JSR-01336; No of Pages 9

Journal of Safety Research xxx (2016) xxx–xxx

Contents lists available at ScienceDirect

Journal of Safety Research

journal homepage: www.elsevier.com/locate/jsr

1Q1 Off-road truck-related accidents in US mines

2Q2 S.R. Dindarloo, a,⁎ J. Pollard, b E. Siami-Irdemoos c

F
3 a
Department of Mining and Nuclear Engineering, Missouri University of Science and Technology, MO, USA
4 b
Workplace Health Branch, Pittsburgh Mining Research Division, NIOSH, PA, USA

O
5 c
Department of Geoscience and Geological and Petroleum Engineering, Missouri University of Science and Technology, MO, USA
6

O
7 a r t i c l e i n f o a b s t r a c t

8 18

R
Article history: Introduction: Off-road trucks are one of the major sources of equipment-related accidents in the U.S. mining in-
9 Received 17 January 2016 dustries. A systematic analysis of all off-road truck-related accidents, injuries, and illnesses, which are reported 19
10 Received in revised form 2 May 2016 and published by the Mine Safety and Health Administration (MSHA), is expected to provide practical insights 20
11

P
Accepted 12 July 2016 for identifying the accident patterns and trends in the available raw database. Therefore, appropriate safety man- 21
12 Available online xxxx
agement measures can be administered and implemented based on these accident patterns/trends. Methods: A 22
17
16
15
14
13
hybrid clustering-classification methodology using K-means clustering and gene expression programming 23
38 Keywords:
39 Off-road mining trucks
(GEP) is proposed for the analysis of severe and non-severe off-road truck-related injuries at U.S. mines. Using
D 24
40 Fatalities and injuries the GEP sub-model, a small subset of the 36 recorded attributes was found to be correlated to the severity 25
41 MSHA level. Results: Given the set of specified attributes, the clustering sub-model was able to cluster the accident re- 26
42 cords into 5 distinct groups. For instance, the first cluster contained accidents related to minerals processing 27
E
K-means clustering
43 Genetic programming mills and coal preparation plants (91%). More than two-thirds of the victims in this cluster had less than 28
44
Q3 Classification 5 years of job experience. This cluster was associated with the highest percentage of severe injuries (22 severe 29
T

accidents, 3.4%). Almost 50% of all accidents in this cluster occurred at stone operations. Similarly, the other 30
four clusters were characterized to highlight important patterns that can be used to determine areas of focus 31
32
C

for safety initiatives. Conclusions: The identified clusters of accidents may play a vital role in the prevention of se-
vere injuries in mining. Further research into the cluster attributes and identified patterns will be necessary to 33
determine how these factors can be mitigated to reduce the risk of severe injuries. Practical application: Analyzing 34
E

injury data using data mining techniques provides some insight into attributes that are associated with high ac- 35
curacies for predicting injury severity. 36
© 2016 National Safety Council and Elsevier Ltd. All rights reserved. 37
R

48
46
45
R

47
49 1. Introduction U.S. mining facilities. Reportable illnesses include any illness or disease 65
that may have resulted from work. The database of these reports is 66
O

50 Analysis of workplace injuries has been heavily utilized as a means available in the public domain and is provided by the National Institute 67
51 to determine high-risk tasks, prioritize workplace redesign, and deter- for Occupational Safety and Health (http://www.cdc.gov/niosh/mining/ 68
C

52 mine areas of concern for worker safety in many industries including data/default.html). Each entry of the database contains 36 unique attri- 69
53 healthcare, construction, retail and services, and mining (Cato, Olson, butes including: mine id, mining method, accident date, degree of inju- 70
54 & Studer, 1989; Drury, Porter, & Dempsey, 2012; Mardis & Pratt, 2003; ry, accident classification, mining equipment, employee's experience 71
N

55 Moore, Porter, & Dempsey, 2009; Pollard, Heberger, & Dempsey, 2014; and activity, and a narrative briefly explaining the accident. Previous 72
56 Schoenfisch, Lipscomb, Shishlov, & Myers, 2010; Turin, Wiehagen, mining research has examined the injury and fatality causes associated 73
U

57 Jaspal, & Mayton, 2001; Wiehagen, Mayton, Jaspal, & Turin, 2001). with maintenance and repair, haulage vehicles, ingress and egress from 74
58 While many industries would require injury records from individual mobile equipment, operating underground and surface mining mobile 75
59 companies or insurance providers to perform an analysis, mining is equipment, and other mining tasks (Drury et al., 2012; Moore et al., 76
60 uniquely suited for a more comprehensive injury analysis. An important 2009; Pollard et al., 2014; Reardon, Heberger, & Dempsey, 2014; Turin 77
61 feature of U.S. mining is the accessibility of injury records. The Mine et al., 2001; Wiehagen et al., 2001). Traditional injury data analysis 78
62 Safety and Health Administration requires all mine operators and con- uses counts and cross-tabulations as a means to determine trends in in- 79
63 tractors to file a Mine Accident, Injury and Illness Report (MSHA Form juries. While this typically yields useful information, more sophisticated 80
64 7000-1) for all reportable accidents, injuries, and illnesses incurred at data mining techniques may allow for more improved classification of 81
injuries through identification of injury patterns. 82
⁎ Corresponding author at: 226 McNutt Hall, 1400 N. Bishop Ave., Rolla, MO 65409, USA. Clustering and classification are the two widely used methods of 83
E-mail address: srd5zb@mst.edy (S.R. Dindarloo). data mining for the purpose of pattern recognition. Clustering is 84

http://dx.doi.org/10.1016/j.jsr.2016.07.002
0022-4375/© 2016 National Safety Council and Elsevier Ltd. All rights reserved.

