Method for sensing closing time of injector using artificial neural network and method for controlling injector using the same

Abstract

A method for sensing a closing time of an injector using an artificial neural network may include: sensing, by a controller, a voltage generated by an injector; performing, by the controller, a preprocess to derive an input matrix using variation characteristics of the voltage; and performing, by the controller, a closing time prediction to derive a closing time of the injector by an artificial neural network model including an input layer including the input matrix, a hidden layer, and an output layer.

Claims

1. A method for sensing a closing time of an injector using an artificial neural network, the method comprising: sensing, by a controller, a voltage generated by an injector; performing, by the controller, a preprocess to derive an input matrix using variation characteristics of the voltage; and performing, by the controller, a closing time prediction to derive a closing time of the injector by an artificial neural network model including an input layer including the input matrix, a hidden layer, and an output layer, wherein the variation characteristics of the voltage are a half-life constant of the voltage, and the half-life constant is a value calculated as: $k=-\frac{t}{{\log }_2\frac{V_t}{V_0}}$ where, V.sub.t is a voltage value V for each measurement point, V.sub.0 is a voltage value at an initial measurement point, k is a half-life constant, and t is a time at a measurement point.

2. The method of claim 1, wherein performing the preprocess comprises: calculating half-life constants at a plurality of measurement time points of the voltage in a specific section; deriving an approximation polynomial for changes in the calculated half-life constants in accordance with the time; and deriving the input matrix by normalizing respective coefficients of the approximation polynomial.

3. The method of claim 2, wherein in deriving the approximation polynomial, the coefficients of the approximation polynomial are derived using a normal equation utilizing linear algebra.

4. The method of claim 2, wherein in performing the closing time prediction, the hidden layer derives a first preparation matrix by multiplying the normalized input matrix by a first weight matrix and adding a first bias matrix to the multiplied matrix, and the hidden layer derives a first resulting matrix by substituting a transfer function of the following equation for the first preparation matrix, $a^1=\frac{2}{1+e^{-2n^1}}-1$ where, a.sup.1 is the first resulting matrix, and n.sup.1 is the first preparation matrix.

5. The method of claim 4, wherein in performing the closing time prediction, the output layer calculates a normalized closing time of the injector by multiplying the first resulting matrix by a second weight matrix and adding a bias value to the multiplied matrix, and the output layer calculates a final closing time of the injector by de-normalizing the calculated closing time of the injector.

Description

DRAWINGS

(1) In order that the disclosure may be well understood, there will now be described various forms thereof, given by way of example, reference being made to the accompanying drawings, in which:

(2) FIG. 1 is a diagram schematically illustrating the configuration of an injector control system in which a method for sensing a closing time of an injector and a method for controlling an injector using the sensing method are performed;

(3) FIG. 2 is a flowchart of a method for sensing a closing time of an injector;

(4) FIG. 3 is a block diagram explaining the method for sensing a closing time of an injector as illustrated in FIG. 2;

(5) FIG. 4A is a diagram illustrating a voltage section used to sense a closing time of an injector, FIG. 4B is a diagram illustrating a waveform K composed of a half-life constant k calculated at each measurement point, and FIG. 4C is a diagram illustrating the result of curve fitting with fifth-order polynomials with respect to the waveform K of FIG. 4B;

(6) FIG. 5 is a block diagram illustrating an operation logic including an artificial neural network model composed of an input layer, a hidden layer, and an output layer;

(7) FIG. 6 is a flowchart of a method for controlling an injector using the method for sensing a closing time of an injector illustrated in FIG. 2;

(8) FIG. 7 is a flowchart of a method for learning a closing time of an injector being executed in the method for controlling an injector as illustrated in FIG. 6;

(9) FIGS. 8A to 8B are graphs illustrating injector drive characteristics;

(10) FIG. 9 is a graph showing the relationship between an injector opening period that is a period in which a fuel is actually injected from an injector, current being applied to the injector, and a voltage being generated from the injector; and

(11) FIG. 10 is a graph illustrating changes in a fuel amount m for each section in accordance with an injector operation time T.sub.i.

(12) The drawings described herein are for illustration purposes only and are not intended to limit the scope of the present disclosure in any way.

DETAILED DESCRIPTION

(13) The following description is merely exemplary in nature and is not intended to limit the present disclosure, application, or uses. It should be understood that throughout the drawings, corresponding reference numerals indicate like or corresponding parts and features.

