Control device

Abstract

A processor of a control device calculates a state variable at a second time on the basis of a state variable (previous value) of a vehicle at a first time, a control input of the vehicle at a second time after the first time, and an observation value (sensor input) at the second time detected by a sensor of the vehicle. The processor approximates a state space model composed of a state equation related to a state variable and an observation equation related to an observation value with a neural network model. The processor determines the reliability of the state variable at the second time on the basis of an index (Mahalanobis distance) based on the control input at the second time and the observed value at the second time.

Description

TECHNICAL FIELD

[0001] The present application relates to a control device. BACKGROUND

[0002] Due to the strengthening of exhaust gas restrictions, high-precision powertrain control is required, and in order to achieve development efficiency of the powertrain control constituted by the MAP control in the past, control development using a neural network model is being promoted. The neural network model is a mathematical model that simulates the structure of the human brain neural circuit. PRIOR ART DOCUMENT PATENT DOCUMENT

[0003] Patent Document 1: Japanese Patent Application Publication No. 2021-124419 SUMMARY PROBLEMS TO BE SOLVED BY THE INVENTION

[0004] Control using a neural network model requires learning using teacher data, and the neural network model ensures the accuracy of the operation result by acting within the learned region. However, in the case where the neural network model acts outside the learned region, the accuracy can deteriorate. In the above-described Patent Document 1, it is determined whether the model can learn based on the reliability of the teacher data, but the reliability of the operation result of the learned model is not disclosed in Patent Document 1.

[0005] An object of the present application is to provide a control device capable of evaluating the reliability of a calculated value of a state variable of a vehicle that is not detected by a sensor. TECHNICAL MEANS FOR SOLVING THE PROBLEM

[0006] In order to achieve the above object, the control device of the present application has a processor that performs the following processing: calculates a state variable of a vehicle at a second time after a first time, based on the state variable of the vehicle at the first time, a control input of the vehicle at the second time, and an observation value at the second time detected by a sensor of the vehicle, approximates a state space model constituted by a state equation related to the state variable and an observation equation related to the observation value, using a neural network model, and determines the reliability of the state variable of the vehicle at the second time, based on an index based on the control input at the second time and the observation value at the second time. EFFECT OF THE INVENTION

[0007] According to the present application, it is possible to evaluate the reliability of a calculated value of a state variable of a vehicle that is not detected by a sensor. The above-mentioned problems, configurations, and effects are made clear by the following description of the embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1is a whole schematic configuration view of the internal combustion engine control system of the first and second embodiments of the present application. Figure 2 is a block diagram showing a hardware configuration example of the ECU of the first and second embodiments of the present application. Figure 3 is a view showing an example of the neural network model of the first embodiment of the present application. Figure 4 is a flowchart showing an example of the algorithm of the extended Kalman filter of the first embodiment of the present application. Figure 5 is a view showing an outline of the Mahalanobis distance of the first embodiment of the present application. Figure 6 is a view showing an example of the actual behavior of the Mahalanobis distance of the first embodiment of the present application. Figure 7 is a state quantity estimation and Mahalanobis distance calculation block diagram of the first embodiment of the present application. Figure 8 is a flowchart showing the state quantity estimation and Mahalanobis distance calculation of the first embodiment of the present application. Figure 9A is a schematic view showing a configuration example of the aftertreatment system of the second embodiment of the present application. Figure 9B is a view showing the relationship between the equivalence ratio of exhaust gas and the output of the air-fuel ratio sensor of the second embodiment of the present application. Figure 9C is a view showing the relationship between the equivalence ratio of exhaust gas and the output of the post oxygen sensor of the second embodiment of the present application. Figure 10A is a view showing the tendency of H2O (water), CO (carbon monoxide), CO2 (carbon dioxide), H2 (hydrogen), and O2 (oxygen) with respect to the equivalence ratio of the second embodiment of the present application. Figure 10B is a view showing the tendency of HC (hydrocarbon) and NOx (nitrogen oxide) with respect to the equivalence ratio of the second embodiment of the present application. x Figure 11 is a view showing the main reaction process of the three-way catalyst (cerium-based) used in the aftertreatment system of the second embodiment of the present application. Figure 12 is a view showing the tendency of the purification rate of the three-way catalyst with respect to the equivalence ratio of exhaust gas above the catalyst activation temperature of the second embodiment of the present application. Figure 13 is a view showing the output operation of the post oxygen sensor 22 with respect to the equivalence ratio upstream of the catalyst and the equivalence ratio downstream of the catalyst of the second embodiment of the present application. Figure 14 ​is a graph illustrating hysteresis of output characteristics of the post oxygen sensor of the second embodiment of the present application. Figure 15 is a graph illustrating time changes in the NOx concentration and HC concentration downstream of the catalyst of the second embodiment of the present application. x Figure 16 is a graph illustrating a relationship between the catalyst deterioration degree and the oxygen storage capacity of the three-way catalyst of the second embodiment of the present application. Figure 17 is a graph illustrating a relationship between the oxygen storage ratio of the three-way catalyst and the NOx purification rate of the three-way catalyst of the second embodiment of the present application. x Figure 18 is a graph illustrating a comparison result of the equivalence ratio upstream of the catalyst, the equivalence ratio downstream of the catalyst, the oxygen storage ratio, and the output behavior of the post oxygen sensor when a new catalyst and a deteriorated catalyst of the second embodiment of the present application are used. Figure 19 is a block diagram of the catalyst temperature estimation and the Mahalanobis distance calculation of the second embodiment of the present application. Figure 20 is a flowchart showing the catalyst temperature estimation and the Mahalanobis distance calculation of the second embodiment of the present application. Figure 21 is a graph showing an example of the neural network model of the second embodiment of the present application. DETAILED DESCRIPTION

[0009] Hereinafter, a mode for carrying out the present application will be described with reference to the accompanying drawings. In the present specification and drawings, the same symbols are assigned to constituent elements having substantially the same functions or configurations, and overlapping descriptions will be omitted.

[0010] [First Embodiment] Figure 1 is a block diagram showing the entire outline configuration of the internal combustion engine control system 100.

[0011] The internal combustion engine control system 100 includes an internal combustion engine 1 and an ECU (Electronic Control Unit) 28 attached to the internal combustion engine 1. The ECU 28 is electrically connected to various sensors, actuators, and the like that constitute the internal combustion engine 1, and is an example of an internal combustion engine control device that controls the internal combustion engine 1.

[0012] ​​The internal combustion engine 1 includes a flow rate sensor 2 (an air flow rate sensor), a turbocharger 3, an air bypass valve 4, an intercooler 5, a supercharging temperature sensor 6, a throttle valve 7, an intake manifold 8, a supercharging pressure sensor 9, a flow enhancement valve 10, an intake valve 11, a valve opening and closing phase sensor 12, an exhaust valve 13, a valve opening and closing phase sensor 14, a fuel injection valve 15, a spark plug 16, a knock sensor 17, a crank angle sensor 18, an exhaust bypass valve 19, an upstream air-fuel ratio sensor 20 of a catalyst, an exhaust purification catalyst (three-way catalyst) 21, a rear oxygen sensor (downstream oxygen sensor of a catalyst) 22, an EGR (Exhaust Gas Recirculation) pipe 23, an EGR cooler 24, an EGR valve 25, an exhaust temperature sensor 26, a differential pressure sensor 27, and a downstream NO x sensor 29. In addition, the rear oxygen sensor 22 can be replaced with a downstream (rear) air-fuel ratio sensor of a catalyst (not shown).

[0013] An intake temperature sensor (not shown) is attached to the flow rate sensor 2 provided in the intake flow path of the internal combustion engine 1. The intake temperature sensor measures the intake air temperature.

