Neural networks for estimating pollutant emissions from vehicles

Abstract

To estimate pollutant emissions (P_ESTIMATED, P) at a tailpipe (125) of a vehile, a first neural network (210) of a computer (200) processes a first set of multiple engine parameters E_ALPHA to obtain data EO_ESTIMATED that represent the estimated content of multiple chemical compounds in the gas at the engine outlet (115). A second neural network (220) processes (420) a second set of multiple engine parameters E_BETA and EO_ESTIMATED to obtain data that represents a preliminary estimate of the content of multiple compounds of the gas at the tailpipe (125): B_ESTIMATED. A multiplier unit (230) of the computer (200), multiplies the preliminary catalyst-out data B_ESTIMATED by a pre-defined factor vector to obtain pollution estimate data (P_ESTIMATED), according to the degradation of the catalytic converter system (120).

Description

P-DELPHI-460 / WO- 1 -NEURAL NETWORKS FOR ESTIMATING POLLUTANT EMISSIONS FROM VEHICLESTechnical Field

[0001] In general, the disclosure relates to on-board monitoring the emission of vehicles and to equipment that performs such monitoring. More particularly, the disclosure relates to computer systems, methods, and computer-program products that use neural networks to provide an estimate of the pollutant emissions.Background

[0002] Internal combustion engines are ubiquitous to move vehicles, such as road vehicles, rail vehicles, vessels, and so on. Combustion engines use fuels, such as gasoline (gasoline / petrol engines, i.e., spark-ignition engines), diesel fuel (diesel engines), hydrogen (H2 engines), or other liquid or gaseous fuels.

[0003] However, due to the combustion it is inevitable that the engines also produce exhaust gas. The exhaust gas contains a mixture of substances that are chemical compounds. Most of these compounds act as pollutants, and the emission of such pollutants to the atmosphere should be minimized. Prominent examples for such pollutants are nitrogen oxides (that are usually known as NOx).

[0004] The skilled person is able to provide a catalytic converter as a technical tool that seeks to purify the exhaust gas by eliminating pollutants before emission. Simplified, a catalytic converter is a chemical reactor with a catalyst. Gases from the engine are processed so that the share of potential pollutants is minimized. The skilled person is familiar with various converter concepts, and the three-way catalytic converter (TWC) is just an example. However, an ideal catalytic converter that would remove all pollutants is not available. In practice, some pollutants can pass through real catalytic converters.

[0005] On-board monitoring (OBM) equipment allow the measurement (or the estimation) of the share of such pollutants in the exhaust gas.

[0006] Knowledge of pollutant data can also be used to control the engine. Such data allows,P-DELPHI-460 / WO- 2 - for example, to run the engine with settings to minimize the output of pollutants. As the vehicles move with constantly changing speed, the pollution may change within a few seconds. The control would have to be adaptive to such changes.

[0007] However, there is a discrepancy in the availability of sensors for obtaining pollution data. Equipment to exactly identify and measure pollutants can be provided in laboratory settings (test benches, test stands, etc.). Vehicles are regularly inspected in such settings. For vehicles that are moving, a portable emissions measurement system (PEMS) can be used, but such systems are available during testing only.

[0008] Testing standards are available, and the skilled person knows them by acronyms such as Worldwide Harmonized Light Vehicles Test Procedure (WLTP), Worldwide Harmonized Light Vehicles Test Cycle (WLTC), Real Driving Emissions (RDE), or others.

[0009] A current version is outlined by the EU Commission in Regulation (EU) 2024 / 1257 of the European Parliament and of the Council, published 24 April 2024.

[0010] Mathematical approaches to estimate pollution can be based on pre-defined rules, physical equations, and / or machine-learning (ML) processing. However, there are constraints regarding the calibration, the availability of training data (for ML), and so on. Further, the chemical reactions in the converter depend on the temperature within the converter. The converter needs some time to warm up. The pollution at the exit of a cold converter is different from the pollution of a warmed-up converter. The mathematical approaches would have to take such effects into account.

[0011] Under real-world driving conditions of passenger cars (and / or commercial vehicles such as cargo vehicles or medium / high duty vehicles), controlling the engine requires estimating pollution data while the vehicle moves. The time-interval for pollution data is relatively short, for example, set every second. Details for emission monitoring are available in emission standards.

[0012] Such a real-time estimation would allow a controller computer to interfere with the operation of the engine, should the estimated pollution reach certain threshold values.P-DELPHI-460 / WO- 3 -

[0013] As already mentioned, NOx is not the only pollutant. But providing pollutant-specific sensors would add complexity and weight to the vehicle and increase its overall fuel consumption.

[0014] Using sensors may have further constraints. For example, (i) a particular sensor may not yet be ready to provide data when the vehicle starts moving; (ii) sensors change over time (sensor drift) and would have to be calibrated in regular intervals; and (iii) sensors may be modified without authorization (tampering, etc.).Summary

[0015] The disclosure refers to a computer-implemented method to estimate pollutant emission at a tailpipe of vehicles according to claim 1. The exhaust line from the engine to the tailpipe comprises a catalytic converter system that is arranged between an engine outlet of the engine and the tailpipe. The method-executing computer comprises neural networks and a multiplier unit.

[0016] A first neural network processes a first set of multiple engine parameters to obtain data that represent the estimated content of multiple chemical compounds in the exhaust gas at the engine outlet (estimated engine-out data).

[0017] A second neural network processes a second set of multiple engine parameters and the estimated engine-out data to obtain data that represents a preliminary estimate of the content of multiple compounds of the exhaust gas at the tailpipe (preliminary catalyst-out data).

[0018] A multiplier unit multiplies the preliminary catalyst-out data by a pre-defined factor vector to pollution estimate data, according to the degradation of the catalytic converter system.

[0019] Optionally, the preliminary catalyst-out data and the pollution estimate data are vectors with V elements each. The vector elements correspond to V chemical compounds of the exhaust gas at the tailpipe. The pre-defined factor vector has V factors that are specific to the chemical compounds, and at least one chemical compound is NOx.P-DELPHI-460 / WO-4 -

[0020] Optionally, the computer obtains the NOx-specific factor and an ammonia (NH3)- specific factor from an NOx value and an NH3 value that are measured by sensors at the tailpipe and the pollution estimate data for NOx and for NH3.

[0021] A computer-implemented method to train a first neural network and a second neural network of a computer is also disclosed. There is a purpose to estimate pollutant emissions at a tailpipe of a vehicle. After training, the first neural network is adapted to process a first set of multiple engine parameters to obtain data that represent the estimated content of multiple chemical compounds in the exhaust gas at the engine outlet (estimated engine-out data), and the second neural network is adapted to process a second set of multiple engine parameters and the estimated engine-out data to obtain data that represents a preliminary estimate of the content of multiple compounds of the gas at the tailpipe (preliminary catalyst-out data). A multiplier unit is adapted to multiply the preliminary catalyst-out data by a pre-defined factor vector to obtain pollution estimate data, according to the degradation of the catalytic converter system. The method to train the first and second neural networks is performed with historical data from a reference catalyst with a standardized age (or performed with data from a reference catalyst with non-standardized age, wherein the data has been normalized to standardized age), so that historical data that represent degradation of the catalytic converter system is used to identify the predefined factor vector of the multiplier unit.

