1. Introduction

Modern combustion applications are challenged to lower their emissions and comply with emission restrictions [1,2]. Exhaust aftertreatment systems are commonly used to achieve this goal. In the case of nitrogen oxides the SCR (selective catalytic reduction) system is widely used in both stationary and mobile applications, since it is highly suitable to lower fuel consumption and operational costs [3]. Within SCR systems, an aqueous urea solution with 32.5%_wt urea, also referred to as AUS32 in the ISO-22241 standard, is injected into the exhaust. Ammonia is then formed by thermal decomposition and hydrolysis as the reactant to catalytically reduce NO_x. To achieve high conversion rates of the pollutant while preventing ammonia slip it is of great importance to distribute the urea homogeneously across the exhaust pipe system upstream of the catalyst. Extractive scanning diagnostic methods (e. g. with FTIR (Fourier transform infrared) or mass spectrometers) are currently used in SCR system design to measure spatial ammonia distributions behind the catalyst and gather the upstream urea distribution from this. Due to the long measurement period of such devices, highly stable operating conditions are essential for high-quality measurements. Increasingly important unsteady operating conditions such as engine start or load changes cannot be surveyed. To overcome this drawback, an alternative fast non-intrusive in situ method for ammonia distribution diagnostics is demanded.

Optical absorption-based ammonia concentration diagnostics have been performed previously in various spectral regions [4–7]. Absorption paths in the range of meters allowed for very low detection limits of 0.5 ppm_V and 1.8 ppm_V·m/√Hz in laboratory environments. Two-photon LIF (laser induced fluorescence) offers an in situ possibility for 2-D imaging [8] but has not yet reached detection limits sub 100 ppm_V and is very demanding to apply in an industrial environment. Furthermore, absorption methods have already been combined with tomographic algorithms for imaging 2-D water concentration or temperature distributions using either scanning systems [9] or simultaneous multichannel acquisition [10]. The method presented in this work combines the advantages of in situ TDLAS (tunable diode laser absorption spectroscopy) for ammonia concentration measurement with tomographic methods to make 2-D diagnostics directly in the exhaust of an industrial system possible. To enable fast distribution measurements a multichannel system is preferred to a scanning system.

2. Theory and experimental design

2.1 Absorption spectroscopy

Absorption spectroscopy is a line-of-sight (LOS) method. Assuming a homogeneous path, the absorbance $O{D}_e(\nu )$ of a propagating light beam at frequency $\nu$ is expressed by the Beer-Lambert relation:

A summation at a specific frequency $\nu$ of all contributing species $i$ within the gas matrix and each relevant transition $j$ at a specific frequency $\nu$ yields the combined absorbance probed by a narrowband laser. By tuning the laser across a wavelength region of interest, the absorbance line-area of one or more transitions is measured. Using the ideal gas law, the mole fraction of the species in question is derived from these measurements. The selection of the appropriate wavelength range to be probed and of the utilized laser for the measurement conditions within the SCR process can be found in a previous publication [13].

2.2 Heavy duty engine setup and exhaust aftertreatment

A straight 6-cylinder diesel engine (BR-1600, MTU Friedrichshafen GmbH) was used in the experiments. These engines are typically used off-highway in agricultural machines, construction machines, trains and power generators.

The SCR-System was based on a Bosch Denoxtronic 2.2. The PWM-controlled (pulse width modulation) dosing valve was located 1.8 m upstream the catalyst, followed by a mixing element. A fully extruded vanadium SCR catalyst with a diameter of 275 mm and a total volume of 24.5 dm³ was used, built of two separate bricks of 1/3 and 2/3 of the total catalyst length. The catalyst exhibited 300 cpsi (cells per square inch), 0.305 mm wall thickness, and a vanadium content of 1.7%_wt.

