JP2020056378A - Control device of internal combustion engine - Google Patents

Abstract

To provide a control device of an internal combustion engine which can reflect information obtained by learning up to an update to the control of the internal combustion engine even if a learnt model is updated.SOLUTION: This control device of an internal combustion engine comprises: a parameter value output part 221 for outputting a value of an output parameter by using a learnt model; a difference output part 222 for outputting a difference between the value of the output parameter and an actual value of the output parameter by using a difference model; an addition output part 223 for outputting a sum of an output value of the parameter value output part and an output value of the difference output part; an engine control part 224 for controlling the internal combustion engine on the basis of an output value of the addition output part; a learning part for acquiring a difference between the value of the output parameter and an actual measurement value of the output parameter as right answer data, and learning a difference model with a group of a value of an input parameter of the difference model and the right answer data as teacher's data; and an update part for updating the learnt model on the basis of the update data acquired from a computer which is installed outside a vehicle.SELECTED DRAWING: Figure 5

Description

  The present disclosure relates to a control device for an internal combustion engine.

  2. Description of the Related Art Conventionally, a technique for controlling an internal combustion engine of a vehicle using a learned model learned by machine learning is known (for example, see Patent Document 1). In particular, in each model described in Patent Literature 1, when a plurality of input parameters related to the operation of the internal combustion engine are input, the flow rates of the intake gas, the exhaust gas, and the EGR gas are output using the neural network.

JP 2012-112277 A

  In a control device using such a learned model of machine learning, the learned model may be updated by a service campaign, a recall, or the like for the purpose of further improving various control functions. In this case, for example, the control device mounted on the vehicle acquires update data for updating the learned model from a computer (for example, an external server) installed outside the vehicle via wireless communication or the like. Thereafter, the control device rewrites the learned model using the update data, that is, updates the learned model.

  As a control device using such a learned model, a control device that updates the learned model by autonomously re-learning the learned model may be considered. In the control device that autonomously re-learns the learned model, the learned model is updated based on the vehicle-specific teacher data acquired during driving. For this reason, the updated learned model is a model in which characteristics specific to the vehicle such as the use situation and the use environment of the vehicle are reflected.Accordingly, according to such a learned model, it is possible to more appropriately control the internal combustion engine. It becomes possible.

  By the way, in the control device that performs such re-learning, it is assumed that the learned model is updated by the above-described service campaign or recall. In this case, the learned model in which the characteristics unique to the vehicle are reflected by the re-learning up to that time is rewritten to a standard learned model uniformly created by a manufacturer or the like. As a result, the updated trained model does not reflect the vehicle-specific characteristics reflected by the previous re-learning. When machine learning such as a neural network is used, since the model is a black box model, the vehicle-specific characteristics reflected in the updated trained model by re-learning up to that time Difficult to reflect. Therefore, in order to reflect the vehicle-specific characteristics in the updated trained model, it is necessary to re-learn from the beginning again based on the vehicle-specific teacher data acquired during driving after the update.

  The present disclosure has been made in view of the above-described problem, and its purpose is to obtain a learning model even if a learned model is updated based on update data obtained from an external computer. An object of the present invention is to provide a control device for an internal combustion engine capable of reflecting information in control of the internal combustion engine.

  The gist of the present disclosure is as follows.

  (1) When a value of an input parameter is input, a parameter value output unit that outputs a value of an output parameter using a learned model, and when a value of an input parameter is input, the parameter value is output using a difference model. A difference output unit that outputs a difference between a value of the output parameter output by the value output unit and an actual value of the output parameter; an output value of the parameter value output unit and an output value of the difference output unit are input; Then, an addition output unit that outputs the sum of the input output value of the parameter value output unit and the input output value of the difference output unit, and controls the internal combustion engine based on the output value of the addition output unit An engine control unit that obtains a difference between a value of the output parameter output by the parameter value output unit and an actually measured value of the output parameter as correct data, and obtains the difference model A learning unit that learns the difference model, using a set of the input parameter value and the correct data as teacher data, and obtains update data for updating the learned model from a computer installed outside the vehicle. An update unit that updates the learned model based on the update data.

