KR20220149222A - Apparatus and method for predicting engine pressure - Google Patents

Abstract

The present invention is to provide an apparatus and a method for predicting an engine pressure, which can estimate the pressure of an engine without using an expensive pressure sensor. A pressure predicting method of an engine pressure predicting apparatus according to the present invention comprises a step of configuring input/output data between an engine block vibration value and a cylinder pressure value for a crank angle; and a step of defining a loss function for a combustion unit to learn ranges of the crank angle.

Description

Engine pressure prediction device and method {APPARATUS AND METHOD FOR PREDICTING ENGINE PRESSURE}

The present invention relates to an engine pressure prediction apparatus and method.

In general, an engine is an engine that converts thermal energy into mechanical work. An engine is a reciprocating engine that moves a piston by igniting and detonating a mixture of fuel and air in a cylinder. The engine includes a cylinder block and a cylinder head coupled to each other to form a combustion chamber. If the performance of the engine can be monitored in real time, the engine can be controlled stably. And in order to monitor the performance of the engine in real time, the temperature, rotation speed, or pressure of various parts constituting the engine is measured. If the pressure of the combustion chamber can be monitored in real time, it is possible to stably control a knocking phenomenon, which is a representative unstable element of the engine. Knocking phenomenon is when a mixture of fuel and air is compressed, and when the mixture is ignited and combusted before the appropriate explosion time, unburned gas is generated, and the generated unburned gas spontaneously ignites and burns explosively. say metal When the knocking phenomenon occurs, the piston, cylinder block, valve, etc. may be damaged, and thus the output may be reduced and the lifespan may be shortened. However, when the pressure of the combustion chamber is high as in a marine engine, an expensive pressure sensor capable of measuring the high pressure must be used, so that the cost increases. And when the pressure of the combustion chamber is high, since the pressure sensor can be easily damaged, the replacement cycle is short. For this reason, there is a disadvantage in that the cost is further increased.

Japanese Patent Laid-Open No. 2020-186683 Learning apparatus and estimation apparatus Korean Patent Application Laid-Open No. 10-2019-0089004 Apparatus and method for diagnosing engine health

An object of the present invention is to provide an engine pressure prediction apparatus and method for estimating engine pressure without using an expensive pressure sensor.

A pressure prediction method of an engine pressure prediction apparatus according to an embodiment of the present invention comprises: configuring input/output data between a vibration value of an engine block with respect to a crank angle and a pressure value of a cylinder; and defining a loss function for learning of the combustion unit for the section of the crank angle.

In an embodiment, configuring the input/output data may include: acquiring the vibration value of the engine block from an acceleration sensor; obtaining a pressure value of the cylinder from a pressure sensor; and obtaining angle information using the crank angle.

In an embodiment, the step of configuring the input/output data comprises synchronizing a time-based first vibration signal and a first pressure signal with an encoder signal of the crank angle, and a second vibration signal and a second vibration signal based on the crank angle and cycle. It may further comprise expanding into a two-dimensional matrix with a pressure signal.

In an embodiment, the configuring of the input/output data may further include learning N (natural number) dimensional data of the crank angle using the vibration value and the pressure value arranged in N input neurons. .

In an embodiment, the defining of the loss function may include learning the combustion pressure inside the cylinder by applying a model weight learned model to the section of the crank angle.

In an embodiment, the step of defining the loss function comprises: calculating a weight of the loss function for each output node; and calculating an error between the target value multiplied by the weight and the output value node.

In an embodiment, the method may further include determining whether a learning stop condition is satisfied.

In an embodiment, the method may further include updating the learning model when the learning stop condition is not satisfied.

An engine pressure prediction apparatus according to an embodiment of the present invention includes an acceleration sensor for measuring an engine vibration value; a memory device for storing the training model; and an artificial neural network learner that receives the vibration value and the learning model and predicts the pressure value of the engine by performing artificial neural network learning, wherein the learning model includes: the vibration value of the engine block with respect to the crank angle and the pressure of the cylinder compose input and output data between values; And it is characterized in that it is generated by defining a loss function for learning the combustion unit for the section of the crank angle.

The engine pressure prediction apparatus and method according to an embodiment of the present invention may estimate the pressure value of the vibration value of the acceleration sensor through artificial neural network learning.