[0014] The turbocharger 3 is composed of a compressor 3a provided with compressor blades facing the intake flow path, and a turbine 3b provided with turbine blades integrally linked to the compressor 3a so as to rotate and facing the exhaust flow path. The compressor 3a and the turbine 3b are rotatably supported inside the turbocharger 3. The turbine 3b converts the energy possessed by the exhaust gas from the internal combustion engine 1 into rotational energy. The compressor 3a connected to the turbine 3b compresses the intake air flowing in from the intake flow path using the rotational energy of the turbine 3b.

[0015] In the intake flow path of the internal combustion engine 1, the intercooler 5 is provided downstream of the compressor 3a of the turbocharger 3, and cools the intake air that has been adiabatically compressed by the compressor 3a and has risen in temperature. The supercharging temperature sensor 6 is installed downstream of the intercooler 5, and measures the temperature (supercharging temperature) of the intake air that has been cooled by the intercooler 5.

[0016] The throttle valve 7 is provided downstream of the intercooler 5, throttles the intake flow path, and controls the amount of intake air flowing into the cylinder of the internal combustion engine 1. The throttle valve 7 is composed of an electronically controlled butterfly valve whose valve opening degree can be controlled by the ECU 28. The intake manifold 8 in which the supercharging pressure sensor 9 is installed communicates with the downstream of the throttle valve 7.

[0017] In addition, it is also possible to integrate the intake manifold 8 provided downstream of the throttle valve 7 and the intercooler 5. In this case, since it is possible to reduce the volume from the downstream of the compressor 3a to the cylinder, it is possible to improve the responsiveness and controllability of acceleration and deceleration.

[0018] To prevent excessive rise in pressure from the downstream of the compressor 3a to the upstream portion of the throttle valve 7, the air bypass valve 4 is provided in the intake passage on an intake bypass passage connecting the upstream and downstream of the compressor 3a. For example, in the case where the throttle valve 7 is abruptly closed in the supercharged state, the air bypass valve 4 is opened according to the control of the ECU 28, whereby the compressed intake air of the downstream portion of the compressor 3a is countercurrent to the upstream portion of the compressor 3a through the bypass passage. As a result, by immediately lowering the supercharging pressure, it is possible to prevent a phenomenon called surging, and to prevent damage to the compressor 3a, to be precise.

[0019] The flow enhancement valve 10 is provided downstream of the intake manifold 8, and enhances the flow turbulence in the cylinder by creating a bias flow in the intake air to the cylinder. In the case of the exhaust gas recirculation combustion described later, by closing the flow enhancement valve 10 by the ECU 28, it is possible to promote the turbulent combustion, and to stabilize the combustion generated in the combustion chamber.

[0020] The intake valve 11 and the exhaust valve 13 each have a variable valve mechanism for continuously varying the phase of valve opening and closing. On the variable valve mechanisms of the intake valve 11 and the exhaust valve 13, a valve opening and closing phase sensor 12, 14 for detecting the opening and closing phase of the valve is assembled, respectively. In the cylinder of the internal combustion engine 1, a fuel injection valve 15 of a direct injection type is provided, which injects fuel pressurized by a fuel pump not shown into the combustion chamber of the cylinder. In addition, the fuel injection valve 15 can be of an intake port injection type which injects fuel into the intake passage. Alternatively, the fuel injection valve 15 can be provided with a plurality of fuel injection valves 15 of the direct injection type or the intake port injection type for each cylinder.

[0021] On the cylinder of the internal combustion engine 1, a spark plug 16 is assembled, which exposes an electrode portion in the cylinder, and ignites the combustible mixture with a spark. A knock sensor 17 is provided in the cylinder block, and detects the presence or absence of knock by detecting the vibration of the cylinder block due to the combustion pressure vibration generated in the combustion chamber. A crank angle sensor 18 is assembled in the vicinity of the crankshaft, and outputs a signal corresponding to the rotation angle of the crankshaft to the ECU 28.

[0022] In the exhaust pipe of the internal combustion engine 1, the upstream air-fuel ratio sensor (catalyst upstream air-fuel ratio sensor 20) is disposed downstream of the turbine 3b of the turbocharger 3, that is, upstream of the three-way catalyst (exhaust purification catalyst 21) having an oxygen storage capacity, which is provided in the exhaust pipe. Then, the upstream air-fuel ratio sensor (catalyst upstream air-fuel ratio sensor 20) detects the exhaust gas composition from the exhaust gas, that is, the catalyst upstream air-fuel ratio of the exhaust gas, and outputs a signal indicating the catalyst upstream air-fuel ratio to the ECU 28. The exhaust purification catalyst 21 is an example of a three-way catalyst, which is disposed downstream of the catalyst upstream air-fuel ratio sensor 20. The exhaust purification catalyst 21 purifies harmful exhaust gas components such as carbon monoxide, nitrogen compounds, and unburned hydrocarbons in the exhaust gas by a catalytic reaction. As described above, the downstream oxygen sensor 22 is disposed downstream of the exhaust purification catalyst 21. The downstream oxygen sensor 22 detects the amount of oxygen contained in the exhaust gas after the exhaust purification by the exhaust purification catalyst 21. Hereinafter, the exhaust purification catalyst 21 will also be referred to as a three-way catalyst or a catalyst.

[0023] The turbocharger 3 has an air bypass valve 4 and an exhaust bypass valve 19. In order to prevent excessive rise in pressure from the downstream of the compressor 3a to the upstream portion of the throttle valve 7, the air bypass valve 4 is disposed on a bypass flow path connecting the upstream and the downstream of the compressor 3a. In the case where the throttle valve 7 is abruptly closed in the boosted state, the air bypass valve 4 is opened in accordance with the control of the ECU 28, whereby the compressed intake air of the downstream portion of the compressor 3a is reversed to the upstream portion of the compressor 3a through the bypass flow path. As a result, by immediately reducing the boost pressure, it is possible to prevent a phenomenon called surging, and to surely prevent damage to the compressor 3a.

[0024] The exhaust bypass valve 19 is disposed on an exhaust bypass flow path connecting the upstream and the downstream of the turbine 3b in the exhaust flow path. The exhaust bypass valve 19 is an electrically driven valve capable of freely controlling the valve opening degree with respect to the boost pressure by the control of the ECU 28. When the opening degree of the exhaust bypass valve 19 is adjusted by the ECU 28 on the basis of the boost pressure detected by the boost pressure sensor 9, a part of the exhaust gas passes through the bypass flow path, whereby it is possible to reduce the energy applied to the turbine 3b by the exhaust gas. As a result, the exhaust bypass valve 19 can adjust the boost pressure to the target pressure.

[0025] The EGR pipe 23 communicates the exhaust gas flow path downstream of the exhaust purification catalyst 21 with the intake gas flow path upstream of the compressor 3a, and recirculates the exhaust gas branched from the downstream of the exhaust purification catalyst 21 to the upstream of the compressor 3a. The EGR cooler 24 provided in the EGR pipe 23 cools the branched exhaust gas. The EGR valve 25 is provided in the EGR pipe 23 between the EGR cooler 24 and the upstream of the compressor 3a, and controls the flow rate of the exhaust gas recirculated to the upstream of the compressor 3a. In addition, the exhaust gas temperature sensor 26 that detects the exhaust gas temperature upstream of the EGR valve 25, and the differential pressure sensor 27 that detects the differential pressure upstream and downstream of the EGR valve 25 are provided in the EGR pipe 23.

[0026] The ECU 28 has a function of controlling the operation of the engine 1, and is described later. Figure 2 The ECU 28 has a function of controlling the operation of the engine 1, and is described later. The ECU 28 has a function of controlling the operation of the engine 1, and is described later.

