In the age of connected vehicles, large amounts of data can be collected while driving through a variety of on-board sensors. The information collected can be used for various types of data-driven analytics that can be of great benefit to both vehicle owners, e.g., to reduce costs by means of predictive maintenance, and to society as a whole, e.g., to optimize mobility behavior. Prior to any real-world data analysis, an investigation and characterization of the available data is of utmost importance in order to evaluate the quality and quantity of the data and to set the right expectations. In this paper, we focus on the data exploration and characterization step, which is necessary to avoid inconsistencies in the collected parameters and to enable valid, data-driven modeling. The proposed data exploration considers both the frequency of samples and their values for all monitored parameters. A specific cross-provider data comparison is performed to compare values collected for the same vehicle at the same time from diferent fleet monitoring data providers. The study is applied to a real-world use case with months of data from dozens of vehicles deployed in the waste collection service managed by SEA, Soluzioni Eco Ambientali, in Italy. The analyzes uncover unexpected behaviors in the measurements and lead to their early identification, bringing great benefits to the company operating the fleet by improving data collection and enabling a safe modeling phase.
hicles is opening a broad range of applications for
machine learning (ML) techniques to process this data [
collection company, which is tracking its fleet by means of diferent data providers. 2. An exploratory analysis of the recorded maintenance interventions to enable a future predictive maintenance solution.
presents the related works, Section 3 describes the realworld use case under analysis, and Section 4 focuses on vehicle-tracking data exploration. Finally, Section 5 draws conclusions and future works.
In recent years, the advent of the Internet of Things and
its rapid difusion has facilitated real-time
communication and data exchange between technological devices
that can be used in a wide variety of scenarios. When
the technologies involved are located on a vehicle (e.g.
car, truck), the term Internet of Vehicles (IoV) is used
[
Many infrastructures and diferent technologies have
been implemented to support large-scale, real-time, and
reliable information services. Basically, for this purpose, We analyze real-world fleet data of garbage collection
a three-tier architecture [
The GPS data which reflects the movement of vehi- sites, located in diferent cities in the north and center
cles can be traced to simulate the mobility model in IoV. of Italy. An unexpected failure of one of these vehicles
Authors in [
In many works, data is collected and managed with a also irregular in their timings, which leads to the presclear purpose in mind. However, in other cases, compa- ence of a variety of anomalies. Among all the available nies would like to gain general insights of the information data sources, we can identify the following types of data they collected, and then drive the decisions on how to in the data lake. efectively reach cost reductions and higher operational eficiency. In such cases, an exhaustive exploration of the available data is required, allowing the company to possibly introduce early changes in the collection process.
In this way, our work difers from the above-mentioned ones because our focus is on the data exploration phase, to highlight noteworthy behaviors or anomalies of interest for the data owner. Furthermore, all the abovementioned works have been conducted on data collected by a single provider. Instead, in our use case, we analyze and compare three diferent data providers, each with its own specific sensors. A single parameter can be either specific to a single provider, or it can be monitored by several providers (each with its peculiarities). For this reason, we focus on a data comparison step, not covered in previous related works, to highlight the benefits and weaknesses of each data provider. • Tracking. This data describes the trips each truck performs. We have 3 providers (named P1, P2, and P3 in the following), each tracking different features. The only common measures are GPS position, odometer value, total time of engine activity, and average speed. Some providers also include other measures such as acceleration and estimated fuel consumption. In addition to the variability on the tracked features, diferent providers also have diferent policies on the sampling of the data, hence leading to very diverse daily measure rates. The raw datasets collected by the providers are time series of a total of 117880 samples of 38 tracked features for 35 vehicles for P1, 68611 samples of 68 features for 8 vehicles for P2, and 91072 samples of 17 features for 15 vehicles for P3. • Maintenance. Among the goals of the current exploratory analysis, we would like to assess the feasibility of a future predictive model aiming to predict failures based on the historical maintenance interventions. To this aim, in the data lake Invoices 7 15 8 7 4 1 1 1 6 4 2 2 2 2 2 2 1 7 6 4 4 4 4 3 3 2 2 2 2 1 1 1 1 0 0 3 2 1 0 0 P1 P2 P3 4. Exploratory analysis we analyzed the list of all the repairs of the dif- P3 on the rows) supply information across the 40 vehiferent components of the trucks, for which the cles (on the columns). The heatmap shows the average corresponding invoices are collected. We have a daily number of samples, for each vehicle (column) and total of 120 invoices associated with 36 vehicles. provider (row). The first row reports the number of invoices for each truck, each corresponding to the maintenance interventions.