  According to the present disclosure, even when a learned model is updated based on update data obtained from an external computer, an internal combustion engine capable of reflecting information obtained by learning up to that time in control of the internal combustion engine. A control device can be provided.

FIG. 1 is a schematic configuration diagram of an internal combustion engine and an ECU that controls the internal combustion engine according to the first embodiment. FIG. 2 is a diagram illustrating an example of a neural network. FIG. 3 is a diagram illustrating a specific example of a neural network in the learned model according to the first embodiment. FIG. 4 is a diagram illustrating a specific example of a neural network in the difference model according to the first embodiment. FIGS. 5A and 5B are block diagrams illustrating the control of the internal combustion engine according to the first embodiment. FIG. 6 is a schematic configuration diagram of a control system for an internal combustion engine according to the second embodiment.

  Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings. In the following description, similar components are denoted by the same reference numerals.

<First embodiment>
≫Configuration of internal combustion engine≫
FIG. 1 is a schematic configuration diagram of an internal combustion engine and an electronic control unit (ECU) that controls the internal combustion engine according to the first embodiment. FIG. 1 shows an internal combustion engine 100 including an engine body 1, an intake manifold 4, and an exhaust manifold 5. As shown in FIG. 1, the engine main body 1 includes a fuel injection valve 3 disposed toward the inside of the combustion chamber 2 of each cylinder. The intake manifold 4 is connected to an outlet of a compressor 7 a of an exhaust turbocharger 7 via an intake duct 6, and an inlet of the compressor 7 a is connected to an air cleaner 9. A throttle valve 11 which is opened and closed by a throttle actuator 10 is arranged in the intake duct 6, and an intercooler 12 for cooling intake air flowing through the intake duct 6 is arranged around the intake duct 6.

  On the other hand, the exhaust manifold 5 is connected to an inlet of an exhaust turbine 7b of the exhaust turbocharger 7, and an outlet of the exhaust turbine 7b is connected to an exhaust purification catalytic converter 14 via an exhaust pipe 13. In the example shown in FIG. 1, an oxidation catalyst 15 and a particulate filter 16 are accommodated in the exhaust purification catalytic converter 14 in order from the upstream side. The exhaust manifold 5 and the intake manifold 4 are connected to each other via an exhaust gas recirculation (hereinafter, referred to as “EGR”) passage 17, and an EGR control valve 18 is disposed in the EGR passage 17. Further, an EGR cooler 19 for cooling the EGR gas flowing in the EGR passage 17 is disposed in the EGR passage 17. Each fuel injection valve 3 is connected to a common rail 21 via a fuel supply pipe 20, and this common rail 21 is connected to a fuel tank 23 via a fuel pump 22.

  As shown in FIG. 1, the intake manifold 4 is provided with an intake air temperature sensor 24 for detecting the intake air temperature in the intake manifold 4, and the exhaust manifold 5 is used to detect the exhaust gas temperature in the exhaust manifold 5. Temperature sensor 25 is disposed for the purpose. The intake pipe 8 is provided with an air flow meter 8a for detecting a flow rate of air flowing through the intake pipe 8. Further, the engine body 1 has a water temperature sensor 26 for detecting the temperature of the engine cooling water (hereinafter, simply referred to as “water temperature”), and a temperature (hereinafter, referred to as lubricating oil) for lubricating the frictional sliding portion of the engine body 1. And an oil temperature sensor 27 for detecting the “oil temperature”).

  The throttle actuator 10 is provided with a throttle opening sensor 10a for detecting the opening of the throttle valve 11 (hereinafter, referred to as "throttle opening"). The EGR control valve 18 is provided with an EGR valve opening sensor 18a for detecting the opening of the EGR control valve 18 (hereinafter, referred to as “EGR valve opening”).

A gas concentration sensor 28 for detecting the gas concentration in the exhaust gas and an air-fuel ratio sensor 29 for detecting the air-fuel ratio of the exhaust gas are disposed in the exhaust pipe 13. Gas concentration sensor 28, for example, be a NO _x sensor which detects a NO _x concentration in the exhaust gas. As the gas concentration sensor 28, in addition to of the NO _x sensor, depending on the parameters used in the learned model and the differential model described below, for example, detecting the HC concentration in the exhaust gas, CO concentration, CO ₂ concentration, respectively HC These sensors such as a sensor, a CO sensor, and a CO ₂ sensor can be used as appropriate.
<< Configuration of ECU >>

  The electronic control unit (ECU) 200 includes a storage unit 210, a control unit 220, an input port 230, and an output port 240 connected to each other by the bi-directional bus 201.

