Predictive maintenance has become an important tool to avoid unplanned downtime of modern vehicles. The exchanged data between Electronic Control Units (ECU) is simultaneously increasing with the functionality. A large number of in-vehicle signals are provided and facilitate the monitoring of physical component aging processes. In this work, we generated a training dataset and observed aging of a selected powertrain component. First, we preprocessed in-vehicle signals to generate a time equidistant signal database. Furthermore, the signals were segmented in various time periods and subsequently aggregated to statistical features. Second, the signal associated aging information were synchronized to an equal time frame. We investigated several signals preselection approaches to predict an aging-value for the powertrain component with machine learning methods. These approaches di er in the count of selected in-vehicle signals for the aging-value prediction. Our results show that in-vehicle signals can be used to predict powertrain component aging. The quality of estimation di ers with respect to the selected regression methods. In this work we present an approach to narrow down the prediction quality of di erent preselection approaches for the estimation of a powertrain component aging.
Various amounts of time-resolved data is recorded in the life-cycle of vehicles.
This data is transmitted from Electronic Control Units (ECU) via an Control
Area Network (CAN) bus of the vehicle. The complexity of modern vehicles
grows rapidly. Many components in the vehicle communicate with each other. A
reliable diagnosis of an potential aging of a component is complex.
Predictive maintenance in a commercial mobility context let the customer know
the current status of his vehicle(s). There is no extensive de nition of predictive
maintenance. Hence, it is de ned in various ways according to its use in
literature [
In this section, we provide background information to predict the aging degree of
a selected powertrain component by using di erent regression methods. We use
machine learning algorithms to create a model of the aging degree. The training
set's target value (ground truth) is given by observing the fouling of the Exhaust
Gas Recirculation (EGR) cooler components in certain intervals in a workshop.
Data-driven diagnostics is applied in the automotive domain to analyze vehicle
components and support manufacturer and vehicle driver decision making. For
example, the On-board Diagnostics (OBD) system monitors fault diagnosis of
vehicle components and noti es the driver regarding the possible malfunction of
vehicle component. This is initially designed to keep the vehicle emissions within
statutory thresholds [
Besides the monitoring of emission limits with OBD systems, other authors use machine learning algorithms to apply data-driven diagnostics in the automotive domain. Machine learning algorithms within the context of regression problems generate a functional dependency between input data and target values, without explicit being programmed for it. Training data based on in-vehicle signal logging is used to t a machine learning model. The learned model is subsequently applied to make predictions on new datasets.
The data-driven diagnostics for component aging prediction presented by
several papers di er in the source and amount of data used for training the model.
On one hand, models are trained with data from special sensors. For that
purpose some authors use vibration sensors or acoustic emissions [
Di erent machine learning algorithms are used for fault diagnosis. The input
data is labeled concerning a fault state or normal state. Support Vector
Machines (SVM) are used to predict the fault state [
EGR Component The speci c vehicle component investigated in this work is analyzed with the help of in-vehicle date from several vehicle prototypes with diesel engines. To keep the legal pollutant emission limit a recirculation of exhaust gas is applied in these prototypes.
At an ideal combustion of sulfur-containing diesel fuel carbon dioxide (CO2), water (H2O) and sulfur dioxide (SO2) are released as resulting products. The air conditions are uctuating strong locally. Due to a non-ideal combustion, nitrogen oxides (NOx), carbon monoxide (CO), hydrocarbons (HC) and particulate matter (PM) are created. The EGR valve controls the recirculating exhaust gas of the engine back into the intake tract.
Nitrogen oxides (NOx) emissions can be reduced by increasing the EGR rate.