Among the 40 vehicles, the first 17 share tracking data from multiple providers at the same time, and are described by more than 30 samples per day per vehicles in some cases, with the average being less than 5 samples per day per vehicle. The number of maintenance interventions (invoices) ranges from zero to 15 in the 9 months under analysis.
In the following, we discuss the results of the exploratory analysis on such data.
and results. The dataset of interest in the current work includes information of 40 vehicles from May 2021 to January 2022. Even though the data providers’ REST APIs have been directly used to collect the data at regular and synchronous intervals during the whole period of time, the timestamps of some samples refer to periods sensibly far from the range (May 2021 - January 2022).
Furthermore, significant portions of the original data are 4.1. Multi-provider data comparison duplicated, i.e., diferent calls to the provider APIs return the very same data, also with identical timestamps. Some of these anomalous repeated data are associated with a P1 subsequent period with no collected data. Nevertheless, the overall trend of the vehicles in these situations sug- Location_references(4) gests they where still moving without being tracked. For Alarm_id this reason, we dropped all duplicated values down, since we assumed they are only the result of irregularities of the providers’ update systems. In the same way, we then Engine_working_hour also removed all samples outside of the considered time cuOrrednotm_septeered range. Performance_stats(5) GPS_coordinates
Since at each API call, all variables are returned by the Guidemode_distance(4) data provider, but not all variables efectively recorded a Fuel_Acoltnitusudmeption P3 change in their value, a feature selection step has been P2 used to remove irrelevant features. After removing duplicate records, we also removed all those variables that we felt would not be helpful for an upcoming data-driven analysis. These variables were identified according to the following criteria: (i) a missing value ratio greater than Figure 2: Venn diagram of the high-level semantic aggrega50%, (ii) limited variance (e.g., variables whose values are tion of the features from 3 tracking data providers. always constant).
As a first indicator of the remaining information car- Since each provider has diferent variables, we exedinality, we defined the average number of samples per cuted a semantic association to identify related features vehicle per day as a synthetic and high-level measure. from diferent providers. The resulting features with Figure 1 shows how the diferent providers (P1, P2, and the corresponding associations are presented in Figure 2, Sep 2021
Oct
Nov
Dec where the Venn diagram shows how the semantic groups From Figure 1 we note that 10 vehicles share the P1are shared between the 3 providers. and P3-provided odometer and engine working time. In
The features in common to all the providers are the Figure 3 we report the comparison of such variables for odometer, the speed, and the geographical position of 3 overlapping trucks (namely v13, v16, and v17) in the the vehicles. The working time of the engine is available period between August 2021 and November 2021, after a in 2 out of 3 providers. The odometer and the engine simple conversion of units of measure (P1 uses kilometers working time are of particular interest in our analysis and hours, while P3 meters and minutes). since they give a measure of the mechanical "age" of the P1 presents a temporal gap in the collected samples, of truck, and despite some possible diferences on the ini- about one week from September 1st, for all the reported tial ofset among diferent providers, they are supposed trucks. Since the last measure for each vehicle before to be sampled coherently among them all. Other vari- the gap and the first one after drastically difer, we can ables, such as the current speed, fuel consumption, and assume that within this period the trucks were still acperformance statistics, are also potentially significant to tive. This assumption is proven by the measures from trace the degradation of the vehicles. However, even if P3, confirming that it was associated with a lack of inforthey share the same high-level semantics, we noticed mation from the provider itself. The issue is common to that their values were so diferent that each provider ac- all the data of this source, and it was not reported to the tually tracks a variable on its own, hence they cannot be knowledge of the company before the current analysis. merged together. Moreover, this period, as highlighted by the lower plot, is also associated with a drastic lowering in the value 4.1.1. Comparative analysis on monotonic features of the engine time. This can correspond to a sort of internal reset of the sensors, which may have made the The odometer and the engine working time are two mono- registered values start from a default position, slightly tonic variables shared among diferent data providers (P1 diferent from any truck. In this situation, a fix for this and P3). Both variables are constant when the vehicle error can be obtained by adding the first value from P3 of is turned of, whereas both must be strictly increasing the correspondent vehicle to all measures of P1. With this while driving. correction, each couple of curves after the gap becomes comparable.