The accompanying drawings below are provided to help understanding of the present embodiment, and provide embodiments together with detailed description.
1 is a schematic diagram of a typical engine.
2 is a diagram illustrating an engine pressure prediction apparatus 200 according to an embodiment of the present invention.
3 is a flowchart exemplarily illustrating a method of predicting an engine pressure according to an embodiment of the present invention.

Hereinafter, the contents of the present invention will be described clearly and in detail using the drawings to the extent that those of ordinary skill in the art can easily implement it.

1 is a schematic diagram of a typical engine. Referring to FIG. 1 , the engine may include a cylinder block 110 and a cylinder head 120 coupled to each other. The combustion chamber 130 may be formed by the cylinder block 110 and the cylinder head 120 . In more detail, most of the space of the combustion chamber 130 may be formed in the cylinder block 110 . The cylinder head 120 may be coupled to one surface of the cylinder block 110 to seal one side of the combustion chamber 130 . A valve mechanism for opening and closing the intake passage and the exhaust passage may be installed in the cylinder head 120 . Also, a water jacket or spark plug serving as a coolant passage may be installed in the cylinder head 120 .

The cylinder block 110 and the cylinder head 120 may be coupled by a plurality of coupling members 140 . That is, one end side of the coupling member 140 may be supported and fixed to the cylinder block 110 , and the other end side may be supported and fixed to the cylinder head 120 .

A cylinder liner 150 may be installed on the inner circumferential surface of the cylinder block 110 forming the combustion chamber 130 . A piston 161 may be installed on the cylinder liner 150 to enable linear reciprocating motion. The piston 161 may be inserted into the cylinder liner 150 through the other side of the combustion chamber 130 . The other side of the combustion chamber 130 may be sealed by the piston 161 . As the mixed gas in which fuel and air are mixed is introduced into the combustion chamber 130 formed between the piston 161 and the cylinder head 120 and is compressed and exploded, the piston 161 may reciprocate linearly. The crankshaft 165 connected to the piston 161 may rotate through the connecting rod by the linear reciprocating motion of the piston 161 . Most of the rotational force of the crankshaft 165 is transmitted to the power transmission unit side through the flywheel, and a part of the rotational force may be transmitted to a valve, a pump, a generator, or a fuel supply device.

In general, when monitoring the pressure of the combustion chamber 130 in real time, the control of the engine including the knocking phenomenon may be stable.

Recently, in order to replace an expensive piezo type pressure sensor, research has been actively conducted by measuring cylinder and engine block vibration using a low-cost accelerometer sensor. For example, a technology for predicting the maximum combustion chamber pressure by learning the knock sensor measurement result predicts the cylinder pressure using an accelerometer sensor. In other words, it is a technology to obtain cylinder pressure as a transfer function using an accelerometer sensor. In addition, a technique for predicting the pressure inside the engine cylinder through the correlation between the vibration of the engine block and the pressure sensor using a neural network algorithm is also introduced.

In the implementation of cylinder pressure in the engine cycle, it is not clear to secure a high correlation between vibration and pressure due to errors in the transfer function and combustion stability. In order to construct a transfer function between the pressure sensor and the acceleration sensor, an artificial neural network is used for alternative measurement. However, unlike time series data of an artificial neural network that learns by sampling changes in time flow in general, engine cycles have a repetitive pattern based on angles. There is a possibility of overfitting for the combustion core section where the similarity is not high in all sections and the section in which combustion does not occur among the sections in which combustion does not occur with high similarity in all pattern sections. This makes it difficult to train the model by selecting the interval.

In the engine pressure prediction method according to an embodiment of the present invention, an artificial neural network (eg, Deep Neural Network (DNN)) algorithm using data of a specific angular range for input/output neurons using an acceleration sensor of an engine cylinder is configured, It is possible to secure high correlation for each angle and secure robustness to accuracy.

The engine pressure prediction method according to an embodiment of the present invention includes preprocessing of an angular range through synchronization of an engine speed signal for arranging on an angular basis, and constructs a loss function for ensuring accuracy of a combustion unit having a high pressure. can The present invention can secure the accuracy of engine pressure prediction by predicting the pressure value for a specific angle in repeated engine cycles using the information of the loss function.

2 is a diagram illustrating an engine pressure prediction apparatus 200 according to an embodiment of the present invention. Referring to FIG. 2 , the engine pressure prediction apparatus 200 may include an artificial neural network learner 210 and a memory device 220 .

