Method for detecting misfires and estimating hydrogen concentration

Abstract

Method (100) of detecting misfires and estimating the volumetric concentration of hydrogen in the exhaust gases of an internal combustion engine, the method comprising the following steps: - acquiring (10) a set of input data, - identifying (20) misfire events using a neural network (21), - calculating (40), for each misfire event identified, the volumetric concentration of hydrogen in the exhaust gases of the internal combustion engine, - comparing (80) the regulated limit values of the volumetric concentration of hydrogen with the previously calculated volumetric fuel concentration values, and - in the event that at least one calculated hydrogen concentration value is higher than a respective hydrogen concentration limit value, implementing (90) a recovery action on the internal combustion engine.

Description

[0001] METHOD FOR DETECTING MISFIRES AND ESTIMATING HYDROGEN

[0002] CONCENTRATION

[0003] D E S C R I PTI O N

[0004] Technical field of the invention

[0005] The present invention relates to a method of detecting fuel misfires in one or more cylinders of an internal combustion engine for the purpose of estimating the concentration of unburned fuel in the exhaust gases of the engine. The method is particularly suitable for internal combustion engines fueled by hydrogen, but not limited to such fuel.

[0006] Background art

[0007] Motor vehicles typically operate by using an internal combustion engine to convert the energy in a fuel, such as gasoline or diesel, into mechanical energy to propel the motor vehicle and then provide motion to the vehicle's wheels. Unfortunately, fossil fuels are expensive and contribute to environmental pollution. Because of these drawbacks, increasing attention is being paid to reducing fuel consumption and pollutants emitted by automobiles and other highway vehicles.

[0008] To alleviate some of these drawbacks, hydrogen-fueled internal combustion engines have been proposed that do not produce polluting emissions except for negligible quantities of nitrogen oxides. However, such engines require special precautions to ensure proper operation.

[0009] In fact, hydrogen ignition is strongly influenced by the air / hydrogen ratio: in particular, ignition is difficult at low load, in the case of lean mixtures, and it may happen that ignition does not occur at all in a cylinder, causing the well-known "misfire" phenomenon.

[0010] Misfire is essentially a lack of combustion in one or more cylinders.

[0011] It should be detected for the following reasons:

[0012] - safety of gas-fueled engines, especially hydrogen, without an oxidation device in the exhaust line,

[0013] - reduction in power, irregular torque delivery and increased emissions,

[0014] - misfire detection is required by regulatory authorities,

[0015] - possible degradation or damage to engine components.

[0016] There are several known misfire detection strategies, based on the processing of specific parameters: for example, the rotary motion of the crankshaft, the instantaneous pressure of the exhaust gases, the ionization signal in the combustion chamber (detected by a spark plug), the trend of the engine torque, the trend of the pressure in the combustion chamber, optical methods, the signal from the knock sensor.

[0017] The most commonly used known strategy is based on the identification of misfires through post-processing of the crankshaft rotation signal. The crankshaft is equipped with a phonic wheel, for example, of the 60-2 tooth type; a position sensor and a revolution sensor (often combined) identify the position of the crankshaft and its rotation speed by monitoring the rotational motion of the phonic wheel that has a singularity at the two missing teeth. Specifically, the temporal data related to the presence / absence of the teeth of the phonic wheel are post-processed to calculate the variables that will be compared with the threshold values. The usual approach is to "post-process" the tone wheel signal: for example, comparing the acceleration variables and their time derivative ("jerk") with calibratable parameters, dependent on the engine operating point and the misfiring cylinder. Consequently, a misfire event is recognized when both of these variables are below predetermined thresholds.

[0018] Although these strategies require a relatively simple architecture and have a proven detection efficacy, each method works efficiently only in some conditions and there is no method that reliably identifies misfires in all different engine conditions. Furthermore, and consequently, the known methods require a heavy effort for their calibration: they require, in fact, multiple algorithmic logics depending on the engine operating point and the specific cylinder that misfires.