(14) Hereinafter, a method for sensing a closing time of an injector and a method for controlling an injector using the same in accordance with exemplary forms of the present disclosure will be described with reference to the accompanying exemplary drawings.

(15) FIG. 1 is a diagram schematically illustrating the configuration of an injector control system in which a method for sensing a closing time of an injector and a method for controlling an injector using the sensing method according to one form of the present disclosure are performed.

(16) Referring to FIG. 1, the injector control system includes a fuel tank 150, a fuel pump 140, a rail 130, a pressure sensor 132, an injector 120, an engine 110, and a controller 100.

(17) The fuel tank 150 is filled with a fuel being used for an internal combustion engine, and the fuel pump 140 pumps the fuel, contained in the fuel tank 150, to the rail. On the rail 130, the pressure sensor 132 for sensing an internal pressure is deployed, and a separate regulator valve (not illustrated) and a return line (not illustrated) are formed.

(18) The fuel pumped to the common rail 130 is distributed to the injector 120, and the injector 120 is deployed corresponding to each cylinder to inject the fuel to combustion chambers of the engine 110.

(19) The controller 100 may set a required fuel injection amount in accordance with a driving condition, for example, an engine rpm and an acceleration pedal signal, and may control an injection command time of the injector corresponding to the set required fuel injection amount.

(20) In one form of the present disclosure, the controller 100 controls the operation of the injector by applying current to the injector 120 for a required injection command time, senses a fuel injection pressure through reception of a pressure signal from the pressure sensor 132, and senses a voltage formed in the injector 120. Further, to be described later, the controller 100 senses a closing time of the injector 120 by variation characteristics of the voltage generated by the injector 120 and an artificial neural network model.

(21) The controller 100 senses the closing time of the injector 120, and can operate an actual fuel injection amount through an opening time, the closing time, and the fuel pressure.

(22) In addition, to be described later, the controller 100 may correct the required injection command time based on the sensed closing time of the injector, and may correct the fuel injection amount of the injector 120 more accurately.

(23) Hereinafter, referring to FIGS. 2 and 3, a method for sensing a closing time of an injector being performed by the controller 100 of FIG. 1 will be described.

(24) FIG. 2 is a flowchart of a method for sensing a closing time of an injector according to one form of the present disclosure, and FIG. 3 is a block diagram explaining the method for sensing a closing time of an injector as illustrated in FIG. 2.

(25) As illustrated in FIG. 2, the method for sensing a closing time of an injector in accordance with the present disclosure includes sensing a voltage generated by an injector (S10), performing preprocessing to derive an input matrix using variation characteristics of the voltage (S20), and performing closing time prediction to derive a closing time of the injector by means of an artificial neural network model including an input layer composed of the input matrix, a hidden layer, and an output layer (S30).

(26) First, the controller 100 senses the voltage generated by the injector (S10). As illustrated in FIG. 9, when an injection control is ended, a reverse voltage signal is naturally generated from the injector 120. The controller 100 senses the reverse voltage signal using a sensor (not illustrated) or the like.

(27) Next, the controller 100 performs the preprocessing to derive the input matrix using the variation characteristics of the voltage (S20).

(28) During the preprocessing, the controller 100 first selects a voltage section used to sense the closing time of the injector. In FIG. 4A, the selected voltage section is indicated by a rectangle indicated by a thick line. The voltage section is selected as a section in which a probability that the closing time of the injector exists in the entire voltage section is relatively high, and the range of the voltage section can be changed to be reduced in accordance with a calculation capability of an ECU or based on data on the closing time of the injector accumulated by learning. In this case, the calculation amount for sensing the closing time of the injector can be reduced. In the case of the reverse voltage, the measurement value has a negative (−) value, and by substituting a positive (+) value for the negative (−) value, inverted data shown in FIG. 4A can be obtained.

(29) If the voltage section to be used to sense the closing time of the injector is selected, measurement points are set in plural predetermined sections in the corresponding section, a half-life constant is calculated using the voltage values measured at the respective measurement points.

(30) The half-life constant indicates a half-life of the voltage values at the respective measurement points with respect to a voltage value V0 measured at a first start point of a waveform, and the half-life constant may be calculated by the following equation 1,

(31) $\begin{array}{cc}k=-\frac{t}{{\log }_2\frac{V_t}{V_0}}& \left(1\right)\end{array}$

(32) wherein V.sub.t is a voltage value V for each measurement point, V.sub.0 is a voltage value V at an initial measurement point, k is a half-life constant, and t is a time at a measurement point.