Please cite this article as: Dindarloo, S.R., et al., Off-road truck-related accidents in US mines, Journal of Safety Research (2016), http://dx.doi.org/
10.1016/j.jsr.2016.07.002
 2 S.R. Dindarloo et al. / Journal of Safety Research xxx (2016) xxx–xxx

severe and non-severe) were filtered for further analysis. This dataset in- 131
cluded 125 severe records that affected 140 employees. These severe in- 132
juries consisted of 88 fatalities and 52 permanent disabilities. These 133
records were analyzed using Minitab (Minitab Inc., State College, Pennsyl- 134
vania), MATLAB (MathWorks, Inc., Natick, Massachusetts), Rapidminer 135
(RapidMiner, Inc., Cambridge, Massachusetts) and GenExprotools 136
(Gepsoft Limited, Bristol, UK) to determine factors associated with the 137
highest counts of severe injuries. 138

2.2. GEP-clustering modeling 139

The objective of clustering is to discriminate between dissimilar data 140

by dataset partitioning (clustering). As an unsupervised data mining 141
technique, the aim of clustering is to split a heterogeneous dataset 142

F
Fig. 1. GP tree representation of Coshx/(C1sinx.). into several more homogenous groups. The optimization task is to max- 143
imize the similarity between the in-cluster members and dissimilarity 144

O
between the out-cluster members. K-means clustering is used to parti- 145
85 among the unsupervised methods of pattern recognition while the clas- tion a large, highly variable dataset such that like data are grouped to- 146
86 sification is a supervised learning method. By an unsupervised method, gether. As an example, one is given a set of n data points in d- 147

O
87 one means that the data analyzer does not have any prior hypothesis or dimensional space (Rd) and an integer k. The goal is to determine a set 148
88 pre-specified models for the data, but wants to understand the general of k points in Rd, called centers, so as to minimize the mean squared dis- 149
89 characteristics or the structure of the high-dimensional data. A super- tance from each data point to its nearest center (Kanungo et al., 2002). 150

R
90 vised method means that the investigator wants to confirm the validity Let X = {xi} , i = 1 , … , n,be the d-dimensional observations which are 151
91 of a hypothesis/model or a set of assumptions, given the available data clustered into a set of k clusters, C = {ck, K = 1, … , K}. The K-means clus- 152

P
92 (Jain, 2010). Clustering and classification are also called un-labeled tering finds a partition such that the squared error between the empir- 153
93 and labeled, respectively. In pattern recognition, data analysis is con- ical mean of a cluster and the points in the cluster is minimized. Let μk be 154
94 cerned with predictive modeling: given some training data, the predic- the mean of the cluster ck. The squared error between μk and the points
D 155
95 tion task is to find the behavior of the unseen test data. This task is also in cluster ck is defined as shown in Eq. (1). The goal of K-means cluster- 156
96 referred to as learning. Often, a clear distinction is made between learn- ing is to minimize the sum of the squared error over all k clusters as 157
97 ing problems that are (i) supervised (classification) or (ii) unsupervised shown in Eq. (2). (For a detailed review of the theory and background 158
E
98 (clustering), the first involving only labeled data (training patterns with of K-means clustering see: Halkidi et al. 2001, Fraley and Raftery 2002, 159
Q5
99 known category labels), while the latter involves only unlabeled data Jain, 2010.) 160
T

100 (Duda, Hart, & Stork, 2001; Jain, 2010). Clustering and classifications
X 
101 are performed using differing algorithms but may be used together to J ðck Þ ¼ xi −μ 2 ð1Þ
k
C

102 improve prediction accuracy. xi ∈ck

162
103 The aim of this research was to gain a better understanding of the
163
104 factors associating with severe injuries (fatalities and permanent dis- K X 
X 
E

105 abilities) in U.S. mining by employing data mining techniques. Cluster- J ðcÞ ¼ xi −μ 2 ð2Þ
k
106 ing and classification were employed for a comprehensive analysis of k¼1 xi ∈ck
R

107 off-road truck-related accidents and injuries reported to MSHA during 165
108 a 13-year period (2000–2012). Gene expression programming was Genetic programming (GP) creates a functional relationship be-
109 used for classification, allowing all injury attributes to be considered tween inputs (attributes) and outputs to predict the occurrence of the 166
R

110 and tested to determine which were associated with the highest predic- output based on the properties of the attributes. Genetic programming 167
111 tion accuracies. The most explanatory attributes were selected among can be represented as a hierarchically structured tree comprising func- 168
O

112 the available 36 unique attributes in the MSHA database. Then, K- tions and terminals. Fig. 1 illustrates a simple representation of a GP 169
113 means clustering was used as a means to identify similarity/dissimilar- tree for the function Coshx/(C1sinx) . The tree reads from left to right 170
114 ity between the accident records using the selected attributes for the and from bottom to top. Mimicking the Darwinian principle of survival, 171
C

115 purposes of pattern recognition in the raw data. It should be noted the fittest solutions (smallest error) are chosen to generate a population 172
116 that the goal of this study was not to establish cause–effect relationships of new offspring programs for the next generation (Koza, 1992). In the 173
N

117 between accident attributes and outcomes, but to: (a) use data mining
118 to systematically identify important attributes from MSHA incident re- 25
No. of Fatalities and Disabilities

119 ports that are highly associated with the outcomes of accidents (classi-
U

120 fication), and (b) recognize patterns in the accidents (clustering) given 20
121 a set of work-related attributes.
15
122 2. Materials and methods

10
123 2.1. MSHA injury data

124 A dataset comprised of 13 years (2000–2012) of Mine Accident, In- 5

125 jury and Illness Reports was selected beginning with 1/1/2000
126
Q4 (NIOSH, 2014). From this dataset, records of severe injuries (fatalities 0
127 and permanent disabilities) and non-severe injuries associated with 2000 2003 2006 2009 2012
128 off-road trucks were selected. The NIOSH code “minemach-44, all acci- Calendar Year
129 dents related to off-road mining trucks” was identified to select the re-
130 cords of interest in this study. A total of 5,831 records of injuries (both Fig. 2. Time series of off-road truck-related severe injuries at US mines (2000–2012).