[0019] Due to the low ignition energy and high flame propagation, misfire detection in hydrogen-fueled internal combustion engines is considered a safety-relevant feature. In fact, the regulations concerning the specifications of a vehicle fuel system incorporating the compressed hydrogen storage system provide two safety limits, in terms of hydrogen concentration. In particular, the following conditions must be respected, dictated by specific regulations which in the case of stationary engines (NRMM) are: "Regulation No 134 Addendum 133 - Part III":

[0020] - not exceed 8% of the volumetric concentration of hydrogen in the exhaust gas at any time,

[0021] - not exceed 4% average by volume, of the hydrogen concentration in the exhaust gas, during any rolling time interval of 3 seconds.

[0022] Unfortunately, in the case of hydrogen, a sensor for the concentration of unburned fuel in the exhaust line is not yet available.

[0023] There is therefore a need to define a method for detecting misfires and estimating the hydrogen concentration in the exhaust gases of an internal combustion engine that is free from the above-mentioned drawbacks.

[0024] Summary of the Invention

[0025] In order to substantially solve the above-mentioned technical problems, an object of the present invention is a method of detecting misfire and estimating the volumetric concentration of fuel, in particular hydrogen, in the exhaust gases of an internal combustion engine, the method being based on detecting misfire of fuel in a cylinder of an internal combustion engine and estimating the concentration of hydrogen in the exhaust gases of the engine itself. The misfire detection is a neural networkbased algorithm.

[0026] Accordingly, according to the present invention, a method of detecting misfires and estimating the volumetric concentration of fuel in the exhaust gases of an internal combustion engine is provided having the characteristics set forth in the independent claim, attached to the present description.

[0027] Further preferred and / or particularly advantageous embodiments of the invention are described according to the characteristics set forth in the attached dependent claims.

[0028] Brief description of the drawings

[0029] The invention will now be described with reference to the attached drawings, which illustrate some non-limiting examples of its implementation, in which:

[0030] - figure 1 is a logic diagram of the method of detecting misfire and estimating the volumetric concentration of fuel in the exhaust gases of an internal combustion engine according to a preferred embodiment of the invention,

[0031] - figure 2 is a first detail of the logic diagram of figure 1,

[0032] - figure 3 is a graph of the volumetric concentration of hydrogen in the exhaust gases, and

[0033] - figure 4 is a second detail of the logic diagram of figure 1.

[0034] Detailed description

[0035] By way of example and not limitation, the method of detecting misfire and estimating the concentration of hydrogen in the exhaust gases of an internal combustion engine will now be described with reference to the above figures. It should be noted that the method, specifically applied to hydrogen fuel, can also be used for other fuels, modifying the fuel parameters and the safety requirements related to the presence of unburned fuels in a way that is obvious to a technician in the sector.

[0036] With particular reference to Figure 1, the method 100 according to the present invention is based on the following steps:

[0037] - acquiring 10 a set of input data,

[0038] - identifying 20 misfire events by means of a neural network,

[0039] - acquiring 30 the misfire events, previously detected,

[0040] - calculating 40, for each misfire event, the volumetric concentration of hydrogen - instantaneous and averaged on a 3-second moving average - in the exhaust gases of the internal combustion engine, as a function of the air / fuel ratio,

[0041] - acquiring 50 the volumetric concentrations of hydrogen previously calculated,

[0042] - implementing 60 the regulation with the safety requirements related to the presence of unburned fuel in the exhaust gases of the internal combustion engine,

[0043] - acquiring 70 the limit values of the volumetric concentration of hydrogen from the previously implemented regulation,

[0044] - comparing 80 the regulated limit values of the volumetric concentration of hydrogen with the volumetric concentration values of hydrogen previously calculated,

[0045] - in case at least one calculated hydrogen concentration value is higher than a respective hydrogen concentration limit value, implement 90 a recovery action provided for by the regulation, for example, an immediate shutdown of the internal combustion engine.