Finally, the upper plot also shows another anomaly concerning the odometer tracking of v17. The traces from the two providers have significantly diferent trends that continue diverging as time increases. Vehicle v17 traveled with an average speed of 83 / and 23 /, for P1 and P3, respectively. All providers’ average speed distributions are peaked at around 60 /, as shown in Fig. 4, but the calculated average values are on the tails of both distributions.
In this situation, as for other trucks monitored by more than one data provider, we experienced that information from diferent providers can reveal significant diferences in the reported values.
An important analysis to perform in order to understand efectively the data collected is an exploration of the time period in which the vehicles are monitored. How long and in which manner the data is collected is in fact a useful information for the company that can highlight important diferences between the various providers. Below, for each provider we have reported a graph representing the time on the x axis, and the diferent trucks monitored by the provider on the y axis. In this way, for each truck we can easily see the start and end date of collection and whether there were any anomalies in the monitored period (e.g. missing data or changes in the daily data rate). 4.2.1. P1 provider The graph shown in Fig 5 shows all trucks monitored by provider P1. The diferent color shades of the points represent a diferent sampling frequency on a given day (the darker the point, the more measurements were collected on the same day for that truck). This graph shows that the number of daily measurements collected is almost the same and that all trucks were monitored during the period from May 2021 to January 2022. The anomalies are highlighted by the red rectangles in the figure. The ifrst anomaly that is detected is common to all trucks, and is a lack of data at the beginning of 2021/09 (September).
This behavior can be explained by a period of company closure, or by a general failure in the data collection process, afecting all trucks. Other anomalies are instead related to specific trucks, and show rather long periods (even several months) of lack of data. In general, to see if in all of these situations we don’t have data because the truck was actually idled or because there was a problem in the collection, we need to look at the cumulative values collected from variables such as the odometer. If the value of this variable in the last sample collected before the period without data is equal to the value of the first sample collected after the period without data, it means that the truck actually did not travel in that period, so no information was lost. Otherwise, if the two values are different, it means that the truck moved but we don’t have data that monitored its operations during that period. So there was an error in the data collection phase.
