=Paper=
{{Paper
|id=Vol-2578/DARLIAP9
|storemode=property
|title=Machine Learning Supported Next-Maintenance Prediction for Industrial Vehicles
|pdfUrl=https://ceur-ws.org/Vol-2578/DARLIAP9.pdf
|volume=Vol-2578
|authors=Sachit Mishra,Luca Vassio,Luca Cagliero,Marco Mellia,Elena Baralis,Riccardo Loti,Lucia Salvatori
|dblpUrl=https://dblp.org/rec/conf/edbt/MishraVCMBLS20
}}
==Machine Learning Supported Next-Maintenance Prediction for Industrial Vehicles==
https://ceur-ws.org/Vol-2578/DARLIAP9.pdf
Machine Learning Supported Next-Maintenance Prediction
for Industrial Vehicles
Sachit Mishra Luca Vassio Luca Cagliero
Politecnico di Torino Politecnico di Torino Politecnico di Torino
Turin, Italy Turin, Italy Turin, Italy
sachit.mishra@studenti.polito.it luca.vassio@polito.it luca.cagliero@polito.it
Marco Mellia Elena Baralis Riccardo Loti
Politecnico di Torino Politecnico di Torino Tierra Spa
Turin, Italy Turin, Italy Turin, Italy
marco.mellia@polito.it elena.baralis@polito.it rloti@tierratelematics.com
Lucia Salvatori
Tierra Spa
Turin, Italy
lsalvatori@topcon.com
ABSTRACT time, oil pressure, temperature, engine speed). Domain experts
Industrial and construction vehicles require tight periodic main- can thus monitor vehicle state in order to understand which
tenance operations. Their schedule depends on vehicle character- maintenance actions need to be performed. In recent years, the
istics and usage. The latter can be accurately monitored through significance advances of Internet of Things (IoT) and Big Data
various on-board devices, enabling the application of Machine analytics technologies have fostered the development of smart
Learning techniques to analyze vehicle usage patterns and design predictive maintenance solutions [15]. Specifically, the analysis
predictive analytics. This paper presents a data-driven application of CAN bus data by means of data mining and machine learning
to automatically schedule the periodic maintenance operations techniques allow us to (i) predict the future vehicle usage by
of industrial vehicles. It aims to predict, for each vehicle and date, means of classification and regression techniques (e.g., [7, 10]),
the actual remaining days until the next maintenance is due. Our (ii) aggregate vehicles with similar characteristics using cluster-
Machine Learning solution is designed to address the following ing techniques, (e.g., [1, 4]), and (iii) identify malfunctioning of
challenges: (i) the non-stationarity of the per-vehicle utilization specific vehicle components (e.g., [6, 15]).
time series, which limits the effectiveness of classic scheduling Optimizing maintenance activities of fleet vehicles is a priority
policies, and (ii) the potential lack of historical data for those in several industrial processes [2]. In fact, technological applica-
vehicles that have recently been added to the fleet, which hinders tions related to activity planning and resource management are
the learning of accurate predictors from past data. Preliminary crucial for efficiently handling logistics [14]. This paper proposes
results collected in a real industrial scenario demonstrate the a machine learning approach to support the smart planning of
effectiveness of the proposed solution on heterogeneous vehi- the fleet maintenance operations in a industrial context.
cles. The system we propose here is currently under deployment, Data-driven vehicle maintenance planning has already been
enabling further tests and tunings. addressed using various optimization methods. For example,
in [11] the authors have applied Genetic Algorithms to plan
the maintenance of geographically distributed assets by consid-
1 INTRODUCTION ering routing constraints and travel time to reach the assets. The
Fleets of industrial and construction vehicles are subject to pe- authors in [12] have presented a dynamic optimization method to
riodic maintenance. These vehicles are of different models and plan maintenance of heavy vehicles by jointly scheduling main-
types. Thus, they require to plan actions of different type and tenance operations and production activities, whereas in [8] the
frequency. According to the current vehicles’ workload, mainte- authors propose a data-driven simulation framework for planning
nance schedule often changes. For example, some vehicles could snow removal projects considering weather and truck-related
remain unused for a relatively long period of time, then be moved data acquired by real-time sensors. All the aforesaid strategies
to a construction site, and keep working at full capacity for many are possible if accurate predictions of next maintenance events
days or weeks. The heterogeneity of industrial vehicle usage are available.