The artificial neural network learner 210 may be implemented to receive a pressure prediction model and a vibration value, and to estimate a pressure value corresponding to the vibration value based on artificial neural network learning. Here, the pressure prediction model may be generated by learning the transfer functions of the pressure sensor and the acceleration sensor. In an embodiment, the crank angle may be measured from an acceleration sensor related to the engine. In an embodiment, the artificial neural network learner 210 may be implemented in hardware, software, or firmware.

In general, input data for prediction is input to a model through an input layer, and target data to be predicted is calculated as an output layer. At least one hidden layer exists between the input layer and the output layer. The hidden layer plays a key role in building a nonlinear model. Each layer consists of a plurality of nodes. The connection between each node is made by a multi-layer perceptron (MLP) of perceptrons corresponding to nerve cells of the artificial neural network. Each perceptron receives a signal from each node in the previous layer, applies a weight and a bias to add it, and then determines the activation degree of each node through an activation function. After that, the neural network is constructed by passing the calculated final result as an input signal to the node of the next layer. By comparing the result calculated through this series of processes and the measured actual value (label), the model is updated by itself in a direction to minimize the error with the actual value. This learning method is called supervised learning. Since the purpose of the cylinder pressure prediction model is to accurately predict the actual pressure, model learning proceeds through supervised learning.

In general, the accuracy of a model built through supervised learning is related to the similarity with the configuration of the training condition and the test condition, the accuracy of the training data, the configuration of the model and the preprocessing method of the data, and the optimization. It can be varied by algorithm and learning rate. Therefore, in order to develop an accurate model, a quantitative analysis of the effect of the above-described variables on the prediction accuracy of the model is required. In particular, since the prediction accuracy of the test condition by the change of the learning condition is greatly affected by the nonlinearity of the variables constituting the condition, the learning condition should be configured for a close examination.

The memory device 220 may be implemented to store a pressure prediction model. In an embodiment, the memory device 220 may include a nonvolatile memory device. Here, the nonvolatile memory device includes a NAND flash memory, a vertical NAND flash memory, a NOR flash memory, a resistive random access memory (RRAM), and a phase-change memory. PRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), spin transfer torque random access memory (STT-RAM), and the like.

3 is a flowchart exemplarily illustrating a method of predicting an engine pressure according to an embodiment of the present invention. Referring to FIG. 3 , the method of predicting the pressure of the engine may proceed as follows.

Input/output data for using the DNN of engine block vibration may be configured (S110). In order to configure input/output data, engine block vibration (Vibblock) generated during engine combustion may be measured from an acceleration sensor (Accelerometer sensor) capable of measuring vibration in the engine block. For artificial neural network learning, the cylinder pressure (Pcyl) may be measured with a cylinder pressure sensor (S111). The measurement position of the engine block vibration can be adopted in consideration of the correlation to the pressure signal according to the acquisition environment.

Pcyl(출력)의 DNN 학습이 수행될 수 있다(S112). 이러한 학습을 통하여 엔진의 크랭크 각도 마다 수행되는 진동에 의한 가중치를 합산함으로써, 각도 기준의 실린더 압력을 재구성하는 DNN 모델이 생성될 수 있다.Synchronized with the encoder signal of the crank angle of the engine, the time-based signal first vibration signal (Vibblock--(t)), the first pressure signal (P-cyl-(t)) is the crank angle [-360,360] and cycle ( A signal based on the number of cycles) may be expanded into a two-dimensional matrix [N x M] with a second input signal (Vibblock(θ, number of cycles)) and a second pressure signal (Pcyl(θ, number of cycles)). The N-dimensional values mean the size of the angle vector selected by varying the importance of resolution and angle according to the user's criteria. The N-dimensional values can be displayed on a constant angular basis according to the resolution of the obtained Vibblock (input) and Pcyl (output) angles. The M-dimensional values mean a number obtained by dividing data synchronized with the obtained crank angle by the number of cycles. Since the M-dimensional values are the number of selected cycle data samples to be used for training the DNN model, the data acquired according to the experimental conditions of the engine may be selected and used to prevent noise and overfitting. Crank angle (θ) N-dimensional data input Vibblock (input) using Vibblock and Pcyl placed on N neurons

DNN learning of Pcyl (output) may be performed (S112). Through this learning, a DNN model that reconstructs the cylinder pressure based on the angle can be generated by summing the weights by vibrations performed for each crank angle of the engine.