[0046] In particular, and also with reference to figure 2, the input data set may include, in the case of an eight-cylinder engine and to optimize the "accuracy / computational cost" trade-off:

[0047] - four values of the crankshaft acceleration 11, each corresponding to a respective cylinder,

[0048] - four values of the derivative 12 of the crankshaft acceleration ("Jerk"), each corresponding to a respective cylinder,

[0049] - the value of the required driving torque 13, and

[0050] - the identification number 14 of a specific cylinder (in the example of the eight-cylinder engine, from 1 to 8). Advantageously, the input dataset can already be available in the Engine Management System (EMS).

[0051] A misfire event can then be detected through a data-driven approach: given a matrix of measurements and predefined training settings, the calibration process can be automated. As better described below, training can be performed, preferably offline, and the parameters of the neural network are provided as calibrations.

[0052] Still referring to figure 2, the neural network 21 used to detect 20 misfire events is organized by means of input layers 22, intermediate layers 23 and output layers 24.

[0053] An input layer 22 stores the input dataset, previously described. For example, the neural network 21 can be organized with an input layer comprising 10 neurons. The number 10 corresponds to the number of data blocks (e.g., 4 accelerations, 4 jerks, 1 torque, 1 cylinder ID). The activation function is a hyperbolic tangent that normalizes the data within the range -1 1.

[0054] At least one output layer 24 provides the probability 25 of misfire. In the proposed example, the neural network 21 has a single output layer comprising a neuron. The activation function, in this case, is a sigmoid function used to introduce non-linearity in the model and to normalize the values within the range 0 1.

[0055] Finally, the neural network has a plurality of intermediate layers 23 designed to improve the accuracy of the neural network itself. However, a very high number of intermediate layers would require excessive computational effort and, therefore, the number of intermediate layers must be chosen by optimizing the accuracy I computational effort tradeoff. In the proposed example, four intermediate layers were selected, each equipped with seven neurons. The activation function is, also in this case, the hyperbolic tangent that normalizes the data within the range -1 -?1.

[0056] The misfire probability 25 is then compared 27 with a predetermined misfire probability threshold value 26 and if the estimated value is higher than the threshold value, the misfire event is acquired 30.

[0057] As mentioned, to take full advantage of the performance of the neural network, it needs to be properly trained. The neural network training is done offline and is based on misfire data that can be collected using a misfire generator in the engine control unit. The misfire generator can simulate misfire events by deactivating the spark plug or fuel injection at a predetermined frequency on a cylinder-by-cylinder basis. Therefore, the misfire generator data is used to train the neural network, through supervised learning techniques, and create the correlation between input data and output results.

[0058] In particular and by way of example, the generated data can be divided into a first group of data (training subset) which is used for training the neural network and into a second group of data (validation subset) which is used to validate the training achieved by the neural network. Preferably, the training subset contains approximately 80% of the data, while the validation subset contains the remaining 20% of the data. The selection between training data and validation data is carried out in a random manner.

[0059] Training is performed by splitting the data into small blocks (e.g., 32) to facilitate learning and setting a predetermined number of repetitions (e.g., 50) and a learning rate.

[0060] It is evident that within the distribution of the data set, the data containing no misfires are dominant over misfire events. Such a class imbalance could significantly alter the performance of the neural network once trained. This happens because the loss function algorithm will be more likely to assign weights to classify the event as no misfire rather than to recognize misfire phenomena. Such an imbalance can be fixed by acting on the implementation of the "Focal loss" cost function, a mathematical function whose value is minimized during the training process. For this purpose, a factor is introduced to the aforementioned cost function, useful for obtaining a lower weighting of frequent events (no misfires) and a higher weighting of rare events (misfires). In other words, the factor y reduces the influence of frequent events on the loss function, and emphasizes the role of rare events.

[0061] Returning to Figure 1, once the misfire event has been identified, the next step of the method according to the present invention is to calculate 40 the volumetric concentration of hydrogen in the exhaust gases of the internal combustion engine. This step is essential, since a hydrogen concentration sensor is not currently available.

[0062] The hydrogen concentration estimate is made based on the following assumptions:

[0063] - hydrogen is considered an ideal gas,

[0064] - there are no partial misfire phenomena,

[0065] - no nitrogen oxide (NOx) production, and - no exhaust gas recirculation (EGR).

