Printer Friendly

Real-time Sensing of Particulate Matter in a Vehicle Exhaust System.

INTRODUCTION

Particulate filters are now commonplace throughout the world for emissions control in automotive, light, and heavy duty diesel vehicle applications for the reduction of particulate matter (PM) in the vehicle exhaust. Global regulations have continued to tighten the allowable PM emissions as shown in Figure 1, while simultaneously broadening the scope of the regulation to include gasoline fuelled vehicles. Tier 3 regulations which phase in from 2017 to 2025 allow only a small fraction of the PM and [NO.sup.x] emissions allowed by the Tier 2 standards. Simultaneous reductions in engine-out emissions and improvements in aftertreatment performance are required to meet the new regulations.

As with other vehicle emissions-related devices, diagnosis of the particulate filter is typically required. The diagnostic threshold limits continue to be reduced making diagnosis accuracy, repeatability, and sensitivity increasingly important. Figure 2 illustrates how the On-Board Diagnostic (OBD) thresholds are decreasing thereby necessitating an increase in the required diesel particulate filter (DPF) efficiency of a threshold filter, making the pass/fail diagnostic window narrower.

Several different techniques or methods of diagnosing DPF's have been developed, each with its own unique characteristics and capabilities [2]. One such method of diagnosing the particulate filter is through the use of a PM sensor that is placed downstream of the filter and detects PM that has passed through the DPF. PM emissions that are higher than the normal range would be indicative of a failed or malfunctioning particulate filter or engine system. PM sensor designs are typically either an accumulating style sensor (in which soot accumulates on an internal pair of electrodes) or are based on measuring the flow of charged particles. A general description, principle of operation and performance summary of various styles of PM sensors is provided by Khalek et. al. in [3].

The focus of this paper is on the development of improved algorithms for interpreting and controlling the output of an accumulating type, resistive style PM sensor as shown in Figure 3. The electrodes for this style sensor are closely spaced on a non-conductive substrate. The carbonaceous PM particles are conductive and cause the resistance between the electrodes to decrease as the particles deposit onto the electrode surface. The equivalent electrical circuit for the sensor is shown in Figure 4. The sensing element includes a bias resistor (which is in parallel with the substrate and soot resistance), an integrated heater near the sensing electrodes, and integrated resistive temperature device (RTD) element sensing the temperature of the sensor/soot deposit.

Control electronics monitor this resistance and periodically trigger a sensor regeneration event based on reaching a pre-defined resistance threshold. Sensor regeneration is achieved by heating the PM deposit to a temperature above its oxidation point using an electric heater contained within the sensor element. The time between regenerations is termed the "cycle time" or "response time" and is correlated to the average soot flux occurring in the exhaust pipe during that time period. For a detailed description of this specific sensor construction, principle of operation, control electronics and performance characteristics please see Husted et. al [4].

OBD regulations specify that the diagnostic methods used for verifying DPF efficiency are capable of detecting a failed or degraded filter which causes the vehicle to exceed a threshold value during a defined test cycle. Summary information regarding emissions and OBD requirements for passenger cars, light duty, heavy duty and off-highway vehicles is available on the Delphi website [5, 6]. This threshold PM emissions level is correlated to the response time of the PM sensor such that a response time less than a minimum value is interpreted as an underperforming filter. Varying driving conditions will naturally cause changes in the amount of particulate matter generated by the engine and subsequently emitted at the tailpipe. A model of engine-out soot flux, minimum DPF efficiency and corresponding tailpipe soot flux can be used to generate an expected PM Sensor response time for the varying driving conditions, and thus a threshold response time for the diagnostic. Alternative strategies are also being developed using a virtual pre-DPF PM sensor in conjunction with a conventional post-DPF PM sensor [1].

OVERVIEW OF PM SENSOR OPERATION AND DESCRIPTION OF NEW ALGORITHMS FOR INTERPRETING SENSOR SIGNAL

In the absence of soot, there is a very high electrical resistance between the electrodes of sensing element (see Fig. 5). As soot accumulates on the surface of the sensor between the electrodes, the soot particles begin to overlap and bridge the gap between the electrodes (kerf). Because the soot particles are electrically conductive, the resistance between the electrodes decreases, as does the effective sensor resistance, and is measured by the control electronics. The authors have found that it is beneficial to analyze and interpret the sensor signal in the conductance domain rather than the resistance domain to simplify computations (example will be provided later in the paper). Equivalent sensor conductance can be expressed by eq. (1):

[G.sub.sensor] = [G.sub.bias] + [G.sub.substrate] + [G.sub.soot] (1)

The sensor's operational cycle consists of two parts:

* Collection and detection of particulate matter during the sensing time period.

* Sensor "regeneration" in which the collected soot is oxidized at high temperature and thus removed from the sensing portion of the sensor.

During the sensing time period, there are two distinct phases of the sensor's electrical conductance behavior: a "deadband" phase when the sensor's conductance does not noticeably change even though the sensor surface is collecting soot, and an "active" phase when the sensor's conductance increases noticeably in response to increasing soot loading. The remainder of the time which defines a single complete "cycle" is the regeneration time in which the soot deposit is removed from the sensing element. Definition of a single cycle is depicted in Figure 6. The sensor's conductance reading representing the conclusion of the deadband is chosen at a point where the conductance trace clearly is responding to soot accumulation. The conclusion of the cycle and onset of regeneration is limited only by the amount of soot the regeneration process is able to handle.