has indeed prompted the need for tracking their utilization and Contribution. This work presents an application of regres-
automating the schedule of maintenance operations [3]. sion techniques to predict for a given vehicle when the next
In the context of industrial vehicle management, the advent maintenance will be due. Specifically, it predicts the number of
of CAN bus technology has enabled the design of data-driven days left to the next maintenance operation, based on the series
decision-making processes [5]. The CAN bus provides access to of past daily utilization levels and on the current time of usage
various signals describing the vehicle usage state (e.g., working left to the next maintenance. The application faces three of the
main issues related to fleet maintenance planning: (i) Vehicle
Copyright © 2020 for this paper by its author(s).
Published in the Workshop Proceedings of the EDBT/ICDT 2020 Joint Confer- heterogeneity: The industrial vehicles in a fleet are commonly
ence (March 30-April 2, 2020, Copenhagen, Denmark) on CEUR-WS.org. rather heterogeneous, in terms of number, type, and frequency of
Use permitted under Creative Commons License Attribution 4.0 International (CC the necessary maintenance operations. This makes the planning
BY 4.0)
of these activities particularly challenging and time-consuming
v
for fleet managers. (ii) Non-stationarity of the utilization series: least half of the usage in one cycle ( T2 ) is already available. (iii)
v
Vehicle usage pattern levels are rather irregular, whereas main- New: If the vehicle has been used for less than T2 seconds since
tenance actions rather frequent. Seasonal trends may depend the beginning of the data acquisition phase.
on vehicle type, model, and context of use. (iii) Lack of data for
new vehicles: When a vehicle is added to the fleet, usage data 2.1 Error computation
is typically not available. This hinders the learning of accurate To effectively support fleet managers in planning periodic vehicle
machine learning models, which requires historical data to train maintenance, our prediction system is tailored to a specific goal
reliable predictors. that we encode by considering specific error function definition.
To handle heterogeneous fleets, we train a separate regression We define three errors for each vehicle v: the daily error Ev (t),
model per vehicle. Each regressor analyzes the vehicle usage v
the global error EGlobal , and the Mean Residual Error EvM RE (D).
e
patterns and the current time to maintenance of a specific vehi-
The daily error counts on each day t the gap between the
cle. We conduct the analysis on 24 industrial and construction
predicted and actual values of the next day of maintenance:
vehicles of different models. To handle the non-stationarity of
the analyzed series, we incorporate the historical usage levels
Ev (t) = Dv (t) − DvP r edict (t) (2)
in the predictive models and train both linear and non-linear
models. Finally, to overcome the lack of data related to new ve- The global error is a mean of the daily errors over all the N v
hicles, we combine the regressor outcomes achieved on similar samples related to the vehicle under analysis, i.e.,
vehicles. The presented application is complementary to existing ÍN v v
optimization-based planning strategies, e.g., [8, 11], providing v t =1 E (t)
EGlobal = (3)
the fleet management system with specific hints on future vehi- Nv
cle usage states. The results, achieved in real industrial scenario, The global error combines the daily errors together, but it
show substantial improvements achieved by applying non-linear does not consider nor weight the time that is left for the next
regression models compared to classical statistics-based or linear maintenance. In other words, an error of 1 day when we are close
models. In light of the achieved results, the data owner (collecting to the maintenance (e.g., Dv (t) = 1 ) is considered as equal as an
telematics data from real industrial vechicles) has decided to put error when we are far from the maintenance (e.g., Dv (t) = 100).
the present application under deployment, thus enabling further In order to solve this issue, the mean residual error is considered.