[0022] Other embodiments and aspects of the invention are recited in the appended claims.Brief Description of the Drawings

[0023] FIG. 1 illustrates a block diagram of a vehicle with an engine and a catalytic converter system;

[0024] FIG. 2 illustrates a system overview of a vehicle engine and an estimation computer;

[0025] FIG. 3 illustrates a time diagram for network phases: collecting historical data, training the networks, and operating the networks;

[0026] FIG. 4 illustrates a flow-chart diagram of a computer-implemented method toP-DELPHI-460 / WO-5 - estimate pollutant emission at a tailpipe of a vehicle;

[0027] FIG. 5 illustrates a time diagram for the availability of data;

[0028] FIG. 6 illustrates a diagram for two numerical criteria over different architectures;and

[0029] FIG. 7 illustrates a generic computer.Detailed DescriptionVehicle

[0030] FIG. 1 illustrates a block diagram of vehicle 100 with engine 110, catalytic converter system 120, and estimation computer 200. As already mentioned, on-board monitoring (OBM) equipment 350 allows the measurement (orthe estimation) of the share P of pollutants in the exhaust gas (i.e., in the exhaust gas that remains at the tailpipe-end of the exhaust line).

[0031] As symbolized by an arrow to 350, engine 110 and catalytic converter system 120 can be regarded as a technical system by which the state is communicated to a user. The user can be, for example, a driver 191 (or another person) who is located in the vehicle, or an inspector (or other person) 192 who is located outside the vehicle.

[0032] OBM equipment 350 can further evaluate the measurement or the estimation from estimation computer 200. Depending on thresholds or other pre-defined data, the technical states can be identified, for example,• as the differentiation between normal and abnormal pollution, optionally with intermediate states (for that uncertainty is allowed);• as the detection of sensor malfunction (such as sensor drift or un-authorized manipulation such as tampering).

[0033] Obtaining such and other state information requires the quantitative estimation of chemical compounds that act as pollutants: P_ESTI MATED (the letter P in the cloud symbol).P-DELPHI-460 / WO- 6 -

[0034] The description concentrates on how estimation computer 200 uses neural networks (FIG. 2) to estimate the pollutant emissions. Estimation computer comprises a plurality of modules (210, 220, 230 in FIG. 2) that perform the steps of the computer- implemented method in FIG. 4. The neural network has been trained to have network weights (see FIG. 3 for details), and neural networks provide prediction data.Notation

[0035] To identify data, the specification uses acronyms that can be concatenated by underscores

[0036] An acronym at the left position identifies a system component or data that is representative of the component, such as, for example:• E stands for data that represents one or more engine parameters of engine 110, "parameter data" in short, that are being processed by networks 210 and 220. E is a subset of E_LONGLIST, which stands for a list of parameters that can be applied. E comprises all parameters of E_SHORTLIST, which stands for a list of (mandatory) parameters that have to be applied;• EO stands for the engine outlet;• A, B differentiate the outputs (or output pipes) of optional converter subsystems inside the catalytic converter system 120;• CS stands for data from the converter system 120; and• P stands for the quantity of pollutants at the tailpipe of the vehicle, and so on.

[0037] For convenience, some of the acronyms are mentioned in FIG. 1 as well.

[0038] An acronym at the next position can differentiate measurement data (MEASURED) that the computer receives from sensors from estimation data (ESTIMATED) that the computer obtains by calculation.

[0039] An acronym at the right position can indicate a chemical compound, such as NOx (nitrogen oxides), CO2 (carbon dioxide), etc. PM stands for particulate matter. For simplicity, the acronym PM is used as if the particulate matter would be a chemical compound.P-DELPHI-460 / WO- 7 -

[0040] The computer can process data by vectors. Different vectors can have different numbers of vector elements. For example, the vector E has multiple parameters that are vector elements and relate to the engine. For example, engine parameters can be E_PRAIL (i.e., the pressure at a common rail), E_TEMP (i.e., a temperature), among others. The description provides a more detailed list below.

[0041] In a further example, the pollution estimate P_ESTI MATED (a vector) is further differentiated into estimates P_ESTIMATED_NOx, P_ ESTIMATED_CO2, P_ESTIMATED_PM, and so on.

[0042] Data that represent physical phenomena or the like are usually associated with measurements units (no matter if ESTIMATED or MEASURED). For vehicles, the units can be related to distances. For example, a typical value for P_NOx could be 40 mg / km (40 mg of NOx pollution per kilometer). When skilled persons set up neural networks in computers, they usually adapt data by harmonizing data, removing measurement units, harmonizing processing rates to sampling rates, and so on. The description will therefore leave out such well-known implementation details.

[0043] The specification uses the term "gas" (in singular) for a mixture of different chemical compounds that are in gaseous form, such as NOx, CO2, and so on. The word "gas" in combination with the system component stands for the gas at the output of a particular component, e.g., "gas EO" is the gas at engine outlet EO.Engine types

[0044] The parameters E_*** are differentiated according to engine types. As internal combustion engines can be differentiated according to the fuel, the description differentiates engine types as gasoline, hydrogen, and diesel. Some parameters are common for all types, some parameters are typical for particular engine types.Engine parameters

[0045] Engine parameters E | gasoline / hydrogen for gasoline and hydrogen engines can comprise (E_LONGLIST | gasoline / hydrogen):P-DELPHI-460 / WO- 8 - • Engine Speed *• Engine Load *• Total Injected Fuel *• Absolute Spark angle *• Split Injection Fueling and Pulse Angle (1 to N injections depending on combustion mode)• Rail Pressure• Turbine Flow *• CAM Phaser Intake Position• CAM Phaser Exhaust Position• Intake Total Air Mass *• WRAF Lambda *• Switch O2 rear sensor• TWC O2 Storage Capacity• Manifold Pressure *• Atmospheric Pressure• Swirl Position• Coolant Temperature• Ambient Temperature• Intake Temperature *• Inlet Turbine Temperature *• TWC1 Input / Output Temperature *, BETA only• TWC2 Input / Output Temperature *, BETA onlyThe asterisk * identifies mandatory engine parameters that belong to E_SHORTLIST | gasoline / hydrogen. Additionally, there are two parameters for BETA only.

[0046] Engine parameters (E | diesel) for diesel engines can comprise (E_LONGLIST | diesel ):Engine Speed *Engine Load *P-DELPHI-460 / WO- 9 - • Total Injected Fuel *• Injected fuel without Post injection• Pilot Injection Fueling and Timing• Main Injection Fueling and Timing• After Injection Fueling and Timing• Post Injection Fueling and Timing• Intake Total Air Mass *• EGR inert rate *• Turbine Flow *• Rail Pressure• Manifold Pressure *• Atmospheric Pressure• Swirl Position• Air / Fuel Ratio *• Coolant Temperature• Ambient Temperature• Intake Temperature *• Inlet Turbine Temperature *• DOC Input / Output Temperature *, BETA only• SCR Input / Output Temperature *, BETA onlyThe asterisk * identifies mandatory engine parameters that belong to E_SHORTLIST | diesel. Additionally, there are two parameters for BETA only.