It is worthwhile to notice that, to-date, "cycle time" (also referred to as "response time" or "sensing time"), i.e., time interval between regenerations (or strictly the combined time of deadband and active zones) is used for the DPF diagnostic, since the "cycle time" clearly correlates with the soot flux rate in the exhaust pipe (see e.g., [7]). For the purposes of this paper, "cycle time" excludes the sensor regeneration time as a matter of convenience.

In order to more fully understand the behavior of the sensor, simulations of sensor response to particle deposition were conducted [3]. Figure 7 represents the sensor response ([G.sub.s]) to various rates of particle deposition, representing corresponding exhaust system soot flux rates. For each deposition rate, 20 identical sensors were simulated, and the results combined into one plot. Clearly, even in the absence of ever-present changes in the exhaust gas temperature, velocity, etc., there is variation in the sensor response. In particular, neither the conclusion of the deadband nor the onset of regeneration is uniquely identified within a single particle deposition rate selected for simulation. This is due to the randomness of the deposition of individual soot particles on the sensing element, which, in turn, generates variation in the sensor conductance as the particles unite to form conductance paths across the sensor electrodes.

In addition, as has been observed by others (e.g., [8]), the sensor's output signal can be corrupted by soot re-entries into the exhaust stream (blow-offs), and sudden signal jumps associated with unusually large particles strikes. Also, it has been observed that deposit instabilities can exist that cause intermittent fluctuations in sensor conductance. Consequently, the unprocessed "raw" sensor conductance provides, in practice, the "cycle time" upon which a potentially less reliable estimate of soot flux is correlated. In short, further improvement in the data processing and interpretation of the sensor output signal, in the light of more stringent regulations, is highly desirable.

The first, clearly desired improvement, is to remove parasitic soot blow-offs and sudden, large particle signal steps corrupting the conductance trace. This substantially improves reliability of cycle time-based diagnostic whenever these anomalies occur. Moreover, it pre-conditions the signal for subsequent real-time soot concentration estimate and possibly would allow for implementation of PM sensor-based diagnostics assessing transient engine-out particulate emissions. The technique, which quite successfully handles the sensor's conductance signal anomalies associated with large particle strikes, blow-offs, and soot deposit instabilities is presented next, and thereafter the newly developed PM-controller algorithms and their performance is discussed.

Anomalies Filter and Large Particle Sizer

Study of large particle impacts on the shape of the sensor conductance trace indicates that only particles of a size on the order of the distance between sensing electrodes have a substantial impact on the sensor conductance. Depicted in Figure 8 are the sensor resistance traces (upper plot) and corresponding conductance traces (lower plot) which represent a simulation of periodic random deposition of a large particle (with size as listed in the legend of the figure) during the baseline random deposit of normal size particles, plus simulation of only normal-size particles (with no large particle strikes). The sensor resistance plot illustrates the dramatically different effect that identically sized large particle strikes create in the resistance domain. When the same data is plotted in the conductance domain, the steps caused by the large particles are substantially equal, regardless of the status (conductance) of the sensor. This illustrates the reasoning behind the authors' decision to process the sensor signal in the conductance domain. Clearly, the simulation indicates that large particles representing sizes of up to 50% of kerf width do not substantially corrupt sensor traces as the sensor response is practically identical to those calculated in the absence of large particles. Furthermore, the simulated analysis of soot deposition progress, in the presence of periodic but occasional large particle strikes, indicates that the volume of all particles accumulated at the sensor surface at the onset of regeneration remains nearly constant if large particles are not greater than approximately 80% of kerf width (see Fig. 9). This suggests that larger than normal particles, as long as they do not approach the size of the kerf, induce an effective increase in the sensor conductance representing true incremental growth of the soot deposit. When the large particle dimension approaches the size of the kerf, an individual large particle strike can create a step-like change in sensor response that is not proportional to the increase in the mass of the deposit, and thus corrupt the process of translating cycle time into the soot flux estimate.

Substantial statistical separation of what is considered to be normal soot accumulation versus large particle strikes, allows for identification of the critical threshold in the conductance time differential indicative of large particle strike events (see [9]). The filter's logic is based on the assumption that the differential between two subsequent readings of the sensor conductance signal cannot be larger than a certain pre-defined level if compared to the previous differential, or if compared to a mean (or median) of some arbitrary selected number of previous readings; otherwise it is flagged as being an anomaly. Once the violation of a certain calibrated threshold level is detected, the large particle mass is estimated based on the magnitude of the differential, and stored in the event accumulator, available for the calculation of PM mass for the duration of the vehicle drive cycle. The amount of error introduced by large particle strikes is dependent not only on the size of the particle, but also the position into the overall cycle at which the particle strikes. If it strikes near the end of an otherwise normal cycle, the effect is minimal; but if it strikes during the deadband or initial portion of the active zone it is much more pronounced. The example shown in Figure 9 indicates over 50% error for particles approaching the size of the kerf (see the diving slope of the data points representing volume of all particles). The large particle filter algorithm reduces this error to less than 15% under the same conditions.