[0066] The estimation of the volumetric concentration of hydrogen starts from the general oxidation reaction of a hydrocarbon as a function of the equivalent ratio equal to the equivalent air / fuel ratio. It should be noted that this formulation is valid for lean and / or stoichiometric mixtures applying the specific coefficients of hydrogen to the chemical formula of the generic fuel: we obtain the following formula, with the air / fuel equivalence ratio as an explicit parameter (for lean and / or stoichiometric combustion,

[0067] Therefore, the hydrogen concentration is calculated as a mole fraction as a function of the misfire event rate in a given moving time window:

[0068] Explaining from formula (3) the volumetric concentration of hydrogen in the exhaust gases (XH2,exh) you get:

[0069] (4)

[0070] If this method is applied to a fuel other than hydrogen, for example methane, if necessary the formulation valid for rich mixtures may also be used

[0071] The graph in figure 3 shows the volumetric concentration of hydrogen (XH2,Exh) depending on the air / fuel ratio and of the cylinder that loses ignition. From the graph it can be deduced that the concentration value will decrease as the air / fuel ratio increases This result should not be surprising: in fact, the value of the air / fuel ratio determines the relative quantity of fuel and air. If this value is high (a lot of air in the mixture in fuel ratio), the concentration of hydrogen in the exhaust gases will be lower. Conversely, if the value of the air / fuel ratio is low < 1, i.e. little air inside the mixture in fuel ratio) the concentration of hydrogen released into the exhaust will be higher. Furthermore, the concentration value of hydrogen at the exhaust may depend on the specific cylinder in which the misfire event occurs: this dependence may be due to the shape of the intake ducts that supply a different quantity of air to the different cylinders or to the precision of the injector of the respective cylinder. Normally this dependence is of little importance and can be neglected in numerical processing.

[0072] Using these formulas, the method according to the present invention continues with the step of acquiring 50 the volumetric concentration of hydrogen in the exhaust gases of the internal combustion engine and, in particular, the instantaneous volumetric concentration and the average volumetric concentration over a moving time interval of 3 seconds, as required by the reference regulation.

[0073] The method must then verify whether the safety requirements imposed by the current regulations are met.

[0074] In fact, due to the low ignition energy and the high flame propagation, the detection of misfires in the cylinders of hydrogen internal combustion engines is considered a safety-relevant feature.

[0075] In accordance with Regulation No. 134 Addendum 133 - Part III - "Specifications of a fuel system of a vehicle incorporating the compressed hydrogen storage system", safety limits have been identified, in terms of hydrogen concentration. In particular, the following limits shall be considered:

[0076] - not to exceed 8% volumetric concentration at any time, and

[0077] - not to exceed 4% average by volume during any moving time interval of 3 seconds.

[0078] Having acquired 60 the current regulation and acquired 70 the limits of the hydrogen concentration from the previously implemented regulation, the method proceeds with the comparison between the acquired limit values and the calculated current values.

[0079] With reference to figure 4, a first logic block 85 performs the comparison between the instantaneous limit value 81 of the volumetric concentration of hydrogen (8%) and the instantaneous current value 83 of the volumetric concentration of hydrogen previously calculated. A second logic block 86 instead performs the comparison between the average limit value 82 of the volumetric concentration of hydrogen on a moving average of 3 seconds (4%) and the current average value 84 of the volumetric concentration of hydrogen on the same moving average of 3 seconds. Therefore, to comply with the safety requirements, it is necessary to ensure that the instantaneous concentration of hydrogen does not exceed 8% and that the concentration on a moving average of 3 seconds does not exceed 4%. A third logic block 89 (block of the "OR" type) checks whether the output 87 from the first logic block 85 signals the exceeding of the first instantaneous concentration limit threshold equal to 8% or whether the output 88 from the second logic block 86 signals the exceeding of the second average concentration limit threshold (on a moving average of 3 seconds) equal to 4%.