[0047] The sets E can be further differentiated into sets ALPHA, BETA. For example, E_ALPHA stands for a first set of engine parameters in a first vector, and E_BETA stands for a second set of engine parameters in a second vector. Most of the engine parameters can be in either or both sets. Feeding different data to different neural networks can be advantageous. This simplifies training (phase 2) and operation (phase 3)

[0048] E_ALPHA are the parameters that go into a first neural network (210 in FIG. 2) and E_BETA are the parameters that go into a second neural network (220 in FIG. 2).P-DELPHI-460 / WO- 10 -

[0049] E_ALPHA can be subset of E_BETA, as some parameters are common in both data sets. E_LONGLIST are given for E_ALPHA but the longlists also indicate "BETA only" parameters.

[0050] For simplicity, the figure does not differentiate engine type. By way of example, the description refers to the gasoline engine type. The skilled person can adapt the example to other types. For example, one of the engine parameters is data that represents the timing of the ignition sparks.

[0051] Having explained writing conventions, the description now turns to more details, especially for the computer.System overview

[0052] FIG. 2 illustrates a system overview to a vehicle engine 110 and to estimation computer 200. Simplified, the figure can be viewed as a matrix: In "columns" it illustrates a gas flow (here identified by location: gas EO, gas A, and gas B) and of corresponding data (EO_ESTI MATED, A_ESTIMATED, B_ ESTIMATED) from left to right. In "rows" it illustrates technical equipment (1**) at the top with its computer equivalents (2**) below.

[0053] Internal combustion engine 110 has engine outlet 115 (e.g., exhaust manifold) that is symbolized here by an arrow. The gas at engine outlet 115 is called "EO". EO may comprise NOx, CO (carbon monoxide), HC (hydrocarbons), CH4 (methane), CO2 (carbon dioxide), O2 (oxygen), PM, and other chemical compounds.

[0054] Catalytic converter system 120 (such as a system that uses an oxidation catalyst for gasoline engines) receives EO from engine outlet 115. System 120 "cleans" that gas (optionally, to intermediate gas A), and forwards gas B to tailpipe 125. The skilled person is familiar with the chemical reactions in such systems. For example, gas B may still comprise chemical compounds that were present prior to cleaning (such as NOx, CO, HC, CH4, CO2 etc.) in differing amounts, along with new compounds that result from the reactions. For example, gas B may comprise NH3 (ammonia) that was not present in EO (at least not in significant amounts).P-DELPHI-460 / WO- 11 -

[0055] Advantageously, converter system 120 has two converter sub-systems, 120-A and 120-B, that are arranged in a cascade with pipe 121 between the sub-systems. Both sub-systems 120-A and 120-B can be of the above-mentioned three-way catalyst (TWC) type.

[0056] Sub-system 120-A is located nearer to engine 110, and sub-system 120-B is located nearer to tailpipe 125. Sub-system 120-A provides gas A, while sub-system 120-B provides gas B.

[0057] Tailpipe 125 outputs gas B to the environment (cloud symbol).

[0058] Gases EO, optionally A, and B comprise NOx and other pollutants, with the amount of pollutants becoming smaller as the gases pass through the converter system. The letter "P" stands for pollutant data that the skilled person can quantitatively estimate.

[0059] Exhaust line 101 is the collective term for engine outlet 115, system 120, and tailpipe 125.

[0060] Engine 110 is controlled by engine controller 300 (i.e., a computer, with a dotted arrow that symbolizes the controlling). Parameter data E stands for data that represent the operation of the engine. E can be set by engine controller 300 and / or can be measured by sensors that are associated with engine controller 300. E is illustrated with bold lines from the controller to estimation computer 200. To control engine 110, engine controller 300 does also communicate at least some of E to and from engine 110. For simplicity, that data communication is not illustrated. In other words, as the skilled person is able to provide engine parameter data E, the description will discuss the processing of E by estimation computer 200.

[0061] With an overall goal to reduce pollutants P, estimation computer 200 receives and processes data, such as the following:• E engine parameters can be received from engine controller 300, without the need to generate that data (E in E_ALPHA and E_BETA).• CS (i.e., data from various points within catalytic converter system 120)

[0062] Estimation computer 200 provides the estimate P_ESTI MATED of the pollutantP-DELPHI-460 / WO- 12 - emission at the tailpipe 125 of the vehicle. As an ideal estimation (P_ESTI MATED = P, or P_ESTIMATED = P_MEASURED) is not available, the description explains how the estimation P_ESTI MATED can be approximated to P. In embodiments, P_ESTIMATED can be a vector with separate share values (for NOx, CO, ..., NH3 etc.)

[0063] Estimation computer 200 uses neural network 220 ("converter network"). Optionally, converter network 220 comprises neural networks 220-A and 220-B in a cascade. As data at the outputs of these networks 220-A and 220-B can only be estimates, the description uses the acronym ESTIMATED, i.e., A_ESTIMATED and B_ESTIMATED, both being vectors for different chemical compounds.

[0064] The description occasionally uses the term "sub-network" in the sense that both networks 220-A and 220-B in combination can be viewed as a single neural network with B_ESTI MATED as its output.

[0065] With details to be explained, the topology of neural networks 210 and 220 (optionally with 220-A, 220-B) in estimation computer 200 corresponds to the cascaded structure of the hardware illustrated in FIG. 2.

[0066] Neural network 210 is an "engine network" that corresponds to engine 110.Converter network 220 corresponds to converter system 120, optionally with subnetworks corresponding to the sub-systems 120-A and 120-B. In view of the location, sub-system 120-A is the "engine-side network", and sub-system 120-B is the "tailpipe-side network".

[0067] Due to that cascading, estimation computer 200 not only provides P_ESTI MATED, but can also provide estimations of intermediate substances:• The estimation EO_ESTIMATED is the estimate of EO that occurs at engine outlet 115 of engine 110. For example, estimation computer 200 can estimate the share of some substances at EO 115, with the corresponding data here called EO_ESTI MATED, or specific to substances such as EO_ESTIMATED_NOx, EO_ESTIMATED_CO, and so on.• The estimation A_ESTIMATE is the estimate of gas A at pipe 121 (between theP-DELPHI-460 / WO- 13 - sub-systems 120-A and 120-B). This, too, may be specific to substances such as A_ESTIMATED_NOx, A_ESTIMATED_CO, and so on.• The estimation B_ESTI MATED is the preliminary estimate of gas B at tailpipe 125. Similarly, B_ESTIMATED may be specific to substances as the vector for B_ESTIMATED_NOx, B_ESTIMATED_CO, etc.

[0068] Engine controller 300 can use EO_ESTIMATED, P_ESTIMATED, and other data, such as A_ESTIMATED or B_ESTIMATED if available, as feedback, as illustrated by the dashed arrows at the bottom of box 200. With such differentiated availability of estimation data, engine controller 300 can control engine 110 accordingly.Network Inputs / Outputs

[0069] For convenience of illustration, FIG. 2 shows inputs IN and outputs OUT of the neural networks. The data flow is that of the vehicle in movement.Sensors

[0070] Exhaust line 101 can be equipped with multiple sensors. By way of example, FIG. 2 only shows sensor 140 at tailpipe 125, which may be a NOx sensor to obtain NOx_MEASURED, a NH3 sensor to obtain NH3_MEASURED, and / or another sensor to measure other substances.Multiplier to compensate converter degradation

[0071] FIG. 2 also shows that estimation computer 200 uses a multiplier unit 230 to account for the fact that converters degrade over time. Using multiplier unit 230 simplifies the training for the neural networks. Degradation may be modelled by multiplication, which is less complex than processing vectors from the input to the output of the neural networks.