A particle blow-off condition accounts for another anomaly causing an error effect, one in which a step-like decrease of the measured sensor conductance occurs due to particles becoming dislodged from the sensor surface. While large particle strikes are expected to be rare, unusual events, it is expected that small blow-offs are a part of almost every monitoring cycle. Kinetics of the soot particle impacts and the tangential component of the exhaust gas velocity can be expected to, at least occasionally, lead to soot re-entry into the exhaust stream with the rate most likely affected by the adhesion of the newly-settled soot particles to the substrate. However, if all negative differentials of the conductance were "flagged-out" as blow-offs, electronic noise would be misinterpreted as minute blow-offs and, therefore, create erroneous corrections. Consequently, a threshold value defining a single blow-off event must be established with the consideration of the presence of electronic noise (see [9]).

The action of the large particle / blow-off conductance corrector algorithm was captured during dyno experiments and is shown in Figure 10, where the raw sensor conductance represented by the blue trace is depicted together with the filter-corrected purple trace. The corrected (reconstructed) conductance trace is subsequently used to extract soot flux information and also used to trigger the conclusion of the deadband and onset of the regeneration. Note that the cycle without large particle strike or discernable blow-offs remains unaltered.

For some not yet fully identified reasons, the impact of a large particle is sometimes followed by a subsequent partial separation. Alternatively, vibration of a delaminated "flake" is recorded as alternating strikes of large particle and blow-off events. These types of events lead to an oscillating conductance trace that can corrupt the extraction of soot accumulation information. Nearly coincidental detection of blow-offs and large particle strikes of similar severity, however, can potentially be used as a warning signal to initiate the sensor regeneration procedure (cycle abort).

In the summary, filtering large particles and blow-offs provides the following functionality:

* Detection of large particle strikes

* Large particle mass sizer and accumulator

* Correction of errors in sensor conductance caused by large particles

* Detection of PM blow-offs from sensor

* Correction of errors in sensor conductance caused by blow-offs

* Detection of simultaneous large particle and blow-off events signaling instability of sensor output which may initiate, if necessary, the sensor cycle abort procedure

At the filter output, the corrected conductance trace provides much improved information (when these anomalies exist in the signal) for the core algorithms which then extract soot flux information from the evolving soot deposit as described in the following section.

Basis for Estimation of PM Mass, Flux and Concentration

As a result of the algorithm described in the previous section, the filtered and corrected conductance signal now carries higher quality information regarding the soot flux (mass per unit area per unit time) in the exhaust system. New algorithms which have been developed and embedded in the PM Sensor Development Controller consist of several independent segments, each of which supports the estimation of the soot mass (or mass flow rate) and soot concentration for various portions of the sensor cycle. This is in contrast to existing production algorithms which provide response time as the sensor output parameter. Having the sensor output in engineering units that correspond to the emissions regulations will simplify the interpretation of the sensor signal, reduce vehicle diagnostic calibration time and provide direct correlation to the pollutant being controlled. As was stated earlier, an individual full sensing cycle includes a regeneration period at the end of each "active cycle", which conditions the sensor for the next cycle. The identification of the start and conclusion of each time sector provides the set of gates triggering the logic of individual segments of the overall algorithm. Consequently, the new algorithms provide information earlier in the sensor cycle, thereby allowing diagnostic decisions to be made sooner and more frequently.

The corrected conductance signal also provides the opportunity to extract real-time, instantaneous soot flux, concentration, and mass flow rate during the active zone of the cycle. Cumulative soot mass, mass flow rate, flux and concentration are all related via exhaust velocity, pipe cross-sectional area and time.

Triggered Mass Estimators (1)

Since the cycle starts at the conclusion of the regeneration procedure with a clean sensing surface and is concluded when triggered by a specific conductance reading, one may assume that a specific sensor conductance is represented by the same mass of soot accumulation, [M.sub.s], for steady state exhaust gas flow, i.e., constant velocity and temperature.

Now, since the soot concentration distribution over the cross-section of the exhaust pipe should be uniform for a properly located sensor, the mass of soot, [m.sub.s]. accumulated over the time At at the sensing element should be proportional to the amount of soot, [M.sub.s], measured in the full exhaust flow. The time period At is defined as the cycle time.

Conversion of time into soot mass, [M.sub.s], at constant exhaust velocity, u, and soot concentration, [[rho].sub.s], may be described by the following equation:

[M.sub.s] = [[rho].sub.s]xux[A.sub.pipe]x[DELTA]t (2)

where [DELTA]t is time between two arbitrarily defined time points, and [A.sub.pipe] is the cross-section of the exhaust pipe.

Exhaust velocity may be obtained directly from the engine controller, or calculated using exhaust mass flow rate and temperature (to adjust for density variation) along with pipe area. In the above equation, if velocity is not constant, a fair estimate of the total soot mass passing through the exhaust would be provided by replacing velocity with the mean velocity, u, over the At time period, namely:

[M.sub.s] = [[rho].sub.s]xux[A.sub.pipe]x[DELTA]t (3)

In other words, one may assume that if all variables in equation (3) are constant, a laboratory grade instrument monitoring cumulative soot mass in the exhaust, always reports same soot mass, [M.sub.s,],, once a specific PM sensor electric conductance heralds the end of the cycle and triggers regeneration, since a specific sensing element soot mass, [m.sub.s], triggers the end of the cycle.