  The storage unit 210 can include, for example, a ROM (Read Only Memory) and a RAM (Random Access Memory). The storage unit 210 stores various programs used for processing by the control unit 220, a learned model, a difference model, and various data (for example, various parameters, teacher data, various thresholds, and the like).

  The control unit 220 can be, for example, a processor having a CPU and its peripheral circuits. The control unit 220 can execute various controls of the vehicle by executing various programs stored in the storage unit 210.

  As shown in FIG. 1, the control unit 220 includes a parameter value output unit 221, a difference output unit 222, an addition output unit 223, an engine control unit 224, a learning unit 225, and an update unit 226. . Each of these units included in the control unit 220 is a functional module implemented by a program executed on a processor of the control unit 220.

  The input port 230 includes an air flow meter 8a, a throttle opening sensor 10a, an EGR valve opening sensor 18a, an intake temperature sensor 24, an exhaust temperature sensor 25, a water temperature sensor 26, an oil temperature sensor 27, a gas concentration sensor 28, an air-fuel ratio sensor 29 Output signals of a torque sensor 51 for detecting an output torque (hereinafter, referred to as “torque”) of the internal combustion engine 100 and a knocking sensor 52 for detecting the presence or absence of knocking are input via the corresponding AD converters 231. You. The input port 230 receives, as a signal for detecting an engine load, an output voltage of a load sensor 53 a that generates an output voltage proportional to the amount of depression of an accelerator pedal 53 via a corresponding AD converter 231. Is done. The input port 230 receives, as a signal for calculating the engine rotation speed and the like, an output signal of the crank angle sensor 54 that generates an output pulse every time the crankshaft rotates, for example, by 15 °. As described above, the output signals of various sensors necessary for controlling the internal combustion engine 100 are input to the input port 230.

  Further, to the input port 230, update data for updating the learned model via a wireless communication or the like is input from a computer (for example, an external server) installed outside the vehicle.

  The output port 240 is electrically connected to each control component such as the fuel injection valve 3, the throttle actuator 10, the EGR control valve 18, and the fuel pump 22 via a corresponding drive circuit 241.

  ECU 200 outputs a control signal for controlling each control component from output port 240 based on output signals of various sensors input to input port 230, and controls internal combustion engine 100. Therefore, the ECU 200 functions as a control device of the internal combustion engine 100.

の Overview of Neural Network≫
In the present embodiment, the learned model and the difference model use a neural network. First, a neural network used in the learned model and the difference model according to the present embodiment will be described with reference to FIG. FIG. 2 shows an example of a neural network. The circles in FIG. 2 represent artificial neurons. In a neural network, this artificial neuron is usually referred to as a node or unit (referred to herein as a node). In FIG. 2, L = 1 indicates an input layer, L = 2 and L = 3 indicate a hidden layer, and L = 4 indicates an output layer. In FIG. 2, x ₁ and x ₂ indicate nodes in the input layer (L = 1) and output values from the nodes, and y indicates each node in the output layer (L = 4) and the output values from the nodes. Is shown. Similarly, z ₁ , z ₂ and z ₃ of the hidden layer (L = 2) indicate output values from each node of the hidden layer (L = 2), and z ₁ and z _{2 of the} hidden layer (L = 3). Indicates an output value from each node of the hidden layer (L = 3). The number of hidden layers can be one or any number, and the number of nodes in the input layer and the number of nodes in the hidden layer can also be any number. In this embodiment, the number of nodes in the output layer is one.

At each node of the input layer, the input is output as it is. On the other hand, the output values x ₁ and x ₂ of each node of the input layer are input to each node of the hidden layer (L = 2). At each node of the hidden layer (L = 2), the total input value u is calculated using the corresponding weight w and bias b. For example, the total input value u _k calculated at the node indicated by z _k (k = 1, 2, 3) of the hidden layer (L = 2) in FIG. Number of nodes).

