professor at the Department 1121
of Computer and Systems, Mansoura University. His research interests 1122
1067 OWAIS ASIF was born in Wah Cantt, Pakistan, include artificial intelligence, software engineering, control systems, and 1123
1068 in 1994. He received the B.S. degree in electri- e-learning technology. 1124
1069 cal engineering from the Federal Urdu University
1070 of Arts, Science and Technology, Islamabad,
1071 Pakistan, in 2017, and the M.S. degree in elec-
1072 trical engineering from COMSATS University
1073 Islamabad, Wah Campus, in 2021. His research
1074 interests include prognostics and system health
1075 management, development of data-driven-based
1076 algorithms for contemporary research problems
1077 especially for multivariate time series as well as image data, and RUL KYUNG-SUP KWAK (Life Senior Member, 1125
1078 estimation for improving the efficiency of critical industrial machines. IEEE) received the Ph.D. degree from the Univer- 1126
sity of California. He was with Hughes Network 1127
Systems and the IBM Network Analysis Center, 1128
USA. He was a Professor with Inha University, 1129
South Korea. He was also the Dean of the Graduate 1130
1079 SAJJAD ALI HAIDER received the B.S. degree School of Information Technology and Telecom- 1131
1080 in from COMSATS University Islamabad (CUI), munications and the Director of the UWB Wireless 1132
1081 Islamabad, in 2005, the M.S. degree in embedded Communications Research Center. In 2008, he was 1133
1082 systems and control engineering from Leicester an Inha Fellow Professor (IFP). He is currently an 1134
1083 University, U.K., in 2007, and the Ph.D. degree Inha Hanlim Fellow Professor and a Professor with the School of Information 1135
1084 from Chongqing University, China, in 2014. He is and Communication Engineering, Inha University. His research interests 1136
1085 currently working as an Assistant Professor at the include UWB radio systems, wireless body area networks and u-health 1137
1086 Department of Electrical and Computer Engineer- networks, and nano and molecular communications. He received the Offi- 1138
1087 ing, CUI, Wah Campus, Pakistan. His research cial Commendations for achievements of UWB radio technology research 1139
1088 interests include embedded systems, control and development from the Korean President in 2009. In 2006, he was the 1140
1089 systems, and machine learning. President of the Korean Institute of Communication Sciences (KICS) and 1141
the Korea Institute of Intelligent Transport Systems (KITS) in 2009. 1142

1090 SYED RAMEEZ NAQVI received the M.Sc.

1091 degree in electronic engineering from The Uni-
1092 versity of Sheffield, U.K., in 2007, and the Ph.D.
1093 degree in computer engineering from the Vienna
1094 University of Technology, Austria, in 2013. He is
1095 currently an Assistant Professor with the Depart- S. M. RIAZUL ISLAM (Member, IEEE) was with 1143
1096 ment of Electrical and Computer Engineering, the University of Dhaka, Bangladesh, as an Assis- 1144
1097 COMSATS University Islamabad, Wah Campus. tant Professor, and a Lecturer with the Depart- 1145
1098 Since his Ph.D., he has published more than ment of Electrical and Electronic Engineering, 1146
1099 35 research articles in international journals and from 2005 to 2014. In 2014, he worked at 1147
1100 symposiums of high reputation, and carried several funded research projects the Department of Solution Laboratory, Samsung 1148
1101 with various research and development organizations. His experience as an Research and Development Institute Bangladesh 1149
1102 academic has led him to believe that there is a considerable gap between (SRBD), as a Chief Engineer. From 2014 to 2017, 1150
1103 academia and industry in Pakistan, especially in the application of contem- he was a Postdoctoral Fellow at the Wireless 1151
1104 porary methods, such as data analytics, in real-life problems. To address this Communications Research Center, Inha Univer- 1152
1105 disparity, he has co-founded Convex Solutions Pvt. Ltd., and established a sity, South Korea. From 2017 to 2022, he worked as an Assistant Professor 1153
1106 consortium of peers in academics, mainly responsible for developing cur- at the Department of Computer Science and Engineering, Sejong University, 1154
1107 riculum for training on artificial intelligence and business analytics for young South Korea. He is currently a Senior Lecturer in computer science at the 1155
1108 professionals and graduate students having diverse academic backgrounds. University of Huddersfield, U.K. His research interests include the Internet 1156
1109 His research interests include asynchronous logic, computer architecture, of Things (IoT), applied artificial intelligence (AI), and 5G/6G. 1157
1110 hardware optimization, and artificial intelligence. 1158

95440 VOLUME 10, 2022