Predicting future vehicle usage based on the analysis of CAN bus data is a popular data mining application. Many of the usage indicators, like the utilization hours, are non-stationary time series. To predict their values, recent approaches based on Machine Learning combine multiple data features describing engine status, travels, and roads. While most of the proposed solutions address cars and trucks usage prediction, a smaller body of work has been devoted to industrial and construction vehicles, which are usually characterized by more complex and heterogeneous usage patterns. This paper describes a real case study performed on a 4-year CAN bus dataset collecting usage data about 2 250 construction vehicles of various types and models. We apply a statistics-based approach to select the most discriminating data features. Separately for each vehicle, we train regression algorithms on historical data enriched with contextual information. The achieved results demonstrate the efectiveness of the proposed solution.
Vehicle usage prediction is an established data mining problem,
which has found application in both industrial and academic
contexts. Since approximately one third of the energy usage is
due to transportation [
Thanks to the advent of Controlled Area Network (CAN)
standards and Internet-of-Things technologies, vehicles are nowadays
equipped with smart sensors and tracking systems that capture
and transmit high-resolution and multivariate time series data
regarding fuel consumption, vehicle movements (e.g., accelerations
and drifts), engine conditions (e.g., oil temperature), and route
characteristics (e.g., slope). A relevant research efort has been
devoted to analyzing CAN bus data by means of supervised
Machine Learning techniques [
Due to the peculiar characteristics of the environment (e.g.,
construction sites, rural areas) and to their context of use,
industrial and construction vehicles show fairly heterogeneous and
hardly predictable usage patterns. Hence, a smaller body of work
has been devoted to predicting their future usage. Most of the
presented works in this field targeted very specific challenges.
For example, the works presented in [
This paper addresses the automatic prediction of the daily utilization hours of industrial and construction vehicles belonging to multiple types and models. It presents a real case study performed on a 4-year real dataset collecting CAN bus data of 2 239 industrial vehicles with various characteristics working in construction sites placed all over the world. The aim of the study is to help site managers to properly schedule short-term fleet management and maintenance actions (e.g., schedule refueling). To address utilization hours prediction, we apply regression models trained on past vehicle data. Training data consist of CAN bus data (engine rpm, oil temperature, fuel level) enriched with contextual information (e.g., day of the week, season, location). Since the analyzed time series shows non-stationary and rather heterogeneous usage trends (independently of vehicle type and model), we train regression models separately on each vehicle. To tailor prediction models to the most discriminating vehicle characteristics, we apply a statistics-based approach to select the most relevant data features. This reduces the bias due to the presence of uncorrelated variables and thus allows to focus the model training on the most salient information.
The rest of the paper is organized as follows. Section 2 describes the dataset. Section 3 formalizes the problem addressed in the paper and describes the methodology we adopt. Section 4 summarizes the main experimental results, while Section 5 draws conclusions and discusses the future research directions. 2
This study focuses on analyzing telematics data acquired from industrial vehicles used in heterogeneous contexts (e.g., construction sites). The data was provided by Tierra S.p.A. Tierra is a company that provides telematics solutions for tracking vehicles of multiple vendors.
Data description. The study encompasses a quite large set of heterogeneous industrial vehicles. Overall, we analyzed data related to 2 239 vehicles belonging to 10 diferent types and located in 151 diferent countries spread all over the world. The dataset collects approximately 4-year data (from January 2015 to September 2018). For each vehicle we consider diferent data characteristics. The main classes of data features are enumerated below.
• CAN bus information (e.g., Engine ON/OFF, CAN paramet
ric messages, Diagnostic Messages, and status reports); • Vendor information (e.g., unit/asset info, maintenance ser
vices); • Information provided by embedded devices installed on
board (e.g., digital inputs report); • Contextual information – Spatial information (geographical location of the vehicle,
region, country) – Temporal information (time stamp, day of the week, holiday/working day – depending on the country, week of the year, month of the year, season, year)
CAN messages are generated by the vehicle sensors and the Machine Control Systems at a high frequency (up to 100 Hz) and gathered by a controller, where they are collected and preprocessed. The system then sends an aggregated report to a centralized server every 10 minutes. For each vehicle, the report contains a set of data features describing the engine and vehicle statuses, e.g., fuel level, engine oil pressure, engine coolant temperature, engine fuel rate usage, speed, working hours, percent load, digging press, pump drive temp, oil tank temperature, etc.