Then, the total input value u _k is converted by activation function f, is output from the node indicated by z _k of the hidden layer (L = 2), as the output value _{z k (= f (u k} )) . On the other hand, the output values z ₁ , z ₂ and z ₃ of each node of the hidden layer (L = 2) are input to each node of the hidden layer (L = 3). At each node of the hidden layer (L = 3), the total input value u (Σz · w + b) is calculated using the corresponding weight w and bias b. The total input value u is similarly converted by the activation function f, and output as output values z ₁ and z ₂ from each node of the hidden layer (L = 3). In the present embodiment, a sigmoid function σ is used as the activation function.

On the other hand, the output values z ₁ and z ₂ of each node of the hidden layer (L = 3) are input to the nodes of the output layer (L = 4). At the output layer node, the total input value u (ｕz · w + b) is calculated using the corresponding weight w and the bias b, or the total input value u (Σz · w) is calculated. In the present embodiment, the identity function is used as the activation function at the output layer node. Therefore, the total input value u calculated at the output layer node is directly output from the output layer node as the output value y Is output as

≫Learning in neural networks≫
In the present embodiment, the value of each weight w and the value of the bias b in the neural network are learned using the backpropagation method. This backpropagation method is well known, and therefore, the outline of the backpropagation method will be briefly described below. Since the bias b is a kind of the weight w, the bias b is assumed to be one of the weights w in the following description.

Now, in a neural network as shown in FIG. 2, if the weight at the input value u ^(L) to the node of each layer of L = 2, L = 3 or L = 4 is represented by w ^(L) , the error function E The differentiation by the weight w ^(L) , that is, the gradient ∂E / ∂w ^(L) is expressed by the following equation (1).

Here, since z ^(L-1) · ∂w ^(L) = ∂u ^(L) , if (∂E / ∂u ^(L) ) = δ ^(L) , the above equation (1) becomes: It can be expressed by the following equation (2).

ここで、z^(L)＝ｆ（ｕ^(L)）と表すと、上記（３）式の右辺に現れる入力値ｕ_k ^(L+1)は、次の（４）式で表すことができる。

Here, if u ^(L) fluctuates, the error function E fluctuates through a change in the total input value u ^{(L + 1)} of the next layer, so δ ^(L) is expressed by the following equation (3). (K) is the number of nodes in the L + 1 layer).

Here, assuming that z ^(L) = f (u ^(L) ), the input value u _k ^{(L + 1)} appearing on the right side of the above equation (3) can be expressed by the following equation (4). .

Here, the first term on the right side (∂E / ∂u ^{(L + 1)} ) of the above equation (3) is δ ^{(L + 1)} . Equation (3) of the second term on the right side ^{_{(∂u k (L + 1)}} / ∂u (L)) , from equation (4) can be expressed by the following equation (5).

Therefore, δ ^(L) can be expressed by the following equation (6) from the above equations (3) to (5).

That is, ^once δ ^{(L + 1)} is obtained, δ ^(L) can be obtained.

By the way, teacher data including a certain input value x and correct data t for the input value x has been obtained. If the output value from the output layer for this input value x is y, the square error becomes an error function. when being used, the square error E is calculated by ^{E = (y-t) 2} /2. At the node of the output layer (L = 4) shown in FIG. 2, the output value is y = f (u ^(L) ). In this case, δ ^{(L) at} the node of the output layer (L = 4 ⁾ Is given by the following equation (7).

By the way, in the present embodiment, as described above, since f (u ^(L) ) is an identity function, f ′ (u ^(Ll) ) = 1. Therefore, δ ^(L) = y−t, and δ ^(L) can be obtained.

Once δ ^(L) is obtained, δ ^(L-1) of the preceding layer can be obtained using the above equation (6). In this way, the δ of the previous layer is sequentially obtained, and using the values of δ, the differential of the error function E for each weight w, that is, the gradient ∂E / ∂w ^(L) Can be requested.