[0080] If at least one calculated hydrogen concentration value is higher than a respective limit value of hydrogen concentration, the method provides for the step of performing 90 an immediate shutdown of the internal combustion engine.

[0081] The method described above, in addition to providing a valid solution for the detection of cylinder misfires and the estimation of the hydrogen concentration in the exhaust gases, presents further undoubted advantages:

[0082] - the implementation of the relative algorithm does not require a specific control unit;

[0083] - the algorithm can be customized for the end user's needs in any control unit, the accuracy of the results obtainable being linked only to the resolution of the crankshaft position sensor;

[0084] - the method is scalable for any type of internal combustion engine and for any fuel used;

[0085] - the method can also be used as a "calibration service". Data analysis and calibration parameters can be obtained directly from the described method in the face of a test campaign on the end user's application;

[0086] - as already mentioned, currently in the case of hydrogen there is no unburned fuel concentration sensor available in the exhaust line and, therefore, this method is essential and indispensable for hydrogen applications. In any case, even in cases where a concentration sensor is present, this method will still be indispensable and mandatory for sensor diagnostics.

[0087] In addition to the embodiment of the invention as described above, it is to be understood that there are numerous other variations. It is also to be understood that such embodiments are exemplary only and do not limit the scope of the invention, its applications, or its possible configurations. On the contrary, while the above description allows the skilled person to practice the present invention at least according to one exemplary embodiment thereof, it is to be understood that many variations of the described components are possible without departing from the scope of the invention as defined in the appended claims, which are interpreted literally and / or according to their legal equivalents.

Claims

C LA I M S1. Method (100) of detecting misfires and estimating the volumetric concentration of hydrogen in the exhaust gases of an internal combustion engine, the method comprising the following steps:- acquiring (10) a set of input data,- identifying (20) misfire events using a neural network (21),- calculating (40), for each misfire event identified, the volumetric concentration of hydrogen in the exhaust gases of the internal combustion engine,- comparing (80) the regulated limit values of the volumetric concentration of hydrogen with the previously calculated volumetric fuel concentration values, and- in the event that at least one calculated hydrogen concentration value is higher than a respective hydrogen concentration limit value, implementing (90) a recovery action on the internal combustion engine, wherein the neural network (21) is trained by using a factor (y) of a cost function which reduces the weight of misfire events absence and increases the weight of misfire events.

2. Method (100) according to claim 1, wherein the neural network (21) comprises:- an input layer (22) that stores the input data set,- intermediate layers (23), configured to improve the accuracy of the neural network (21), the number of which optimizes the accuracy / computational effort trade-off, and- at least one output layer (24) which provides a probability (25) of misfire.

3. Method (100) according to claim 2, wherein a misfire event is acquired (30) if the probability (25) of misfire is greater than a predetermined threshold value (26).

4. Method (100) according to claim 2 or 3, wherein the activation function of the input layer (22) and the intermediate layers (23) is a hyperbolic tangent, while the activation function of the at least one output layer (24) is a sigmoid function.

5. Method (100) according to any of the previous claims, wherein the neural network (21) is trained offline using a set of data produced by a misfire generator and through supervised learning techniques.

6. Method (100) according to claim 5, wherein the training is carried out by dividing the data into blocks of reduced dimensions and setting a predetermined number of repetitions and a learning rate.

7. Method (100) according to any of the previous claims, wherein the volumetric concentration of hydrogen in the exhaust gases calculated is the instantaneous one and the average one on a 3-second moving average.

8. Method (100) according to any of the previous claims, wherein the step of comparing (80) the regulated limit values of the volumetric hydrogen concentration with the previously calculated volumetric hydrogen concentration values includes:- a first logic block (85) which performs the comparison between the instantaneous limit value (81) of the volumetric hydrogen concentrationand the previously calculated instantaneous current value (83) of the volumetric hydrogen concentration,- a second logic block (86) which performs the comparison between the average limit value (82) of the volumetric hydrogen concentration on the moving average of 3 seconds and the average current value (84) of the volumetric hydrogen concentration on the same moving average of 3 seconds.

Images