JP6914084B2 - Knock control device, knock matching method and knock matching program - Google Patents

Description

The present invention relates to a knock controller, a knock adaptation method and a knock adaptation program.

Conventionally, there is a knock control device that detects knocking that occurs in an internal combustion engine such as a vehicle engine and controls the ignition timing of the engine based on the knocking detection state. The knock control device determines the presence or absence of knocking based on a knock model that models the characteristics of the waveform shape of the knock signal indicating knocking (see, for example, Patent Document 1).

Japanese Unexamined Patent Publication No. 2009-203842

However, in the conventional technique, since the process of creating the knock model is entrusted to the operator, the matching accuracy of the generated knock model may decrease. Therefore, when the presence or absence of knocking is determined using such a knock model, the knocking detection accuracy may decrease.

The present invention has been made in view of the above, and an object of the present invention is to provide a knock control device, a knock matching method, and a knock matching program capable of improving the detection accuracy of knocking.

In order to solve the above-mentioned problems and achieve the object, the knock control device according to the present invention includes an acquisition unit, an image generation unit, a model generation unit, and a determination unit. The acquisition unit acquires a knock signal whose waveform changes according to knocking generated in the internal combustion engine. The image generation unit generates a knock image by converting the knock signal acquired by the acquisition unit into a frequency domain at regular intervals. The model generation unit generates a knock model showing the knocking characteristics based on the knock image generated by the image generation unit. The determination unit determines the presence or absence of the knocking based on the knock model generated by the model generation unit.

According to the present invention, the knocking detection accuracy can be improved.

FIG. 1 is a diagram showing an outline of processing performed by the knock control device according to the embodiment. FIG. 2 is a block diagram showing a configuration of a knock control device according to an embodiment. FIG. 3 is a diagram showing the processing contents of the model generation unit. FIG. 4 is a diagram showing the processing content of the weight value determination process. FIG. 5 is a diagram showing the processing contents of the first processing. FIG. 6 is a diagram showing the processing contents of the second processing. FIG. 7 is an explanatory diagram of the classifier information. FIG. 8 is a flowchart showing a processing procedure of the conforming process executed by the knock control device according to the embodiment. FIG. 9 is a flowchart showing a processing procedure of the determination process executed by the knock control device according to the embodiment. FIG. 10 is a block diagram showing a configuration of a knock control device according to a modified example of the embodiment. FIG. 11 is an explanatory diagram of the classifier information according to the modified example. FIG. 12 is a block diagram showing a configuration of a knock control device according to a modified example of the embodiment.

Hereinafter, embodiments of the knock control device, the knock adaptation method, and the knock adaptation program disclosed in the present application will be described in detail with reference to the accompanying drawings. The present invention is not limited to this embodiment.

First, the outline of the processing performed by the knock control device according to the embodiment will be described with reference to FIG. FIG. 1 is a diagram showing an outline of processing performed by the knock control device 1 according to the embodiment. FIG. 1 shows an internal combustion engine 10 and a knock sensor 11 that detects knocking of the internal combustion engine 10.

The internal combustion engine 10 is, for example, an engine of a vehicle or the like, and is a device that obtains power from combustion gas generated by burning fuel such as gasoline in a cylinder. The knock sensor 11 is a sensor that detects vibration generated by driving the internal combustion engine 10, and outputs a knock signal whose waveform changes according to knocking generated in the internal combustion engine 10. Note that knocking refers to a state in which abnormal vibration or abnormal noise is generated due to abnormal combustion of fuel residue remaining in the cylinder after combustion, for example, at the next ignition timing.

Further, FIG. 1 shows knock images Pa and Pb obtained by converting a knock signal into a frequency domain at regular intervals (for example, every cycle). The knock image Pa is used for the knock model generation process of the knock control device 1, and is, for example, an image of a sensor signal obtained during an operation test of the internal combustion engine 10.

Further, the knock image Pb is used for the knocking determination process of the knock control device 1, and is, for example, an image of a sensor signal detected from the internal combustion engine 10 when the vehicle actually travels on a public road or the like. In the following, the knock image Pa may be described as a learning image Pa, and the knock image Pb may be described as a determination image Pb.

The knock control device 1 according to the embodiment uses the learning image Pa to generate a knock model showing the characteristics of knocking, and uses the knock model to determine the presence or absence of knocking in the determination image Pb.

Here, the conventional knock control device will be described. In the conventional knock control device, the process of creating the knock model is entrusted to the operator. Specifically, conventionally, the operator has extracted the characteristics of the waveform shape of the knock signal from the knock image and used it for creating the knock model.

For this reason, conventionally, the knocking characteristics extracted by the skill level of the operator and the like vary, and there is a risk that the matching accuracy of the knock model may decrease. Therefore, conventionally, when the presence or absence of knocking is determined using such a knock model, the knocking detection accuracy may decrease. In addition, since the knocking feature must be extracted from each knock image, the man-hours of the operator increase.