Based on acquisition time and number of acquired samples we derive the daily utilization hours for each vehicle.
Data preparation. To prepare CAN bus data for the Machine Learning process, we apply the following preparation steps: (i) Data cleaning, to handle missing values, make data formats uniform, and verify the absence of data inconsistencies. This is particularly important since vehicles operate in remote regions where the sudden absence of connectivity may afect data collection. (ii) Normalization, to normalize the values of continuous features in order to make them comparable with each other, (iii) Aggregation, to aggregate feature values on a daily basis. (iv) Enrichment, to enrich CAN bus data with multiple-level and multi-faceted contextual information, and (v) Transformation, to tailor input data to a relational data format.
Vehicles are characterized by a unique identifier (the Vehicle id) and are classified based on the type of construction vehicle (e.g., refuse compactor, single drum roller, tandem roller, coring machine, paver, recycler, cold planner, and grader). Each type is then split into several models (i.e., a type subcategory). For example, the dataset contains vehicles of 44 diferent models of refuse compactors, 65 models of single drum rollers, 10 models of recyclers, etc. Finally, the dataset contains data for multiple units for each model. CAN bus data has been enriched with contextual information, to allow the exploitation of seasonality and short-term periodic trends to enhance prediction accuracy.
For example, for most of the vehicles located in the northern hemisphere, the number of days in which they were unused was maximal in December and January due to Christmas holidays and unfavourable weather.
Data characterization. Data characterization is instrumental to discover similarities and diferences among vehicle usage patterns. We focus on the number of hours a vehicle is active each day, for the whole considered time period. We remove the days where we did not record any usage. In a Cumulative Distribution Function (CDF), a curve value F (x ) indicates the fraction of days where the number of daily utilization hours are less than or equal to x . Figure 1(a) shows the empirical CDF of vehicle usage. The plot highlights the heterogeneity of the vehicle usage distributions across diferent types. For instance, graders and refuse compactors are used more than 6 hours per day in median, whereas the coring machines showed opposite usage patterns, with a median usage of less than one hour. Some vehicle types expose a long tail in the CDF, meaning that they are sometimes working up to 24 hours per day.
Within specific vehicle types and models, vehicles still show highly variable usage patterns. For example, Figure 1(b) shows the boxplots of the utilization hours for all the 44 models of type refuse compactor (that is the mostly used vehicle type).
Models are sorted in ascending order according to their median utilization. The boxplots display the full range of variation (from minimum to maximum), and the first, second and third quartiles.
Values with + marker are classified as outliers (deviation of more than 1.5 times interquartile range from the first and third quartiles). The plot confirms a large variance of the utilization hours across vehicle types. Figure 1(c) deepens the analysis into specific vehicle units of the refuse compactor model. For each vehicle, we analyze the utilization hour series across single units. The results indicate that even within the same model, units have very diferent usage patterns. Finally, Figure 1(d) plots the weekly utilization hours series for five vehicle units at random. Daily patters are even more uncorrelated and noisy. Despite the units being of the same type and model, usage patterns shows non-stationary and uncorrelated trends.
This large variability suggests us to address the prediction problem by training a per-vehicle regression model. Building a model for a vehicle type or model would results in a too generic approach. (a) Cumulative Distribution Functions of number of daily utilization hours per vehicle (b) Boxplots of number of daily utilization hours for diferent models of a single type. Inactive days are removed. vehicle type (all refuse compactors) (c) Boxplots of number of daily utilization hours for single units of a specific model (d) Time series of weekly utilization hours for 5 diferent single vehicles of a specific of a refuse compactor type. model of refuse compactor.
This section formalizes the problem we address in this study. We describe the methodologies to generate per-vehicle training datasets, filter uncorrelated features from training data, and train the regression models.
Problem statement. Let Htx be the daily utilization hours of vehicle x on day t . The classical univariate series forecasting problem entails predicting the utilization hours on the n1e,.x.t.,dHayxHtx+1 based on the series of historical values Htx , Htx− t −w +1 within a time window of size w. Since utilization hours are likely to be temporarily correlated with each other, we consider them as a function f of the most recent values within a specified time window TW – hereafter denoted as training window. Formally speaking,
Htx+1 = f (Htx , Htx−1, . . . , Htx−w +1) where Htx+1 is the value of the target variable and f(·) is the prediction function we need to define.