(33) The graph of FIG. 4B is obtained using data on the half-life K calculated at the respective measurement points.

(34) Next, the controller 100 derives a polynomial function similar to a waveform K by performing multi-order polynomial regression with respect to the waveform K of the half-life constant illustrated in FIG. 4B. In the present form, as illustrated in FIG. 4C, an error least function of the waveform K is estimated exemplarily using a fifth-order polynomial, and as a result, the following equation 2 is obtained. However, the present disclosure is not limited to the fifth-order polynomial regression, but corresponds to the error least function of the waveforms K. In the case where the ECU has sufficient calculation capability, the ECU may use polynomial regressions higher than fifth-order polynomial regression.
k=I.sub.1x.sup.5+I.sub.2x.sup.4+I.sub.3x.sup.3+I.sub.4x.sup.2+I.sub.5x+I.sub.6  (2)

(35) As described later, respective coefficients I.sub.1 to I.sub.6 constituting the equation 2 become input variables constituting an input matrix of an artificial neural network (ANN) model that predicts the closing time.

(36) Meanwhile, in the polynomial regression, as an algorithm estimating the respective coefficients I.sub.1 to I.sub.6, an optimization algorithm may be generally used. However, due to the realistic limit of the calculation capability of the ECU, it may be difficult to estimate coefficients through performing of the optimization algorithm in some cases. In these cases, the input variables constituting the input matrix can be calculated using the following normal equation 3 utilizing leading algebra.
(A.sup.TA).sup.−1A.sup.Tk=I   (3)

(37) In this case, A exemplifies a multi-order term matrix with respect to values (time) on an x axis of FIG. 4B, and if the time t has a value in the range of 1 to 70, the input variables can be calculated by a 6×70 matrix in the following equation 4.

(38) $\begin{array}{cc}\left[\begin{array}{ccccc}{A}_{11}& A_{12}& A_{13}& \mathrm{.Math.}& A_{16}\\ A_{21}& A_{22}& A_{23}& \mathrm{.Math.}& A_{26}\\ A_{31}& A_{32}& A_{33}& \mathrm{.Math.}& A_{36}\\ \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}\\ A_{70,1}& A_{70,2}& A_{70,3}& \mathrm{.Math.}& A_{70,6}\end{array}\right]& \left(4\right)\end{array}$

(39) In the case of using matrix A in the above equation 4, an input matrix I composed of the respective coefficients I.sub.1 to I.sub.6 can be calculated by the following equation 5 (here, k is a half-life constant at each measurement point).

(40) $\begin{array}{cc}{\left[\stackrel{6\times 70\left(A^T\right)}{\left[\begin{array}{ccccc}{A}_{11}& A_{12}& A_{13}& \mathrm{.Math.}& A_{1,70}\\ A_{21}& A_{22}& A_{23}& \mathrm{.Math.}& A_{2,70}\\ A_{31}& A_{32}& A_{33}& \mathrm{.Math.}& A_{3,70}\\ \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}\\ A_{61}& A_{62}& A_{43}& \mathrm{.Math.}& A_{6,70}\end{array}\right]}\stackrel{70\times 6\left(A\right)}{\left[\begin{array}{ccccc}{A}_{11}& A_{12}& A_{13}& \mathrm{.Math.}& A_{16}\\ A_{21}& A_{22}& A_{23}& \mathrm{.Math.}& A_{26}\\ A_{31}& A_{32}& A_{33}& \mathrm{.Math.}& A_{36}\\ \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}\\ A_{70,1}& A_{70,2}& A_{70,3}& \mathrm{.Math.}& A_{70,6}\end{array}\right]}\right]}^{-1}\stackrel{6\times 70\left(A^T\right)}{[\begin{array}{ccccc}{A}_{11}& A_{12}& A_{13}& \mathrm{.Math.}& A_{1,70}\\ A_{21}& A_{22}& A_{23}& \mathrm{.Math.}& A_{2,70}\\ A_{31}& A_{32}& A_{33}& \mathrm{.Math.}& A_{3,70}\\ \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}\\ A_{61}& A_{62}& A_{43}& \mathrm{.Math.}& A_{6,70}\end{array}]}\stackrel{70\times 1\left(k\right)}{\left[\begin{array}{c}{k}_{11}\\ k_{21}\\ k_{31}\\ k_{41}\\ k_{51}\\ k_{61}\\ k_{71}\\ k_{81}\\ \mathrm{.Math.}\\ k_{70,1}\end{array}\right]}=\stackrel{6\times 1\left(I\right)}{\left[\begin{array}{c}{I}_1\\ I_2\\ I_3\\ I_4\\ I_5\\ I_6\end{array}\right]}& \left(5\right)\end{array}$