From the above-described step S110, when the number of engine cylinders and the combustion timing change, the correlation between the engine block vibration for a specific angle and the data inside the cylinder pressure can be secured without changing the artificial neural network model.

The generated DNN model may be updated in the pressure prediction device (S120).

Thereafter, a loss function for constructing a DNN for main learning of the combustion unit corresponding to the high pressure may be generated (S130). The in-cylinder combustion pressure can be learned through the application of the DNN model learned by the model weight to the section of the angular range. In this case, a weight loss function for each output node may be calculated (S131). Here, the weight of the loss function is calculated as follows.

이며, m번째 표본의 가중치는 각도가

이며, m번째 목표 데이터값에서, m번째 사이클 내 목표 압력 데이터의 최대값으로 나눈 값이다.Here, the angle in the target pressure data is

, and the weight of the mth sample is

, and is a value obtained by dividing the m-th target data value by the maximum value of the target pressure data in the m-th cycle.

The calculated weight matrix is multiplied by each element of the predicted value matrix and the target value matrix of DNN, and is used for commonly used loss statistical functions such as MSE (Mean Squared Error), MAE (Mean Absolute Error), RMSE (Root Mean Squared Error), etc. It is applicable (S132). Through this, the learning error of the combustion unit having a high pressure is further increased, so that the learning amount can be calculated.

The accuracy of the maximum pressure value Pmax, which is one of the main indicators of the cylinder pressure, and the pressure value for the main combustion part of the Pmax value will be secured by improving the model contribution to the combustion part angular range, which is a section with a relatively large error in the above-described step S130. can

Thereafter, it may be determined whether the learning stop condition is satisfied (S140). If the learning stop condition is satisfied, the pressure prediction method will be terminated. In general, when all learning of EPOCH, which means the number of times the entire data is learned once, is completed, or when an early stop condition is specified due to an error in the verification data, the learning may be stopped. In an embodiment, when EPOCH = 100 is specified and all 100 learning times are completed, the learning stop condition may be satisfied. In another embodiment, although EPOCH = 100 is specified, the learning stop condition may be satisfied when the error increase of the verification data is 5 or more times due to the early stop condition = 5. On the other hand, if the learning stop condition is not satisfied, step S120 will be entered.

On the other hand, the contents of the present invention described above are only specific examples for carrying out the invention. The present invention will include not only concrete and practically usable means, but also technical ideas, which are abstract and conceptual ideas that can be utilized as future technologies.

200: engine pressure prediction device
210: artificial neural network learner
220: memory device

Claims (9)

In the pressure prediction method of the engine pressure prediction device,
composing input/output data between the vibration value of the engine block and the pressure value of the cylinder with respect to the crank angle; and
and defining a loss function for learning of the combustion unit for the section of the crank angle. The method of claim 1,
The step of composing the input/output data includes:
obtaining the vibration value of the engine block from an acceleration sensor;
obtaining a pressure value of the cylinder from a pressure sensor; and
and obtaining angle information with the crank angle. 3. The method of claim 2,
The step of composing the input/output data includes:
Expanding a time-based first vibration signal and a first pressure signal into a two-dimensional matrix with a second vibration signal and a second pressure signal based on the crank angle and cycle by synchronizing with the encoder signal of the crank angle Way.
4. The method of claim 3,
The step of composing the input/output data includes:
The method further comprising the step of learning N (natural number) dimensional data of the crank angle using the vibration value and the pressure value arranged in N input neurons. The method of claim 1,
Defining the loss function comprises:
and learning the combustion pressure inside the cylinder by applying a model weight learned model to the section of the crank angle. 6. The method of claim 5,
Defining the loss function comprises:
calculating a weight of the loss function for each output node; and
and calculating an error between a target value multiplied by the weight and an output value node. The method of claim 1,
The method further comprising the step of determining whether the learning stop condition is satisfied. 8. The method of claim 7,
The method further comprising the step of updating the learning model when the learning stop condition is not satisfied. an acceleration sensor that measures the vibration value of the engine;
a memory device for storing the training model; and
and an artificial neural network learner that receives the vibration value and the learning model and predicts the pressure value of the engine by performing artificial neural network learning,
The learning model comprises input/output data between the vibration value of the engine block and the pressure value of the cylinder with respect to the crank angle; and engine pressure learning apparatus, characterized in that it is generated by defining a loss function for learning of the combustion unit for the section of the crank angle.

Images