[0072] Multiplier unit 230 applies a vector-element wise multiplication. The elements of B_ESTI MATED are multiplied with factors, resulting in P_ESTI MATED, for example:• P_ESTIMATED_NOx = FACTOR_NOx * B_ESTIMATED_NOx• P_ESTIMATED_CO = FACTO R_CO * B_ESTIMATED_CO

[0073] The factors are not only specific to the chemical compounds, but also a function of the mileage. For a converter unit with a nominal performance (e.g., at around 7.000P-DELPHI-460 / WO- 14 - km mileage), the factors would be 1. For aged converter units, pollution will increase with mileage, and the factors will be greater than 1. In other instances, factors can be smaller than 1.

[0074] As multiplier unit 230 belongs to networks 210 and 220, it does not have to be trained. .

[0075] The factors can be obtained according to different strategies. The skilled person can use the O2 storage capacity estimation strategy for gasoline engines and can use the selective catalyst reduction (SCR) efficiency estimation for diesel engines.

[0076] While network 220 could be trained for converters with different degradation stages separately by "age group", the use of the multiplier allows processing data to a preliminary value, B_ESTI MATED, and to correct that value later on with factors that take degradation into account. In other words, training can be performed with historical data that are obtained from converters that have a standardized lifetime (or mileage), but data from all possible lifetimes does not have to be collected.Phases to operate the neural networks

[0077] FIG. 3 illustrates a time diagram for network phases:• collecting historical data (phase 1, references -1),• training the networks (phase 2, references -2), and• operating the networks (phase 3 references -3)

[0078] The progress of time (over relatively long time intervals, such as weeks, months, or years) is illustrated from left to right with phases:• Phase 1 involves collecting historical measured data (e.g. E_MEASURED, A_MEASURED, B_MEASURED, P_MEASURED, etc.) for training the networks. Data may come from multiple sources such as different engines (110-1, 110-1', 110-1", and many others) and converters (120-1, 120-1', 120-1", and many others). The skilled person can obtain the data from multiple vehicles (i.e., multiple engines and converters). In other words, the historical data serves as reference data for training and such reference date would not have to beP-DELPHI-460 / WO- 15 - limited to a single source.• Phase 2 involves training the networks 210-2 and 220-2 with the collected historical data. The historical data are applied to the input IN and to the output OUT of the networks. For example, sub-network 210-A with E_MEASURED at the input IN, EO_MEASURED instead of EO_ESTI MATED at the output OUT; subnetwork 220-A with E_MEASURED and EO_ESTIMATED at the input IN, A_MEASURED at the output OUT; sub-network 220-B accordingly, with historical E at the further input IN. Measured data at OUT serves as ground truth.• Phase 3 uses the trained networks in vehicles as networks 210-3 and 220-3 (with sub-networks) when the method is being performed. Factors Fl, F2, and F3 symbolize the application of different numerical factors in a multiplication operation of the networks.

[0079] Training the networks in phase 2 can be repeated. Over time, the networks provide estimations P_ESTIMATED that become closer to reality P_MEASURED. In other words, a neural network that is trained regularly becomes more effective over time.

[0080] In contrast, the repeated use of converters over time makes them less effective.Performance deterioration / degradation is a normal behavior. During training, such a degradation can be disregarded. This is possible because the effects of deterioration can be processed independently from training by simple multiplication at multiplier unit 230.

[0081] Training in phase 2 can be performed for standardized converters and / or standardized data, such as for converters in vehicles with rated 7.000 km mileage (i.e., the total number of kilometers that a vehicle with a particular engine and with a particular converter has traveled since it was first driven).

[0082] Using the networks in phase 3 is illustrated for a single converter 120-3 of a particular vehicle. FIG. 3 symbolizes performance deterioration cloud symbols of with increasing sizes. This acknowledges that because the converter deteriorates, the share of pollutants would rise over time. For example, a new converter for a new vehicle would output less pollution than the same converter for the same vehicleP-DELPHI-460 / WO- 16 - after, for example, 50.000 km mileage.

[0083] This description, however, is simplified; the share of some chemical compounds may increase, while the share of a different chemical compound may decrease. But with the multiplication factors that are compound-specific, the numeric output of the last network (e.g., network 120-B) can be corrected by simple element-wise multiplication and the networks do not have to be adapted to performance deterioration. As the figure symbolizes, factors Fl, F2, and F3 are used to process a preliminary result from networks 210-3 and 220-3. The adaptation to the deterioration of the converter is arranged by the factors, but the networks remain the same without any required re-training.

[0084] FIG. 3 is simplified, and the skilled person understands that the training phase can comprise validating and testing the trained networks. For example, once the weights are identified during initial training, the trained networks can be validated against a test database with MEASURED data. Such a step can confirm that neural networks perform well on unseen data (i.e., data that has not been used in training phase 2), ensuring their generalizability.

[0085] Training comprises calibrating the neural networks. This involves iterative learning to minimize error and improve prediction accuracy. The skilled person is able to provide databases or other data storage for the training phase.

[0086] Training can be performed for different networks differently with different training data. During training, engine network 210-2 receives historical data that has been collected for engines 110-1, 110-1', etc. Converter network 220-2 receive historical data that has been collected for converter systems 120-1, 120-1', etc. If converter network 220 comprises two or more sub-networks (such as 220-A and 220-B), training may be performed with separately collected data as well.

[0087] The network independence is noted: different designs of converters do not impact the chemical compounds at EO. Therefore, the method is performed with networks that have been trained for particular implementations of the engine and of the converter system.P-DELPHI-460 / WO- 17 -

[0088] It is possible, however, to train converter network 220 with historical data obtained from two or more converter systems that are, for example, (i) a converter system with nominal mileage (e.g., 7.000 km) and (ii) a deteriorated (aged) converter system.

[0089] Collecting training data in phase 1 can comprise collecting data for different standardized conditions, such as for WLTC and RDE. Data can be collected for different environments, for example, for ambient temperatures from -7°C to 40°C and relative humidity from 5% to 80%. The geographical altitude can be taken into account as well.Training details

[0090] During training in phase 2, input vectors with historical data from phase 1 are fed into the neural networks, 210-2, 220-2. This initiates feedforward propagation through the network layers to compute a preliminary estimated output value (e.g., EO_ESTIMATED_NOx, A_ESTIMATED_NOx, A_ESTIMATED_CO, etc.). The error between the MEASURED_NOx (historical data from phase 1) and ESTIMATED_NOx is then calculated. Using this error, the weights of the artificial neural network (ANN) are adjusted via the backpropagation algorithm. Through recursive iterations, the ANN learns and stores the correct emissions concentrations values within its weights. This offline learning procedure is stopped once the error falls below a predetermined threshold. At the end of this process, phase 2 ends, and the neural network is prepared to receive new inputs and predict the NOx value in real time when the vehicle is in operation (phase 3).Validation and performance comparison:

[0091] After completing the training phase 2, trained networks can be evaluated with further data to compare measurements (e.g., P_MEASURED_NOx) with estimations (e.g., P_ESTIIVIATED_NOx) at standard conditions (e.g., temperature 23°C, and 50% relative humidity, altitude approximately 100 meters). The error between measured and estimated data was 1%. The skilled person would use a newly manufactured exhaust line for that data just becomes available, practically this can be the exhaust line in a new vehicle .P-DELPHI-460 / WO- 18 -

[0092] Validation was performed for other chemical compounds as well, such as CO and HC.The error at 6% was slightly higher, but acceptable.