Please cite this article as: Dindarloo, S.R., et al., Off-road truck-related accidents in US mines, Journal of Safety Research (2016), http://dx.doi.org/
10.1016/j.jsr.2016.07.002
 S.R. Dindarloo et al. / Journal of Safety Research xxx (2016) xxx–xxx 3

2 217 86

Occupational Fatality
1275 Non Fatal, days lost
No days lost
Non Occupational Fatality
Not specified

F
O
4251

O
Fig. 3. Distribution of off-road truck-related accident types at US mines (2000–2012). Only, two Non-Occupational Fatalities occurred.

R
174 next step, some genetic operations, namely mutation and crossover, will shapes (parse trees). In GEP the individuals are also nonlinear enti- 194

P
175 generate new offspring from the fittest programs of the previous gener- ties of different sizes and shapes (expression trees), but these 195
176 ation. The operator selects a random node in a tree and replaces it with complex entities are encoded as simple strings of fixed length (chro- 196
177 another node or subtree. The new offspring will be evaluated with the mosomes). The chromosomes in this study are the attributes of each
D 197
178 error or fitness function. The process continues until reaching a record of injuries (60 attributes and 5831 records). Unlike the parse 198
179 predefined threshold in terms of the best fit or error. In GP, thousands tree representation in conventional GP, GEP uses a fixed length of 199
180 of solutions (computer programs) are generated and evolved consecu- character strings to represent solutions to the problems, which are 200
E
181 tively based on the Darwinian principle of survival. The search for the afterwards expressed as parse trees of different sizes and shapes. 201
182 solution starts with a population of completely randomly generated These trees are called GEP expression trees (ETs). One advantage of 202
T

183 programs (solutions) from a predefined set of available functions the GEP technique is that the creation of genetic diversity is extreme- 203
184 (e.g., arithmetic functions) and terminals (independent variables). All ly simplified as genetic operators work at the chromosomal level. 204
C

185 programs are measured against a fitness function (e.g., root mean Another GEP strength is its unique, multi-genic nature that allows 205
186 square error in a regression problem) and the best ones survive and the evolution of more complex programs composed of several sub- 206
187 are bred to the next generation. programs. The GEP algorithm begins with an initial population of 207
E

188 Proposed by Ferreira (2001), gene expression programming chromosomes, which are randomly generated, linear strings of a 208
189 (GEP) is a development of the conventional GP. As with GP, in GEP fixed length. Then, the linear chromosomes are expressed as ETs 209
R

190 the main steps include: the function set, terminal set, fitness func- and the fitness of each individual is evaluated based on a predefined 210
191 tion, control parameters, and stop criteria. The fundamental differ- fitness function. The individuals are then selected, according to fit- 211
192 ence between GP and GEP resides in the nature of the individuals. ness, to form a new generation—i.e., the higher the fitness value, 212
R

193 In GP the individuals are nonlinear entities of different sizes and the more chance an individual has to be selected. The selected 213
O

5 11
Underground Operations
10
C

25
Surface at Underground
N

Surface Mines
Mill or Preparation Plant
U

Other

89
Fig. 4. Distribution of mine operation type for severe injuries related to off-road trucks at US mines (2000–2012).

Please cite this article as: Dindarloo, S.R., et al., Off-road truck-related accidents in US mines, Journal of Safety Research (2016), http://dx.doi.org/
10.1016/j.jsr.2016.07.002
 4 S.R. Dindarloo et al. / Journal of Safety Research xxx (2016) xxx–xxx

a) 11
32 > 20
10 - 20
5 - 10
<5
Not Specified

14
63

F
20

O
O
b)

R
3500 70

3000 60

P Number of Severe Injuries

Number of other Injuries

2500 50

2000
D 40
E
1500 30 Other Injuries
Severe Injuries
T

1000 20
C

500 10

0 0
E

<5 5 - 10 10 - 20 > 20
Job Experience
R

Fig. 5. a Distribution of job experience (years) for severe injuries related to off-road trucks at US mines (2000–2012). b Distribution of job experience (years) for severe injuries vs. other
R

injuries, related to off-road trucks at US mines (2000–2012).

O

214 individuals are also subjected to reproduction with modification, to the same developmental process: expression of the genomes, se- 217
215 through genetic operators like crossover, mutation, and rotation. lection, and reproduction. The process is repeated for a certain num- 218
216 The individuals of this new generation are, in their turn, subjected ber of generations or until a solution has been found. 219
C

4
22
N

17
Getting on/off truck
Handling supplies/materials
U

Idle (Lunch, Break, etc.)

13
Machine maintenance/repair
Operating truck
Other

27
57

Fig. 6. Distribution of worker activity at the time of off-road truck-related severe injury in US mines (2000–2012).

Please cite this article as: Dindarloo, S.R., et al., Off-road truck-related accidents in US mines, Journal of Safety Research (2016), http://dx.doi.org/
10.1016/j.jsr.2016.07.002
 S.R. Dindarloo et al. / Journal of Safety Research xxx (2016) xxx–xxx 5

220 The GEP algorithm was used for classification of the off-road truck- the highest fitness or accuracy. Hence, all available independent vari- 226
221 related injury dataset. The objective was to identify the most explanato- ables included in the incident reports, including both the categorical 227
222 ry (independent) variables (attributes) to be used in the subsequent and numerical variables, were initially used in the GEP model building 228
223 clustering sub-model. One advantage of GEP over conventional classifi- as explanatory variables. The injury severity was converted into a binary 229
224 cation algorithms (such as support vector machines and artificial neural variable with: class 0 for those severe truck-related incidents that re- 230
225 networks) is that it can select the best explanatory features to achieve sulted in either fatalities or permanent disabilities, and class 1 for all 231

F
O
O
R
P
D
E
T
C
E
R
R
O
C
N
U

Fig. 7. The three GEP expression trees (Sub-ET 1, Sub-ET 2, and Sub-ET 3) for classification of the injury severity.