Along with the ammonia spectrometer, an AVL AMA 4000 exhaust gas analyzer was used to measure additional gas concentrations of exhaust species (not NH₃) during operation. The extraction points of the AMA 4000 system were located 2.6 m upstream and 0.8 m downstream the catalyst (see Fig. 1). The TDL-spectrometer was used in between the two catalyst bricks and directly behind the catalyst. Additional ammonia analysis results from an AVL SESAM i60 multi component FTIR spectrometer gathered prior to this campaign will be compared with the TDLAS system’s ammonia concentration measurements later in this work.

 

Fig. 1 Schematic of the internal combustion engine and the test rig configuration. 1&2: turbochargers; 3: intercooler; 4: waste gate; 5: EGR cooler; 6: EGR valve; 7: SCR dosing unit; 8: mixing; 9: catalyst bricks.

Download Full Size | PDF

TDLAS measurements were performed at four different operating points (OP): At a fixed engine speed of 1300 rpm, the injected amount of AUS32 was varied according to Table 1. The dosing factor $d$ is the ratio of the actual amount of ammonia in the injected AUS32 and the stoichiometric value necessary for complete NO_x reduction. The dosing itself is performed automatically by the test rig. For OP4 the engine was operated at a higher load and thus higher gas temperatures occur.

Table 1. Engine parameters at the used operating points

View Table

2.3 Sensor setup

The spectrometer consists of a stainless steel ring that was inserted between two flanges of the exhaust piping. It is 25 mm thick with an inner diameter of 275 mm compatible to the engine’s SCR exhaust system (see Fig. 2). It provides eight direct absorption paths within the flow with path lengths between 247 mm and 275 mm. All eight channels form a measurement plane located 25 mm downstream of the catalyst surface. The spectrometer ring is additionally equipped with 11 thermocouples (mineral insulated thermocouples type K, Ø 0.5 mm, locations shown in Fig. 2) and a pressure sensor (Keller PAA-35XHTC digital pressure transducer 0 – 10 bar) to provide temperature and pressure measurements for the data evaluation.

 

Fig. 2 Schematic of the spectrometer setup.

Download Full Size | PDF

The laser source is a fiber-coupled distributed feedback (DFB) diode laser (nanoplus GmbH) operating at a nominal wavelength of 2201 nm and a maximum output power of 2 mW ex-fiber [13]. To supply light to all eight channels at the same time, a 1 × 8 single-mode fiber optic splitter employing a planar light-wave circuit design was used. For each channel’s emitting side an aspheric beam collimator with a diameter of 1.25 mm fused directly to the end of the fiber was mounted via a stainless steel tube flush with the inner surface of the exhaust pipe. The beams exit the collimators with a divergence of approximately 0.2°.

On the opposite side fused silica lenses (also mounted flush with the inner surface of exhaust pipe) with an aperture of 4.5 mm and a focal length of 10 mm were used to focus each beam into multi-mode fibers with a core diameter of 910 µm. The fibers are housed within a stainless steel tube similar to the SM side. Directly at the end of the MM fiber, a Hamamatsu double-extended InGaAs photodiode with a typical cutoff frequency of 6 MHz is mounted. The output of each detector was amplified by variable gain transimpedance amplifiers (Femto DHPCA100). The amplifier’s bandwidth at the gain used in the experiments in combination with the photodiode’s impedance was 1.8 MHz. The amplified signal was recorded phase-locked by a National Instruments PXI system equipped with two four-channel 4 MSamples/s 16-bit PXIe-6124 data acquisition modules.

2.4 Data processing

The laser modulation settings and data acquisition parameters were identical for all measurements. The laser was current modulated for each scan with a 90/10 asymmetric triangle shape (90% up-ramp and 10% down-ramp) at a modulation frequency of 5,040 Hz. The up-ramp covering a spectral range of 2 cm⁻¹ was evaluated to obtain the concentrations.