In addition to that the implemented system uses a cooler to decrease the gas
temperature. Therefore the peak temperature can be reduced. Another bene t
is the possibility to control and reduce NOx emissions from diesel engines by
decreasing the combustion temperature [
The EGR cooler fouling consists of HC and PM deposits. Due to these deposits,
the ow resistance rises. To reduce the quantity of deposits an additional
catalyst in the EGR line can be implemented [
Data Description The following explanations describe the two data sources used for further analysis and model prediction. The rst data source includes the in-vehicle signals from several prototypes. The prototypes itself were used for road trails especially designed to collect data related to the EGR aging process. The second data source is provided through data collected during workshop visits. This includes the measured degree of the EGR cooling system aging that was measured in certain intervals. The two data sources were originally not time synchronized. The in-vehicle signals from the rst source are recorded from the internal vehicle network (CAN bus) of each prototype. The data is recorded in form of time series containing series of tuples with measurement value and associated timestamps. As described in section 1, the recorded in-vehicle signals are independent and have not the same timely resolution. Owing to this, the signals were synchronized by means of their related time-stamps. We use time-series data with time resolution of 100 ms in time-equidistant form. For some binary signals (e.g. binary status signals) it is necessary not to interpolate between the signal values. For this reason the time stamps of the signals are changed to the given 100 ms grid. In order to train a machine learning model, all data vectors have to be the same size. Thus, the signal length of all analyzed signals has to be the same. At the recording start the engine could be idle and in-vehicle signals are not transmitted yet. After turning o the engine, the recording device is still working for some time and collecting all remaining information on the CAN bus. To secure the same value length over several engine starts, all in-vehicle signals have to be cut according a trigger signal. The trigger signal should be accessible on every prototype over the whole analysis period (see g. 1). The second data > h t signal 1 signal 2 trigger signal source provides the aging degree of each prototype's EGR cooling system. The unevenly spaced time series (target value information) were interpolated to the same 100 ms grid as mentioned above. But in this case, the aging e ect occurs not only while measuring, but during the whole vehicle usage. At the given prototypes the fouling of EGR cooling system increase due to their usage. The aging degree is de ned as mass ow ratio m_ r between the air mass ows of di erent states. While measuring the mass ow ratio the driving state has to be constant, especially the engine speed and torque. For this reason the measurements are operated in certain intervals workshops. The air mass ow is quanti ed while the EGR valve is in a closed and opened state. The equidistant and 100 ms synchronized aging degree is used as target value (ground truth) for further analysis.
m_r = m_fo m_fc (1) m_r : mass ow ratio, degree for EGR cooler aging m_fo : mass ow, EGR valve open m_fc : mass ow, EGR valve closed 3
In this section, we present results for the aging degree prediction of the prototypes. We choose di erent preselection approaches to determine prediction relevant signals (physical and data-driven approach). The modeled aging degrees were compared to the measured ones. The analyzed signals were segmented according di erent time periods (10 minutes to 15 hours). We used the root-meansquare error (RMSE) as performance measure to evaluate the models performance. The RMSE indicates how well the average modeled aging estimation of the respective model is compared to the target value. The smaller the RMSE, the better the model aging estimation.
Data Preselection The in-vehicle signals were preselected regarding three
different approaches. The rst one is the physical approach. In Section 2.2 the aging
degree is given as a function of the mass ow. In a previous work we show results
concerning selecting the right in-vehicle signals to monitor this powertrain
component aging in dynamic working conditions for each prototype. Furthermore an
optimal time period for the data aggregation is given in [
The extended physical approach includes the preselection of relevant signals according the theoretical information given in Section 2.1. The in-vehicle signal preselection for the extended physical approach is supplemented by signals of working conditions and internal cooling temperatures. In total we preselected ten signals. The data-driven approach selects all in-vehicle signals, which are present on all prototypes. As a reason of having di erent engine con gurations and recording loggers the signal selection distinguishes. After the signal intersection there are still more than 130 signals valid for analysis the aging degree.
Data Segmentation In addition to the preselection, the data is segmented in various time periods from 10 minutes till 15 hours. In each time period an associated target value is calculated concerning the given input data. Although the measured value is not as highly time-resolved as the time period, we interpolate the target value for the given time period as long as the aging process between two measurements is continuous. Thus, for each time period the ground truth is provided as aging value and we calculate a modeled aging value with di erent approaches by the given aggregated input data.
Data Aggregation In order to aggregate the data, we evaluate di erent statistical features for each of the signals used in the given segmentation. The selected statistical features are: arithmetical mean, 25th and 75th percentile and the standard deviation of the values in each time period. The statistical features of each data segmentation were used to train a multiple linear regression, Bayesian linear regression and Random Forest regression model.