Let F be the set of CAN bus and contextual features stored in the relational dataset (according to the data description in Section 2), and let Ftx be the value of an arbitrary feature F ∈ F (e.g., engine oil pressure) associated with vehicle x on day t . We can extend the classical forecasting problem to a multivariate context by considering not only the historical values of the utilization hours themselves, but also those of the other features in F . Formally speaking, the problem can be formalized as follows: Htx+1 = f (Htx , Htx−1, . . . , Htx−w +1, Ftx , Ftx−1, . . . , Ftx−w +1, . . .) According to our preliminary data exploration, most vehicles are used only for few days a week (e.g., vehicles of type refuse compactor were used 36% of the days in 2017). For this reason, we investigate two variants of the multivariate problem: • Predict the utilization hours on the next day (next-day scenario, in short), • Predict the utilization hours on the next working day, i.e., the next day on which the vehicle will be used at least 1 hour (next-working-day scenario).
Training data generation. Given dataset D, target vehicle x , and training window TW , in this phase we build a per-vehicle training dataset Tx by applying the sliding window approach. More specifically, a smaller window SW slides over the entire training window TW to capture interesting temporal correlations among close utilization hours. Each slide generates a record in the training dataset. For instance, if we set the window size w = |SW | to 7 to capture weekly periodicity, we obtain |TW | − 7 training samples, i.e., we can have up to |TW | − 7 windows containing the next targeted day and the previous 7 days. In a nutshell, each record contains a target value of utilization hours on arbitrary day t + 1, and the feature values associated with each of the previous 7 days, i.e., t − 6, . . ., t .
Statistics-based feature selection. Since the considered vehicles
show variable usage patterns, the correlation between the
features in the training dataset and the target feature Htx+1 is likely
to change from vehicle to vehicle. To tailor prediction models
to the most discriminating features, we adopt a statistics-based
approach to filter training features prior to regression model
learning. Given a vehicle x , we compute the autocorrelation
function [
Figure 2 shows an example of autocorrelation function associated with the utilization hours of a specific refuse compactor unit. Obviously, the autocorrelation value is maximal with lag equal to 0, i.e., comparing the series with itself. In the example, the autocorrelation function is able to capture a weekly periodicity (i.e., high autocorrelation value at l =7, 14, 21, . . .). Similarly, the day after (l = 1, 8, 15, . . .), and before (l = 6, 13, . . .) also exhibits a high correlation with the target day. Other days instead are marginally correlated.
To exploit this correlation, we pick the K lags l ∈ [1, w] with maximal autocorrelation value, which correspond to the K days that are mostly correlated with the target day. Then, we select the features in the training dataset corresponding to these days. Formally speaking, let t ∗, t ∗∗, . . . be the subset of selected days. The problem is reformulated as follows:
Htx+1 = f (Htx∗ , . . . , Htx∗∗ , . . . , Ftx∗ , Ftx∗∗ , . . .)
Regression model learning. To finally build the regression model
f (·), we rely on standard approaches. Specifically, we focus on
the following algorithms: (i) Linear Regression (LR), (ii) Lasso
Regression, (iii) Support Vector Regression (SVR), and (iv)
Gradient Boosting (GB). The last approach is an ensemble method (i.e.,
an ensemble of decision tree models) [
This section summarizes the most relevant results obtained by applying the explained methodology on the Tierra dataset. We conducted an extensive experimental campaign to (i) analyze the impact on prediction outcome, (ii) test multiple algorithm configurations to identify the best settings in diferent scenarios, and (iii) estimate the prediction errors to get confidence intervals for the estimations, and (iv) use the best obtained models on vehicles of diferent models and types. Due to the lack of space, hereafter we will report a selection of the most significant results.
The experiments are performed on an Intel(R) Core(TM) i78550U CPU with 16 GB of RAM running Ubuntu 18.04 server. 4.1
To evaluate the performance of the proposed approach we apply a
hold-out validation. We adopt two established strategies for time
series forecasting, i.e., sliding window and expanding window
methods [
The expanding window strategy difers from the sliding window one because the training window at Step (1) is not fixed-size, but it includes all the preceding days in the original dataset (see Figure 3).