Prior to data evaluation, 5000 scans were averaged resulting in a total temporal resolution of 1 Hz. To correct for transmission fluctuations and background emissions on the raw signal, a fourth order polynomial was used. This polynomial correction was mathematically linked with the model of a multi-line Voigt shape using a simultaneous optimization method for fitting the measured spectrum. By coupling the polynomial shape and the line shape the solutions were limited to physically reasonable solutions. Doppler and pressure broadening within the Voigt line shape were calculated based on the HITRAN2012 database [14] using the temperatures and pressures measured during the experiments. Foreign broadening of the absorption lines was calculated assuming air as the ambient gas. However, in exhaust applications the gas matrix differs from pure air and the uncertainty of coefficients at the chosen spectral transition are not documented. This adds a systematic error which is estimated to be approximately 10%. Nevertheless, this method is preferred to leaving the collisional width of the Voigt shape as a fitted parameter, because fitting at low SNR would lead to higher statistical variations in concentration. By relying on possibly improper coefficients and accepting a systematic error which could be diminished by calibration the processing stays comprehensible and more stable. In addition to this systematic error a general uncertainty of 10-20% has to be considered due to the documented uncertainties of line strengths when using the given spectroscopic parameters from HITRAN2012 [14] for this ammonia transition. Measuring line strengths in a laboratory gas cell is one work in progress and would also minimize the overall uncertainty of the presented ammonia sensing device.

2.5 Data filtering

The level of electromagnetic noise in an industrial environment can be orders of magnitude higher than in spectroscopic laboratories. Due to an attenuation of approximately −25 dB by the fiber splitter, the remaining light intensity per channel was very low. To compensate for this low intensity, the gain of the transimpedance amplifiers was set to a rather high value of 10⁵ V/A causing all electromagnetic noise in the amplifier’s bandwidth of up to 1.8 MHz to be amplified as well. To counter the influence of this noise, recorded signals were low-pass filtered before processing. In absorption spectroscopy the Savitzky-Golay filter [15] has been proven to significantly lower noise levels while preserving the shape of absorption lines. The filter’s influence on the absorption line was tested by processing the same signal with and without filtering. To perform this, an almost noise-free signal was recorded by removing the splitter and using the full laser power on only one channel. Using the full laser power allowed for a reduction of the amplification to only 10³ V/A. The signal was recorded during engine operation at OP1 and processed without any filtering, while averaging only 50 scans. This signal was compared to the results of processing the same signal including filtering (see Fig. 3). While the relative deviation of the calculated concentrations was 0.1%, the SNR almost doubled. Since the line shape function is fitted using calculated line broadening, this implies a preservation of the Voigt line shape and not only line area. Based on these results, all further data processing applied a Savitzky-Golay filter to each data set (336 pixels per evaluation window) with a window size of 15 pixels and a 5th-order polynomial.

 

Fig. 3 Comparison of single channel data processed with and without Savitzky-Golay (SG) filtering (window width 15 pixels, polynomial order 5). Data were recorded for OP1, averaging 50 scans, resulting in a 100 Hz resolution. Calculated concentrations are 262 ppm_V with a SNR = 14 without and 261 ppm_V with a SNR = 25 with filtering.

Download Full Size | PDF

2.6 2-D reconstruction via limited data absorption tomography

Within the cross section of the spectrometer, an unknown spatial distribution of ammonia is present for a certain point in time which shall be determined. For this purpose, the measurement plane is probed by eight laser beams transecting the spectrometer’s cross section yielding a total amount of eight path-averaged concentration measurements. Deriving a two-dimensional concentration distribution from these measurements calls for a tomographic reconstruction algorithm that can cope with the severe limitation of the available data.

At first, the measurement plane was discretized into triangular pixels by means of Delaunay triangulation [16] with the help of the MATLAB pde toolbox as in Fig. 4. This method is preferred to the discretization into rectangular pixels, because the measurement plane is of circular shape and discretization errors can be minimized. Furthermore, the concentrations would generally be assumed to be homogeneous within a single rectangular pixel. In contrast to this, we only reconstruct the concentration values of the triangle nodes.

 

Fig. 4 Discretization of the standardized measurement plane in a circular shaped exhaust pipe. The dots mark the 2129 triangle nodes, at which ammonia concentrations are reconstructed. Additionally the 8 laser beams are shown as red lines.