[0027] [Example of hardware configuration of ECU] Figure 2 is a block diagram showing an example of the hardware configuration of the ECU 28.

[0028] The ECU 28 includes an input circuit 201, an input / output port 202, a RAM 203, a ROM 204, a CPU 205, a throttle valve drive circuit 206, a fuel injection valve drive circuit 207, and an ignition output circuit 208.

[0029] The output signals from various sensors are input to the input circuit 201. In the example shown in Figure 2 In the example shown in FIG. 2, the output signals of the throttle valve sensor of the throttle valve 7, the flow rate sensor 2, the supercharging temperature sensor 6, the supercharging pressure sensor 9, the valve opening / closing phase sensors 12, 14, the knock sensor 17, the crank angle sensor 18, the catalyst upstream air-fuel ratio sensor 20, and the rear oxygen sensor 22, and the like are examples of the output signals input to the input circuit 201 of the ECU 28. The signals input to the input circuit 201 are sent to the input / output port 202.

[0030] The signal transmitted to the input / output port 202 is stored in the RAM 203, and is subjected to arithmetic processing by the CPU 205. A control program describing the content of the arithmetic processing is written in advance in the ROM 204, and is executed by the CPU 205. The ROM 204 records programs, data, and the like necessary for the operation of the CPU 205, and functions as an example of a computer-readable non-transitory storage medium that stores the programs executed by the ECU 28.

[0031] The control signal calculated by the CPU 205 according to the control program is output to each device such as the throttle valve drive circuit 206, the fuel injection valve drive circuit 207, the ignition output circuit 208.

[0032] The throttle valve drive circuit 206 outputs a drive signal for controlling the opening and closing drive of the throttle valve 7 to the throttle valve 7.

[0033] The fuel injection valve drive circuit 207 outputs a drive signal for controlling the opening and closing drive of the fuel injection valve 15 at the fuel injection timing to the fuel injection valve 15.

[0034] The ignition output circuit 208 outputs a drive signal for controlling the ignition of the spark plug 16 at the ignition timing to the ignition output circuit 208.

[0035] <Neural network model> Figure 3 An outline of the neural network is shown. The neural network model is a mathematical model that simulates the structure of the human brain neural circuit, and the weight and bias are set for each neuron constituting the model.

[0036] In addition, a function called an activation function is defined in the neuron. A logic function, a slope function, and the like are appropriately set in the activation function. One layer is formed by a plurality of neurons, and an intermediate layer is provided between the input layer and the output layer. By increasing the number of neurons and the number of layers of the intermediate layer, a more complex input-output relationship can be approximated. There is a trade-off relationship between the approximation accuracy and the model size, and a balance point that satisfies both requirements is selected. For example, by setting the frequency of the exhaust sensor signal on the input layer and setting the catalyst deterioration diagnosis on the output layer, and performing machine learning (supervised) on the weight and bias of each neuron, the input-output relationship can be approximated. Error backpropagation can be applied to the machine learning algorithm.

[0037] <Extended Kalman filter> Figure 4is a flowchart showing an example of an extended Kalman filter algorithm. For the extended Kalman filter, which is one of the constituent elements of the present application, the algorithm and the application method of the present control are described. The algorithm of the extended Kalman filter and the application method of the present control are described. In the extended Kalman filter, consideration is made on the premise of state, output equations including system noise Q and observation noise R defined by equations (1), (2).

[0038] [Equation 1]

[0039] The above nonlinear equation is piecewise linearly approximated as follows.

[0040] [Equation 2]

[0041] Here, F k (k) and H k (k) are Jacobian matrices with respect to state variables of f and h, respectively.

[0042] [Equation 3]

[0043] The extended Kalman filter algorithm is composed of a prediction step and a filtering step. In the prediction step, the internal state variable vector and the covariance matrix are updated with the following equation from the input variable and the system noise.

[0044] [Equation 4]

[0045] Next, in the filtering step, the Kalman gain specified by the updated covariance matrix and the observation noise is calculated.

[0046] [Equation 5]

[0047] The internal state variable vector and the covariance matrix are updated with the following equation using the Kalman gain and the observation data.

[0048] [Equation 6]

[0049] The time step is updated and returned to equation (6). By repeating the above steps, it is possible to estimate the internal state variable that is difficult to directly measure based on the measured value.

[0050] <Mahalanobis distance> Figure 5A summary chart showing the Mahalanobis distance. The Mahalanobis distance has the following characteristics: even if the distance from the center point is the same, it is easy to determine normal values or abnormal values by combining the variances in multiple dimensions. The Mahalanobis distance is calculated by the following equation. x represents the input data vector as shown in Equation (7). In addition, μ represents the mean vector as shown in Equation (8).

[0051] [Equation 7] x = (x1, x2,... x n ) T ... (7)

[0052] [Equation 8] μ = (μ1, μ2,... μ n ) T ... (8)

[0053] Equation (9) shows the variance-covariance matrix. E is the expected value. The Mahalanobis distance D M .

[0054] [Equation 9]

[0055] [Equation 10]

[0056] Figure 6 An example of the actual behavior of the Mahalanobis distance. In intervals (a) and (c), the control input u exists in the interpolation region, and in interval (b), it exists in the extrapolation region. In addition, in interval (c), the sensor used in the extended Kalman filter is abnormal, which is an interval in which an abnormal sensor value is input. Next, the behavior of the Mahalanobis distance for each interval will be described. In interval (a), the control input u is in the interpolation region, and the sensor is operating normally, so the Mahalanobis distance is small, and the estimation accuracy is good. In interval (b), since the control input u is in the extrapolation region, the Mahalanobis distance increases, and the estimation accuracy deteriorates compared to interval (a). In interval (c), the control input u is in the interpolation region, but since the sensor used in the extended Kalman filter outputs an abnormal value, the Mahalanobis distance increases, and the estimation accuracy deteriorates compared to interval (a). In this way, when the control input or the sensor is operating abnormally, the reliability of the estimation value of the state observer can be determined by the increase in the Mahalanobis distance.

[0057] Next, the following equation is used Figure 7The configuration of the present embodiment will be described. First, arbitrary input values input from the ECU are input to the extended Kalman filter 701. The input values include control values and sensor values. The input values are used to estimate the state quantity. Here, since the state quantity is a differential value, using a prescribed time scale value and a last value state quantity, an integral calculation is performed to find the current state quantity. Also, the same kind of input values as the extended Kalman filter 701 are input to the Mahalanobis distance calculation section 702. The input values are used to calculate the Mahalanobis distance of the current input. In the comparison section 703, a prescribed threshold value and the calculated Mahalanobis distance are compared. In the case where the current Mahalanobis distance is greater than the threshold value, 0 is output, and this is considered to be outside the learning range or a sensor failure. In the case where the current Mahalanobis distance is less than the threshold value, 1 is output, and this is judged to be within the learning range or a sensor normal. Then, the interpolation result is recorded. In the switching section 704, in the case where 0 is input, it is considered that the extended Kalman filter 701 is unable to perform a normal estimation operation in the learning range or in a sensor failure, and a prescribed steady value is output. In the case where 1 is input, it is considered that the extended Kalman filter 701 is able to perform a normal estimation operation in the learning range or in a sensor normal, and the estimated state quantity is output. Next, using Figure 8 The flowchart of the present embodiment will be described. In step 801, control input and sensor input values from the ECU are received. In step 802, the Mahalanobis distance is found using the control input and sensor input values from the ECU. In step 803, a comparison is made between the threshold value prescribed in step 802 and the calculated Mahalanobis distance. In the case where it is less than the threshold value, step 804 is entered, and in the case where it is greater than the threshold value, step 805 is entered. In the case where step 804 is advanced to, since the extended Kalman filter is operating in the learning range or in a sensor normal, the control input and sensor input values from the ECU are used to estimate the state quantity (differential value) by the extended Kalman filter, and an integral calculation is performed using a prescribed time scale value and a last value state quantity to find the current state quantity.