tests, tunings, and extensions. It is the mean of the daily errors over specific days. In particular
This paper is organized as follows. Sections 2 formalizes the we want to compute the average only for specific values of Dv ,
problem, while Section 3 describes the dataset and its prepara- contained in a set D.e D e consists of a selection of days that are
tion phase. Section 4 presents the data-driven methodologies. closer to the maintenance operation, for each maintenance cycle.1
Section 5 summarizes the main experimental results. Finally, Sec- EvM RE is computed as follows:
tion 6 draws conclusions and summarizes the future research Í v
v e E (i)
agenda. e = i:D (i)∈D
EvM RE (D) (4)
i : Dv (i) ∈ De
2 PROBLEM STATEMENT
The idea behind EvM RE is that fleet managers are mainly in-
Given an arbitrary industrial or construction vehicle v, our goal terested in getting accurate predictions when the vehicles are
is to predict when the next maintenance operation for v will be towards the end of their maintenance cycle, i.e., when mainte-
due. Let N v be the number of days for which historical data about nance operations need to be scheduled soon. Therefore, our main
v usage is available and let T v be the allowed usage times (in objective is to minimize EvM RE (D).
e
seconds) for v between two consecutive maintenance operations.
The period from one maintenance operation to the next one will 3 DATA PREPARATION
be hereafter denoted as a cycle. The count of the number of days
The application presented in this paper has been developed and
left to the next maintenance of vehicle v varies day by day. Let
tested on real vehicle data provided by Tierra S.p.A.2 , a company
Dv (t) be the series of the aforesaid daily counts. Our aim is to
that provides IoT solutions for monitoring vehicles of multiple
predict Dv (t), where t denotes the current day. The series used
vendors. The dataset consists of historical usage of 24 heteroge-
to drive the prediction are enumerated below. For each vehicle v:
neous vehicles acquired over a 4 year period (from January 2015
• U v (t): series of the daily utilization of vehicle v. to September 2019). For each vehicle, we consider the informa-
• C v (t): series of the counts of the number of days already tion coming from the CAN bus. Onboard sensors and Machine
passed from the last maintenance operation. Control Systems generate messages for CAN at a frequency of
• Lv (t): series of the utilization times left to the next main- approximately 100 Hz. Each message is collected by a controller
tenance operation. On an arbitrary day t, it is computed which processes it, periodically generates a summary report, and
as follows: sends it to a cloud server [7].
To prepare vehicle data for the present study, the input CAN
t −1
Õ bus data goes through a series of steps: (i) Data Cleaning, (ii) Nor-
Lv (t) = T v − U v (i) (1) malization, (iii) Aggregation, (iv) Enrichment and (v) Transforma-
i=t −C v (t ) tion. A more detailed description of each of the above-mentioned
In the following, we define three categories of vehicles accord- steps is given below. Data cleaning entails properly handling
ing to the amount of historical data that is currently available: missing values and inconsistent values. Data normalization al-
(i) Old: If at least one maintenance cycle has already been com- lows us to scale the values of the utilization times to a uniform
pleted since data acquisition has started. (ii) Semi-new: If the first e = {1, . . . , 29}.
1 We have considered the last 29 days per cycle, i.e., D
maintenance cycle has not been completed yet, but data about at 2 https://www.tierratelematics.com/
50000 400
v1 v1
v2
40000 v2
300
30000
Dv (t)
U v (t)
200
20000
100
10000
0 0
0 20 40 60 80 0 200 400 600 800
t t
Figure 1: Daily utilization in seconds (U v ) in function of Figure 2: Target variable number of days left to the next
the day in the series (t) for two sample vehicles. maintenance (Dv ) with respect to day in the series (t).
Many cycles are shown for two sample vehicles.
value range (e.g., from 0 to 1) thus avoiding to introduce bias in
regression model learning [13]. Data aggregation, enrichment
150
and transformation aim at providing predictive algorithms with v1
an appropriate set of features describing the usage patterns by 125 v2
aggregating data at the desired time granularity. For each vehicle
100
we transformed the raw CAN bus data to produce the input fea- Dv (t)
tures described in Section 2. Specifically, in our case of study, we 75
primarily focus on daily-usage time series U (t), i.e., the amount
50
of time each vehicle worked on each day.