The assumption that the soot mass accumulation, [m.sub.s], at the sensing surface represents identical fraction of the cumulative mass at the exhaust [m.sub.s] is true only in principle, and in reality is influenced by numerous additional factors. Thus, the soot mass [m.sub.s] triggering the end of the cycle will be modified by exhaust gas velocity and temperature which affects soot electrical conductivity. Consequently, calibration constant [m.sub.s] requires temperature and velocity-based modifications. Thus the equation delivering soot concentration from the sensor cycle time can be expressed as follows:

[[rho].sub.s] = [M.sub.s](u,[T.sub.Soot])/ux[A.sub.pipe]x[DELTA]t (4)

where [M.sub.s] is a velocity- and temperature-dependent calibration constant, and [T.subsoot] is the soot/sensor surface temperature.

The sensor functionality can be enhanced by defining additional time periods (again based on conductance thresholds to trigger the end of the time period) which are shorter than the cycle time, but long enough to provide a statistically reliable calibration constant equivalent of [M.sub.s]. The earliest reliable trigger point is the end of the deadband, but other points along the senor conductance trace between conclusion of the deadband and conclusion of the cycle may be selected and will provide a mid-cycle soot concentration estimate (see [10]).

It must be noted at this point, that the formula (4) delivers the estimate of average soot concentration over the time period [DELTA]t, and thus, only periodically provides the soot concentration estimate. This is emphasized in Figure 11, which provides an example of the algorithm's performance in giving the estimate of soot concentration at the end of the deadband and the conclusion of each sensor cycle.

It is important to note, that an aborted cycle induced by a soot deposit instability (2) event precludes an end-of-cycle soot concentration estimate, and thus mid-cycle estimates are the only alternative since the cycle is abruptly shortened and the onset of regeneration occurs much earlier than otherwise expected. Calibration of a mid-cycle equivalent of mass [M.sub.s] can be structured in steps covering the range between the conclusion of the deadband and the onset of regeneration. Creating a densely populated calibration table, and thus providing much more frequent update of the soot estimate, would not substitute for instantaneous, real-time readings of the soot concentration. This is because the equivalent mass is by definition an average. For that reason, the separate segment of the algorithm is dedicated to the direct extraction of the instantaneous soot flux, mass rate and concentration [10].

Real-time Mass Estimator

Clearly, real-time instantaneous reading of soot concentration is impossible when employing sensor cycle time as the indicator of soot concentration. In short, the soot concentration is not reported until the conclusion of the sensor cycle, and the estimate only provides the average soot concentration over the cycle time period. The intention of the real-time output is to provide information on instantaneous basis so that rapid changes on output, such as those expected on transient engine operation, can be observed. This may be particularly useful to vehicle calibrators so that a better understanding of transient PM emissions can be gained, even when the vehicle is not connected to or equipped with emissions measurement instrumentation. Also, the integral of the real-time output facilitates delivery of continuously updated cumulative soot mass as described in a later section of this paper.

The principle of operation of the real-time algorithm subroutine relies on the sensor conductance differential measured on a short (e.g., one second) loop time. Study of the simulation results presented in Figure 7 above reveals that the conductance response is second order relative to particle deposition rate, as defined in eqn. (5).

simulated particle depostion rate = k [d.sup.2]G/[dt.sup.2] (5)

The constant, k, is an exhaust gas velocity-dependent value that is empirically determined. The second derivative of conductance, G, provides a good representation of the combined effects of simulated particle deposit film thickness build along with the bridging effect that particles have as the deposit forms on the sensor electrode. This is applicable only in the early stages of soot film build, which comprises the majority of the active zone of operation of the sensor.

In a physical sensor operating with the existing control electronics there are additional factors involved which contort the second order relationship found on the ideal sensor. Through experimentation is has been found that sensor conductance can be represented as a first order relationship with soot flux using a non-constant scaling factor, k', that is dependent on position within the cycle (conductance) and exhaust gas velocity, as given in eqn. (6). A gain-scheduling technique can then be used to extract soot flux from the conductance differential.

soot flux = k' dG/dt (6)

Real-time soot concentration is obtained by dividing soot flux by exhaust velocity. As with the previously discussed algorithms, corrections for temperature are also required. Reliability of such an approach is substantially increased when the sensor conductance signal is filtered from large particle strikes and blow-off events as described earlier. Correcting for blow-off events estimates the real deposit build-up, while keeping track of the cumulative mass of large particles provided by the large particle accumulator allows for adding that soot mass balance to the cumulative soot mass.

As indicated by the analysis of the performance of the real-time algorithm discussed below, the instantaneous sample relative error in the reading of soot concentration appears large, but the cycle mean remains on target and thus provides an excellent tool when calculating soot accumulation over any given period of time, even in the case of engine transients. Sensor sampling volume, as a percentage of total exhaust flow, is very small and this, along with the stochastic nature of soot particle deposition on the sensor electrodes, contributes to the real-time output variation.