[0093] FIG. 4 illustrates a flow-chart diagram of a computer-implemented method 400 to estimate pollutant emission (pollution estimate data (P_ESTIMATED)) at the tailpipe of vehicles. Method 400 corresponds to the above-introduced phase -3.

[0094] As already explained forthetop of FIG. 2, exhaust line 101 from engine 110 to tailpipe 125 comprises the catalytic converter system 120 that is arranged between engine outlet 115 of engine 110 and tailpipe 125. The flowchart summarizes the activities by estimation computer 201, in FIG. 2 as well.

[0095] In step processing 410, a first neural network 210 (the engine network) processes a first set of multiple engine parameters E_ALPHA to obtain data EO_ESTIMATED, which represent the estimated content of multiple chemical compounds in the gas at engine outlet 115.

[0096] In step processing 420, a second neural network 220 (the converter network) processes a second set of multiple engine parameters E_BETA and EO_ESTI MATED to obtain data B_ESTIMATED, which represent a preliminary estimate of the content of multiple compounds of the gas at tailpipe 125. Step 420 can be performed by substeps for the sub-networks (network 220-A leading to A_ESTI MATED, network 220-B leading to B_ESTIMATED).

[0097] In step 430, multiplier unit 230 multiplies B_ESTIMATED by a pre-defined factor vector F to pollution estimate data:P_ESTIMATED = B_ESTIMATED * FIn other words, pollution estimate data P_ESTIMATED is obtained / calculated as the product of B_ESTIMATED and F.

[0098] The factor F has been established in advance as a vector for multiple chemical compounds according to the degradation of catalytic converter system 120. The degradation can be related to the mileage, so that the factors can be obtained by aP-DELPHI-460 / WO- 19 - predefined mapping from mileage data.

[0099] As factor F is not processed by the networks, the networks do not have to be trained to take degradation into account.

[0100] A computer program product that, when loaded into a memory of a computer system and executed by at least one processor of the computer system, causes the computer system (e.g., computer 200) to perform the steps of computer-implemented method 400.Synchronization

[0101] FIG. 5 illustrates a time diagram for the availability of data for different neural networks 210, 220-A, and 220-B. By way of simplified example, the networks are shown with boxes for which the width corresponds to calculation time. Figure 5 is simplified to equally sized boxes, but the calculation times can be different. Horizontal lines to the left of the boxes show data availability and data buffering.

[0102] Input data (e.g., engine parameter data E or converter system data CS) is simplified here to digits 1, 2, 3 (input data to network 210, such as engine parameters or measurement data from sensors), 4 (input to network 210-A, such as engine parameters or measurement data), and 5 (input to network 210-B).

[0103] Time-point tl indicates availability for data 1, t2 for data 3 and 4, t3 for data 5.

[0104] Network 210 can start processing data when data 1, 2, and 3 are available, or after t3 at the earliest. Network 220-A can start processing data 4 and the output from the cascade-predecessor network 210 at t4, and network 220-B can start processing data 5 and the output from network 220-A at t5.

[0105] FIG. 5 thereby illustrates data synchronizing. That ensures, for example, that fuel injection data (e.g., the identification of tl as the ignition time-point) with other engine control parameters and emissions concentrations are aligned. Synchronizing allows to establish system causality.

[0106] As modern engines are computer-controlled, an engine controller 300 known as anP-DELPHI-460 / WO- 20 - ECU operates combustion events by performing fuel injection events at predetermined timings and triggering ignition events for spark ignition gasoline / H2 engines. Hence engine controller 300 can provide data regarding the timing of such events (1 in FIG. 5), that can be part of E.

[0107] The ignition and the combustion that immediately follows is responsible to produce torque and also chemical compounds, among them pollutants. Due to different viscosities and other inherent properties, the chemical compounds would arrive at a sensor with a delay. In the simplified example, that could be data 5, which becomes available at t3.

[0108] Data for the compounds need to be related to the event. This is possible by taking the delay into account. Data can be buffered. In the example, data 5 is being buffered without any change in the content of the data, from t3 until t5 when processing starts in network 220-B. Data 5 remains related to the event at tl. When network 220-B starts processing at t5, the sensor for data 5 will provide different data, but by considering the delay, data from t3 is processed.

[0109] In other words, the gas propagation speed through the converter is considered. The content of the gas does not change, but the sensor may be located at the end of a tube. Hence, data that represents the content at the input of a tube becomes available when the gas has reached the end of the tube. With known propagation speed, the data can be harmonized in view of time.

[0110] Time-point t_RESULT marks the availability of the results, here at the output B_ESTI MATED of network 220-B. The calculation time it takes multiplier unit 230 to operate is neglected.

[0111] The skilled person is able to implement such data-preprocessing by computation techniques. Buffering data is just an example for such techniques. Filtering data is possible as well, for example, to filter out data that a sensor provides the "wrong" points in time. For example, a sensor may provide data 5 at t4 and at t5 but the filter takes data from t3, which corresponds to the ignition event, and blocks data from t4 and t5.P-DELPHI-460 / WO- 21 - Harmonizing sampling rates

[0112] The networks 210 and 220 process input data (i.e. the vector E with its elements, as well as measured concentrations of pollutants etc.) at a sampling time that corresponds to the engine cycle events. For example, a gasoline engine running at 3,000 RPM would have 1,500 ignition events per minute (per cylinder), 25 events per second. For a 4-cylinder engine at 3,000 RPM, knowing that an engine cycle has two revolutions (i.e., a four-stroke engine), there are 3000 / 2*4 = 6000 ignition events per minute, i.e., 100 Hz.

[0113] For convenience of explanation, it can be assumed that the sampling frequency is 10 Hz (i.e., sampling time of 100 milli seconds). Other frequencies can be used as well, as long as the frequency fits to the sampling theorem (Shannon).

[0114] If a sensor provide data with longer sampling times (> 100 milli seconds) or shorter sampling times (< 100 milli seconds), the data can be harmonized to 100 milli seconds. Techniques are available for the skilled person.

[0115] Further, it is possible to synchronize buffering and processing such that t_RESULT occurs every 100 milli seconds. FIG. 5 illustrates this approach with two vertical lines with that time-distance.Processing in parallel, data pipelining

[0116] Using a constant sampling frequency allows to distribute the computation for the neural networks to multiple processor cores, symbolized by dashed lines to core numbers 01, 02, and 03 in FIG. 5..

[0117] When core 01 has handed over the result (i.e., vector A to network 120-B), core 01 can process data in network 120-A for a further ignition event. In other words, data for a consecutive second ignition event can be processed while the processor processes data for the first event.Different sensor types

[0118] The skilled person can implement reading and merging data from various raw gas analyzers or sensors, among them sensors that apply FTIR (Fourier TransformP-DELPHI-460 / WO- 22 - Infrared Spectroscopy). Sensors can be embedded on the vehicles (as Portable emissions measurement system (PEMS)), that provide detailed emission concentration data, at least for collecting training data.

[0119] Some of the sensors are available in laboratory only (training only), some of the sensors are built in to every commercially available vehicle.