[0058] In the case where step 805 is entered, since the extended Kalman filter is operating outside the learning range or in a sensor failure, the diagnosis result is recorded, and is used as sensor failure or relearning data. In step 806, a prescribed value such as an estimated value calculated by a physical formula, an output value of a MAP prescribed by an operating point, or the like is output in place of the value.

[0059] [Second Embodiment] In the second embodiment, the estimation of the temperature of a catalyst will be described as an example.

[0060] [Configuration Example of Post-processing System] Next, referring to Figures 9A-9CAn example of a configuration of an exhaust purification system 110 that purifies exhaust gas of an internal combustion engine will be described.

[0061] Figure 9A is a schematic diagram showing an example of a configuration of the exhaust purification system 110.

[0062] As described above, in the exhaust purification system 110, a three-way catalyst is used as an exhaust purification catalyst (exhaust purification catalyst 21). A catalyst-upstream air-fuel ratio sensor 20 is provided at an upstream portion of the three-way catalyst, and an after-oxygen sensor 22 is provided at a downstream portion. The catalyst-upstream air-fuel ratio sensor 20 and the after-oxygen sensor 22 are connected to a control device such as an ECU 28. The ECU 28 is able to measure the catalyst-upstream air-fuel ratio of exhaust gas flowing into the three-way catalyst via the catalyst-upstream air-fuel ratio sensor 20, and is able to detect the amount of oxygen contained in exhaust gas after purification of the catalyst via the after-oxygen sensor 22.

[0063] Figure 9B is a graph showing the relationship between the equivalence ratio (= stoichiometric air-fuel ratio / air-fuel ratio) of exhaust gas and the output of the catalyst-upstream air-fuel ratio sensor 20. Figure 9B The horizontal axis of is the equivalence ratio, and the vertical axis is the output of the catalyst-upstream air-fuel ratio sensor 20.

[0064] As shown in Figure 9B , it is shown that the more the equivalence ratio increases (in other words, the richer the exhaust gas is), the more the output of the catalyst-upstream air-fuel ratio sensor 20 (air-fuel ratio sensor output) decreases. The ECU 28 converts the catalyst-upstream air-fuel ratio sensor signal into the catalyst-upstream air-fuel ratio in accordance with the relationship between the equivalence ratio of the exhaust gas and the output of the catalyst-upstream air-fuel ratio sensor 20. Thereby, the ECU 28 is able to detect the catalyst-upstream air-fuel ratio with high accuracy in a wide range from the lean state to the rich state of the exhaust gas.

[0065] Figure 9C is a graph showing the relationship between the equivalence ratio of exhaust gas and the output of the after-oxygen sensor 22. Figure 9C The horizontal axis of is the equivalence ratio, and the vertical axis is the after-oxygen sensor output.

[0066] The after-oxygen sensor output is also read as the after-oxygen sensor voltage, which changes in accordance with the electromotive force accompanying the concentration difference between the oxygen concentration in the exhaust gas and the oxygen concentration in the air. The after-oxygen sensor output substantially indicates the minimum electromotive force in the lean condition, and indicates the maximum electromotive force in the rich condition. Therefore, in catalyst control, the after-oxygen sensor output has the characteristic of sharply changing at the stoichiometric air-fuel ratio (equivalence ratio 1.0). The ECU 28 is able to detect the amount of oxygen in exhaust gas discharged downstream of the three-way catalyst by capturing the timing of the change in the after-oxygen sensor output.

[0067] <Chemical species concentration of exhaust gas> Figure 10A and Figure 10B is a graph showing the tendency of the equivalence ratio with respect to the concentration of the chemical species of the exhaust gas.

[0068] Figure 10A is a graph showing the tendency of the equivalence ratio with respect to H2O (water), CO (carbon monoxide), CO2 (carbon dioxide), H2 (hydrogen), and O2 (oxygen).

[0069] Figure 10B is a graph showing the tendency of the equivalence ratio with respect to HC (hydrocarbon), NO x (xide), and O2 (oxygen). Figure 10A and Figure 10B The horizontal axis is the equivalence ratio, and the vertical axis is the concentration of the chemical species of the exhaust gas.

[0070] As shown in Figure 10A , the composition of the combustion gas of the hydrocarbon fuel is divided by the stoichiometric air-fuel ratio, and CO (carbon monoxide) and H2 (hydrogen) increase on the rich side, and O2 (oxygen) increases on the lean side. On the other hand, as shown in Figure 10B , NO x (xide) shows a maximum value on the slightly lean side of the stoichiometric air-fuel ratio, and shows a tendency to decrease on the lean side and the rich side thereof. HC (unburned hydrocarbon) is an unburned fuel component that is emitted without being burned, and shows a minimum value in the stoichiometric air-fuel ratio, and there is a tendency for the amount of HC that is emitted without being completely burned to increase when excessively lean or rich is divided by the stoichiometric air-fuel ratio.

[0071] As shown in Figure 10A and Figure 10B , even under the stoichiometric air-fuel ratio condition in which the fuel and air (oxygen) are not supplied in excess or deficiency, a certain amount of CO (carbon monoxide) and NO x (xide) is emitted in the combustion gas at a high temperature, which cannot be converted into H2O (water) and CO2 (carbon dioxide). Therefore, it is necessary to appropriately purify the exhaust gas by the aftertreatment system 110.

[0072] <Reaction process of three-way catalyst> Figure 11 is a graph showing the main reaction process of the three-way catalyst (cerium system) used in the aftertreatment system 110. In the reaction process shown in Figure 11 , the notation of the coefficient is omitted.

[0073] The three-way catalytic reaction process is mainly composed of an oxidation reaction, a NO x reduction reaction, and an oxygen storage and release reaction. In the oxidation reaction, CO, H2, and HC generated under rich conditions or high temperature conditions react with oxygen to generate harmless CO2 and H2O. Unburned hydrocarbon (HC) contains components such as methane, propane, ethylene, butane, and the like, and reacts at different rates, respectively.

[0074] NO x The reduction reaction is mainly represented by the reaction of CO and NO to produce CO2 and N2.

[0075] In the oxygen storage and release reactions, oxygen (O2) storage and release, as well as the respective oxidation and reduction reactions of HC, CO, and NO, occur via Ce (cerium), the catalyst material. Specifically, cerium dioxide (CeO2) reacts with CO and HC to produce CO2 and H2O, while cerium trioxide (Ce2O3) reacts with NO to produce N2. The balance between the simultaneously generated CeO2 and Ce2O3 determines the oxygen storage ratio φ of the three-way catalyst. Specifically, when all Ce2O3 in the catalyst is CeO2, it cannot react with NO, and NO purification is not possible.

[0076] Thus, to properly maintain the purification rate of the three-way catalyst, the balance between CeO2 and Ce2O3, that is, the oxygen storage ratio φ, must be maintained at a predetermined value. Since all of the aforementioned reaction processes are strongly dependent on catalyst temperature, the ECU 28 must appropriately manage the catalyst temperature so that it reaches or exceeds the activation temperature early after the internal combustion engine 1 is started.

[0077] In addition, in the system shown in this embodiment, a cerium-based three-way catalyst is used, but the present invention is not limited thereto. Even if a catalyst of another material showing a similar effect is used, the same effect can be achieved without changing the structure of the invention by adjusting the constants of the control model. Figure 11 In addition to the reaction mechanisms shown, water-gas shift reaction and the like may be used. The ECU 28 can also adjust the control model constants to cope with these reaction mechanisms.