25
3.1 Preliminary data exploration 0
0 500000 1000000 1500000 2000000
Figure 1 plots part of the series of daily utilization seconds, i.e., Lv (t)
U v (t), for two sample vehicles. Curves show that vehicle utiliza-
tion patterns are rather heterogeneous. Vehicle v 1 has a daily
utilization of about 20 000-30 000 seconds, with few days without Figure 3: Utilization seconds left to maintenance (Lv ) vs.
usage every 10-15 working days. On the other hand, vehicle v 2 the number of days to maintenance (Dv ) for a single cycle
remains almost unused for several weeks (from t=0 to t=40) and of two vehicles.
then suddenly changes its usage pattern.
After a fixed time amount of usage (we have considered T v =
2 000 000 seconds), every vehicle needs to go under maintenance.
Notice that we do not know whether maintenance operations 4 METHODOLOGY
have actually been performed or how long they take (from T v In this section we present the data mining process designed
on). Figure 2 shows two examples of target series Dv (t), with to address the problem under analysis. We propose different
many shown cycles. When Dv (t) reduces to zero, the vehicle methodologies for vehicles when they are recognized as new,
goes to maintenance. Then a new maintenance cycle starts, the semi-new, or old (see Section 2). For each old vehicle we have
number of days left to maintenance is maximal, and it mono- enough data in order to train a prediction model on its own past.
tonically decreases (one day for each day passed) until the next Semi-new and new vehicles will be separately handled.
maintenance operation is carried out. Notice how v 1 has a first For each vehicle v we generate a relational dataset containing
longer cycle (221 days), while the others are more constant and the historical utilization series U v (x) [13]. More specifically, each
homogeneous, with length between 65 and 105 days. record corresponds to a different day t and consists of a set
In Figure 3 we show the number of days left to maintenance of attributes denoting the past utilization levels (in seconds).
Dv (t) with respect to the number of utilization seconds left for Given a window size W , the attributes include the values U v (x)
the next maintenance Lv (t). The functions seems to have a con- [t − W ≤ x ≤ t − 1]. Along with the utilization level series, the
stant rate when Lv (t) is closer to zero, reflecting that for most of attributes include the current time left until the next maintenance,
the time the utilization rate is relatively constant and different i.e., Lv (t), and the target variable, i.e., the number of days left to
from zero. However, there are some vertical steps, corresponding maintenance, i.e., Dv (t).
to consecutive days when the utilization was null. This confirms Many machine learning techniques work better with more
that the presence of low- or zero-utilization periods has a serious data, hence we desire to increase the number of records per
impact on the target variable. Thus, predicting the correct target vehicle. Since we do not know when vehicle actually had the
value could be challenging. Hopefully, it is unlikely to see long maintenance done, we can shift the time reference, i.e., chang-
periods of zero-utilization in the days approaching the deadline. ing the first starting day t = 0, without introducing errors. We
This reinforces the motivations behind using EvM RE (D) e as ref- randomly re-sampled multiple times the time reference starting
erence error metric, considering values relatively close to the from different time points within the training data and build the
maintenance. utilization series.
4.1 Approaches record of the training dataset.
We apply three different methods: (i) a baseline model, relying on Support Vector Regressor (SVR): It is among the most effec-
simple estimate, (ii) a univariate model, whose predictions rely on tive solutions to address regression and classification problems.
a single variable, and (iii) a multivariate model, where the model Given a multidimensional training data representation, it finds an
considers multiple series values. These are standard methods for hyper-plane separating points belonging to different target value
time series forecasting. Novelty relies in the categorization of the ranges. According to the kernel function used to derive the hy-
vehicles and data engineering. perplane, the predictive model can be either linear or non-linear
(e.g., polynomial, sigmoid, rbf). Due to the high computational
4.1.1 Baseline algorithm: With this baseline approach we sim- complexity of non-linear kernels, in the remaining of the paper
ply predict when the next maintenance will be due assuming that we focus on linear SVR (LSVR).
the utilization is constant and equal to the average utilization in Random Forest regressor (RF): It is an established ensemble
the past. Hence, we compute the average utilization of vehicle v method combining the predictions of multiple decision trees. De-
in the past (training set of size Tt r ain ). Then we use it for predict- cision trees are the most popular non-linear mapping functions
ing the number of remaining days until the next maintenance is between non-predictive and predictive variables. They rely on
due. The average utilization is defined as follows: tree-based structures. The Random Forest Regression averages
ÍTt r ain the predictions made by various decision tree models, which are
i=1 U v (t) trained on different bootstraps (i.e., samples of the training data
AV G v = (5)
Tt r ain with replacement).