[0120] As mentioned, FIG. 2 shows tailpipe sensor 140 by way of example.Noise

[0121] To enhance the robustness of the neural networks (i.e., of the models they obtain during training, phase 2), it is possible to apply noise to some input data. In other words, the skilled persons measure the numeric range of input data, such as concentration data from sensors, and can provides variations for the training data within certain ranges. To deal with quantization effects arising from the sensor providing data in digital form, the dithering data is applicable as well.Network architecture

[0122] The networks such as networks 210, 220-A, 220-B can be implemented as ANNs. The architecture of the network is defined, for example, by:• the number of network layers among them hidden layers,• the way the layers are interconnected,• the way nodes are implemented, such as artificial neurons,• the signal processing at the inputs,• the selection of the activation function, and• the implementation of backpropagation algorithms that are performed by the network to estimate gradients during training.

[0123] FIG. 6 illustrates a diagram for two numerical criteria (AIC the top graph, FPE the graph below with values at the ordinate) over different architectures (D, Hl, H2, M) at the abscissa.

[0124] It is well known that the architecture of such ANNs can be optimized and that different optimization criteria are available.P-DELPHI-460 / WO- 23 -

[0125] For the neural networks that are described herein, the criteria had been the Akaike Information Criterion (AIC) and the Final Prediction Error (FPE) criteria. The optimal architecture corresponds to the minimal values of AIC and of FPE.

[0126] The architecture can be defined by• D (the number of inputs),• Hl (the number of neurons in the first hidden layer),• H2 (the number of neurons in the second hidden layer),• M (the number of neurons in the output layer), and• the number of layers and the number of neurons in each layer.

[0127] As labelled "optimum architecture", AIC and FPE reach minima value at (D, Hl, H2, M) = (23, 35, 20, 6).OBM

[0128] Having explained how estimation computer 200 estimates the pollutant emission P_ESTI MATED at the tailpipe, the description closes with discussing some aspects for the operation of OBM 350.

[0129] For NOx, measurement data is available from NOx-sensor 140 and estimated data is available as P_ ESTIMATED_NOx (output of multiplier 230).

[0130] As mentioned above, estimations may not correspond to reality. OBM can evaluate P_ESTIMATED (for NOx and for other pollutants) in view of absolute or relative thresholds.

[0131] An absolute threshold may be set by a standard (i.e., 60 mg / km), and a threshold relation may be set for the purposes of detecting sensor malfunction. For example, if P_MEASURED_NOx * 2.5 > P_ ESTIMATED_NOx is detected, an information can be given to the driver, and a state transition for the NOx-sensor 140 from normal to abnormal can be determined and recorded in a log file.

[0132] Defining other pre-defined rules is possible as well. P_MEASURED_NOx = 0 is an indicator for abnormal sensor operation.P-DELPHI-460 / WO- 24 - Generic computer

[0133] FIG. 7 illustrates an example of a generic computer device that may be used with the techniques described here. FIG. 7 is a diagram that shows an example of a generic computer device 900 and a generic mobile computer device 950 that may be used with the techniques described here. Computing device 900 is intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. Generic computer device may 900 correspond to computer 200 of FIG. 2. Computing device 950 is intended to represent various forms of mobile devices, such as personal digital assistants, cellular telephones, smart phones, driving assistance systems or board computers of vehicles, and other similar computing devices. For example, computing device 950 may be used as a frontend by a user (e.g., an operator of a blast furnace) to interact with the computing device 900. The components shown here, their connections and relationships, and their functions, are meant to be exemplary only, and are not meant to limit implementations of the inventions described and / or claimed in this document.

[0134] Computing device 900 includes a processor 902, memory 904, a storage device 906, a high-speed interface 908 connecting to memory 904 and high-speed expansion ports 910, and a low speed interface 912 connecting to low speed bus 914 and storage device 906. Each of the components 902, 904, 906, 908, 910, and 912, are interconnected using various busses, and may be mounted on a common motherboard or in other manners as appropriate. The processor 902 can process instructions for execution within the computing device 900, including instructions stored in the memory 904 or on the storage device 906 to display graphical information for a GUI on an external input / output device, such as display 916 coupled to high speed interface 908. In other implementations, multiple processors and / or multiple busses may be used, as appropriate, along with multiple memories and types of memory. Also, multiple computing devices 900 may be connected, with each device providing portions of the necessary operations (e.g., as a server bank, a group of blade servers, or a multi-processor system).P-DELPHI-460 / WO- 25 -

[0135] The memory 904 stores information within the computing device 900. In one implementation, the memory 904 is a volatile memory unit or units. In another implementation, the memory 904 is a non-volatile memory unit or units. The memory 904 may also be another form of computer-readable medium, such as a magnetic or optical disk.

[0136] The storage device 906 is capable of providing mass storage for the computing device 900. In one implementation, the storage device 906 may be or contain a computer- readable medium, such as a floppy disk device, a hard disk device, an optical disk device, or a tape device, a flash memory or other similar solid state memory device, or an array of devices, including devices in a storage area network or other configurations. A computer program product can be tangibly embodied in an information carrier. The computer program product may also contain instructions that, when executed, perform one or more methods, such as those described above. The information carrier is a computer- or machine-readable medium, such as the memory 904, the storage device 906, or memory on processor 902.

[0137] The high speed controller 908 manages bandwidth-intensive operations for the computing device 900, while the low speed controller 912 manages lower bandwidth-intensive operations. Such allocation of functions is exemplary only. In one implementation, the high-speed controller 908 is coupled to memory 904, display 916 (e.g., through a graphics processor or accelerator), and to high-speed expansion ports 910, which may accept various expansion cards (not shown). In the implementation, low-speed controller 912 is coupled to storage device 906 and low- speed expansion port 914. The low-speed expansion port, which may include various communication ports (e.g., USB, Bluetooth, Ethernet, wireless Ethernet) may be coupled to one or more input / output devices, such as a keyboard, a pointing device, a scanner, or a networking device such as a switch or router, e.g., through a network adapter.

[0138] The computing device 900 may be implemented in a number of different forms, as shown in the figure. For example, it may be implemented as a standard server 920, or multiple times in a group of such servers. It may also be implemented as part of aP-DELPHI-460 / WO- 26 - rack server system 924. In addition, it may be implemented in a personal computer such as a laptop computer 922. Alternatively, components from computing device 900 may be combined with other components in a mobile device (not shown), such as device 950. Each of such devices may contain one or more of computing device 900, 950, and an entire system may be made up of multiple computing devices 900, 950 communicating with each other.

[0139] Computing device 950 includes a processor 952, memory 964, an input / output device such as a display 954, a communication interface 966, and a transceiver 968, among other components. The device 950 may also be provided with a storage device, such as a microdrive or other device, to provide additional storage. Each of the components 950, 952, 964, 954, 966, and 968, are interconnected using various buses, and several of the components may be mounted on a common motherboard or in other manners as appropriate.

[0140] The processor 952 can execute instructions within the computing device 950, including instructions stored in the memory 964. The processor may be implemented as a chipset of chips that include separate and multiple analog and digital processors. The processor may provide, for example, for coordination of the other components of the device 950, such as control of user interfaces, applications run by device 950, and wireless communication by device 950.