Please cite this article as: Dindarloo, S.R., et al., Off-road truck-related accidents in US mines, Journal of Safety Research (2016), http://dx.doi.org/
10.1016/j.jsr.2016.07.002
 6 S.R. Dindarloo et al. / Journal of Safety Research xxx (2016) xxx–xxx

232 other injuries. Two-thirds of the records (3,888) were used for training metal, non-metal, stone, sand & gravel operators, as well as, coal and 293
233 the algorithm and the remaining (1,943) for model testing. The fitness non-coal contractors). The overall classification accuracies of 64.55% 294
234 function was set as the sensitivity/specificity to achieve maximum clas- and 65.65% were obtained for the training and testing datasets, respec- 295
235 sification accuracy. K-means clustering was used to group similar inju- tively. The classification accuracies for the severe injury class and the all 296
236 ries therefore maximizing the similarity of records in each cluster. All other injuries class in the testing dataset were 63.04% and 64.69%, re- 297
237 data (both severe and non-severe) were analyzed together and separat- spectively (see Table 2). 298
238 ed into clusters. In order to examine the accuracy of the GEP classification algorithm, 299
the whole database was modeled with the widely-used traditional 300
239 3. Results method of logistic regression for binary classification (Cox, 1958). In a 301
binary logistic regression, the dependent variable is a two-class categor- 302
240 3.1. MSHA injury data overview ical variable. Here, the two classes were severe and non-severe acci- 303
dents. The objective of the regression was to classify the dataset using 304
241 A brief analysis of the MSHA injury data was performed to give an the available attributes. Similar to GEP, the logistic regression builds a 305
242 understanding of the types of data available in the injury records. A tem- relationship between the (categorical) dependent variable and (a com- 306

F
243 poral illustration of off-road truck-related fatalities and disabilities (Fig. bination of categorical, ordinal, and numerical) independent variables 307
244 2) shows fluctuations in the fatality and disability counts over the years (attributes). A logistic regression model assigns to each set of attributes 308

O
245 with a clear decreasing trend. The distribution of accident types is a number between 0 and 1. This output can be interpreted as the prob- 309
246 shown in Fig. 3. An analysis of the type of mine operation found that ability of a set of attributes belonging to each class. For instance, an out- 310
247 most severe injuries occurred at surface mines that use strips, quarries, put value equal to 0.3 means that the accident (given its attributes) is a 311

O
248 or open pits as shown in Fig. 4. Employee job experience was also exam- non-severe one (class 0) with a probability of 70% and a severe accident 312
249 ined as shown in Fig. 5a. Frequencies of severe versus other injuries (class 1) with a probability of 30%. Using the same training, testing, and 313
250 among victims of different job experiences are illustrated in Fig. 5b. attribute sets as the GEP, the results of logistic regressions were obtain- 314

R
251 Nearly half of the affected employees had less than 5 years of job expe- ed for the whole dataset and are compared with the GEP results in 315
252 rience. Work activity at the time of injury was also examined as shown Table 2. GEP had a superior performance compared to logistic regression 316

P
253 in Fig. 6. resulting in a more accurate classification of both severe and non-severe 317
254 Most of the severe injuries were sustained while the employee was accidents. 318
255 operating the off-road truck. In the 56 severe accidents (57 victims) D
256 that occurred while operating a truck, 46 employees sustained fatal in- 3.3. Results of K-means clustering 319
257 juries and 11 employees were permanently disabled. The four main
258 causes of these accidents (as identified in the report of the MSHA inves- Using the GEP sub-model's output, the resulting clusters from the K- 320
E
259 tigation) were: (i) losing control of the truck, (ii) berm/dump failure, means clustering were defined based on their most dominant attribute, 321
260 (iii) unsafe/careless actions, and (iv) truck/component mechanical fail- which discriminated them from the other clusters while maintaining 322
T

261 ure. In the category of “unsafe/careless actions,” MSHA identified that the highest similarity inside the cluster. Therefore, the clusters were de- 323
262 the underlying root causes were either associated with poor safety fined based on their attributes as specified in Table 3a. The optimal 324
C

263 training, management and enforcement with managerial causes, or fail- number of clusters was identified using the GAP Statistics method 325
264 ure of the employees to obey safety regulations (personal issues). In (Tibshirani, Walther, & Hastie, 2001). Five clusters were created as de- 326
265 nearly all of the 26 accidents (affecting 27 victims) that occurred during fined in Table 4. The number of incidents in each cluster and the average 327
E

266 the “maintenance/repair” activity, either the employees' failure to re- job experience of employees at the time of the accident are provided in 328
267 gard safety regulations or a component/equipment malfunction/failure Tables 3a, 3b, 3c, and 3d along with the dominant subunit and activity, 329
R

268 caused the injuries. A major cause of fatalities in this category was
269 failure to block off/lockout the truck/bed. This was either due to the Table 1 t1:1
270 employees' unsafe actions or the mechanical malfunction of the equip-
R

GEP tree key. t1:2

271 ment. In the “other” category, 17 fatal accidents resulted in 15 fatalities t1:3

272 and 7 permanent disabilities. The main causes of this category were: Symbol Definition
O

273 (i) the operator(s) exited the cab while operating the equipment and
274 (ii) unsafe/careless minor repairing attempts.
C

LT2A if x<y, then x, else y

275 3.2. Results of gene expression programming
3Rt x1/3
LT2G if x<y, then (x+y), else arctan(x*y)
N

276 Gene expression programming utilized all available variables in the

Functions

277 incident records to determine those that best predicted injury severity. GOE2E if x>=y, then (x+y), else (x*y)
278 The important GEP parameters (number of chromosomes, head size, OR1 if x<0 OR y<0, then 1, else 0
U

279 and number of genes) were determined through a grid search algorithm OR2 if x>0 OR y>0, then 1, else 0
280 to achieve the maximum classification accuracy. A grid search is an ex- Avg2 avg(x,y)
281 haustive search in the pre-specified domains of the parameters of inter-
282 ests. In this case, using the grid search, the best results (highest
283 accuracies) were achieved when the three parameters were set to 34, Exp Exponential
284 11, and 5, respectively. The resulting best model is shown in the GEP ex- Atan Arctan
285 pression tree in Fig. 7 with the key in Table 1. The two classes are “se- X2 X2
286 vere” and “other injuries.” Results of this programming elicited five
d0 Subunit
287 attributes that best explained the injury severity. These attributes
Explanatory
Variables