The real-time algorithm is deactivated for the duration of the sensor regeneration and deadband time period. For convenience, during this deactivation the real-time output reports the average concentration experienced near the end of the last cycle, rather than reporting nothing. This provides a visual indicator on how the previous cycle ended, and is a reasonable predictor of the ongoing soot concentration under steady state conditions. An example of the real-time soot estimator behavior is depicted in Figure 12.

As indicated by the results presented in Figure 12 for steady state conditions, the extrapolated signal covering the regeneration and deadband time periods provides a good estimate of the ongoing soot concentration. However, the reported projection might be deceiving if during that period of time a transient occurred.

Deadband Upper-Bound Mass Estimator

Beginning at the onset of the deadband, it is possible to create an upper-bound estimator which provides a maximum value of the average soot concentration at that point in the sensor cycle with the intention to signal that the system is "alive and well" and that the cumulative soot mass and average soot concentration occurring thus far during the deadband is below the maximum number calculated by the estimator (see Fig. 13). As described previously (see equations 2 through 4), at any constant velocity and temperature, the exhaust soot mass necessary to trigger the end of deadband threshold is known. By dividing this exhaust soot mass by the deadband time, the average soot flux and concentration during the deadband period are also known (taking into consideration average velocity and temperature). The upper-bound estimator utilizes this relationship to calculate the soot flux and concentration as if the deadband exit threshold had been reached at the current elapsed time into the deadband (a "fixed" mass divided by a variable, and increasing, elapsed time). This produces a characteristic decay of the upper-bound estimator, until the deadband threshold trigger is finally reached, at which point the upper-bound estimate and deadband estimate are identical as demonstrated in Figure 13. Essentially, during the deadband the PM sensor system cannot report what the cumulative soot mass or concentration is, but it can report what it is not. DPF diagnostic decisions can be made using these upper bound estimators since they demonstrate that the system is not exceeding certain threshold levels.

This feature is especially useful when soot concentrations are very low and the duration of the deadband zone stretches over many minutes or hours. In these cases, diagnostic decisions can be made much sooner compared to waiting for a cycle time result. The other threshold-based algorithms for that period are in a stand-by mode and do not provide any information. Also, the extrapolation of previous cycle performance becomes less reliable due to the significant lapse of time since the end of the previous cycle. The upper-bound estimator concludes its action at the exit of the deadband zone, and always at the conclusion represents the end-of-deadband cumulative soot mass, flux and concentration estimates as is depicted in Figure 13. Note also that this upper bound estimator will remain effective for stop/start micro-hybrid engine operation since the data on the exhaust flow rate (which drops to zero upon engine stop) is utilized in the computations of cumulative mass and concentration.

Cumulative Mass Estimators

Cumulative cycle mass is obtained by adding the real-time cumulative mass to the deadband mass as shown in the bottom plot of Figure 14. In this plot, the extrapolation of the mass flow rate of the previous cycle is used to estimate the cumulative mass during the sensor regen and deadband period. This extrapolation is then corrected at deadband exit (indicated by the vertical steps in the algorithm output). Note that on the first sensor cycle, there is no previous cycle, and therefore the extrapolation remains at zero cumulative mass until the first deadband exit. As expected, the extrapolated data is not entirely accurate, but the cumulative mass from deadband exit through the end of each cycle tracks the instrumentation data quite well. Of course, the upper bound estimate could be used in place of the extrapolated data during the deadband periods to provide a continuous upper limit to cumulative mass. Average flux and average concentration can also be provided in addition to cumulative soot mass during the cycle.

Shown in the middle plot is the engine run cycle cumulative soot mass algorithm output along with the cumulative data from the instrumentation. The algorithm sums the current cycle data with the cumulative mass from all previous sensor cycles during the current engine run cycle. This estimate also correlates well with the instrumentation measurements. Utilizing information on vehicle driving distance (either from an odometer reading or by integrating vehicle velocity), an emissions rate (mass per distance, i.e. mg/mi) may be calculated throughout the vehicle drive cycle.

ALGORITHM PERFORMANCE

The algorithms were tested under a variety of conditions to evaluate their performance. Testing was conducted at the Delphi Customer Technology Center in Auburn Hills, Michigan using a steady-state engine dynamometer equipped with a GM Duramax 6.6L diesel engine. The exhaust system was equipped with a diesel particulate filter along with a valving system to allow a portion of the exhaust flow to bypass the DPF thereby subjecting the PM Sensor to higher soot concentrations. The engine controls were modified to allow production of higher-than-normal soot concentrations and independently variable exhaust flow and temperature. Again, the reference standard used for soot concentration was the AVL Micro Soot Sensor (MSS), whose output was recorded synchronously with the PM Sensor controller outputs. In post-processing, when assessing performance of time -based algorithms, the MSS output was averaged over the individual portion of each cycle representing the deadband and full cycle periods.

Only two different sensors were used during this data collection, so part-to-part variation is not represented in the data. Sensor-to-sensor variation of this sensor design was studied and reported by Husted et.al. in [4]. Note that this analysis presented here represents performance of the mass algorithm over a very wide range of operating conditions. Results cannot be compared with time-based performance of this sensor in the previous publication. The range of conditions tested includes soot concentrations from 0.5 to 10 mg/[m.sup.3], velocities from 5 to 45 m/s, and temperatures from 100 to 350[degrees]C. It must be noted that there are some data points included in the summary where abnormalities exist, either because of disabled instrumentation or because of detected soot deposit instability resulting in the early abort of the cycle not reflected in correction of the soot estimate (mid-cycle calibration did not yet exist). The following section elaborates on these abnormalities. It is expected that the results will improve when these abnormalities are corrected and removed.