(41) In the case of calculating the input matrix I using a waveform in which the extraction time (value on the x axis) is constant, by using the above-described method, the value of (A.sup.TA).sup.−1A.sup.T on the left side of equation 5 can be pre-calculated, and thus the calculation amount can be reduced during the multi-order polynomial regression type estimation. Further, the calculation amount is always constant, and thus a stable control becomes possible in comparison with the optimization algorithm method.

(42) If the input matrix I composed of the coefficients I.sub.1 to I.sub.6 of the polynomial is calculated through the multi-order polynomial regression, the input matrix I is normalized using the following equation 6.

(43) $\begin{array}{cc}\left(\mathrm{Norm}.\right)I_{i,j}=\frac{\left(\mathrm{old}\right)I_{i.j}-I_{\mathrm{minj}}}{I_{\mathrm{maxj}}-I_{\mathrm{minj}}}& \left(6\right)\end{array}$

(44) Here, (Norm.) I.sub.i,j is a normalized input matrix I in the corresponding learning feature, and (Old) I.sub.i,j is an input matrix I calculated by the equation 5 in the corresponding learning feature.

(45) Further, values of I.sub.max and I.sub.min are determined from a maximum-minimum value table for each learning feature obtained through plural times of learning. For example, if data on n×6 input matrix I are obtained through n times of learning, the values of I.sub.max and I.sub.min in each learning feature are obtained as follows from the corresponding data.

(46) $\left[\begin{array}{ccccc}{I}_{11}& I_{12}& I_{13}& \mathrm{.Math.}& I_{1,6}\\ I_{21}& I_{22}& I_{23}& \mathrm{.Math.}& I_{2,6}\\ I_{31}& I_{32}& I_{33}& \mathrm{.Math.}& I_{3,6}\\ \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}& \mathrm{.Math.}\\ I_{0,1}& I_{0,2}& I_{0,3}& \mathrm{.Math.}& I_{0,6}\end{array}\right]\left[\begin{array}{ccccc}{I}_{\min 1}& I_{\min 2}& I_{\min 3}& \mathrm{.Math.}& I_{\min 6}\\ I_{\max 1}& I_{\max 2}& I_{\max 3}& \mathrm{.Math.}& I_{\max 6}\end{array}\right]$

(47) As described above, by deriving the normalized input matrix Norm.I, the performing of the preprocessing (S20) is ended.

(48) If the performing of the preprocessing (S20) is ended, the controller 100 performs closing time prediction to derive the closing time of the injector 120 by the artificial neural network model including an input layer composed of the calculated input matrix Norm.I, a hidden layer, and an output layer.

(49) FIG. 5 illustrates an operation logic including an artificial neural network model composed of an input layer, a hidden layer, and an output layer. The artificial neural network (ANN) model is a mathematical model having a target to express several characteristics of brain function with computer simulations. The artificial neural network indicates a general model having a problem solution capability, in which an artificial neuron (node) forming a network in combination with a synapse changes the synapse combination strength through learning.

(50) In the artificial neural network, there are supervised learning that becomes optimized to a problem through an input of a supervisory signal (correct answer) and unsupervised learning that does not require the supervisory signal. If there is a clear solution, the supervised learning is used, whereas in the case of data clustering, the unsupervised learning is used. As a result, in order to reduce all dimensions, through the artificial neural network, it is possible to frequently obtain a better response with a relatively small calculation amount with respect to a problem that is unable to be linearly separated due to data with multi-dimensional amount, such as images or statistics. Accordingly, the artificial neural network is applied to various fields, such as pattern recognition or data mining. The artificial neural network may be configured using a special computer, but is implemented by an application software in most general computers.

(51) Basically, the artificial neural network model is composed of an input layer, a hidden layer, and an output layer. In the block diagram of FIG. 5, a calculation order in accordance with such three-stage layers is illustrated.