[0141] Processor 952 may communicate with a user through control interface 958 and display interface 956 coupled to a display 954. The display 954 may be, for example, a TFT LCD (Thin-Film-Transistor Liquid Crystal Display) or an OLED (Organic Light Emitting Diode) display, or other appropriate display technology. The display interface 956 may comprise appropriate circuitry for driving the display 954 to present graphical and other information to a user. The control interface 958 may receive commands from a userand convert them for submission to the processor 952. In addition, an external interface 962 may be provide in communication with processor 952, so as to enable near area communication of device 950 with other devices. External interface 962 may provide, for example, for wired communication in some implementations, or for wireless communication in other implementations,P-DELPHI-460 / WO- 27 - and multiple interfaces may also be used.

[0142] The memory 964 stores information within the computing device 950. The memory 964 can be implemented as one or more of a computer-readable medium or media, a volatile memory unit or units, or a non-volatile memory unit or units. Expansion memory 984 may also be provided and connected to device 950 through expansion interface 982, which may include, for example, a SIMM (Single In Line Memory Module) card interface. Such expansion memory 984 may provide extra storage space for device 950, or may also store applications or other information for device 950. Specifically, expansion memory 984 may include instructions to carry out or supplement the processes described above, and may include secure information also. Thus, for example, expansion memory 984 may act as a security module for device 950, and may be programmed with instructions that permit secure use of device 950. In addition, secure applications may be provided via the SIMM cards, along with additional information, such as placing the identifying information on the SIMM card in a non-hackable manner.

[0143] The memory may include, for example, flash memory and / or NVRAM memory, as discussed below. In one implementation, a computer program product is tangibly embodied in an information carrier. The computer program product contains instructions that, when executed, perform one or more methods, such as those described above. The information carrier is a computer- or machine-readable medium, such as the memory 964, expansion memory 984, or memory on processor 952 that may be received, for example, over transceiver 968 or external interface 962.

[0144] Device 950 may communicate wirelessly through communication interface 966, which may include digital signal processing circuitry where necessary. Communication interface 966 may provide for communications under various modes or protocols, such as GSM voice calls, SMS, EMS, or MMS messaging, CDMA, TDMA, PDC, WCDMA, CDMA2000, or GPRS, among others. Such communication may occur, for example, through radio-frequency transceiver 968. In addition, short-range communication may occur, such as using a Bluetooth, WiFi, or other such transceiver (not shown). In addition, GPS (Global Positioning System) receiver module 980 mayP-DELPHI-460 / WO- 28 - provide additional navigation- and location-related wireless data to device 950, which may be used as appropriate by applications running on device 950.

[0145] Device 950 may also communicate audibly using audio codec 960, which may receive spoken information from a user and convert it to usable digital information. Audio codec 960 may likewise generate audible sound fora user, such as through a speaker, e.g., in a handset of device 950. Such sound may include sound from voice telephone calls, may include recorded sound (e.g., voice messages, music files, etc.) and may also include sound generated by applications operating on device 950.

[0146] The computing device 950 may be implemented in a number of different forms, as shown in the figure. For example, it may be implemented as a cellular telephone 980. It may also be implemented as part of a smart phone 982, personal digital assistant, or other similar mobile device.

[0147] Various implementations of the systems and techniques described here can be realized in digital electronic circuitry, integrated circuitry, specially designed ASICs (application specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which may be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0148] These computer programs (also known as programs, software, software applications or code) include machine instructions for a programmable processor, and can be implemented in a high-level procedural and / or object-oriented programming language, and / or in assembly / machine language. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus and / or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machine instructions and / or data to a programmable processor, including a machine-readableP-DELPHI-460 / WO- 29 - medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal used to provide machine instructions and / or data to a programmable processor.

[0149] To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0150] The systems and techniques described here can be implemented in a computing device that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a client computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network ("LAN"), a wide area network ("WAN"), and the Internet.

[0151] The computing device can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.

[0152] A number of embodiments have been described. Nevertheless, it will be understood that various modifications may be made without departing from the spirit and scopeP-DELPHI-460 / WO- 30 - of the invention.

[0153] In addition, the logic flows depicted in the figures do not require the particular order shown, or sequential order, to achieve desirable results. In addition, other steps may be provided, or steps may be eliminated, from the described flows, and other components may be added to, or removed from, the described systems. Accordingly, other embodiments are within the scope of the following claims.References-1 collecting phase-2 training phase-3 using phase100 vehicle101 exhaust line110 engine115 engine outlet with pipe (to the converter system)120 converter system (with engine side and tailpipe side), single-catalyst or multi-catalyst 121 pipe125 tailpipe140 tailpipe sensor191 driver192 inspector210 engine network220 converter network (220-A, 220-B for sub-networks)300 engine controller350 on-board monitoring (OBM) equipmenttl ... time-points

Claims

1. P-DELPHI-460 / WO- 31 -CLAIMS1. Computer-implemented method (400) to estimate pollutant emissions (P_ESTI MATED, P) at a tailpipe (125) of vehicles, wherein the exhaust line (101) from the engine (110) to the tailpipe (125) comprises a catalytic converter system (120) that is arranged between an engine outlet (115) of the engine (110) and the tailpipe (125), the method (400) comprising:by a first neural network (210) of a computer (200), processing (410) a first set of multiple engine parameters (E_ALPHA) to obtain data (EO_ESTI MATED) that represent the estimated content of multiple chemical compounds in the gas at the engine outlet (115), referred to as estimated engine-out data (EO_ESTI MATED) hereinafter;by a second neural network (220) of the computer (200), processing (420) a second set of multiple engine parameters (E_BETA), and the estimated engine-out data (EO_ESTI MATED) to obtain data that represents a preliminary estimate of the content of multiple compounds of the gas at the tailpipe (125), referred to as preliminary catalyst-out data (B_ESTI MATED) hereinafter;by a multiplier unit (230) of the computer (200), multiplying (430) the preliminary catalyst-out data (B_ESTI MATED) by a pre-defined factor vector (F) to obtain pollution estimate data (P_ESTIMATED) according to the degradation of the catalytic converter system (120).

2. Computer-implemented method (400) according to claim 1, wherein the preliminary catalyst-out data (B_ESTIMATED) and the pollution estimate data (P_ESTIMATED) are vectors with V elements each that correspond to V chemical compounds of the exhaust gas at the tailpipe (125), and wherein the pre-defined factor vector (F) has V factors that are specific to the chemical compounds, wherein at least one chemical compound is NOx.P-DELPHI-460 / WO- 32 - 3. Computer-implemented method (400) according to claim 2, wherein the computer (200) obtains a NOx-specific factor (F_NOx) and a NH3-specfic factor (F_NH3) from(i) a NOx value and NH3value (P_MAESURED_ NOx, P_MEASURED_NH3) that are measured by sensors (140) at the tailpipe (125) and(ii) the pollution estimate data for NOx (P_ESTIMATED_NOx) and for NH3(P_ESTIMATED_NH3).

4. Method according to claim 1, that is applied to estimate pollutant emissions (P_ESTIMATED) for gasoline or hydrogen engines, wherein processing (410) the first set of multiple engine parameters (E_ALPHA) comprises the first neural network (210) processing one or more engine parameters that are selected from the following (E_LONGLIST_ALPHA |gasoline / hydrogen):• Engine Speed *• Engine Load *• Total Injected Fuel *• Absolute Spark angle *• Split Injection Fueling and Pulse Angle• Rail Pressure• Turbine Flow *• CAM Phaser Intake Position• CAM Phaser Exhaust Position• Intake Total Air Mass *• WRAF Lambda *• Switch O2 rear sensor• TWC O2 Storage Capacity• Manifold Pressure *Atmospheric PressureSwirl PositionCoolant TemperatureP-DELPHI-460 / WO- 33 - Ambient Temperature• Intake Temperature *• Inlet Turbine Temperature *.