When the gradient ∂E / ∂w ^(L) is obtained, the value of the weight w is updated using the gradient ∂E / ∂w ^{(L) so} that the value of the error function E decreases. That is, learning of the value of the weight w is performed. When a batch or mini-batch is used as teacher data, a sum-of-squares error E expressed by the following equation (8) is used as the error function E. Here, N is the total number of teacher data, i is a natural number equal to or less than N (i = 1, 2,..., N), and y _i and t _i are the output value and the correct answer data for the input value x _ki , respectively. Show.

On the other hand, if the learning sequentially calculates the square error is carried out, as error function E, the square error E = (y-t) of the above ^2/2 is used.

≫Overview of trained model≫
The outline of the learned model in the present embodiment will be described. First, an example of the input parameters used in the learned model in the present embodiment will be described. The input parameters of the learned model in the present embodiment are ignition timing, fuel injection amount, fuel injection timing, opening / closing timing of intake valves and exhaust valves of the internal combustion engine (hereinafter referred to as “In-Ex VVT timing”), throttle opening. , EGR valve opening, intake air temperature, water temperature, oil temperature, and engine speed.

  Next, an example of a method of acquiring the value of each input parameter will be described. The ignition timing, the combustion injection amount, the fuel injection timing, and the In-Ex VVT timing are each obtained from a command value of the ECU 200. The throttle opening, the EGR valve opening, the water temperature, the intake air temperature, and the oil temperature are obtained from the output values of the throttle opening sensor 10a, the EGR valve opening sensor 18a, the intake temperature sensor 24, the water temperature sensor 26, and the oil temperature sensor 27, respectively. Is obtained. The engine rotation speed is obtained from a value calculated by the ECU 200 based on the output signal of the crank angle sensor 54.

Next, an example of an output parameter used in the learned model in the present embodiment will be described. Output parameters of the learned model in the present embodiment, exhaust temperature, NO _x concentration, HC concentration in the exhaust gas, CO concentration and the CO ₂ concentration, the air-fuel ratio of the exhaust gas, the output torque as well as knocking determination value of the engine 100 At least one of them can be included.

  FIG. 3 shows a specific example of a neural network in the learned model according to the present embodiment. The neural network in the learned model shown in FIG. 3 uses an ignition timing, a fuel injection amount, a throttle opening, and an engine rotation speed as input parameters and torque as an output parameter. In the present embodiment, the neural network of the learned model is composed of P layers (P is an arbitrary integer of 3 or more), and the number of nodes in each hidden layer can be an arbitrary number. In the neural network in the learned model shown in FIG. 3, the input layer (L = 1) has four nodes corresponding to four input parameters. Any number of nodes other than four.

In addition to the example shown in FIG. 3, the neural network in the learned model according to the present embodiment uses, for example, ignition timing, fuel injection amount, throttle opening, EGR valve opening, water temperature, and engine speed as input parameters. , NO _X and HC can be output parameters. The neural network in the learned model according to the present embodiment can use, for example, an ignition timing, a fuel injection amount, a throttle opening, a water temperature, and an engine speed as input parameters and a knocking determination value as an output parameter.

≫Update of trained model≫
In the present embodiment, the update unit 226 of the ECU 200 acquires update data for updating the learned model via the input port 230 from a computer provided outside the vehicle, for example, in a service campaign or a recall. The updating unit 226 updates the learned model by rewriting the learned model stored in the storage unit 210 with a learned model corresponding to the updated data based on the acquired updated data. In the update of the learned model, the number of nodes in the input layer and the hidden layer, the number of hidden layers, the weight, and at least a part of the bias are changed in the neural network of the learned model.

  In the present embodiment, by updating the learned model, for example, it is possible to further optimize the number of nodes of the input layer and the hidden layer, the number of the hidden layers, the weight, the bias, and the like in the neural network of the learned model. Become. As a result, it is possible to improve the prediction accuracy of the output parameters of the learned model.

  When updating the trained model in this way, it is necessary that at least the output parameters of the trained model before the update and the trained model after the update are the same.

問題 Problems of updating the trained model モ デ ル
By the way, the above-described learned model is a standard model uniformly created by a manufacturer or the like, and does not reflect characteristics unique to each vehicle. Therefore, when the value of the output parameter is estimated using the learned model, there is a possibility that an error may occur between the value of the output parameter output from the learned model and the actual value of the output parameter. is there.