(52) First, as illustrated in FIG. 5, values input to the input layer are constructed in the form of a matrix, and the input matrix Norm.I normalized in the performing of the preprocessing is input as an input value of the input layer. Here, as described above, I.sub.1 to I.sub.6 are coefficients of a polynomial obtained through the multi-order polynomial regression of the waveform K. As illustrated in FIG. 5, in the present form, a fifth-order polynomial is used for the multi-order polynomial regression, and the input matrix Norm.I is a 6×1 matrix composed of 6 input variables.

(53) Further, as shown in FIG. 5 and the following equation 7, the hidden layer derives a first preparation matrix n.sup.1 composed of N1 to N6 by multiplying the normalized input matrix Norm.I by a first weight matrix W1 and adding a first bias matrix b1 to the multiplied matrix.

(54) $\begin{array}{cc}\stackrel{\mathrm{first}\mathrm{weight}\mathrm{matrix}W1}{\left[\begin{array}{ccccc}{W}_{11}& W_{12}& W_{13}& \sim & W_{16}\\ W_{21}& W_{22}& W_{23}& \sim & W_{26}\\ W_{31}& W_{32}& W_{33}& \sim & W_{36}\\ \wr & \wr & \wr & \wr & \wr \\ W_{61}& W_{62}& W_{63}& \sim & W_{66}\end{array}\right]}\stackrel{\mathrm{Norm}.I}{\left[\begin{array}{c}{I}_1\\ I_2\\ I_3\\ \wr \\ I_6\end{array}\right]}+\underset{\begin{array}{c}\mathrm{first}\mathrm{bias}\\ \mathrm{matrix}b1\end{array}}{\left[\begin{array}{c}{B}_1\\ B_2\\ B_3\\ \wr \\ B_6\end{array}\right]}=\stackrel{\begin{array}{c}\mathrm{first}\mathrm{preparation}\\ \mathrm{matrix}{N}^1\end{array}}{\left[\begin{array}{c}{N}_1\\ N_2\\ N_3\\ \wr \\ N_6\end{array}\right]}& \left(7\right)\end{array}$

(55) Here, the first weight matrix W1 and the first bias matrix b1 are previously prepared matrix values that are acquired through experiments in a design process before the operation logic is installed in the controller 100 of a vehicle.

(56) Further, the hidden layer derives a first resulting matrix a.sup.1 by substituting a transfer function of the following equation 8 for the first preparation matrix n.sup.1 according to the following equation 8,

(57) $\begin{array}{cc}{a}^1=\frac{2}{1+e^{-2n^1}}-1& \left(8\right)\end{array}$

(58) where, a.sup.1 is the first resulting matrix, and n.sup.1 is the first preparation matrix.

(59) Next, the output layer calculates a normalized injector closing time C.T.sub.normal by multiplying the derived first resulting matrix a.sup.1 by a second weight matrix W2 and adding a bias value b to the multiplied matrix. Here, the second weight matrix W2 and the bias value b are also values acquired through experiments in the design process before the operation logic is installed in the controller 100 of the vehicle, and are values acquired through several times of learning so that the closing time of the injector can be derived from the characteristics of a voltage curve (in the present form, the half-life constant of the voltage).

(60) 0$\begin{array}{cc}\stackrel{\underset{\mathrm{matrix}W2}{\mathrm{second}\mathrm{weight}}}{\left[\begin{array}{ccc}{w}_1& \sim & w_6\end{array}\right]}\stackrel{\begin{array}{c}\mathrm{first}\mathrm{resulting}\\ \mathrm{matrix}{a}^1\end{array}}{\left[\begin{array}{c}{L}_1\\ \wr \\ L_6\end{array}\right]}+b=C.T\mathrm{norm}& \left(9\right)\end{array}$

(61) If the normalized injector closing time C.T.sub.normal is calculated, the controller 100 calculates a final injector closing time C.T by performing de-normalization, as in the following equation 10, using the maximum value C.T.sub.max and the minimum value C.T.sub.min of the closing time of the injector obtained through the learning.
C.T=C.Tnorm×(C.Tmax−C.Tmin)+C.Tmin  (10)

(62) Through the above-described process, the controller 100 derives the closing time of the injector using the artificial neural network.

(63) Hereinafter, referring to FIGS. 6 and 7, a method for controlling an injector, being performed by the controller 100 of FIG. 1, will be described.