Let DvBL (t) be the number of days left to next maintenance Histogram-based gradient boosting (XGB): it is a popular en-
predicted by the Baseline algorithm at time t as: semble method relying on a boosting strategy. It minimizes the
prediction loss by combining many decision tree regressors.
Lv (t)
DvBL (t) = (6)
AV G v Additional models can be straightforwardly added and tested
We will denote the baseline method as BL thoroughout the in the deployed version of the system. Notice that some models
paper. (e.g., Neural Networks) have not been included in this first release
4.1.2 Univariate regression model. We apply a univariate re- due to the lack of a sufficiently large amount of training data.
gression model FU R to predict the number of days left to main-
tenance for a given vehicle v based on the last value of the daily
utilization seconds series Lv : 4.3 Methodology for old vehicles
v v Old vehicles are assumed to have a sufficiently large amount of
DU R (t) = FU R (L (t)) (7)
historical data to train reliable Machine Learning models (see
4.1.3 Multivariate regression model. We extend the univariate Section 2). Thus, separately for each vehicle we train the multiple
regression model to a multivariate context in order to consider regression models described in the previous section. Among the
the temporal correlation between the target variable and the trained models, we select those that minimizes the mean residual
previous series values. error over the last 29 days predicting the maintenance (EvM RE (D) e
with D = {1, . . . , 29}). For each vehicle, we consider the first 70%
e
DvM R (t) = FM R (Lv (t), U v (t − 1), . . . , U v (t − W )) (8) of their samples (N v ) as training set, and the remaining part as
test set.
Unlike the univariate model, the model formalized in Equa-
tion 8 does not consider only the last value of the daily utilization
time but also the most recent values of the historical utilization
series within a size-W window time interval, i.e., from t − W to 4.4 Methodology for new and semi-new
t − 1 (where W is a user-specified parameter). vehicles
To handle new and semi-new vehicles, we need to face the follow-
4.2 Regression algorithms ing issues: (i) The lack of historical usage data, which hinders the
Univariate and multivariate regression models can be solved training of per-vehicle regressors. (ii) The first maintenance cycle
with linear or non-linear models. Linear models are deemed of most vehicles appears to have peculiar characteristics, with
as appropriate whenever usage patterns are quite constant for less usage. Indeed, the mean daily utilization time spent by the
most of the time. Conversely, non-linear models are potentially vehicles within the first cycle (10 676 seconds) is approximately
able to capture more complex, non-stationary usage trends. As 30% lower than in the subsequent cycles (13 792 seconds).
a drawback, the complexity of non-linear models is typically We design ad-hoc strategies to predict Dv for semi-new and
higher than those of linear ones. new vehicles. We consider as training data the utilization series
The deployed system allows fleet managers to select an ac- in the first cycle of many old vehicles. Collecting in the training
ceptable trade-off between the accuracy and complexity based set only usage data related to the first maintenance cycle allows
on empirical evidences. The results of a preliminary performance Machine Learning models to focus on the usage patterns peculiar
evaluation on real vehicle data are reported in Section 5. The to that usage period (which could be significantly different all
models that have already been integrated and tested are briefly from the subsequent ones). We take 70% of the 24 vehicles (i.e.,
described below. A more detailed description is given in [13]. 17 vehicles), and consider their complete first cycle as training
Linear Regression (LR): It is the simplest linear model. It learns set. The first cycle of the remaining 30% of the vehicles (i.e., 7) is
a linear function minimizing the residual sum of squares between considered as test set.