Therefore, the knock control device 1 according to the embodiment automatically generates a knock model from the knock image Pa (learning image Pa) by machine learning or the like.

Specifically, as shown in FIG. 1, the knock control device 1 according to the embodiment first generates a knock model showing knocking characteristics in advance based on the knock image Pa in the learning phase (step S1). The knock model can be generated by using, for example, a discriminator having an algorithm of a convolutional neural network, and this point will be described later with reference to FIGS. 3 to 6.

Subsequently, the knock control device 1 according to the embodiment shifts to the identification phase after generating the knock model, and determines the presence or absence of knocking based on the knock model generated in the learning phase (step S2). Specifically, as shown in FIG. 1, the knock control device 1 first applies the generated knock model to the classifier. Subsequently, the knock control device 1 determines the presence or absence of knocking in the determination image Pb using such a classifier. The algorithm of the classifier used in the discriminating phase is, for example, a convolutional neural network, and the same algorithm used in the learning phase (applying the knock model) is used.

That is, the knock control device 1 according to the embodiment automatically generates a knock model by machine learning or the like, so that the knock model does not vary from worker to worker as in the conventional case, so that the matching accuracy of the knock model is improved. be able to. That is, according to the knock control device 1 according to the embodiment, the knocking detection accuracy can be improved.

Further, since the knock control device 1 according to the embodiment extracts the knocking feature by machine learning, the man-hours of the operator can be reduced. That is, it is possible to improve the efficiency of the fitting work of the knock model.

The knock control device 1 according to the embodiment generates a weight value ω of a discriminator using a convolutional neural network as a knock model, and this point will be described later with reference to FIGS. 3 to 6.

Further, the knock control device 1 according to the embodiment can generate a knock model for each cylinder of the internal combustion engine 10 by using, for example, a CPS (Cylinder Pressure Sensor). Regarding this point, FIGS. 10 and 11 Will be described later using.

Next, the configuration of the knock control device 1 according to the embodiment will be described in detail with reference to FIG. FIG. 2 is a block diagram showing a configuration of the knock control device 1 according to the embodiment.

As shown in FIG. 2, the knock control device 1 according to the embodiment is connected to the knock sensor 11 and the control device 100. In FIG. 2, the knock control device 1 and the control device 100 are separately configured as two devices, but the knock control device 1 and the control device 100 are integrally configured as one device. May be done. First, configurations other than the knock control device 1 will be described.

The knock sensor 11 is attached to the internal combustion engine 10 and generates a knock signal based on, for example, the vibration generated in the internal combustion engine 10. The knock signal is a signal whose waveform changes according to knocking generated in the internal combustion engine 10. The knock sensor 11 outputs the generated knock signal to the knock control device 1.

As the knock sensor 11, for example, a microphone can be used. In such a case, the microphone generates a knock sound generated by knocking as a knock signal. As a result, a knock model based on the knock sound can be generated, so that a knocking determination that captures the human sense of hearing becomes possible.

The control device 100 calculates the engine rotation speed and air load factor (filling efficiency) of the internal combustion engine 10 based on a sensor (not shown) attached to the internal combustion engine 10, and outputs the information to the knock control device 1. .. In FIG. 2, the engine speed, the air load factor (filling efficiency), and the like are output from the control device 100. For example, the detection signal of the sensor is directly output to the knock control device 1 and the engine rotation is performed by the knock control device 1. The speed and air load factor (filling efficiency) may be calculated.

The control device 100 determines the ignition timing of the internal combustion engine 10 from the next time onward based on the determination result of the presence or absence of knocking notified from the knock control device 1, and instructs the internal combustion engine 10 to ignite according to the ignition timing. Outputs an ignition signal.

Next, the knock control device 1 according to the embodiment will be described. The knock control device 1 includes a control unit 2 and a storage unit 3. The control unit 2 includes an acquisition unit 21, an image generation unit 22, a model generation unit 23, and a determination unit 24. The storage unit 3 stores the image information 31 and the classifier information 32.

Here, the knock control device 1 is, for example, a computer having a CPU (Central Processing Unit), a ROM (Read Only Memory), a RAM (Random Access Memory), an HDD (Hard Disk Drive), an input / output port, and various circuits. including.

The CPU of the computer functions as the acquisition unit 21, the image generation unit 22, the model generation unit 23, and the determination unit 24 of the control unit 2, for example, by reading and executing the program stored in the ROM.

Further, at least one or all of the acquisition unit 21, the image generation unit 22, the model generation unit 23, and the determination unit 24 of the control unit 2 are ASIC (Application Specific Integrated Circuit), FPGA (Field Programmable Gate Array), GPU ( It can also be configured with hardware such as Graphics Processing Unit).