(64) FIG. 6 is a flowchart of a method for controlling an injector according to one form of the present disclosure using the method for sensing a closing time of an injector illustrated in FIG. 2, and FIG. 7 is a flowchart of a method for learning a closing time of an injector being executed in the method for controlling an injector illustrated in FIG. 6.

(65) As illustrated in FIG. 6, first, the controller 100 sets a required fuel injection amount in accordance with a driving condition, for example, an engine rpm and an acceleration pedal signal (S100).

(66) If the required fuel injection amount is set at S100, the controller 100 calculates a required injection command time corresponding to the required fuel injection amount (S110).

(67) For this, the controller 100 may calculate the required injection command time directly from the required fuel injection amount based on predetermined characteristic curves and fuel pressure values related to the required injection command time corresponding to the required fuel injection amount.

(68) The above-described characteristic curves are predetermined based on a nominal injector, and are stored in the form of a map in a storage device (not illustrated) provided in the controller 100.

(69) If the required injection command time is calculated at S110, the controller 100 drives the injector 120 by applying current to a solenoid valve of the injector 120 during the required injection command time (S120).

(70) If the injector 120 is driven at S120, the controller 100 senses an actual closing time of the injector 120 at that time (S130).

(71) In order to sense the actual closing time of the injector 120, the controller 100 uses the method for sensing the closing time of the injector, which is described above with reference to FIGS. 2 to 5. If the actual closing time of the injector 120 is sensed using the above-described artificial neural network (ANN), the controller 100 compensates for the required injection command time using the sensed actual closing time (S140).

(72) For example, the controller 100 can correct the required injection command time, calculated at S110, by directly calculating the required fuel injection amount corresponding to the actual closing time of the injector 120 based on the characteristic curve, related to the relationship between the closing time of the injector 120 and the required injection command time, and the fuel pressure value measured by the pressure sensor 132.

(73) If the required injection command time is corrected at S140, the controller 100 controls the injector 120 based on the corrected required injection command time (S150). That is, if the actual injection command time calculated at S140 is longer than the required injection command time calculated at S110, this means that the fuel is being injected in an amount larger than the required fuel injection amount, and thus the controller 100 controls the injection command time to be reduced.

(74) In another form of the present disclosure, the method further includes comparing the required injection command time calculated at S110 with a predetermined specific value, and the method for sensing the actual closing time may be changed based on the result of the comparison.

(75) For example, if the required injection command time is equal to or smaller than the specific value, the required injection command time is corrected at S140 based on the closing time of the injector sensed by the method for sensing the closing time of the injector using the artificial neural network (ANN) illustrated in FIGS. 2 to 5. Referring to FIG. 10, in a ballistic section in which the fuel amount m is abruptly increased even if the injector operation time T.sub.i is slightly changed and in a transient section in which the fuel amount is not greatly changed in accordance with the change in the operation time T.sub.i, it is difficult to accurately sense the closing time by the method for sensing the closing time of the injector in the related art. Accordingly, in the above-described sections, it is more desired to use the method for sensing the closing time of the injector using the artificial neural network.

(76) Meanwhile, if the injection command time exceeds the specific value, the required injection command time is corrected at S140 based on the injector closing time sensed using an inflection point of the voltage generated by the injector.

(77) Referring to FIG. 10, in a linear section in which the fuel amount m is linearly changed in accordance with the operation time T.sub.i (i.e., in a section in which a needle of the injector is fully lifted), it is relatively easy to calculate the closing time of the injector from the voltage curve generated by the injector 120. Accordingly, in the corresponding section, in consideration of the calculation capability of the ECU, it is more desired to sense the closing time using the inflection point of the voltage generated from the injector 120 rather than the artificial neural network. For this, the inflection point is derived through the second derivative of the curve of the voltage generated from the injector 120, and the time point when the corresponding inflection point exists is considered as the closing time point of the injector 120.

(78) FIG. 7 is a flowchart of a method for learning a closing time of an injector being executed in the method for controlling an injector illustrated in FIG. 6. By performing a learning method illustrated in FIG. 7, the characteristic curve being used for the method for controlling the injector illustrated in FIG. 6 can be properly corrected. Accordingly, more accurate fuel injection amount control may be performed through the above-described learning.