5. Method (400) according to claim 4, wherein processing (420) the second set of multiple engine parameters (E_BETA) comprises the second neural network (220) processes the first set (E_ALPHA) as well as the following parameters:TWC1 Input / Output Temperature and TWC2 Input / Output Temperature.

6. Method (400) according to any of claims 4 or 5, wherein the first neural network (210) is processing (410) the first set of multiple engine parameters (E_ALPHA) and the second neural network (220) is processing (420) the second set of multiple engine parameters (E_BETA) such that the first set (E_ALPHA) is a subset of the second set (E_BETA).

7. Method (400) according to any of claims 4 to 6, wherein the first neural network (210) and the second neural network (220) are processing (410, 420) the first set of multiple engine parameters that includes the following parameters (E_SHORT | gasoline / hydrogen):• Engine Speed,• Engine Load,• Total Injected Fuel,• Absolute Spark angle,• Turbine Flow,• Intake Total Air Mass,• WRAF Lambda,• Manifold Pressure,• Intake Temperature,• Inlet Turbine Temperature,• TWC1 Input / Output Temperature, andP-DELPHI-460 / WO- 34 - TWC2 Input / Output Temperature.

8. Method according to claim 1, that is applied to estimate pollutant emissions (P_ESTIMATED) for diesel engines, wherein processing (410) the first set of multiple engine parameters (E_ALPHA) comprises the first neural network (210) processing one or more engine parameters that are selected from the following (E_LONGLIST_ALPHA | diesel):• Engine Speed• Engine Load• Total Injected Fuel• Injected fuel without Post injection• Pilot Injection Fueling and Timing• Main Injection Fueling and Timing• After Injection Fueling and Timing• Post Injection Fueling and Timing• Intake Total Air Mass• EG R inert rate• Turbine Flow• Rail Pressure• Manifold Pressure• Atmospheric Pressure• Swirl Position• Air / Fuel Ratio• Coolant Temperature• Ambient Temperature• Intake Temperature• Inlet Turbine Temperature.

9. Method (400) according to claim 8, wherein processing (420) the second set of multiple engine parameters (E_BETA) comprises the second neural network (220) processing theP-DELPHI-460 / WO- 35 - first set (E_ALPHA) as well as the following parameters:• DOC Input / Output Temperature,• SCR Input / Output Temperature.

10. Method (400) according to any of claims 4 to 6, wherein the first neural network (210) and the second neural network (220) are processing (410, 420) the first set of multiple engine parameters that includes the following parameters (E_SHORT | diesel):• Engine Speed,• Engine Load,• Total Injected Fuel,• Intake Total Air Mass,• EGR inert rate,• Turbine Flow,• Manifold Pressure,• Air / Fuel Ratio,• Intake Temperature,• Inlet Turbine Temperature,• DOC Input / Output Temperature, and• SCR Input / Output Temperature.

11. Method (400) according to any of claims 1 to 10, wherein in the step processing (410) multiple engine parameters (E_ALPHA), the computer uses a first neural network (210) that has been trained to provide the estimated engine-out data (EO_ESTI MATED), with data that represents, in a first set, the content of the following compounds: NOx, CO, HC, CH4, CO2, O2 and that represents, in a second set, the content of particulate matter, PM.

12. Method (400) according to claim 4, wherein in the step processing (420) the multiple engine parameters (E_ALPHA) and the estimated engine-out data (EO_ESTIMATE), the computer uses a second neural network (220) that has been trained, with data thatP-DELPHI-460 / WO- 36 - represents the content of the following compounds: NOx, CO, HC, CH4, CO2, NH3.

13. Method (400) according to any of the claims 1 to 12, wherein the multiplier unit (230) uses a vector with factors that represent the degradation of the catalyst in the catalytic converter system (120), and wherein the factors are specific to the following compounds: NOx, CO, HC, CH4, CO2, NH3.

14. Method (400) according to any of claim 1 to 13, wherein the second neural network (220) is processing (420) the estimated engine-out data (EO_ESTI MATED) to obtain the preliminary catalyst-out data (B_ESTI MATED),by a first sub-network (210-A) that receives estimated engine-out data (EO_ESTI MATED) and provides and intermediate estimate (A_ESTI MATED),by a first sub-network (210-A),wherein both the first sub-network (210-A) and the second sub-network (210-B) also process the multiple engine parameters (E).

15. Method (400) according to any of claims 1 to 13, wherein the processing steps (410, 420) further comprises to process measurement values (C_MEASURED) that are obtained from sensors that are located at the catalytic converter system (120), and wherein the computer performs data pre-processing to align measurement data in time according to propagation delays of the gas components within the exhaust line (101).

16. Method (400) according to claim 15, wherein aligning measurement data in time is performed by buffering data and wherein the network (210, 220) receives the data from the buffer one data (for all vector elements) becomes available.

17. Method (400) according to claim 16, wherein the interval to buffer data corresponds to the propagation time for chemical compounds as they move through the exhaust lineP-DELPHI-460 / WO- 37 - (101) from the engine (110) to sensors (140).

18. A computer program product that, when loaded into a memory of a computer system and executed by at least one processor of the computer system, causes the computer system (200) to perform the steps of a computer-implemented method according to any of the claims 1 to 17.

19. A computer system (200) comprising a plurality of modules that perform the steps of the computer-implemented method according to any of the claims 1 to 17.

20. Computer-implemented method to train a first neural network (210) and a second neural network (220) of a computer (200), to estimate pollutant emissions (P_ESTI MATED, P) at a tailpipe (125) of vehicles, wherein an exhaust line (101) from an engine (110) to a tailpipe (125) comprises a catalytic converter system (120) that is arranged between an engine outlet (115) of the engine (110) and the tailpipe (125), and wherein, after training, the first neural network (210) is adapted to process (410) a first set of multiple engine parameters (E_ALPHA) to obtain data (EO_ESTIMATED) that represent the estimated content of multiple chemical compounds in the gas at the engine outlet (115), referred to as estimated engine-out data (EO_ESTI MATED) hereinafter, and the second neural network (220) is adapted to process (420) a second set of multiple engine parameters (E_BETA) and the estimated engine-out data (EO_ESTIMATED) to obtain data that represents a preliminary estimate of the content of multiple compounds of the gas at the tailpipe (125), referred to as preliminary catalyst-out data (B_ESTIMATED) hereinafter, and wherein a multiplier unit (230) is adapted to multiply (430) the preliminary catalyst-out data (B_ESTIMATED) by a predefined factor vector (F) to obtain pollution estimate data (P_ESTIMATED) according to the degradation of the catalytic converter system (120),the method to train the first neural network (210) and the second neural network (220) being performed with historical data from a reference catalyst with a standardized age, or performed with data from a reference catalyst with non-standardized age, whereinP-DELPHI-460 / WO-38 -the data has been normalized to standardized age, so that historical data that represent degradation of the catalytic converter system is used to identify the predefined factor vector (F) of the multiplier unit (230).

Images