  Therefore, in order to reduce such an error, it is conceivable to re-learn the learned model autonomously based on vehicle-specific teacher data acquired during driving. In such a control device that autonomously re-learns a learned model, the learned model is re-learned in accordance with a vehicle-specific characteristic such as, for example, a use situation and use environment of the vehicle. As a result, the characteristic of the vehicle is reflected in the trained model after the re-learning, so that the prediction accuracy of the output parameter in the trained model after the re-learning can be improved. By using the learned model re-learned according to such a characteristic unique to the vehicle, it becomes possible to more appropriately control the internal combustion engine.

  Specifically, the weight and bias of the neural network of the learned model are updated by re-learning the learned model based on the vehicle-specific teacher data acquired during driving. Thereby, the characteristic peculiar to the vehicle is reflected on the weight and the bias in the learned model after the re-learning.

  By the way, in a control device that re-learns such a trained model, it is conceivable to update the trained model based on update data acquired from an external computer by the above-mentioned service campaign or recall. However, in this case, the trained model in which the characteristics unique to the vehicle are reflected by the re-learning up to that time is rewritten to a standard trained model uniformly created by a manufacturer or the like.

For example, the control device standard learned model A ₁ created uniformly is mounted at the factory of the vehicle by the manufacturer or the like, a case of performing re-learning of the standard learned model A _1. The standard learned model A ₁ is used for controlling an internal combustion engine of a vehicle. In such a control device, the standard learned model A ₁ is updated to a vehicle-specific learned model A ₁ ′ in which characteristics unique to the vehicle are reflected by re-learning based on teacher data acquired during driving. .

However, after that, when update data for updating the trained model is acquired by a service campaign, recall, or the like, the trained model A ₁ ′ unique to the vehicle is updated to the standard trained model A ₂ after the function improvement. Is done. At this time, the vehicle-specific learned model A ₁ ′ having the weight and bias of the neural network reflecting the characteristic of the vehicle is changed to the standard learned model A _{2 in} which the weight of the neural network and the like are uniformly set. Rewritten. In order to control the internal combustion engine according to the vehicle-specific characteristics after the update, it is necessary to re-learn from the beginning again based on the vehicle-specific teacher data acquired during driving.

<< Control method in this embodiment >>
Therefore, in the present embodiment, the storage unit 210 stores, in addition to the learned model, a difference model that outputs a difference between an output parameter value of the learned model and an actual value of an output parameter of the learned model as an output parameter. doing. In the present embodiment, this difference model uses a neural network. The learning unit 225 updates the weight and bias of the neural network of the difference model by performing the supervised learning of the difference model. Then, the engine control unit 224 controls the internal combustion engine based on the sum of the output value of the learned model and the output value of the difference model. Thereby, even when the trained model is updated based on the update data obtained from an external computer, the output value of the difference model in which the weight and bias of the neural network have been updated by the learning up to that point is used. Thus, the internal combustion engine 100 can be controlled. As a result, it is possible to control the internal combustion engine 100 immediately after updating the learned model while reflecting the characteristics unique to the vehicle. Hereinafter, the present embodiment will be described in detail.

≪Difference model≫
First, the input parameters and output parameters of the difference model used in the present embodiment will be described. The input parameters of the difference model can include the input parameters described above as input parameters used in the learned model. That is, the input parameters of the difference model are two of the ignition timing, fuel injection amount, fuel injection timing, In-Ex VVT timing, throttle opening, EGR valve opening, intake air temperature, water temperature, oil temperature, and engine speed. One or more can be included. Preferably, in the difference model, for example, the same input parameters as those of the learned model installed in the storage unit 210 when the vehicle is shipped are used. The method of obtaining the values of these input parameters used in the difference model is the same as the method of obtaining the values of the input parameters used in the learned model.

  The output parameter of the difference model is a difference between the output value of the output parameter of the learned model (that is, the value of the output parameter output by the parameter value output unit 221) and the actual value of this output parameter. For example, when the output parameter of the learned model is torque, the difference model, when the value of the input parameter is input, uses a neural network to output the value of the torque output from the learned model and the actual value of the torque. Output the difference from the value.