288 were: the operation subunit (NIOSH code “subunit”), the month of the d1 Month of the year
289 year in which the accident occurred (NIOSH code “month”), the victim's d2 Activity
290 job experience at the time of the accident (NIOSH code “expjob”), the d3 Job Experience
291 employee's activity at the time of the accident (NIOSH code “mwactiv”),
d4 Operation
292 and the type of operation (NIOSH code “commod,” including: coal, t1:5

Please cite this article as: Dindarloo, S.R., et al., Off-road truck-related accidents in US mines, Journal of Safety Research (2016), http://dx.doi.org/
10.1016/j.jsr.2016.07.002
 S.R. Dindarloo et al. / Journal of Safety Research xxx (2016) xxx–xxx 7

t2:1 Table 2 Table 3b t4:1

t2:2 GEP classification accuracies. Activity code definitions. t4:2

t2:3 Injury classification Accuracy (%) Code Description t4:3

t2:4 GEP Logistic regression 1 Getting on/off truck t4:4

2 Handling supplies/materials t4:5
t2:5 Training Severe 74.68 61.24
3 Idle (Lunch, break, etc.) t4:6
t2:6 Other 64.33 53.62
4 Machine maintenance/repair t4:7
t2:7 Testing Severe 63.04 49.59
5 Operating truck t4:8
t2:8 Other 64.69 51.13
6 Other t4:9

351 operation, and month codes for each cluster. The activity, operation mechanical failure (sudden and inadequate performance), and leaving 395
352 type, and subunit codes in Table 3a are defined in Tables 3b–3d. the driving track. Under non-driving, the factors included: unexpected 396
353 The last column in Table 3a shows the number (percentage) of movement (of the vehicle or part of the vehicle or vehicle's load), falls 397
354 severe accidents (Codes 1 and 2) inside each of the five proposed from vehicle, and hit by other vehicle. Comparison of the results of 398

F
355 clusters. In summary, Tables 3a, 3b, 3c, and 3d show the most dis- these previous studies with the current study reveals that the root 399
356 criminating attribute(s) for each cluster. For instance, in cluster 0, causes of truck-related fatalities have not changed in the past two de- 400

O
357 the operation code 4 (stone operator) is dominant (48%). This cades. A new method to determine factors associated with these injuries 401
358 means almost half of the severe injuries were occurred in the is needed. 402
359 “stone operator” classification. Over two-thirds (72%) of the victims Gene expression programming was found to successfully predict se- 403

O
360 had less than 5 years of “job experience.” Also, the dominant subunit vere accidents in mining based on available injury data. The accuracies 404
361 code was 9 (i.e., mill or preparation plants, see Table 3d), which in- obtained using GEP for both the training and testing data are compara- 405

R
362 cludes over 90% of this cluster's accidents. Similarly, the major activ- ble or superior to those reported in car and traffic accident studies 406
363 ity codes, in cluster 0, were 1 and 5 (see Table 3b) which implied that (Abdelwahab & Abdel-Aty, 2001; De Oña, López, Mujalli, & Calvo, 407
364 the two activities of “getting on/off truck” and “driving off-road 2013; De Oña, Mujalli, & Calvo, 2011; Mujalli & De Oña, 2011). There- 408

P
365 truck” were associated with 72% of the accidents during the period fore, the GEP model's inputs (variables) had the highest explanatory 409
366 2000–2012. In this cluster, a total of 22 fatalities and permanent dis- characteristics, among 36 attributes, for severity estimation. The break- 410
367 abilities have been recorded that account for 3.4% of the total report- down of injuries within the clusters was quite interesting.
D 411
368 ed accidents (including non-severe ones). Minerals processing mills and coal preparation plants essentially 412
formed their own cluster (cluster 0). Therefore, there appears to be a 413
369 4. Discussion and conclusions unique element to these plants that created different truck-related inju- 414
E
ry scenarios from the other subunits. Typically, this subunit includes 415
370 Truck-related fatalities have been previously examined in the litera- mills, coal preparation plants, breaker operations, shops, and yards asso- 416
T

371 ture. An analysis of fatal truck-related accidents during the period ciated with one specific mine. These are non-production locations and 417
372 1995–2006 revealed the three most frequent causes of the haul truck- likely utilize differing types of equipment and have different geogra- 418
C

373 related fatalities as: (i) failure of victims to respect haul truck working phies and layouts from the pits or underground mining locations. Also 419
374 area, (ii) failure to provide adequate berms, and (iii) failure of mechan- of the 653 incidents in this cluster, in the past 13 years about 31% of 420
375 ical components (Md-Nor, Kecojevic, Komijenivic, & Groves, 2008). An- all accidents (203 instances) happened during June–August. The main 421
E

376 other study within the period of 1995–2002 (Kecojevic & Radomsky, reason(s) behind the high accident rates during these months should 422
377 2004) categorized the major causes of fatalities as follows: (i) failure be examined in a case by case manner for all of the mining sites in this 423
R

378 of mechanical components (22%); (ii) lack of and/or failure to obey cluster. Therefore the root cause(s) for the higher accident rates cannot 424
379 warning signals (20%); (iii) failure to maintain adequate berm (13%); be identified for this particular cluster of mining sites. Furthermore, per 425
380 (iv) inadequate hazard training (10%); and (v) failure to recognize Table 3a over 70% of the victims had job experiences in the range of 0– 426
R

381 adverse geological conditions (10%). Kecojevic and Radomsky (2004) 5 years. Therefore, the inexperienced crews in all of the mining sites in 427
382 associated 70% of all fatalities to the above five categories, which are cluster 0 experienced more accidents than other four clusters. The dis- 428
O

383 consistent with the results of studies for periods 1995–2006 and tribution of these inexperienced workers within this cluster as com- 429
384 the current study. Ruff, Coleman, and Martini (2011) studied the pared to the other clusters is not known so determining the relative 430
385 equipment-related fatalities for the period 2000–2007 and found that, rates of accidents was not possible. However, the high number of acci- 431
C