the predicted target value and the expected target value in the
Table 1: E M RE ({1, . . . , 29}) with models trained on all data
60
and models trained in the last 29 days before maintenance, BL
i.e., at times i : D(i) ∈ D e = {1, . . . , 29} LR
40 LSVR
Trained on all data Trained on D = {1, . . . , 29} RF
Improvement(%)
Algorithm
E M R E ({1, . . . , 29}) E M R E ({1, . . . , 29}) XGB
20
BL 20.2 20.2
LR 26.1 10.8
LSVR 13.3 6.1 0
RF 6.9 2.4
XGB 10.9 5.6
−20
4.4.1 Semi-new vehicles. When more than half of the first cy-
cle has been completed, we learn a regression model by combin- −40
0 3 6 9 12 15 18
ing data from the other training vehicles. We apply the following W
strategies:
Baseline: we follow the same approach as for old vehicles. In
practice, we compute AV G v as the average utilization in the Figure 4: Improvement (%) for each algorithm by increas-
first half of the first cycle, i.e, t :Íx ≤t U v (x )≤T v /2 (U v (t)/|t :
Í ing the number of features. W is the window of past usage
Í v v in the time series U v (t).
x ≤t U (x) ≤ T /2)|).
Unified ML model: We create a single model, hereafter denoted
as ModelU ni , on the first cycle data of the training vehicles. In Table 2: Best setting for features and the corresponding
this case, we train a single regression model for all the semi-new mean relative error of the different algorithms.
vehicles by merging data acquired from all the training vehicles
together. The same model is applied to all the test vehicles. Algorithm Best window W E M R E ({1, . . . , 29})
Similarity-based ML model: We pick usage data only for the BL 0 20.2
most correlated vehicle (rather than for all the old vehicles) LR 0 10.8
LSVR 6 5.2
and train vehicle-specific Machine Learning models, denoted RF 18 1.3
as Model Sim , on it. The key idea is to first decide whether each XGB 12 4.2
vehicle is similar to the target semi-new vehicles or not by es-
timating the pairwise correlation between the utilization series
acquired in the first half of the first cycle. Then, we train the (central column) and just in the the last 29 days of the cycles in
regression model only the first cycle data of the selected vehicle. the training data (right hand-side column). E M RE ({1, . . . , 29})
In the current implementation, we estimate the pairwise similar- is the average of the mean residual errors EvM RE ({1, . . . , 29})
ity in terms of point-wise average distance AV G v between the computed over all the test vehicles.
utilization series. However, more advanced similarity measures We found out that by forcing the algorithm to train only on
(e.g., [9]) can be integrated as well. the last 29 days, the error was reduced by 59% in LR, 54% in LSVR,
65% in RF and 48% in XGB. Since BL is not trained, its results
4.4.2 New vehicles. For these vehicles we have very few or do not change. In general, RF presents the best results, with an
none data at all, hence we cannot even compute AV G v . There- average relative error of only 6.9 and 2.4 days, respectively for
fore, the baseline and similarity-based ML models cannot be the two training sets. Second best results are obtained by XGB,
applied. Hence, we apply the Unified ML model (ModelU ni ). No- closely followed by LSVR.
tice that, when dealing with new vehicles, it does not make sense Now we delve into the study of the usage of different features.
to compute EvM RE ({1, . . . , 29}) since when we are approaching Figure 4 shows, for each algorithm, the performance variation
the deadline the vehicle will already be semi-new. Hence, we (percentage) by increasing the window size W. Positive variation
focus on comparing the algorithm performance in terms of the means decrease of the error shown in Table 1 (i.e., performance
v
global error EGlobal . improvement), and vice-versa. W is the window of past usage in
the time series U v (x).W equal to 0 means we are in the univariate
5 EXPERIMENTAL RESULTS case, while W > 0 means we are in the multivariate case. For
We have tested the proposed methodologies on real vehicles of dif- example, W equal to 3 means usage of U (t −1), U (t −2) and U (t −3)
ferent categories (new, semi-new, and old). The experiments were in the regression (other than L(t)). BL is obviously constant, since
performed on a machine with Intel(R) Core(TM) i7-8750H CPU it is not using past values of U . In RF and XGB, adding features
with 16 GB of RAM. To tune the algorithm parameter settings we greatly improve the results. RF and XGB seem to reach a plateau
have performed, separately for each vehicle, a grid search using in performance with more than 15 previous value of utilization
a 5-fold cross validation. Specifically, for RF and XGB we have U . Hence, having more than a couple of weeks of previous data
tuned the maximum tree depth from 3 to 50, and the number of is enough to reach good performance. Respectively, the results
estimators from 10 to 1000. For SVR, we tested the linear kernel improved by 44% and 25% for RF and XGB. In LR the best results
and varied the values of the parameters epsilon (from 0.5 to 2.5) are instead obtained without adding features. Finally, for LSVR
and C (from 0.01 to 100). the results improve by adding up to 6 features, but then decrease
again. We report in Table 2 for each algorithm the final best
5.1 Results for old vehicles results for E M RE ({1, . . . , 29}) and the optimal size of the window
In Table 1 we show the values of E M RE ({1, . . . , 29}) achieved on W . The non-linear algorithms (XGB and RF) reach again the best
the test set by training each algorithm on the whole training data results.