(79) As illustrated in FIG. 7, for the learning, the controller 100 first sets an injector injection time that is a target of learning (S200). The target of learning is the injector injection time requiring the learning in order to reduce the deviation of the closing time of the injector, and the learning of the closing time of the injector in a low flow section (i.e., above-described ballistic section or transient section) that requires a precise control has a high necessity. In one form, if the injector injection time that becomes the object of the learning exceeds the specific time (e.g., in the case of the linear section in which the needle of the injector is fully lifted), the necessity of learning is relatively low, and the learning of the injector closing time to be described later may not be performed.

(80) After the injector injection time that is the target of learning is set, a predicted closing time is calculated, which corresponds to the injector injection time set from the drive characteristic data of the injector prescribing the relationship between the injector injection time for each cylinder and the closing time, being stored in the existing nonvolatile memory (S210).

(81) If the predicted closing time corresponding to the injector injection time that is the target of learning is calculated, the fuel is primarily injected for the predicted closing time, and then a multi-stage fuel injection is performed for each cylinder to secondarily inject the remaining fuel obtained by excluding the fuel amount, consumed during the primary injection, from the entire required injection amount required for the engine driving (S600).

(82) That is, in order to perform the learning, the controller 100 switches the fuel injection mode through the injector 120 to a multi-stage injection mode, and then performs the learning injection to first inject the fuel for the set fuel injection time. Then, the controller 100 injects the fuel corresponding to the remaining injection amount obtained by excluding the injection amount, consumed for the learning injection, from the entire required injection amount to follow the learning injection. Through this, it is possible to perform the fuel injection for a specific injector injection time that is the target of learning.

(83) Next, the controller 100 senses the actual closing time of the injector during the learning injection included in the multi-stage fuel injection by the injector 120 (S230). In one form of the present disclosure, in order to sense the actual closing time of the injector, the controller 100 uses the method for sensing the closing time of the injector using the artificial neural network (ANN) described above with reference to FIGS. 2 to 5. Through this, in the present disclosure, the controller 100 obtains data related to the relationship between the actual injector injection command time for each cylinder that is the target of learning and the closing time.

(84) Next, in order to perform the learning at S240, the controller 100 derives the reliable result between the injector injection command time that is the target of learning and the closing time by repeatedly performing the measurement of the actual closing time of the injector 120 during the learning injection included in the multi-stage injection.

(85) Next, after the learning at S240 is ended, the controller 100 compensates for the injector drive characteristics based on the result of the learning (S250).

(86) For example, the controller 100 determines whether the actual measurement value of the injector closing time obtained by repeatedly performing the learning injection converges to a specific value or range. Further, if the actual measurement value of the injector closing time constantly converges to the specific value, the controller 100 determines whether the number of times of measurement of the injector closing time converging constantly as described above exceeds a predetermined first specific value. If the number of times of measurement of the injector closing time converging constantly to the specific value or range exceeds the first specific value, the controller 100 determines that the reliable relationship between the injector injection time and the closing time, being the target of learning, has been derived, and thus the learning has been completed. The controller 100 performs compensation by applying the relationship between the injector injection time and the closing time, derived through the result of the learning, to the injector drive characteristic curve for each cylinder (S250).

(87) If the actual measurement value of the injector closing time obtained by repeatedly performing the learning injection does not converge to the specific value or range, the controller 100 counts the number of times of injection in a learning mode, and determines whether the number of times of injection in the learning mode exceeds a predetermined second set value. If the number of times of injection in the learning mode exceeds the second set value, the controller 100 determines that the learning of the injector closing time that is the target of the learning has failed, and thus ends the learning of the corresponding injector closing time without performing the learning any more.

(88) According to the method for learning the closing time of the injector according to the present disclosure, it is possible to acquire the data related to the accurate injector drive characteristic related to the specific injector closing time for each cylinder that is the target, by performing the multi-stage injection including the learning injection to inject the fuel for the injector injection time that is the target of the learning in the case of the fuel injection using the injector and by measuring the actual closing time of the injector using the artificial neural network (ANN) in the case of the learning injection.

(89) Through this, the accuracy of the learning for correcting the deviation of the injector closing time can be improved, and thus the correction of the fuel amount deviation between cylinders can also be improved.

(90) While the present disclosure has been described with respect to the specific forms, it will be apparent to those skilled in the art that various changes and modifications may be made without departing from the spirit and scope of the disclosure as defined in the following claims.