386 for mobile equipment, the most frequent fatalities were related to loss dents in the inexperienced worker can justify implementing further 432
387 of control or visibility issues during the operation of the equipment. A training and safety measures. Additionally, over 70% of all off-road 433
N

388 more recent analysis examined 133 fatality reports for the period truck-related accidents happened when the truck operators were either 434
389 1995–2010 using a previously developed coding scheme to determine getting on/off the trucks or when they were operating the trucks. Previ- 435
390 repeating patterns of accidents (Drury et al., 2012). In this work, the au- ous research has identified the safety issues associated with getting on 436
U

391 thors were able to more fully develop the classification patterns previ- and off of mobile equipment (Moore et al., 2009). Thus future safety im- 437
392 ously reported by Wenner and Drury (2000). This scheme was broken provement plans for this cluster of mining operations should focus on 438
393 into driving and non-driving accidents. Under driving, the factors in- the root causes of ingress and egress injuries as well as those injuries 439
394 cluded: loss of control, failure of ground, two-vehicle collisions, sustained during operation of these trucks. 440

t3:1 Table 3a
t3:2 Cluster specifications.

t3:3 Cluster Number of incidents Month of year Operation Code Average experience (years) Subunit code Activity code Number of Severe Injuries

t3:4 0 653 6, 7, 8 (31%) 4 (48%) 0–5 (72%) 9 (91%) 1, 5 (72%) 22 (3.4%)

t3:5 1 371 6, 7, 8 (34%) 1, 4 (54%) 0–5 (51%) 3 (78%) 4 (55%) 0 (0%)
t3:6 2 2,352 1–6 (0%) 1, 4 (55%) 0–5 (51%) 3 (78%) 1, 5 (79%) 58 (2.5%)
t3:7 3 187 1, 4, 10 (33%) 1, 4 (59%) 0–5 (49%) 3 (68%) 6 (62%) 0 (0%)
t3:8 4 2,191 7–12 (0%) 1, 4 (45%) 0–5 (57%) 3 (89%) 5 (52%) 45 (2.1%)
t3:9 Total 5,754 125 (2.17%)

Please cite this article as: Dindarloo, S.R., et al., Off-road truck-related accidents in US mines, Journal of Safety Research (2016), http://dx.doi.org/
10.1016/j.jsr.2016.07.002
 8 S.R. Dindarloo et al. / Journal of Safety Research xxx (2016) xxx–xxx

t5:1 Table 3c Table 4 t7:1

t5:2 Operation code definitions. Cluster definitions. t7:2

t5:3 Code Description Cluster Label t7:3

t5:4 1 Coal operator 0 Accidents related to minerals processing mills and coal preparation plants t7:4
t5:5 2 Metal operator (91%). More than two-thirds of the victims in this cluster had job
t5:6 3 Non-metal operator experiences of less than 5 years. This cluster is associated with the highest
t5:7 4 Stone operator percentage of severe injuries (22 severe accidents, 3.4%). Almost 50% of
t5:8 5 Sand and gravel operator all accidents in this cluster occurred at stone operations.
t5:9 6 Coal contractor 1 In 55% of the cases in this cluster, the victim had been inspecting the truck t7:5
t5:10 7 Non-coal contractor for maintenance/repair and using non-powered hand tools. Most of these
accidents occurred at coal or stone operations. This cluster was not
associated with any severe accidents.
2 No accidents occurred during the first half (January through June) of the t7:6
year. Half of these victims were injured while getting on or off equipment,
441 The interesting pattern in Cluster 1 is that in 55% of all cases, the vic- machines, etc. The average job experience for this cluster is less than 5
442 tim had been inspecting the truck for maintenance/repair with non- years. Also, the percentage of fatalities in this cluster is considerable. The

F
443 powered hand tools. Moreover, over half of the accidents were related cluster contains the highest absolute number of severe injuries, 58
444 to either the coal or stone operators. Although this cluster did not in- victims.

O
3 In this cluster, 62% of all incidents occurred when the victim was either t7:7
445 clude any severe injuries, it is clear that the use of hand tools for vehicle
running or walking. No severe injuries were recorded in this cluster.
446 inspection creates a hazard at mines. Also similar to cluster 0, in this 4 The second highest number of severe accidents occurred in this cluster t7:8
447 cluster the highest numbers of accident were recorded in June–August (45 accidents). Over half of the accidents occurred when equipment

O
448 period. Unlike cluster 0, in cluster 1 the share of the inexperienced em- operators with less than 5 years of experience (57%) were driving dump
449 ployees for all of the accidents is not very high with respect to more ex- trucks (activity code 5).

R
450 perienced employees. Also, a majority of the accidents (78%) took place
451 in either strip or open pit mines.
452 With 2,352 accidents, cluster 2 was the largest cluster in this study. to have the highest number of severe accidents. More in-depth analysis 480

P
453 Interestingly all of the accidents in the past 13 years in this cluster hap- would need to be conducted to determine the key differences between 481
454 pened in the second half of the year. Although it is not within the scope the accidents in clusters 2 and cluster 4 to determine what appears to be 482
455 of this study to analyze the root causes or contributing factors for this a seasonal division in accidents contributing to increased accident se-
D 483
456 unique pattern, it provides a well-justified guideline for conducting fu- verity. More training may be needed to address the injuries in this 484
457 ture, more in-depth analyses to determine activities or tasks associated cluster. 485
458 with high numbers of accidents. Also, similar to cluster 1, in the past This study was limited to data associated with off-road trucks in 486
E
459 13 years most of the accidents in this cluster happened in surface mining. Thus all other injuries that were not associated with off-road 487
460 (strip or open pit) mines. In summary, cluster 2 showed that ingress trucks were not included in this study. It is unclear whether this hybrid 488
T

461 and egress from off-road trucks is still a problem and requires immedi- methodology will have sufficient accuracies when applied to other 489
462 ate attention. types of injuries. Furthermore, the study analyzed only the accident re- 490
C