[0078] Figure 12 This is a diagram illustrating the tendency of the purification rate of the three-way catalyst with respect to the exhaust gas equivalence ratio at or above the catalyst activation temperature. Figure 12 The horizontal axis is the equivalent ratio, and the vertical axis is the catalyst purification rate. Figure 12 Indicates NO when the equivalence ratio changes from lean to rich x , HC, and CO. In addition, the closer the catalyst purification rate is to 100%, the more the components in the exhaust gas are purified, and the components in the exhaust gas are not discharged.

[0079] This graph shows that the purification rate characteristics of the three-way catalyst vary with the theoretical air-fuel ratio (the "control target" in the graph). Under lean conditions, the purification rates of CO and HC are generally maintained above 90%. On the other hand, as the equivalence ratio decreases, the purification rate of NO xThe purification rates decrease. On the rich side, the purification rates of HC and CO show a tendency to decrease as the equivalence ratio increases. Near the stoichiometric air-fuel ratio, the purification rates of NO x , HC, and CO can all reach 90% or more. Therefore, the point near the stoichiometric air-fuel ratio is called the "three-way point". The ECU 28 implements control to maintain the purification rates of the three-way catalyst at a high level by maintaining the equivalence ratio at the stoichiometric air-fuel ratio, which is the three-way point.

[0080] Figure 13 is a graph that illustrates the behavior of the equivalence ratio upstream of the catalyst, the equivalence ratio downstream of the catalyst, and the output of the rear oxygen sensor 22. Figure 13 The horizontal axis of is time, and the vertical axis is the equivalence ratio upstream of the catalyst, the equivalence ratio downstream of the catalyst, and the output of the rear oxygen sensor. Here, centered on an equivalence ratio of 1.0, the behavior of the equivalence ratio downstream of the catalyst (the air-fuel ratio downstream of the catalyst) when the equivalence ratio upstream of the catalyst (the air-fuel ratio upstream of the catalyst) is sharply changed in steps in time on the lean side or the rich side, and the output of the rear oxygen sensor 22 provided downstream of the catalyst are shown. First, as shown in period (a) of the time axis in the left-right direction in Figure 13 , when the equivalence ratio upstream of the catalyst is set exactly to 1.0, a trace amount of oxygen is discharged downstream of the catalyst. Therefore, the rear oxygen sensor output is maintained in an intermediate state, which is a value between its maximum electromotive force and minimum electromotive force. Next, as shown in period (b) in Figure 13 , when the equivalence ratio upstream of the catalyst is decreased from an equivalence ratio of 1.0 and changed in steps to the lean side, the equivalence ratio downstream of the catalyst gradually decreases to the lean side, and the rear oxygen sensor output sharply changes to the side of the prescribed minimum electromotive force value in the lean condition after a relaxation time required for the gradual decrease of the equivalence ratio downstream of the catalyst. Next, as shown in period (c) in Figure 13 , if the equivalence ratio upstream of the catalyst is increased to exceed an equivalence ratio of 1.0 and changed in steps to the rich side, in correspondence thereto, the equivalence ratio downstream of the catalyst starts to gradually increase to the rich side. The rear oxygen sensor output changes to the side of the prescribed maximum electromotive force value in the rich condition after a relaxation time required for the gradual increase of the equivalence ratio downstream of the catalyst. Next, as shown in period (d) in Figure 13 , if the equivalence ratio upstream of the catalyst is again decreased from an equivalence ratio of 1.0 and changed in steps to the lean side, the equivalence ratio downstream of the catalyst again gradually decreases to the lean side. The rear oxygen sensor output, in response to the change in the equivalence ratio upstream of the catalyst, again changes to the side of the prescribed minimum electromotive force value in the lean condition after a relaxation time required for the gradual decrease of the equivalence ratio downstream of the catalyst. Characteristically, the time required for the change of the equivalence ratio downstream of the catalyst from the rich to the lean in this period (d) is shorter than the time required for the change from the lean to the rich in period (c), and has a so-called hysteresis.

[0081] Thus, there is a tendency that the delay time of the catalyst downstream equivalence ratio and the delay time of the output of the rear oxygen sensor 22 differ between the change from lean to rich and the change from rich to lean. This tendency is due to Figure 11 This is due to the difference in reaction rates between the oxygen storage and release reactions of CeO₂ and Ce₂O₃ in the three-way catalyst (cerium-based) described above, which changes from rich to lean and from lean to rich. Furthermore, since the reaction rate of the three-way catalyst also depends on catalyst temperature and exhaust flow rate, the aforementioned hysteresis changes with catalyst temperature and exhaust flow rate.

[0082] Next, refer to Figure 14 , which illustrates the hysteresis of the oxygen sensor characteristics. Figure 14 This is a graph illustrating the hysteresis of the output characteristics of the oxygen sensor. The horizontal axis of the graph is the equivalence ratio, and the vertical axis is the output of the oxygen sensor. Figure 9C As explained above. Meanwhile, the oxygen sensor also uses the catalyst material described above. Therefore, the dynamic characteristics of the oxygen sensor exhibit different hysteresis characteristics for changes from rich to lean and from lean to rich. Specifically, the oxygen sensor signal requires a shorter time to recover from a predetermined maximum electromotive force value under rich conditions to a predetermined minimum electromotive force value under lean conditions than from a predetermined minimum electromotive force value under lean conditions to a predetermined maximum electromotive force value under rich conditions. Furthermore, the behavior of the oxygen sensor signal is affected by changes in the properties of the materials that constitute the oxygen sensor during use and by the sensor's temperature.

[0083] Here, refer to Figure 15 The output of the rear oxygen sensor 22 and the NOx level downstream of the catalyst are described when the internal combustion engine 1 is controlled at an equivalence ratio of 1.0 (theoretical air-fuel ratio) upstream of the catalyst (stoichiometric control) and then resumes operation at the theoretical air-fuel ratio after a period of fuel cut (lean condition). x Time variation of concentration. Figure 15 The output of the rear oxygen sensor 22 and the NO x The graph shows the time variation of the concentration and HC concentration. Figure 15 The horizontal axis is time, and the vertical axis is the equivalence ratio upstream of the catalyst, the output of the rear oxygen sensor, and the NO x Concentration, HC concentration downstream of the catalyst. First, if Figure 15 As shown in the time period (a) along the horizontal time axis in FIG, when the catalyst upstream equivalence ratio is just set to 1.0, the rear oxygen sensor output is maintained at a value intermediate between its maximum electromotive force and minimum electromotive force. Figure 15 As shown in the middle period (b), the operation is shifted from the operation based on the catalyst upstream equivalence ratio of 1.0 to the lean condition based on the fuel cut. Figure 15As shown in the middle time period (c), the operation is restored to the catalyst upstream equivalence ratio of 1.0. At this time, the change of the rear oxygen sensor output from the intermediate state of the catalyst upstream equivalence ratio of 1.0 to the lean side is different from the change from the lean side to the intermediate state of the catalyst upstream equivalence ratio of 1.0 (relaxation time). That is, the oxygen sensor signal changes from the prescribed minimum electromotive force value under the lean condition based on fuel cut to the intermediate state under the stoichiometric control condition (refer to Figure 15 The relaxation time required when the time period (a) changes is longer than the relaxation time required when the output of the oxygen sensor changes from the intermediate state under the stoichiometric control condition to the prescribed minimum electromotive force value under the lean condition. During the relaxation time (delay period) until the output of the post oxygen sensor returns to the intermediate state, NO is observed downstream of the catalyst. x For such a catalyst downstream NO x When the operation based on the catalyst upstream equivalence ratio of 1.0 is resumed after a fuel cut, a so-called rich correction is performed to cope with the increase in concentration: the catalyst upstream equivalence ratio is increased to exceed 1.0, temporarily increasing NO x The catalyst purification rate (refer to Figure 12 ), thereby making NO x However, the rich correction increases NO x There is a trade-off between the catalyst purification rate and the catalyst purification rate of HC and CO (refer to Figure 12 ). Therefore, in the air-fuel ratio control of the internal combustion engine 1, it is necessary to consider the state inside the catalyst and implement the rich correction control of the appropriate correction amount and period. In addition, in order to consider the state inside the catalyst in the air-fuel ratio control of the internal combustion engine 1, for example, the signal of the post oxygen sensor that detects the oxygen state of the exhaust gas downstream of the catalyst can be used as a judgment reference. However, it should be noted that, as mentioned above, the response of the post oxygen sensor 22 has a delay period (see Figure 14 ).