Triggered Mass Estimator Algorithm Performance

A comparison of the individual sensor cycle soot concentration estimates of the new cycle-based mass algorithms to the conventional response-time algorithm was made using data collected early in the algorithm development process. The upper plot in Figure 15 portrays this comparison and indicates that the new algorithms provide an improvement in both mean targeting and variation. Mean error of the distribution was reduced from 17% to 3% and standard deviation was reduced from 26% to 20%. It is evident from the histograms that a significant reduction in the number of highly-errant readings has been achieved, mainly attributable to the filtering of anomalies (large particles and blow-offs). The plot also highlights further areas for improvement in errant readings. Variation reduction and improvement in mean centering provides improved statistical separation between a "good DPF" and a "bad DPF" which may allow diagnostic thresholds to be to be set closer to the limit-bad DPF performance. For a discussion on the impact of sensor accuracy and variation (including quantification of the statistical separation of two populations using DMSS) see Husted et al. [4].

Also shown in the lower plot in Figure 15 is a snapshot of data from a cycle-based mass algorithm analysis plot. This data was taken under constant engine and exhaust system operation and constant exhaust composition. Sampling errors occurred with the AVL MSS instrumentation where intermittent low concentrations were recorded that are not valid (see cycles 890 and 980) These events cause the reported error of the new algorithms to be unjustly large as shown in the bottom portion of that plot. The output of the algorithm remains constant whereas the AVL MSS recording dips substantially. The remainder of the cycles exhibit low error. These recording errors negatively affect the performance statistics and are present at a number of additional sampling points. Work is ongoing to find and fix the cause of the instrumentation sampling errors.

The deadband-based soot concentration algorithm performance is given in Figure 16 for a variety of steady state exhaust conditions. The plot in the upper left of the figure provides the algorithm output (in purple) and the AVL MSS output in blue for over 1600 sensor cycles. As can be observed in the plot, the algorithm output tracks the MSS well under most conditions, with a few outlying points of high error. Along the left column of plots, the relative error (in %) is given, along with the velocity and temperature conditions for each cycle. On the right side of the figure is an expanded relative error chart which shows that most of the data is within a +/- 50% band around nominal (>96% with less than 50% error) on an individual cycle basis, with the majority of point being less than +/- 25% error. Also provided are an absolute error chart, and a plot of the individual deadband times. This performance is considered to be quite good (on par with the full cycle performance) and confirms the usefulness and reliability of the new deadband mass-based estimates.

Figure 17 provides the cycle-based soot concentration algorithm performance, again for the variety of steady state exhaust conditions. The plots are arranged in the same manner as the deadband performance plots in the previous figure. Again, the cycle-based algorithm output tracks the MSS well under most conditions, also with a few outlying points of high error. The cycle-based performance also yields a relative error which is usually within a +/-50% band around nominal (>97% with less than 50% error) on an individual cycle basis, with the majority of point being less than +/-25% error. Of particular note are the statistics shown in the absolute error plot (middle plot on right side). The grand total mean concentration error is 0.005 mg/[m.sup.3] which confirms that the response of this algorithm is extremely well centered with respect to the reference instrumentation. As noted previously, some of the high-error points are related to instrumentation sampling errors. Other high error points are related to soot deposit instabilities which had not been properly characterized in this data set. Further improvement in the performance of these algorithms is expected with ongoing work.

Clearly, for the duration of regeneration and deadband, the only information available is the upper bound limit of soot concentration. The duration of the sensor regeneration time interval is on the order of several tens of seconds, and during that time sensor is not active, thus the only information on the status of the soot concentration comes as the extrapolation of the past cycle. What follows is the upper bound estimator action which starts at the conclusion of sensor regeneration, and provides a very pessimistic result during the early stages of the sensor cycle, whereas the extrapolated estimate provides an expected value assuming nothing has changed. This creates uncertainty when calculating cumulative vehicle drive cycle particulate, for example FTP75 (see, e.g., attempt reported by AVL GmbH [11]).

However, since the regeneration time zone is relatively short, an estimate based in the mean of the soot concentration reported for the certain time period prior to the conclusion of the previous cycle, might be fair enough to provide the estimate of the total soot mass accumulation for the duration of regeneration time, especially if the DPF is not malfunctioning. On the other hand, if the DPF filtration efficiency is compromised, the deadband time interval which follows the regeneration becomes short and consequently reports unusually high soot concentration which is further confirmed by the real-time readings updating soot concentrations estimates once per second.

Real-time Mass Estimator Algorithm Performance

The performance of the real-time algorithm under these same steady state operating conditions is shown in Figure 18. Due to the stochastic nature of the soot deposition on the sensing element, the range of variation on the real-time data is larger than that of the deadband and cycle-based data. As shown in the figure, most of the data (reported on a 1 Hz basis from both the PM sensor and the MSS) is within +/-100% error band (with (>98% with less than 100% error). The median error is less than 5%, indicating that the calibration of the algorithm produces a well-centered error distribution.