463 With 187 accidents, cluster 3 was the smallest cluster in this study. ports within the period January 2000 to December 2012. Data included 491
464 Similar to cluster 1, no severe accidents had been reported since 2000. in this analysis were limited to that provided in the MSHA incident re- 492
465 Over 60% of the accidents were associated with a non-specified activity ports which have been collected, organized, and released to the public 493
E

466 and coded as code 6 (other). Therefore, the activities preceding acci- by NIOSH. Increasing the level of detail provided in incident reports 494
467 dents in this cluster were not within the more frequent accident causing will likely result in improved injury clustering and classifications. It is 495
R

468 activities (e.g., getting on/off truck, handling supplies). In terms of the also important to note that correlation does not equal causation and 496
469 victim job experience this cluster did not differ from clusters 1, 2, and no causal relationships are implied based on the results of this analysis. 497
470 4 (see Table 3a). A more in-depth analysis will be necessary to determine causal factors 498
R

471 Similar to cluster 2, one interesting pattern in cluster 4 is that no ac- for injuries sustained in each cluster. 499
472 cident (severe or non-severe) had been recorded in one half of the year. In conclusion, clustering, classification, and a hybrid methodology of 500
O

473 Unlike cluster 2, all of the accidents in cluster 4 occurred in the first half both—using K-means clustering and gene expression programming—was 501
474 of the year since 2000. Again, this pattern (along with the pattern in shown to be effective when analyzing mining injury data. In particu- 502
475 cluster 2) is a reasonable justification for conducting seasonal studies lar, the GEP classification sub-model had a better performance than 503
C

476 for improving safety in all of the mining sites in this cluster. Thus, the the traditional method of logistic regression in accident type classifi- 504
477 main reasons for zero accidents in the first half of each year and a cation. Furthermore, the identified patterns in the accident database 505
N

478 total of 2,192 in the second half should be identified. Also, nearly 90% using the clustering method are not achievable with the traditional 506
479 of these accidents occurred at surface mines. Cluster 4 was also found injury data analysis which uses counts and cross-tabulations. There- 507
fore, determining the most dominant attributes for specific types of 508
U

injuries (classification) and separating (clustering) the data based 509

on these attributes may allow researchers to better understand the 510
t6:1 Table 3d nature and causal factors of mining injuries. Clustering could likely 511
t6:2 Subunit code definitions.
improve traditional injury analysis methods. Injuries sustained in 512
t6:3 Code Description minerals processing mills and coal preparation plants, and injuries 513
t6:4 1 Underground operations sustained during the operation of trucks, should be analyzed sepa- 514
t6:5 2 Surface at underground rately from larger datasets. The use of non-powered hand tools for 515
t6:6 3 Surface: strip or open pit mining off-road truck inspections, ingress and egress from off-road trucks, 516
t6:7 4 Auger
and newer employees operating dump trucks should be investigated 517
t6:8 5 Culm banks
t6:9 6 Dredge further to determine ways to prevent these severe and non-severe 518
t6:10 7 Other surface accidents. While this analysis was limited to those injuries associat- 519
t6:11 8 Independent shop and yards ed with off-road trucks, it is expected that larger, broader injury data 520
t6:12 9 Mill or preparation plant analyses will also benefit from this hybrid clustering-classification 521
t6:13 10 Office
methodology. 522

Please cite this article as: Dindarloo, S.R., et al., Off-road truck-related accidents in US mines, Journal of Safety Research (2016), http://dx.doi.org/
10.1016/j.jsr.2016.07.002
 S.R. Dindarloo et al. / Journal of Safety Research xxx (2016) xxx–xxx 9

523 5. Practical applications Mine Safety and Health Administration (MSHA) (2014). Summary of selected accidents/ 573
injuries/illnesses reported to MSHA under 30 CFR part 50, mine injury and worktime 574
quarterly and self extracting files, 2000–2014. (Retrieved Dec 25, 2014 from) http:// 575
524 Analyzing the injury data using data mining techniques provides www.msha.gov 576
525 some insight into attributes that are associated with high accuracies Moore, S. M., Porter, W. L., & Dempsey, P. G. (2009). Fall from equipment injuries in U.S. 577
mining: Identification of specific research areas for future investigation. Journal of 578
526 for predicting injury severity. Off-road truck-related injuries continue Safety Research, 40, 455–460. 579
527 to plague the mining industry resulting in fatalities, permanent disabil- Mujalli, R. O., & De Oña, J. (2011). A method for simplifying the analysis of traffic accidents 580
528 ities, and other less severe injuries. Many factors contribute to these in- injury severity on two-lane highways using Bayesian networks. Journal of Safety 581
Research, 42(5), 317–326. 582
529 juries, and many of these are likely preventable. This analysis revealed Pollard, J. P., Heberger, J., & Dempsey, P. G. (2014). Maintenance and repair injuries in US 583
530 that injuries associated with the use of non-powered hand tools for mining. Journal of Quality in Maintenance Engineering, 20(1), 20–31. 584
531 off-road truck inspections, ingress and egress from off-road trucks, Reardon, L. M., Heberger, J. R., & Dempsey, P. G. (2014). Analysis of fatalities during main- 585
532 and newer employees operating dump trucks are areas deserving of at- tenance and repair operations in the U.S. mining sector IIE Transactions on 586
Occupational Ergonomics & Human Factors, 2(1), 27–38. 587
533 tention to determine ways to prevent future injuries. Ruff, T., Coleman, P., & Martini, L. (2011). Machine-related injuries in the US mining in- 588
dustry and priorities for safety research. International Journal of Injury Control and 589
Safety Promotion, 18(1), 11–20. 590
534
Q6 Uncited reference 591

F
Schoenfisch, A. L., Lipscomb, H. J., Shishlov, K., & Myers, D. J. (2010). Nonfatal construction
industry-related injuries treated in hospital emergency departments in the United 592
States, 1998–2005. American Journal of Industrial Medicine, 53(6), 570–580. 593
535 Mine Safety and Health Administration (MSHA), 2014