[0084] Here, the output of the rear oxygen sensor 22 and the NOx downstream of the catalyst are adjusted in the order of stoichiometric control, appropriate rich correction, and excessive rich correction. x The time changes of the concentration and HC concentration are described.

[0085] (Stoichiometric Control) exist Figure 15 In FIG. 1 , the output of the rear oxygen sensor 22 and the NO x Concentration and time variation of HC concentration. Figure 13 As shown, the output of the rear oxygen sensor 22 increases as the condition changes from lean to rich. xConcentration of NO which instantaneously increases during the delay period until the output of the post oxygen sensor 22 recovers x The HC concentration downstream of the catalyst hardly changes, so the emission of HC is prevented.

[0086] (Appropriate Concentration Correction) In Figure 15 , the output of the post oxygen sensor 22 and the time changes of the NOx concentration and the HC concentration downstream of the catalyst when the appropriate concentration correction is performed are indicated by a dotted line. The post oxygen sensor 22 detects the oxygen state of the gas downstream of the catalyst. Therefore, if the appropriate concentration correction is performed, the internal state of the catalyst has already changed to the state in which the oxygen storage capacity is maximum or minimum at the time when the output of the post oxygen sensor 22 reacts. In addition, if the appropriate concentration correction is performed, the NOx concentration downstream of the catalyst becomes low, as shown in x , and the catalyst purification rate of NOx increases, as shown in x , so the emission of NOx is prevented. x

[0087] (Excessive Concentration Correction) In Figure 15 , the output of the post oxygen sensor 22 and the time changes of the NOx concentration and the HC concentration downstream of the catalyst when the excessive concentration correction is performed are indicated by a thick dotted line. The post oxygen sensor 22 detects the oxygen state of the gas downstream of the catalyst. Therefore, if the excessive concentration correction is performed, the internal state of the catalyst has already changed to the state in which the oxygen storage capacity is maximum or minimum at the time when the output of the post oxygen sensor 22 reacts. In addition, if the excessive concentration correction is performed, the NOx concentration downstream of the catalyst becomes low, as shown in Figure 10B , and the catalyst purification rate of NOx increases, as shown in x , so the emission of NOx is prevented. x Figure 12 On the other hand, the HC concentration becomes high, as shown in , and the catalyst purification rate of HC decreases, as shown in x , so HC is emitted. That is, if the control method in which the concentration correction is stopped after the output of the post oxygen sensor 22 reacts is adopted, the timing at which the concentration correction is stopped is too slow for the catalyst, so HC cannot be appropriately prevented. Therefore, in the air-fuel ratio control of the internal combustion engine, the state of the catalyst which cannot be directly observed from the outside needs to be considered, and the concentration correction control for the appropriate period needs to be performed.

[0088] Figure 10B Figure 12 The horizontal axis of the graph of FIG. 7 is the deterioration degree of the catalyst, and the vertical axis is the oxygen storage capacity.

[0089] Figure 16 is a graph of the relationship between the deterioration degree of the catalyst and the oxygen storage capacity (OSC: Oxygen Storage Capacity) of the three-way catalyst. Figure 16 The horizontal axis of the graph of FIG. 7 is the deterioration degree of the catalyst, and the vertical axis is the oxygen storage capacity.

[0090] ​​​Catalyst degradation refers to a state in which the catalytic effect of a three-way catalyst is reduced due to heat or sulfur poisoning in the fuel. Figure 16 As shown, the oxygen storage capacity of the three-way catalyst is Figure 16 The degree of catalyst degradation in the left and right directions is roughly proportional. That is, as catalyst degradation progresses, the oxygen storage capacity of the three-way catalyst tends to decrease.

[0091] The following describes the effect of changes in oxygen storage capacity on the purification function of the three-way catalyst.

[0092] Figure 17 The oxygen storage ratio of the three-way catalyst, that is, the value of the oxygen storage amount of the three-way catalyst at the current moment divided by the above oxygen storage capacity and NO x The relationship between the purification rate and the Figure 17 The horizontal axis is the oxygen storage ratio, and the vertical axis is the NO x Purification rate. Figure 17 In the figure, the solid line shows the NO x The change in purification rate is shown by the dotted line, indicating the NO x Changes in purification rate. When the oxygen storage ratio is small, both the new three-way catalyst and the deteriorated three-way catalyst have a high oxygen storage ratio. In addition, the NO x The purification rate is higher than that of the degraded three-way catalyst. Figure 17 As shown in the figure, when the oxygen storage ratio exceeds the specified value, both the new three-way catalyst and the deteriorated three-way catalyst will produce NO. x The purification rate tends to deteriorate rapidly. Figure 11 As explained, in NO x In the purification process, Ce2O3 in the catalyst is important. When all Ce2O3 reacts and changes into CeO2, it is because Ce2O3 cannot react with NO. NO x Furthermore, if the catalyst deteriorates and the oxygen storage capacity of the three-way catalyst decreases, the value of the oxygen storage ratio relative to the same oxygen storage amount increases as a result.

[0093] Therefore, NO relative to oxygen storage x The purification rate value is Figure 18 The solid line shows the NO of the new three-way catalyst. x The purification rate is compared with Figure 18 The NO of the degraded three-way catalyst is shown by the middle dotted line. x Therefore, in order to reduce NO xTo maintain a high purification rate, it is necessary not only to maintain the exhaust air-fuel ratio at the catalyst inlet at the three-way point, but also to consider the internal state of the catalyst, which cannot be directly observed from the outside, that is, the current oxygen storage capacity and oxygen storage ratio of the three-way catalyst, and to perform appropriate correction control on the exhaust air-fuel ratio at the catalyst inlet to keep the oxygen storage ratio at a level that can obtain the specified NO x In addition, within the prescribed range of the so-called oxygen storage ratio, such as Figure 17 As shown, when the catalyst is degraded, it is represented by a degraded control range 36 , and when the catalyst is new, it is represented by a new control range 35 . The new control range 35 is wider than the degraded control range 36 .

[0094] Next, use Figure 19 The structure of this embodiment is described below. First, control inputs such as the exhaust flow rate and equivalence ratio, and sensor inputs such as the exhaust temperature sensor that detects the exhaust temperature on the downstream side of the catalyst are input from the ECU to the extended Kalman filter 1901. Based on these inputs, the extended Kalman filter 1901 calculates the time differential value of the catalyst temperature. Then, using the specified time scale value and the previous value state quantity, an integral calculation is performed to calculate the current catalyst temperature. In addition, the same type of input value as the extended Kalman filter 1901 is also input to the Mahalanobis distance calculation unit 1902. The input value is used to calculate the Mahalanobis distance of the current input. The comparison unit 1903 compares the calculated Mahalanobis distance with the specified threshold. If the current Mahalanobis distance is greater than the threshold, 0 is output, which is considered to be outside the learning range or the sensor is faulty. If the current Mahalanobis distance is less than the threshold, 1 is output, which is judged to be within the learning range or the sensor is normal. Then, the interpolation results are recorded. If 0 is input to the switching unit 1904, it is assumed that the extended Kalman filter 1901 is operating outside the learning range or the sensor is malfunctioning, and is unable to perform normal estimation calculations, and a predetermined steady-state value is output. If 1 is input, it is assumed that the extended Kalman filter 1901 is operating within the learning range or the sensor is operating normally, and is able to perform normal estimation calculations, and the estimated catalyst temperature is output.