  FIG. 4 shows a specific example of a neural network in the difference model according to the present embodiment. FIG. 4 shows a case where the learned model shown in FIG. 3 is used as the learned model. The neural network in the difference model shown in FIG. 4 uses the ignition timing, the fuel injection amount, the throttle opening, and the engine speed as input parameters, and calculates the value of the torque output from the learned model and the actual value of the torque. The difference is used as an output parameter. In the present embodiment, the neural network of the difference model is composed of a P layer (P is an arbitrary integer of 3 or more), and the number of nodes in each hidden layer can be an arbitrary number. Also, in the neural network in the difference model shown in FIG. 4, the input layer (L = 1) has four nodes corresponding to four input parameters, but according to the number of input parameters, It may have any number of nodes other than four.

学習 Learning method in difference model≫
Next, an outline of a learning method in the difference model according to the present embodiment will be described. In the present embodiment, the learning unit 225 acquires the output parameter value of the learned model from the parameter value output unit 221. Further, the learning unit 225 acquires the values of the input parameters of the difference model and the actually measured values of the output parameters of the learned model from the output values of the various sensors acquired during driving of the vehicle.

Here, the measured values of the output parameters of the learned model are obtained by the following obtaining method. That is, the output parameter is the exhaust temperature of the learned model, NO _x concentration, HC concentration, CO concentration, CO ₂ concentration, the air-fuel ratio, when a torque or knock determination value, the value of the output parameter, respectively exhaust temperature sensor 25, It is obtained from the output value of the gas concentration sensor 28, the air-fuel ratio sensor 29, the torque sensor 51 or the knocking sensor 52.

  Next, the learning unit 225 acquires the difference between the value of the output parameter of the learned model output by the parameter value output unit 221 and the actually measured value of the output parameter of the learned model as correct data. Then, the learning unit 225 learns the difference model using, as teacher data, a set of the value of the input parameter of the difference model and the correct data at the value of the input parameter. Specifically, for example, the learning unit 225 calculates the square error E as described above using the teacher data. Then, the learning unit 225 can update the difference model in the storage unit 210 by learning the weights and bias values of the neural network of the difference model so that the value of the square error E decreases. As a result, the characteristics unique to the vehicle can be reflected on the weight and bias of the neural network of the difference model.

≫Control of internal combustion engine≫
Next, control of the internal combustion engine executed by the control device for the internal combustion engine according to the present embodiment will be described. FIGS. 5A and 5B are block diagrams illustrating the control of the internal combustion engine according to the present embodiment. In the example shown in FIG. 5 (a), the parameter value output unit 221 uses the learned model A _1, the difference output unit 222 is assumed to use a differential model B.

As shown in FIG. 5 (a), the parameter value output unit 221, the value of the input parameters are entered, and outputs the value a ₁ of the output parameter by using the learned model A _1. When the value of the input parameter is input, the difference output unit 222 outputs a difference b between the value of the output parameter of the trained model and the actual value of the output parameter of the trained model using the difference model B. . When the output value a ₁ of the parameter value output unit 221 and the output value b of the difference output unit 222 are input, the addition output unit 223 and the input output value of the parameter value output unit 221 and the input difference output The sum a ₁ + b with the output value of the section 222 is output. The engine control unit 224 controls the internal combustion engine 100 based on the output value a ₁ + b of the addition output unit 223.

Here, as shown in FIG. 5 (b), the updating unit 226 is learned model A ₁ assume a case that has been updated in the learned model A _2. In this case, as shown in FIG. 5, the parameter value output unit 221, the value of the input parameters are entered, and outputs the value a ₂ of the output parameter by using the learned model A _2. When the value of the input parameter is input, the difference output unit 222 outputs a difference b between the value of the output parameter of the trained model and the actual value of the output parameter of the trained model using the difference model B. . When the output value a ₂ of the parameter value output unit 221 and the output value b of the difference output unit 222 are input, the addition output unit 223 and the input output value of the parameter value output unit 221 and the input difference output The sum a ₂ + b with the output value of the section 222 is output. The engine control unit 224 controls the internal combustion engine 100 based on the output value a ₂ + b of the addition output unit 223.