The performance of the real-time algorithm during transients, and performance of cumulative mass algorithm are provided in the following sections.

Real-Time Algorithm Performance during Transients

The "noisy" nature of the instantaneous readings of the real-time algorithm make the signal look unusable, but on average, it in fact provides a reliable indication of soot mass flux (and thus concentration) in the exhaust system, even in the case of transients. Numerical integration of the real-time output provides a reliable cumulative response as described in the following section. An example of the real-time algorithm operation during a step transient in soot concentration is shown in Figure 19.

To verify the accuracy of the real-time output, a comparison was made of the average of the real-time algorithm to the average MSS output during the active zone. Depicted in Figure 20 are the real-time algorithm soot mass flux readings updated once per second and averaged over the duration of individual sensor cycles. The recording shows over 850 sensor cycles of the output acquired in the dyno experiments with the DPF bypass-controlled soot level.

Cumulative Mass Estimator Algorithm Performance

Proven good performance of the cumulative response (integral) of the real time algorithm, in conjunction with the ability to define the upper bound (or extend the estimates from previous cycle) for regeneration and deadband time intervals, results in the ability of fairly efficient monitoring of the cumulative soot mass over extended periods of time (up to and including the full drive cycle). This ability is demonstrated in the middle plot in Figure 21 which shows the cumulative mass estimate by the PM Sensor algorithm (in red) against the MSS cumulative mass (in black) over an extended time period. Soot concentration is shown in the top plot and each sensor cycle is shown in the bottom plot (algorithm and instrumentation). This is particularly useful when operational conditions are such that a low soot concentration is present in the exhaust stream (which is the usual condition) that translates to long sensor cycle times. The real-time output provides not only the current soot concentration level, but also an additional indicator that the sensor system is still active and functioning properly.

SUMMARY / CONCLUSIONS

Several new algorithms have been developed to improve the accuracy and robustness of PM Sensor output and to increase the amount of information that is extracted from the signal. The detection and correction of large particle and blow offs along with the detection and cycle abort for instabilities improve the accuracy and robustness of the sensor output. The additional algorithms which translate sensor output into soot mass, mass flux and concentration offer a new level of functionality not available previously from this sensor. All of these algorithms have been coded into an automotive grade embedded PM Sensor Development Controller and simultaneously operated in real time on an engine dyno to verify their performance. The mass-based algorithms are complementary and use common data streams and signal processing to minimize the incremental computational burden. The overall flow of the new PM Sensor algorithm is shown in Figure 22 for reference.

The additional throughput, RAM and flash memory usage of the new algorithms were accommodated with the existing PM Sensor Development Controller without expansion (this controller is an up-level version of the production controller with slightly more memory and added electrical configuration flexibility). An image of the PM Sensor Development Controller is shown in Figure 23.

Algorithm and calibration robustness have been studied and verified over a fairly wide range of operational conditions (soot concentration, exhaust temperature, and velocity) and offer reasonable performance given the low-cost design of the sensor and controller.

It is envisioned that the new functionality in the mass-based outputs will provide an alternative to the response-time based DPF diagnostic methods. Since the new algorithms report mass, a simple calculation could be made to determine mass per unit distance of vehicle travel (mg/mile for instance) which are the same engineering units as the emissions regulations. This could simplify and reduce the time required for vehicle diagnostics calibration. The real-time algorithms also provide insight into engine and aftertreatment transient behavior.

Further work on calibration improvement is in process to expand operational range and accuracy. Several new sub-routines are being developed to improve overall sensor performance and robustness. The system is also being applied to a non-diesel development vehicle, specifically a Gasoline Direct injection Compression Ignition (GDCI) vehicle running on E10 gasoline. In addition to providing aftertreatment PM emissions efficiency information, it will also be used as a proxy for PM emissions measurement instrumentation prior to the official emissions measurement testing.

Production implementation of the algorithm is under consideration and an evaluation of controller requirements is underway. The increase in processor requirements is relatively minor (as evidenced by the PM Sensor Development Controller performance). No sensor hardware changes were required to achieve the results presented in this paper.

In summary, the newly assembled PM algorithms provide the following services:

* Pre-conditions the sensor electric conductance signal by removing blow-off events and large particle strikes corrupting the ability of reliable signal interpretation.

* Provides the sizing and accounting of large particle strikes for cumulative soot accumulation correction.

* Initiates the cycle abort procedure in the case of the soot deposition instability and generates the estimate of soot concentration at the onset of the instability.

* Delivers instantaneous (one second update) soot concentration, mass rate and/or soot flux.

* Provides mid-cycle, independently calculated soot concentration with the frequency of readings proportional to the soot concentration in the exhaust pipe. The higher the soot concentration, the more often the soot concentration estimate is provided.

* Reports cumulative soot accumulation estimate at the end of each sensor cycle.

* Reports cumulative soot accumulation for the duration of the drive cycle.

* Facilitates data logging of soot concentration and mass flux which may assist in on-road development and testing.

REFERENCES

(1.) Bovi, P., "Particulate Matter Filter Efficiency Monitoring Using A Virtual PM Sensor", SAE OBD Symposium, Indianapolis Indiana, September 2016.