[0095] Next, use Figure 20The flowchart of this embodiment will be described. In step 2001, control inputs and sensor input values ​​are received from the ECU. In step 2002, the Mahalanobis distance is calculated using the control inputs and sensor input values ​​from the ECU. In step 2003, the Mahalanobis distance calculated is compared with the threshold value specified in step 2002. If the Mahalanobis distance is less than the threshold value, the process proceeds to step 2004; if it is greater than the threshold value, the process proceeds to step 2005. If the process proceeds to step 2004, the extended Kalman filter is operating within the learning range or the sensor is operating normally. The catalyst temperature (differential value) is estimated using the control inputs and sensor input values ​​from the ECU. An integral calculation is performed using the specified time scale value and the previous value state quantity to determine the current catalyst temperature. If the process proceeds to step 2005, the extended Kalman filter is operating outside the learning range or a sensor failure occurs. The diagnostic result is recorded and used as sensor failure or relearning data. In step 2006, a substitute value, such as an estimated value calculated using a physical formula or a MAP output value specified by the operating point, is output.

[0096] The main features of the above-mentioned embodiments can also be summarized as follows.

[0097] Control unit (ECU 28, Figure 1 ) of the processor (CPU 205, Figure 2 ) according to the state variables of the vehicle at the first moment (for example, the last value of the catalyst temperature, Figure 19 ), the control input of the vehicle at a second moment after the first moment (such as exhaust flow, equivalence ratio, Figure 19 ) and the observation value at the second moment detected by the vehicle sensor (sensor input: for example, the exhaust gas temperature on the downstream side of the catalyst, Figure 19 ), calculates the state variable at the second moment (e.g., catalyst temperature). The processor uses a neural network model to approximate a state space model consisting of a state equation related to the state variable (e.g., catalyst temperature) (the first equation of equation (1)) and an observation equation related to the observed value (e.g., exhaust gas temperature on the downstream side of the catalyst) (the second equation of equation (1)). The processor determines the reliability (accuracy) of the state variable at the second moment (e.g., catalyst temperature) based on an indicator (e.g., Mahalanobis distance) based on a control input at the second moment (e.g., exhaust gas flow rate, equivalence ratio) and an observed value at the second moment (e.g., exhaust gas temperature on the downstream side of the catalyst). In addition, the observed value is a sensor value related to the state variable (the second equation of equation (1)).

[0098] This allows the reliability of calculated values ​​(estimated values) of vehicle state variables not detected by sensors to be evaluated. Utilizing highly reliable state variables in subsequent processing also improves the reliability of subsequent processing.

[0099] An example of the index is a Mahalanobis distance based on the control input at the second time (e.g., exhaust flow rate, equivalence ratio) and the observation value at the second time (e.g., exhaust temperature on the downstream side of the catalyst). In addition, a value related to the Mahalanobis distance can be used as the index.

[0100] Thus, the reliability of the calculated value of the state variable at the second time can be evaluated while taking into account the variances of the control input and the observation value.

[0101] The processor (CPU 205) determines that the reliability of the state variable at the second time (e.g., catalyst temperature) is low if the Mahalanobis distance is greater than the threshold value (comparing section 1903, Figure 19 ). In addition, in the second embodiment, the processor (CPU 205) outputs 0 from the comparing section 1903 if the Mahalanobis distance is greater than the threshold value, determining that the reliability of the state variable at the second time (e.g., catalyst temperature) is low.

[0102] Thus, if the sensor is normal, interpolation (Mahalanobis distance ≤ threshold value) and extrapolation (Mahalanobis distance > threshold value) can be determined. In the case where the sensor is abnormal (failure, attachment of dirt, etc.), it is likely that the Mahalanobis distance > threshold value.

[0103] The state equation and the observation equation are nonlinear. The processor (CPU 205) calculates the state variable (e.g., catalyst temperature) using the extended Kalman filter 1901 Figure 19 ).

[0104] Thus, even if the state equation and the observation equation are nonlinear, the state variable can be easily calculated by linear approximation.

[0105] In detail, the processor (CPU 205) outputs the state variable at the second time (e.g., catalyst temperature) calculated using the extended Kalman filter 1901 if the Mahalanobis distance is equal to or less than the threshold value (step 2004, Figure 20 ). The processor (CPU 205) does not output the state variable at the second time (e.g., catalyst temperature) calculated using the extended Kalman filter 1901 if the Mahalanobis distance is greater than the threshold value, and outputs (sends or stores) a diagnosis result indicating that the reliability of the state variable at the second time (e.g., catalyst temperature) is low or that the sensor has failed (step 2005). Figure 20 ).

[0106] Thus, since the state variable with low reliability is not used in the processing at the later stage, the decrease in reliability of the processing at the later stage can be suppressed.

[0107] For example, as Figure 21As shown, the neural network model is composed of a first neural network model and a second neural network model. The first neural network model includes: an input layer, which is input with the state variable x(k) at the first moment and the control input u(k) at the first moment; and an output layer, which outputs the state variable x(k+1) at the second moment. The second neural network model includes: an input layer, which is input with the state variable x(k) at the first moment and the control input u(k) at the first moment; and an output layer, which outputs the observation value y(k) at the first moment.

[0108] This makes it possible to estimate state variables and observations without solving state and observation equations. Furthermore, learning improves the reliability of the state variables and observations estimated using the neural network model. By using state variables estimated by the neural network model in powertrain control, high-precision powertrain control can be achieved.

[0109] like Figure 19 As shown, the processor (CPU 205) inputs a state variable at a first moment (e.g., the previous value of the catalyst temperature), a control input at a second moment (e.g., exhaust flow rate, equivalence ratio), and an observation value at the second moment (sensor input: e.g., exhaust temperature on the downstream side of the catalyst) into the extended Kalman filter 1901, and obtains a time-differential value (state quantity) of the state variable from the extended Kalman filter 1901. The processor (CPU 205) integrates the time-differential value (state quantity) of the state variable and calculates the state variable at the second moment (e.g., catalyst temperature) by adding the integrated value to the state variable at the first moment (e.g., the previous value of the catalyst temperature).

[0110] This makes it possible to calculate state variables at high speed.

[0111] When the Mahalanobis distance is less than the threshold, the processor (CPU 205) outputs the state variable (e.g., catalyst temperature) at the second moment calculated using the extended Kalman filter 1901. When the Mahalanobis distance is greater than the threshold, the processor (CPU 205) does not output the state variable (e.g., catalyst temperature) at the second moment calculated using the extended Kalman filter 1901, but outputs a substitute value derived from a physical formula or a map as the state variable (e.g., catalyst temperature) (step 2006, Figure 20 ).

[0112] As a result, for example, subsequent processing using the state variable is not interrupted.

[0113] The processor (CPU 205) performs machine learning using the state variable at the first time point calculated by the extended Kalman filter 1901 (for example, the last value of the catalyst temperature), the control input at the first time point (for example, the exhaust flow rate, the equivalence ratio), and the state variable at the second time point calculated by the extended Kalman filter 1901 (for example, the catalyst temperature) as the teacher data of the first neural network model. In addition, the processor (CPU 205) performs machine learning using the state variable at the first time point calculated by the extended Kalman filter 1901 (for example, the last value of the catalyst temperature), the control input at the first time point (for example, the exhaust flow rate, the equivalence ratio), and the observation value at the first time point (for example, the exhaust temperature on the downstream side of the catalyst) as the teacher data of the second neural network model.