Table 3: Results for semi-new and new vehicles
40
BL
LR Semi-new vehicles New vehicles
Algorithm
30 LSVR E M R E ({1, . . . , 29}) EGl obal
BL 34.9 -
EM RE (D)
RF
e
XGB LR S im 4.9 -
20 LSV R S im 6.2 -
RF S im 2.9 -
X G B S im 5.3 -
10 LRU ni 5.1 27.2
LSV RU ni 8.8 27.8
RFU ni 3.2 30.1
0 X G BU ni 4.2 17.9
{0} {5} {10} {15} {20} {25} {30}
De
to train reliable predictors, (ii) based on data acquired from sim-
Figure 5: Mean relative error analyzed over all the test ve- ilar vehicles, when the vehicle is semi-new (i.e., the first cycle
hicles computed for De ranging from 1 to 29 days. maintenance is partly completed), or (iii) vehicle-independent,
when the vehicle under analysis is new. The achieved results
show that, when a minimal amount of vehicle-related data is
Finally we show in Figure 5 the error, i.e., E M RE (D)
e obtained available, Machine Learning approaches relying on non-linear
with the best configuration, for each algorithm, in each of the last models outperform both naive approaches and linear ML models.
29 days before maintenance. Here D e is a set composed of a single The deployed version of the current system will be further
value, corresponding to the number of days left for the the next extended. Specifically, we plan to enrich regression models using
maintenance. Clearly, the closer to the deadline, the smaller the contextual information (e.g., meteorological data, fleet move-
error. All the algorithms improve the poor performance of the ments) and to design ML supported scheduling strategies.
baseline. Interestingly, RF gets very good results even when we
are 29 days from the deadline, with an average error of only 2.4. REFERENCES
We also analyze the time complexity of the proposed method- [1] Carlos Alonso de Armiño, Miguel Ángel Manzanedo, and Álvaro Herrero.
2019. Studying Road Transportation Demand in the Spanish Industrial Sec-
ologies, in terms of execution time on our machine. The whole tor Through k-Means Clustering. In International Joint Conference SOCO’18-
methodology pipeline includes (i) Data preparation, (ii) Model CISIS’18-ICEUTE’18. Springer International Publishing, Cham, 387–396.
[2] Luis Barreto, Antonio Amaral, and Teresa Pereira. 2017. Industry 4.0 implica-
training, and (iii) Model Testing. Step (ii) was the most time- tions in logistics: an overview. Procedia Manufacturing 13 (2017), 1245–1252.
consuming task for all the algorithms. The average training time [3] T.G. Crainic and G. Laporte. 1998. Fleet Management and Logistics. Springer
on a single vehicle is 30.4 s for XGB and 8.1 s for RF, while BL, US.
[4] Zahid Halim, Rizwana Kalsoom, and Abdul Rauf Baig. 2016. Profiling drivers
LR, and LSVR are faster taking respectively 2.5 s, 3.8 s, and 2.8 s. based on driver dependent vehicle driving features. Applied Intelligence 44, 3
Moreover, the model complexity increases more than linearly (01 Apr 2016), 645–664.
with the number of considered features, i.e., window size W. [5] Karl Henrik Johansson, Martin Törngren, and Lars Nielsen. 2005. Vehicle
Applications of Controller Area Network. Birkhäuser, Boston, MA, 741–765.