Download Full Size | PDF

Calculating a simulated noisy measurement for each laser beam in case the concentration distribution $x$ is known is mathematically spoken a well-posed forward problem – and therefore easy to solve. The underlying model can be expressed by the following equation:

Now, since we can only measure the path-integrated concentration value along the laser beam paths $c$, an inverse and ill-posed set of linear equations needs to be solved to derive the concentration distribution $x$. According to Hadamard [17], a well-posed problem needs to fulfill three criteria: existence, uniqueness and stability. It is obvious that the uniqueness of the solution is not given due to the underdeterminedness ($\text{M}\ll \text{N}$) of the problem, thus $S$ is rank-deficient and cannot be inverted directly. Nevertheless, lowering the discretization to $\text{M}=\text{N}=8$ will not lead to better reconstructions. Instead, Eq. (2) can be augmented to form the so-called normal equation for inversion:

In this work, the Tikhonov algorithm [22] is used for regularization and reconstruction. The main idea is to find an appropriate approximation to the ill-posed problem of Eq. (4) that is well-posed:

It is necessary to determine $\alpha$ according to the noise level and the experiment itself because it influences the reconstruction directly. The optimization of $\alpha$ and the beam arrangement of the eight beams of the spectrometer was performed simultaneously also considering the physical constraints of the instrument rig and the spectrometer ring. Numerous reconstructions based on simulated measurements of artificial concentration distributions were performed. These distributions were generated by convolution of several Gaussian phantoms, which randomly vary in number, position within the cross section, size (FWHM) and amplitude. Comparison of the reconstructed and the initial distributions yielded best results for $\alpha =2$, which will be used throughout this work. At last, $R$ is realized using the discrete Laplacian operator for the triangulated measurement area and reflects the physical assumption of smoothness as described in Daun et. al. [20]

2.7 Challenges in real exhaust processes

Using an optical spectrometer in real exhaust processes leads to several challenges. The optical access to the exhaust gases must be temperature resistant, chemically resistant, and gas tight. Therefore, all TDL spectrometer parts in contact with the exhaust gases are made of glass, stainless steel or ceramic glue exclusively to be compatible with these conditions. During engine operation, access was prohibited for safety reasons and no alignment of the system could be performed. The spectrometer was therefore pre-adjusted and then installed into the engine. The endoscopic detection probes were designed to handle any distortions resulting from beam-steering and thermal deformation of the spectrometer.

Except for the SCR system, there were no other exhaust gas after-treatment systems. All optics were directly exposed to the uncleaned exhaust gases including cold start conditions which means particulates, water vapor, carbon dioxide, and other remaining components from combustion together with liquid urea wetting the wall when overdosing. The latter can become problematic as it quickly lowers transmission when wetting the optic. Since urea mixed with other debris tends to gather at the bottom of the piping, the detection units rather than the sending probes were placed at the lower hemisphere of the piping. After cooling down the engine, the detection units were easily removed, cleaned ultra-sonically, and mounted again in the spectrometer ring without neither the need nor the possibility to realign them.

As a line-of-sight technique, TDLAS suffers from any inhomogeneities of the probe volume along the path. While deviations of number density along the path simply influence the concentration proportional to the length of the section, temperature inhomogeneities have a nonlinear impact. The used absorption transitions at 2200.5 nm have been selected particularly for their small temperature sensitivity [13], minimizing this effect. To measure the spatial temperature distribution a total of 11 thermocouples were positioned within the exhaust pipe as shown in Fig. 2. Thermocouple locations spanned from the center of the pipe to about 90% of the radius. During stable operating conditions, the thermocouple readings were fairly constant and the deviations from each other were within a few Kelvin. Based on this fairly homogeneous temperature distribution, it was assumed that temperatures along a single path were within 5 K of each other for 90% of the path length. The influence of temperature on line strengths, broadening widths, and ideal gas law adds up to a total concentration error of 0.02%/K at 470 K and 0.08%/K at 570 K, respectively. The total error of concentration measurement resulting from temperature inhomogeneities and limited accuracy of the thermocouples should therefore not exceed 1%.