[0114] Thus, with respect to the state variable that is not detected by the sensor, it is possible to perform machine learning using the state variable calculated by the extended Kalman filter.

[0115] The processor (CPU 205) performs machine learning when the Mahalanobis distance is less than or equal to the threshold value, and does not perform machine learning when the Mahalanobis distance is greater than the threshold value.

[0116] Thus, it is possible to perform machine learning using the state variable calculated by the extended Kalman filter, which has high reliability. As a result, the reliability of the state variable and the observation value estimated using the neural network model is improved. By using the state variable estimated by the neural network model in powertrain control, it is possible to achieve high-precision powertrain control.

[0117] In addition, the present application is not limited to the above-described embodiments, and various other applications and modifications can of course be made without departing from the spirit of the present application described in the claims.

[0118] For example, the above-described embodiments are embodiments in which the configuration of the system is described in detail and specifically in order to easily understand the present application, and are not necessarily limited to embodiments having all the configurations described. In addition, with respect to a part of the configuration of the present embodiment, addition, deletion, and substitution of other configurations can be made.

[0119] In addition, the control lines and the information lines represent lines that are considered necessary for explanation, and all the control lines and the information lines are not necessarily represented on the product. In fact, it can be considered that almost all the configurations are connected to each other.

[0120] In addition, each of the above-described configurations, functions, and the like can be implemented by hardware, such as by an integrated circuit design, or by a part or all of them. In addition, each of the above-described configurations, functions, and the like can be implemented by software by a processor interpreting and executing a program that implements each function. The program, table, file, and the like that implement each function can be stored in a memory, a hard disk, an SSD (Solid State Drive), or the like, or an IC card, an SD card, a DVD, or the like.

[0121] In addition, the embodiment of the present application can also be in the following manner.

[0122] (1) A powertrain control device for a vehicle having a sensor that detects a current state, a state observer that includes: an index value setting section that, based on a state variable related to the vehicle detected by the sensor, a control input for controlling the vehicle, and a state estimation quantity estimated by the state observer, calculates an abnormality diagnosis index based on the control input and the sensor detection value, and a determination threshold value as an index for determining the validity of the state estimation quantity; and a neural network model that approximates a state space model provided with the sensor detection object as one of the output variables, the powertrain control device for a vehicle including: a unit that determines the validity of the state quantity estimation result using the state observer based on the abnormality diagnosis index specified by the input of the neural network model and the sensor detection value.

[0123] (2) The powertrain control device includes a means that uses a Mahalanobis distance as the abnormality diagnosis index specified by the control input and the sensor detection value.

[0124] (3) In the powertrain control device described in (1), there is a means that determines the validity of the state quantity estimation result of the state observer based on the Mahalanobis distance.

[0125] (4) In the powertrain control device described in (1), there is a means that implements state quantity estimation using an extended Kalman filter that can consider nonlinearity as the state observer.

[0126] (5) In the powertrain control device described in (1), when the Mahalanobis distance is equal to or greater than a specified value, state quantity estimation using the extended Kalman filter is stopped, it is diagnosed that the state quantity estimation or the sensor is abnormal, and the diagnosis result is notified to the outside or recorded.

[0127] (6) In the powertrain control device described in (1), the neural network model that approximates the state space model includes a transition neural network model that approximates a state space model that uses a previous value as one of the inputs.

[0128] (7) In the power plant control device described in (1), the extended Kalman filter calculates a time differential value of the state quantity, adds it to a previous value, and estimates a current state quantity.

[0129] (8) In the power plant control device described in (1), there is provided a configuration in which, when the Mahalanobis distance is equal to or greater than a predetermined value, a state quantity estimation using the extended Kalman filter is stopped, and a substitute value calculated by a physical formula or a MAP is output.

[0130] According to (1) to (8), it is possible to diagnose the effectiveness and ineffectiveness of the transition neural network model and the sensor with a minimum amount of information and calculation, and to ensure the reliability of the control software. Explanation of symbols

[0131] 1... internal combustion engine, 20... catalyst-upstream air-fuel ratio sensor, 21... exhaust purification catalyst, 22... rear oxygen sensor, 28... ECU, 51... exhaust flow rate calculation section, 52... catalyst-upstream state quantity estimation section, 53... catalyst-downstream state quantity estimation section, 54... correction section, 55... NO x sensor diagnosis section, 56... fuel injection amount correction section, 56a... fuel injection amount correction map, 57... fuel composition discrimination section, 100... internal combustion engine control system, 110... aftertreatment system.

Claims

1. A control device characterized by comprising: a processor that performs processing of: calculating a state variable of a vehicle at a second time after a first time, based on the state variable of the vehicle at the first time, a control input of the vehicle at the second time, and an observation value at the second time detected by a sensor of the vehicle, approximating a state space model constituted by a state equation related to the state variable and an observation equation related to the observation value with a neural network model, determining reliability of the state variable at the second time based on an index based on the control input at the second time and the observation value at the second time.

2. The control device according to claim 1, characterized in that: the index is a Mahalanobis distance based on the control input at the second time and the observation value at the second time.

3. The control device according to claim 2, characterized in that: the processor determines that the reliability of the state variable at the second time is low when the Mahalanobis distance is greater than a threshold value.

4. The control device according to claim 1, characterized in that: the state equation and the observation equation are nonlinear, the processor calculates the state variable using an extended Kalman filter.

5. The control device according to claim 2, characterized in that: the processor performs processing of: outputting the state variable at the second time calculated using the extended Kalman filter when the Mahalanobis distance is below a threshold value, not outputting the state variable at the second time calculated using the extended Kalman filter when the Mahalanobis distance is greater than a threshold value, and outputting a diagnosis result indicating that the reliability of the state variable at the second time is low or that the sensor has failed.

6. The control device according to claim 4, characterized in that: the neural network model is constituted by a first neural network model and a second neural network model, the first neural network model includes: an input layer that inputs the state variable at the first time and the control input at the first time; and an output layer that outputs the state variable at the second time, the second neural network model includes: an input layer that inputs the state variable at the first time and the control input at the first time; and an output layer that outputs the observation value at the first time.

7. The control device according to claim 4, characterized in that: the processor performs processing of: inputting the state variable at the first time, the control input at the second time, and the observation value at the second time to the extended Kalman filter, and acquiring a time differential value of the state variable from the extended Kalman filter, integrating the time differential value of the state variable, and calculating the state variable at the second time by adding an integrated value to the state variable at the first time.

8. The control device according to claim 2, characterized in that: the processor performs processing of: in a case where the Mahalanobis distance is below a threshold value, outputting the state variable at the second time calculated using the extended Kalman filter, in a case where the Mahalanobis distance is greater than a threshold value, not outputting the state variable at the second time calculated using the extended Kalman filter, and outputting a substitute value derived from a physical equation or a map as the state variable.

9. The control device according to claim 6, wherein the processor performs processing of: using the state variable at the first time calculated using the extended Kalman filter, the control input at the first time, and the state variable at the second time calculated using the extended Kalman filter as teacher data of the first neural network model, using the state variable at the first time calculated using the extended Kalman filter, the control input at the first time, and the observation value at the first time as teacher data of the second neural network model, and performing machine learning.

10. The control device according to claim 9, wherein the index is a Mahalanobis distance based on the control input at the second time and the observation value at the second time, the processor performs processing of: in a case where the Mahalanobis distance is below a threshold value, performing the machine learning, in a case where the Mahalanobis distance is greater than a threshold value, not performing the machine learning.

Images