To assess the quality of the prediction outcomes at Steps (5) and (6), we computed the Percentage Error (PE) between predicted and actual utilization hours:
P E = 100 · Ín i=1 |H pir ed − H aictual | Ín
i=1 |H aictual | 4.2
For each algorithm we run a grid search to fit the model to the
analyzed data distribution. The selected settings are summarized
below. More details on the algorithm parameters are provided
by [
• Lasso: α =0.1 • Support Vector Regressor (SVR): kernel = rbf, C = 10, ϵ = 0.1, γ = 1 • Gradient Boosting (GB): learning rate=0.1, n_estimators = 100, max_depth = 1, loss = lad • Baseline - Predict the Moving Average (MA): moving average period = 30. 4.3
We first analyze the efect of varying the number of considered previous days (K ) on the average Prediction Error. Since its efect is influenced by the sliding window size w, we jointly analyzed the two efects. Figure 4 reports results, with one curve per different value of window size w. The results show that • A smart feature selection yields up to 10% improvement in terms of PE. • The optimal number of considered previous days K ranges between 10 and 30. • The more previous days you consider, the more features you add in the training dataset. • Focusing on a limited number of days (< 10) makes prediction models sensitive to noise and data overfitting. • Including a very large number of features in the training dataset may significantly increase the complexity of the learning generation phase. • The more training data you get (i.e., larger w), the more robust the model (except for very small K values, for which the robustness of the generated models is not guaranteed). • Expanding the training window (i.e., increasing |SW | on all past data) performs better, but at the cost of additional computational complexity (not reported - training time grows with data).
To balance model accuracy and complexity, hereafter, we will set K = 20 and w = 140 (unless otherwise specified).
We now compare the performance of diferent regression models. Figure 5 summarize the results in two considered scenarios (Next-day and Next-working-day). As expected, Machine Learning approaches perform better than baseline strategies in both cases. Within each strategy, single and ensemble methods achieve similar percentage errors, with SVR performing comparably to Gradient Boosting. Learning an ensemble of multiple models as such does not yield significant performance improvements.
In the Next-working-day scenario prediction errors are much better - approaching 15% vs. 30% average error in the Next-day scenario. This because removing non-working days simplifies the forecasting problem, being the latter almost randomly present. To exemplify this efect, Figures 6(a) and 6(b) plot the actual and predicted series in the two scenarios. In the Next-workingday scenario the predicted curve appears to fit better the actual series thanks to the absence of less-predictable non-working days. Hence, in the contexts in which holidays/non-working days are known in advance, adapting the raw data to this simpler scenario yields significant advantages in terms of model accuracy. 4.5
We measured the execution time of the applied methodology, which includes (i) Data preparation and feature selection, (ii) Model training, and (iii) Model application. Learning the regression models, i.e., Step (ii), turned out to be the most computationally expensive task. According to the performed experiments, the time spent in accomplishing the other tasks is negligible.
The overall execution time taken by single regression models (trained with the settings recommended in Section 4.2) varied between few seconds for simpler models (MA, LV, LR, Lasso) and few tens of seconds for the most complex ones (SVR). The time spent by ensemble methods (i.e., Gradient Boosting) was approximately one order of magnitude higher than single models. However, as discussed in Section 4.4, combining multiple models did not provide significant performance improvements. 5
The paper describes our experience in using supervised regression techniques to predict industrial vehicle usage based on the analysis of CAN bus and contextual data. The presented case study corroborates previous studies focused on specific construction vehicle types by considering a broader set of vehicle types and models. To efectively cope with a heterogeneous set of vehicles, we selected ad hoc feature sets tailored to each vehicle prior to model learning. The generated models appeared to be quite efective (i.e., with 15% error) in predicting vehicle utilization hours on the next working day. Without any apriori knowledge about the days of idleness, prediction errors on average double, but for many vehicle types and models it was still possible to (a) Error distribution. Scenario: Next-day. (a) Scenario: Next-day: idle days are hard to predict. (b) Scenario: Next-working-day: working hours are easier to predict. accurately forecast non-stationary trends. Future developments of this research will entail the integration of additional contextual information (e.g., weather) and the use of classification models to predict discrete usage levels.
The research leading to these results has been funded by the SmartData@PoliTO center for Big Data and Machine Learning technologies.