Additionally, the pressure has an effect on spectroscopic parameters and the ideal gas law. The combined influence upon ammonia concentrations is less than 0.02%/mbar. The pressure at OP1 between the bricks for example was 1015.08 mbar with a standard deviation of 0.10 mbar during 1 second. Assuming a constant pressure in the cross-section of the spectrometer the total error from averaging 5000 spectroscopic measurements as described earlier is negligible, while the uncertainty from the pressure sensor’s accuracy of 50 mbar was in the range of 1%.

3. Results and discussion

3.1 LOS single path measurement

For OP1 Fig. 5 shows a typical signal of one channel after processing and filtering the signal as described in sections 2.4 and 2.5. The thick line represents the Voigt-fit and the fit residual is located in the lower graph. The measured ammonia concentration is 531 ppm_V and the detection limit based on the ratio of peak absorbance and 1 σ of the fit residual is 64 ppm_V for this case. Other channels showed similar performance with detection limits ranging from 25 to 80 ppm_V depending on their individual signal strength and noise influence. Since the data were averaged for about 1 s per measurement the corresponding time and path related detection limits range from 6.5 to 22 ppm_V·m/√Hz.

 

Fig. 5 Typical single channel evaluation of one of the 8 channels. Scan rate of 5 kHz, averaged 5000 Scans, NH₃ concentration calculated to 531 ppm_V, fitted peak OD_e = 3.65x10⁻³, σ_res = 4.41x10⁻⁴, → detection limit = 64 ppm_V and 17.5 ppm_V·m/√Hz (typical detection limits were 25-80 ppm_V).

Download Full Size | PDF

3.2 Reconstruction of 2-D ammonia inhomogeneities

The simultaneous spectroscopic analysis of all channels provides one vector $c$ of eight individual concentration values per second. The reconstruction of ammonia distributions according to section 2.6 yields a 2D-image for every vector $c$. However, in this work the results of the reconstruction are shown averaged for 2-5 minutes while concentrations were nearly constant.

Results are shown for operating points OP1, OP2 and OP3 directly behind the SCR catalyst in Fig. 6 and operating points OP1 and OP4 in between the bricks in Fig. 7. Moreover, the concentration distributions are averaged and compared to extractive FTIR measurements of ammonia acquired previously. The FTIR measurements were only performed at the operating conditions OP1 and OP2 and at a different location placed 0.5 m downstream of the catalyst (see Fig. 1). Due to cross-sectional extractive sampling, no spatial distribution of ammonia is available. Nevertheless, the extraction employed a perforated pipe through the entire cross section so the extracted probe should give a good assumption of the average ammonia concentration.

 

Fig. 6 Measurements behind catalyst at OP1 in (a) and (b), OP2 in (c) and OP3 in (d). One channel has been removed from reconstruction in (c) and (d) because of liquid urea wetting its optics, which ruined the laser signal. Distribution (b) is based on the same data as (a), but also with that channel removed to show that the algorithm still produces similar results. Mean concentrations are 283 ppm_V, 284 ppm_V, 326 ppm_V and 369 ppm_V, respectively.

Download Full Size | PDF

 

Fig. 7 Measurements in between catalyst bricks at different temperatures of 470 K (OP1, (a)) and 570 K (OP4 (b)). With stoichiometric injection but due to higher temperature and therefore higher catalyst activity a lower ammonia concentration was measured (mean 486 ppm_V compared to 326 ppm_V).