O
Tibshirani, R., Walther, G., & Hastie, T. (2001). Estimating the number of clusters in a data 594
set via the gap statistic. Journal of the Royal Statistical Society, Series B (Statistical Meth- 595
odology), 63(2), 411–423. 596
536 References Turin, F. C., Wiehagen, W. J., Jaspal, J. S., & Mayton, A. G. (2001). Haulage truck dump site 597

O
safety: An examination of reported injuries. (DHHS (NIOSH) publication no. 2001–124, 598
537 Abdelwahab, H. T., & Abdel-Aty, M. A. (2001). Development of artificial neural network information circular 9454)Pittsburgh, PA: U.S. Department of Health and Human Ser- 599
538 models to predict driver injury severity in traffic accidents at signalized intersections. vices, Public Health Services, CDC-NIOSH. 600
539

R
Transportation Research Record, 1746, 6–13. Wenner, C., & Drury, C. G. (2000). Analyzing human error in aircraft ground damage inci- 601
540 Cato, C., Olson, D. K., & Studer, M. (1989). Incidence, prevalence and variables associated dents. International Journal of Industrial Ergonomics, 26(2), 177–199. 602
541 with low back pain in staff nurses. American Association of Occupational Health Nurses Wiehagen, W. J., Mayton, A. G., Jaspal, J. S., & Turin, F. C. (2001). An analysis of serious in- 603
542 Journal, 37, 321–327. 604

P
juries to dozer operators in the U.S. Mining industry. (DHHS (NIOSH) publication no.
543 Cox, D. R. (1958). The regression analysis of binary sequences (with discussion). Journal 2001–126, information circular 9455)Pittsburgh, PA: U.S. Department of Health and 605
544 of the Royal Statistical Society B, 20, 215–242. Health Services, Public Health Services, CDC-NIOSH. 606
545 De Oña, J., López, G., Mujalli, R., & Calvo, F. J. (2013). Analysis of traffic accidents on rural
546 highways using latent class clustering and Bayesian networks. Accident Analysis and D
547 Prevention, 51, 1–10. Saeid R. Dindarloo is a researcher in the area of heavy mining machinery in Mining & Nu- 607
548 De Oña, J., Mujalli, R. O., & Calvo, F. J. (2011). Analysis of traffic accident injury severity on clear Engineering at Missouri University of Science & Technology (Missouri-Rolla). He 608
549 Spanish rural highways using Bayesian networks. Accident Analysis and Prevention, holds Ph.D., M.Sc. and B.Sc. degrees, all in Mining Engineering, from Missouri University 609
550 of Science and Technology, USA, and Amirkabir University of Technology (Tehran 610
E
43(1), 402–411.
551 Drury, C. G., Porter, W. L., & Dempsey, P. G. (2012). Patterns in mining haul-truck acci- Polytechnic), Iran. Dr. Dindarloo has 10 years of research and professional experience. 611
552 dents. Proceedings of the human factors and ergonomics society 56th annual meeting He has extensive experience in mine design, planning, and computer application. His cur- 612
553 rent research interests include: mini safety and health, mining machinery, and mining 613
T

(pp. 2011–2015). Santa Monica, CA: Human Factors and Ergonomics Society.
554 Duda, R., Hart, P., & Stork, D. (2001). Pattern classification (2nd ed.). New York: John Wiley equipment management. 614
555 and Sons. 615
616
C

556 Ferreira, C. (2001). Gene expression programming: A new adaptive algorithm for solving Jonisha P. Pollard M.S., CPE is a Research Engineer at the National Institute for Occupa-
557 problems. Complex Systems, 13(2), 87–129. tional Safety and Health in the Office of Mine Safety and Health Research. She joined 617
558 Jain, A. K. (2010). Data clustering: 50 yrs beyond K-means. Pattern Recognition Letters, NIOSH while pursuing her Master's degree in Bioengineering from the University of 618
559 31(8), 651–666. Pittsburgh in 2007. During her time at NIOSH, she has been involved in numerous research 619
E

560 Kanungo, T., Mount, D. M., Netanyahu, N. S., Piatko, C. D., Silverman, R., & Wu, A. Y. (2002). studies and has published research on knee injuries in mining, the effect of cap lamp light- 620
561 An efficient k-means clustering algorithms: Analysis and implementation. IEEE ing on balance, the effect of kneepads on balance, injuries due to maintenance and repair 621
562 Transactions on Pattern Analysis and Machine Intelligence, 24(7), 881–892. work in mining, locomotion in restricted workspaces and a variety of other topics. Ms. 622
R

563 Kecojevic, V., & Radomsky, M. (2004). The causes and control of loader- and truck-related Pollard enjoys conducting field and laboratory research to positively impact the health 623
564 fatalities in surface mining operations. Injury Control and Safety Promotion, 11(4), and safety of mine workers. 624
565 239–251. 625
R

566 Koza, J. R. (1992). Genetic programming, on the programming of computers by means of nat- Elnaz Siami Irdemoosa obtained her B.Sc. in Mining Engineering and her M. Sc. in Rock 626
567 ural selection. Cambridge, MA: MIT Press. Mechanics from Amirkabir University of Technology (Tehran Polytechnic). She is currently 627
568 Mardis, A. L., & Pratt, S. G. (2003). Nonfatal injuries to young workers in the retail trades a Ph. D. candidate at Missouri University of Science and Technology in the field of Geolog- 628
629
O

569 and services industries in 1998. Journal of Occupational and Environmental Medicine, ical Engineering. Her research interests include underground design and construction,
570 45(3), 316–323. tunneling, construction management, and geophysical method. 630
571 Md-Nor, Z. A., Kecojevic, V., Komijenivic, D., & Groves, W. (2008). Risk assessment for haul 631
572 truck-related fatalities in mining. Mining Engineering, 60(3), 43–49.
C

632
N
U

Please cite this article as: Dindarloo, S.R., et al., Off-road truck-related accidents in US mines, Journal of Safety Research (2016), http://dx.doi.org/
10.1016/j.jsr.2016.07.002