  As described above, in the present embodiment, even when the learned model is updated by the updating unit 226, the difference model is not changed. Further, in the present embodiment, the engine control unit 224 uses the output value of the updated trained model and the output value of the difference model in which the characteristics unique to the vehicle are reflected by the learning by the learning unit 225, and uses the output value of the internal combustion engine 100. Can be controlled. Thus, even when the learned model is updated, the weight and bias of the neural network obtained by learning the difference model up to that time can be reflected in the control of the internal combustion engine 100.

<Second embodiment>
FIG. 6 is a schematic configuration diagram of a control system for an internal combustion engine according to the second embodiment. FIG. 6 shows a vehicle 310 and a server 320 installed outside the vehicle 310. As shown in FIG. 6, the vehicle 310 includes a communication unit 311 and an ECU 312. The ECU 312 has the same configuration as the ECU 200 according to the first embodiment, except that the ECU 312 does not include the learning unit 225.

  In the present embodiment, the ECU 312 of the vehicle 310 acquires, as correct data, a difference between the value of the output parameter of the learned model and the actually measured value of the output parameter of the learned model, for each value of the input parameter of the difference model. Learning data including the input parameters of the difference model and the corresponding correct data is transmitted to the server 320 via the communication unit 311.

  6, the server 320 includes a communication unit 321, a parameter value acquisition unit 322, a data set creation unit 323, and a learning unit 324. The communication unit 321 of the server 320 is configured to be able to communicate with the communication unit 311 of the vehicle 310.

  The parameter value acquiring unit 322 of the server 320 receives the learning data transmitted from the vehicle 310 via the communication unit 321 and acquires the values of the input parameters and the correct answer data in the learning data. The data set creation unit 323 of the server 320 creates a data set indicating the relationship between the value of the input parameter acquired by the parameter value acquisition unit 322 and the corresponding correct data. The learning unit 324 learns the weight of the neural network in the difference model by using a set of the input parameters of the difference model and the corresponding correct data as teacher data according to the data set created by the data set creation unit 323. I do. Next, the learning unit 324 of the server 320 transmits the learned difference model to the vehicle 310 via the communication unit 321.

  The ECU 312 of the vehicle 310 receives the difference model transmitted from the server 320 via the communication unit 311, and rewrites the difference model stored in the storage unit 210 of the ECU 312 with the received difference model. Update the weight of. Engine control unit 224 of ECU 312 of vehicle 310 controls internal combustion engine 100 based on the output value of addition output unit 223 that outputs the sum of the output value of the learned model and the output value of the updated difference model. .

  According to the present embodiment, the server 320 learns the weight and bias of the neural network of the difference model. Therefore, it is not necessary to provide the ECU 312 of the vehicle 310 with a high-performance arithmetic processing device for learning the weight of the neural network in the difference model. As a result, the cost of the vehicle can be reduced.

  In the above-described embodiment, an example in which a neural network is used as the learned model and the difference model has been described, but another machine learning model such as a decision tree may be used. Further, the method of obtaining the values of the respective parameters described above is merely an example, and the values of the respective parameters can be obtained by other methods.

Reference Signs List 1 engine body 10a throttle opening sensor 18a EGR valve opening sensor 24 intake temperature sensor 25 exhaust temperature sensor 26 water temperature sensor 27 oil temperature sensor 28 gas concentration sensor 29 air-fuel ratio sensor 100 internal combustion engine

Claims (1)

When a value of the input parameter is input, a parameter value output unit that outputs the value of the output parameter using the learned model,
When the value of the input parameter is input, using a difference model, a difference output unit that outputs a difference between the value of the output parameter output by the parameter value output unit and the actual value of the output parameter,
When an output value of the parameter value output unit and an output value of the difference output unit are input, a sum of the input output value of the parameter value output unit and the output value of the input difference output unit is output. An addition output section,
An engine control unit that controls the internal combustion engine based on the output value of the addition output unit,
The difference between the value of the output parameter output by the parameter value output unit and the measured value of the output parameter is obtained as correct data, and a set of the value of the input parameter of the difference model and the correct data is used as teacher data. A learning unit for learning the difference model;
An update unit that obtains update data for updating the learned model from a computer installed outside the vehicle, and updates the learned model based on the update data.
A control device for an internal combustion engine, comprising:

Images