Implementation In order to estimate the aging-value, we implement a
multiple linear regression, a Bayesian ridge regression and a Random Forest
regression [
"
N (0; 2) (3) A Random Forest regression is related to the decision tree regression. The decision tree is learning the rules of if-else sequences. At the end of these trees, numerical predictions are calculated for those leaves. The Random Forest regression combine several tree decision under a certain kind of randomization to prevent over tting their training set.
The Figure 2 shows the estimated degree of aging for the three di erent regression methods for a selected prototype. For each sample consisting of statistical features an estimated aging degree is calculated using the mentioned regression methods. For the training of the regression model we use the datasets of three vehicles. The dataset of a fourth and fth vehicle is used for the inter-vehicle validation of our approach. The RMSE is calculated for each regression methods and each data preselection approach. The physical and the extended physical preselection approach deliver a quite precise degree of aging for the selected component. The data driven approach is not as good as the other approaches, especially while using the multiple linear and Bayes regression. Instead of prer a e n i L s e y a B t s e r o F m o d n a R senting the estimated aging degree, we compare di erent preselection approaches and various data segmentations in Figure 3. As described above, the model is trained with the dataset of three prototypes. The model is validated with the dataset of two separate prototypes. The Random Forest regression has a quite similar result for both physical approaches for the two selected prototypes. The Random Forest regression has better predictions than the other two for the given input.
Furthermore, the physical selection approach provides a lower RMSE for the multiple linear and Bayesian regression. The RMSE of the extended physical approach is slightly worse than that. All visible regression methods have in common, that the estimation quality gets signi cant higher for a data segmentation above 2 hours. The data-driven preselection approach is not plotted, because the RMSE of that approach is not as precise as the other approaches.
RMSE for various approaches and regression methods 0.11 0.10 0.09 E0.08 S M R0.07 0.06 0.05 0.04 car [’Prototype 2’] [’Prototype 3’] type physical linear physical random forest physical bayes ext. physical linear ext. physical random forest ext. physical bayes 10min 1h 2h 3h 4h 5h 6h 7h 8h data segregation
9h 10h 11h 12h 13h 14h 15h In this work we analyzed dynamic in-vehicle signals regarding the aging process of the EGR cooling system. Our approach enables the possibility to estimate the aging-value of this structural component using di erent preselected datasets. First, the data from dynamic in-vehicle signals are preprocessed to generate a time equidistant dataset. Thus, it is possible to predict the aging degree by using unevenly spaced time series. Moreover, several vehicles with a various set of signals can be utilized to train the model. For each vehicle, the degree of aging is observed in certain time intervals and a training dataset is generated. This aging degree is interpolated to the same time grid as the in-vehicle signals. We provided di erent data preselection approaches in order to enable the aging estimation on a reduced dataset. Afterwards, we segmented the data in several time periods. The aggregated statistical features of the preselected signals were used to train the models. Figure 2 and 3 show that a prediction of a degree of aging for a selected component is feasible by using the right amount of signals. It can be noted that the quality of the prediction evidently depends on the data segmentation and on the preselected signals used for model training. As shown in Figure 3, the quality of prediction is dependent on the selected data preselection approach. The Random Forest regression nds relevant features also in the bigger dataset for the given vehicle.
The preselection approach of in-vehicle signals can be extended in consideration of each signal's relevance for the physical aging process. In this context, the relevance of each signal can be weighted concerning the physical context and used for the model training. The aim is to train the model with a subset of data to get an optimal performance. Our goal was to deliver component aging indicators for the usage of predictive maintenance. Predictive maintenance comprises not only the prediction of component failures. In this context, predictive maintenance tries to de ne the degree of aging by using in-vehicle signals. In particular, for predicting this aging degree relevant signals have to be identi ed. Besides the determination of the aging degree, predictive maintenance is understood as noti cation system, in which part of the vehicle a possible aging process can be detected. For that a list of relevant in-vehicle signals should be generated.
In the future, the waveform characteristics of various component aging processes can be stored. With the help of this characteristics the relevant in-vehicle signals can be detected. Furthermore, the results of the future approach are the list of in-vehicle signals, which are identi ed to be relevant for a given aging process. This signal list indicates, which component aging occurs in the given dataset.