[6] Q. Li, S. Meng, S. Wang, J. Zhang, and J. Hou. 2019. CAD: Command-Level
5.2 Results for new and semi-new vehicles Anomaly Detection for Vehicle-Road Collaborative Charging Network. IEEE
Access 7 (2019), 34910–34924.
Table 3 shows the results of the above discussed methods for [7] Dena Markudova, Elena Baralis, Luca Cagliero, Marco Mellia, Luca Vassio,
Elvio Amparore, Riccardo Loti, and Lucia Salvatori. 2019. Heterogeneous
semi-new vehicles. RF Sim has the best results among all mod- Industrial Vehicle Usage Predictions: A Real Case. In Workshop Proceedings of
els, meaning that comparing the similarity of average usage can the EDBT/ICDT 2019 Joint Conference (March 26, 2019, Lisbon, Portugal).
slightly improve the results (from 3.2 to 2.9 days in E M RE ). More- [8] Emad Mohamed, Parinaz Jafari, Ming-Fung Francis Siu, and Simaan AbouRizk.
2017. Data-Driven Simulation-Based Model for Planning Roadway Opera-
over, notice how the baseline approach performs badly, with a tion and Maintenance Projects. In Proceedings of the 2017 Winter Simulation
mean relative error of 34.9, a value much higher than all the Conference (WSC ’17). IEEE Press, Article Article 272, 12 pages.
other metrics. This is because the limited amount of past data in [9] Rodica Neamtu, Ramoza Ahsan, Elke A Rundensteiner, Gabor Sarkozy, Ea-
monn Keogh, Hoang Anh Dau, Cuong Nguyen, and Charles Lovering. 2018.
the semi-new test vehicle cannot be trusted to create a simple Generalized dynamic time warping: Unleashing the warping power hidden
predictor based on average usage. in point-wise distances. In 2018 IEEE 34th International Conference on Data
Engineering (ICDE). IEEE, 521–532.
In the last column of Table 3 we also report the results of [10] Federico Perrotta, Tony Parry, and Luis C. Neves. 2017. Application of machine
the above discussed methods for new vehicles. Baseline and learning for fuel consumption modelling of trucks. In 2017 IEEE International
similarity-based models cannot be applied to new vehicle, since Conference on Big Data, Boston. 3810–3815.
[11] Mehdi Rashidnejad, Sadoullah Ebrahimnejad, and Jalal Safari. 2018. A bi-
there is no past data. XGBU ni has the best result among all mod- objective model of preventive maintenance planning in distributed systems
els. Even if the results in terms of error appears poor, we recall considering vehicle routing problem. Computers & Industrial Engineering 120
that the global error accounts also for dates that are far from the (2018), 360 – 381.
[12] Elodie Robert, Christophe Berenguer, Keomany Bouvard, Hoceane Tedie, and
deadline. Even more promising, the global error of XGBU ni is Romain Lesobre. 2018. Joint dynamic scheduling of missions and mainte-
comparable with the baseline mean relative error in the case of nance for a commercial heavy vehicle: value of on-line information. IFAC-
PapersOnLine 51, 24 (2018), 837 – 842.
old vehicles (Table 2). [13] Pang-Ning Tan, Michael Steinbach, Anuj Karpatne, and Vipin Kumar. 2018.
Introduction to Data Mining (2nd ed.). Pearson.
[14] Dieter Uckelmann. 2008. A Definition Approach to Smart Logistics. In Next
6 CONCLUSIONS AND FUTURE WORK Generation Teletraffic and Wired/Wireless Advanced Networking, Sergey Ba-
The paper has presented a Machine Learning application to sup- landin, Dmitri Moltchanov, and Yevgeni Koucheryavy (Eds.). Springer Berlin
Heidelberg, 273–284.
port maintenance planning for fleets of industrial and construc- [15] W. Zhang, D. Yang, and H. Wang. 2019. Data-Driven Methods for Predictive
tion vehicles. It proposes to use regression techniques to predict Maintenance of Industrial Equipment: A Survey. IEEE Systems Journal 13, 3
the remaining days until the next maintenance is due. The trained (Sep. 2019), 2213–2227.
models are (i) vehicle-specific, when sufficient data is available