Download Full Size | PDF

Figure 6 shows reconstructions behind the catalyst for varied injection conditions. At OP2 the injected amount was slightly increased from 100% to 110% ($d=1.1$) of the stoichiometric ratio whereas at OP3 it was set to 200% ($d=2$). After a short time of overdosing, one of the detectors was covered by liquid urea and did not supply reasonable values. The flawed channel was not considered in the reconstructions and only the 7 remaining paths are sketched. This increased the uncertainties of the reconstructed distributions. To estimate the influence of a lost signal, a distribution was derived from the data set of Fig. 6(a), but with 7 instead of 8 channels. The resulting distribution in Fig. 6(b) looks very similar and does not differ significantly in the averaged concentration of 284 ppm_V compared to 283 ppm_V. The FTIR-spectrometer measured 225 ppm_V downstream at this operating point. The distribution reconstruction based on the TDLAS measurements overestimates this by 26%. This is in the range of the uncertainties of this setup of 20-30%, mentioned in section 2.4. Beyond that an overestimation can also be explained by the different measurement positions of the two methods. According to Table 1 the conversion rate at OP1 is at 0.28, leaving about 500 ppm_V NO_x in the flow behind the catalyst. This residue could also react with the NH₃ slowly without a catalyst, which would decrease the concentration the FTIR-spectrometer would measure.

The reconstruction for OP2 in Fig. 6(c) shows a distribution comparable to OP1. The homogeneity is decreased slightly and its mean value rose to 326 ppm_V. Compared to the FTIR measurement of 279 ppm_V there is an overestimation of 17% which is again in the range of the uncertainty. For OP3 ($d=2$) the ammonia distribution was even less homogeneous (see Fig. 6(d)). The average of 369 ppm_V shows a significant but not to scale increase, compared to the increase of injected AUS32. This is caused by extensive wall wetting at OP3. At the rather low gas temperature of 470 K, AUS32 stays liquid instead of evaporating. AUS32 droplets consequently cool the wall and liquid AUS32 adheres to the surface. This process is also consistent with slightly decreased NO_x conversion rates when increasing the injected amount of AUS32 (see Table 1). Despite injecting >80% more AUS32 for OP3 ($d=2$) gaseous ammonia concentrations rose only by approximately 13% compared to OP2 ($d=1.1$). Its spatial distribution however is increasingly inhomogeneous because of the wall wetting. At the elevated temperature of 570 K of OP4 no ammonia slip was observed behind the catalyst, neither by the TDL spectrometer nor the FTIR.

Figure 7 compares the ammonia distribution at different temperatures of 470 K and 570 K between the catalyst bricks. The injection was kept at the stoichiometric ratio ($d=1$). The concentration at OP1 is obviously higher compared to the ones measured behind the catalyst in Fig. 6(a), because less ammonia has been consumed after passing only 1/3 of the active catalyst. With increasing temperature the mean concentration decreased from 486 ppm_V at OP1 to 326 ppm_V at OP4 due to the fact that the catalyst activity strongly improved from 0.28 to 0.96 (see Table 1).

4. Conclusion

A TDL spectrometer with eight simultaneous channels was designed for measuring in situ ammonia distributions in exhaust applications during SCR treatment. The spectrometer has been applied to a 340 kW heavy duty diesel engine without any exhaust cleaning systems except the SCR system. Despite the high particle load and strong CO₂ and H₂O background, it was possible to measure ammonia concentrations in this harsh environment. Measurements were performed either between two catalyst bricks (within the NO_x reduction process) or immediately adjacent to the catalyst exit. Detection limits between 25 ppm_V and 80 ppm_V were achieved while averaging for 1 s. Since the spectrometer’s performance was limited by very low light intensities per channel, a Savitzky-Golay filter was applied to the measured data in order to increase SNR. In measurements without the fiber-optic splitter a detection limit of 100 ppm_V was achieved along with a temporal resolution of 100 Hz.

Measured path-averaged ammonia concentrations were used for a tomographic reconstruction based on Tikhonov regularization. Spatially averaged concentrations compared well to extractive FTIR measurements that were performed 0.5 m downstream of the catalyst. The concentration distributions can be derived at the full sampling frequency of the spectrometer of 1 Hz to resolve large-scale mixing phenomena. This diagnostic system (spectrometer and spatial reconstruction) allows for measuring spatially inhomogeneous ammonia distributions using a line-of-sight technique. Additionally, when injecting large amounts of AUS32, wall wetting was observed in the measurements. Based on the performance of the system during these measurements, the spectrometer can be considered well suited for application in harsh industrial environments.