Sampl963es 12 15 v01 v02 v03 v04 v05 v06 v07 v08 v09 v10 v11 v12 v13 v14 v15 lepvvvv11219706 tav18 v21 v22 v23 v24 v25 v26 v27 v28 v29 v30 v31 v32 v33 v34 v35 2021-05 2021-06 2021-07 2021-08 2021-11 2021-12
2022-01 2021-09 date 2021-10 A similar analysis was performed for the measurements collected by provider P2 and shown in Figure 6. In this case, the number of trucks monitored is much lower than those monitored by P1. Moreover, all the trucks considered by P2 are also monitored by P1. This is useful in order to make a quick cross-check on anomalies between the two providers. In particular, we can easily notice that in the same periods in which data for P1 was missing (both the period common to all trucks and those for individual trucks), we now have measurements collected by P2. This leads us to assume that the previous lack of data were really due to P1-related failures. The only period in which we have no measurements in either providers is between 2021/12 and 2022/01 for vehicle v03. In this case, the only check we can do to verify if the truck was really stopped is to monitor the odometer value before and after the lack of data. we manually excluded all the elements associated with periodical maintenance operations to be done as part of
Regarding the measurements collected by P2, a partic- the security enforcement by law, then all the resulting ular behavior can be noticed in the sampling frequency. elements have been grouped into meaningful categories Until 2021/12 there is a very low sampling frequency, based on the failed components. Looking at the remainequal to around one daily measurement per truck. How- ing invoices, we then observed a significant number of ever, this behavior changes suddenly in the last months operations from the same category carried out on the of collection, where the sampling frequency increases same vehicle within a short time (e.g., the day after). We significantly. This seems to be due to a change in the assumed all these cases describe a situation when the collection pattern from December onward, probably with repair was insuficient to solve the problem, hence not the aim of collecting information with a higher level of de- relevant for a predictive-maintenance analysis. For this tail than the single daily information previously collected. reason, we excluded such data from the analysis. The reThe use of the trucks has not changed since December, sulting, clean number of failures, grouped by categories, what has changed is the amount of information describ- is shown in Figure 8. For each row, besides the total numing the behavior of a truck, with up to 200 samples per ber of failures, also the number of second occurrences, day. i.e., those which are not the first failures for the vehicle and the category, is reported. The failures coming from 4.2.3. P3 provider this second group are particularly interesting when investigating the possibility of building a predictive model on The same temporal analysis has also been performed for top of this data since they can give a preliminary measure the trucks monitored by provider P3. Five of the trucks of the frequencies of the failures. in Figure 7 (from v36 to v40) are monitored only by P3, therefore it is impossible to cross-check them with the Electrical system data of the other providers to try to better understand the Braking system anomalies. Regarding instead the remaining trucks, they Mechanics are all monitored by at least one between P1 and P2 and Engine lubrification system this allows us to cross-check the collected measurements.
In particular, the most evident anomalies in the data Hydraulics collected by P3 are for those five trucks where there are Sensors nHoowmeevaesur,rfeomretnhtes sianmtheetrpuecrkiosdafnrdomin2t0h2e1/s0a5mteo m20o2n1t/h10s,. ExhaPuoswtesrypstlaenmt TSoetcaolnd-occurrence the other providers have collected measurements, so it is 0 5 10 Count of lab1el5ed failures 20 25 fair to assume that the lack of data for P3 is a problem related only to that provider. Figure 8: Clean count of the failures, separated by category.
lection of invoices indicating failures of diferent components on the trucks. With the help of domain experts, The entire process of cleaning and correctly grouping the available invoices is fundamental whether the collected data will be exploited to address a predictive maintenance analysis, building a predictive model capable of estimating the residual life of a truck before a maintenance intervention is needed. The possibility of identifying malfunctions or component issues in advance is in fact a key aspect for automotive companies, which can then intervene properly before a failure occurs. This leads to significant cost savings for the company, since a properly implemented predictive model would allow to avoid breakdowns during vehicle travel and to constantly monitor the condition of the various components, scheduling maintenance interventions only when they are really necessary. From Figure 8, we note that for a few categories we do not have enough data to be able to build a robust predictive model for estimating future failures. This is due to the limited time period in which trucks are monitored, which is not long enough to collect enough maintenance interventions for all categories.
preliminary characterization of real-world fleet data coming from diferent tracking providers. We analyzed and compared the data collected by 3 providers, on 40 vehicles, over 9 months, highlighting issues and challenges. The presence of multi-provider information for the same vehicle, although more expensive, can help recover the frequent issue of missing data. As future work, we plan to extend the current work to integrate data from different sources with the goal of designing a predictive maintenance algorithm, by exploiting the maintenance intervention data. Anomalous information from the vehicle tracing, like the ones we reported in this paper, can lead to unreliable input features and consequently uncertain prediction. Accurate data exploration can instead improve its quality and then overcome the initial cost of implementing redundant monitoring strategies with all the benefits, both in terms of cost and environmental impact reduction, an accurate failure detection could provide.