(2.) Masoudi, M. and Sappok, A., "Soot (PM) Sensors", DieselNet Technology Guide, Sensors for Engine and Emission Control, https://www.dieselnet.com/tech/dpf_soot_sensors.php.

(3.) Khalek, I. and Premnath, V., "Particle Sensor Performance & Durability for OBD Applications & Beyond", SwRI, 19th ETH Conference on Combustion Generated Nanoparticles, Zurich, Switzerland June 30, 2015, http://nanoparticles.ch/archive/2015_Khalek_PR.pdf.

(4.) Husted, H., Roth, G., Nelson, S., Hocken, L. et al., "Sensing of Particulate Matter for On-Board Diagnosis of Particulate Filters," SAE Int. J. Engines 5(2):2012, doi:10.4271/2012-01-0372.

(5.) Worldwide Emissions Standards Reference Booklet for Cars and Light Duty Trucks, Delphi Automotive Systems, 2016, http://delphi.com/docs/default-source/worldwide-emissions-standards/delphi-worldwide-emissions-standards-passenger-cars-light-duty-2016-7.pdf

(6.) Worldwide Emissions Standards Reference Booklet for Heavy Duty Trucks and Off-Highway Vehicles, Delphi Automotive Systems, 2016 http://delphi.com/docs/default-source/worldwide-emissions-standards/2016-2017-heavy-duty-amp-off-highway-vehicles.pdf?status=Temp&sfvrsn=0.03636262961639791

(7.) Ochs, T., Schittenhelm, H., Genssle, A., and Kamp, B., "Particulate Matter Sensor for On Board Diagnostics (OBD) of Diesel Particulate Filters (DPF)," SAE Int. J. Fuels Lubr. 3(1):61-69, 2010, doi:10.4271/2010-01-0307.

(8.) Bender, D., Peyton Jones J., and Harshbarger, D., "Analysis of Particulate Matter Sensor Signals," SAE Technical Paper 2012-01-0871, 2012, doi:10.4271/2012-01-0871.

(9.) Malaczynski, G., Roth, G., "Particulate Matter Sensor Signal Correction", 2016, U.S. Patent Pending.

(10.) Malaczynski, G., Roth, G., "Particulate Matter Detection System and Method", 2016, U.S. Patent Pending.

(11.) Hoepfner, A. and Roduner, C., "PM Sensor Based On-Board Diagnosis of Particulate Filter Efficiency," SAE Technical Paper 2013-01-1515, 2013, doi:10.4271/2013-01-1515.

CONTACT INFORMATION

Gerard W. Malaczynski

Senior Technologist - Advanced Powertrain Controls

Delphi Customer Technology Center - Michigan

3000 University Drive

Auburn Hills, MI 48326-2356

Mail Code: 483-300-321

Phone: (248)836-0436

gerard.w.malaczynski@delphi.com

Gregory T. Roth

Senior Engineering Manager - Advanced Powertrain Controls

Delphi Customer Technology Center - Michigan

3000 University Drive

Auburn Hills, MI 48326-2356

Mail Code: 483-300-321

Phone: (248)732-1890

gregory.roth@delphi.com

ACKNOWLEDGMENTS

The authors wish to thank Donald Johnson, Dave Goulette, Kristen Kirchoff, Dawood Rangwala and Philippe Bovi of Delphi Automotive Systems for their support in the development of the PM Sensor algorithms and in the generation of this paper.

DEFINITIONS/ABBREVIATIONS

DPF - Diesel Particulate Filter

DMSS - Difference of Means divided by Sum of the Standard deviations

ECU - Engine Control Unit

EMS - Engine Management System FIE - Fuel Injection Equipment

kerf - separation distance between sensor electrodes

MSS - Micro Soot Sensor, AVL List (GmbH), used for instrumentation in dyno lab

[NO.sub.x] - Oxides of Nitrogen

OBD - On-Board Diagnostics

PM - Particulate Matter

RTD - Resistance Temperature Detector

soot flux - mass flow rate per unit area per unit time, (mg/[m.sub.2*]s)

Gerard W. Malaczynski and Gregory Roth

Delphi Automotive Systems

(1.) It may be insightful for the reader to reference Figure 22 ("Overall flow of new PM Sensor algorithms") shown in a later section to better understand algorithm flow and function. (2.) In this paper, soot deposit instability is defined as a repetitive increase and decrease of sensor-measured conductance over a short time period and is conceptualized to be caused by an agglomeration of soot or other semi-conductive material that makes intermittent contact with the electrodes.

doi:10.4271/2017-01-1639
COPYRIGHT 2017 SAE International
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 2017 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Malaczynski, Gerard W.; Roth, Gregory
Publication:SAE International Journal of Passenger Cars - Electronic and Electrical Systems
Article Type:Report
Geographic Code:1USA
Date:May 1, 2017
Words:7708
Previous Article:A Tailgate(Trunk) Control System Based on Acoustic Patterns.
Next Article:SmartDeviceLink as an Open Innovation Platform for Connected Car Features and Mobility Applications.
Topics:

Terms of use | Privacy policy | Copyright © 2019 Farlex, Inc. | Feedback | For webmasters