Further, the storage unit 3 corresponds to, for example, a RAM or an HDD. The RAM or HDD can store image information 31, classifier information 32, information on various programs, and the like. The knock control device 1 may acquire the above-mentioned program and various information via another computer or a portable recording medium connected by a wired or wireless network.

The control unit 2 uses the knock image Pa (learning image Pa) to generate a knock model showing the characteristics of knocking, and uses the knock model to determine the presence or absence of knocking in the knock image Pb (determination image Pb). do.

The acquisition unit 21 acquires a knock signal whose waveform changes according to the knocking generated in the internal combustion engine 10. The acquisition unit 21 outputs the acquired knock signal to the image generation unit 22.

The image generation unit 22 generates knock images Pa and Pb by converting the knock signal acquired by the acquisition unit 21 into a frequency domain at regular intervals. Specifically, the image generation unit 22 generates the learning image Pa and the determination image Pb by wavelet transforming the knock signal into the frequency domain for each cycle of the internal combustion engine 10. The learning image Pa and the determination image Pb are, for example, color images represented by three dimensions of RGB values.

Further, the image generation unit 22 is not limited to the wavelet transform, and for example, a short-time Fourier transform or a mel frequency cepstrum coefficient may be used, and any method may be used as long as it can be converted into a frequency domain at regular intervals.

Further, the image generation unit 22 first generates the learning image Pa used by the model generation unit 23, which will be described later, out of the learning image Pa and the determination image Pb. Then, the image generation unit 22 generates the determination image Pb after the knock model is generated by the model generation unit 23.

The image generation unit 22 may generate a learning image Pa by using a knock signal due to vibration when the internal combustion engine 10 is experimentally driven in a laboratory, for example. At this time, the knock signal when knocking is intentionally generated in the internal combustion engine 10 may be used. As a result, the accuracy of the knock model generated by the model generation unit 23, which will be described later, can be improved.

Further, the image generation unit 22 may use a knock signal detected from the internal combustion engine 10 in which the vehicle is actually traveling. The image generation unit 22 stores the generated learning image Pa as image information 31 in the storage unit 3, and outputs the determination image Pb to the determination unit 24.

The model generation unit 23 generates a knock model showing knocking characteristics based on the knock image Pa (learning image Pa) generated by the image generation unit 22. Specifically, the model generation unit 23 generates a weight value ω (see FIG. 5) of a discriminator using a convolutional neural network as a knock model. Here, the processing contents of the model generation unit 23 will be specifically described with reference to FIGS. 3 and 4.

FIG. 3 is a diagram showing the processing contents of the model generation unit 23. FIG. 4 is a diagram showing a determination process of the weight value ω of the classifier. As shown in FIG. 3, the model generation unit 23 performs an averaging process and a weight value determination process when a plurality of (for example, several thousand) learning images Pa are input, thereby performing a classifier using a convolutional neural network. Update the weight value ω of.

The model generation unit 23 may acquire all of a plurality of (for example, several thousand) learning image Pas at once and perform processing. For example, the model generation unit 23 may acquire a plurality of learning image Pas little by little in a repeating loop (mini-batch processing). ) May.

First, the averaging process will be described. The averaging process is a process of generating an average value image obtained by averaging the values of each pixel of the learning image Pa. Specifically, first, the model generation unit 23 sets an arbitrary pixel of the learning image Pa as a central pixel, and sets a small region (for example, a 3 × 3 region) including peripheral pixels in the central pixel. The model generation unit 23 plots the average value of a plurality of pixels included in the set small area on the average value image as the pixel value of the central pixel.

Then, the model generation unit 23 generates an average value image by similarly calculating the average value for all the pixels of the learning image Pa. Then, the model generation unit 23 subtracts and normalizes the average value image in accordance with the subsequent processing, and passes it to the weight value determination processing. The normalization of the average value image is not limited to the subtraction normalization, and other normalization patterns such as division normalization and whitening may be used.

Next, the weight value determination process will be described. When the normalized average value image is input, the weight value determination process performs the initial value setting process, the forward propagation process, the loss error calculation process, and the back propagation process in this order. This is the process of updating the weight value.

Specifically, the model generation unit 23 calculates the loss error using the discriminator of the convolutional neural network, and also performs the process of determining the weight value ω that minimizes the loss error. The algorithm of the discriminator is not limited to the convolutional neural network, and may be another machine learning algorithm such as a recurrent neural network. Here, the weight value determination process performed by the model generation unit 23 will be specifically described with reference to FIGS. 4 to 6.

FIG. 4 is a diagram showing a weight value determination process. FIG. 4 shows the flow of the weight value determination process performed by the model generation unit 23. As shown in FIG. 4, the weight value determination process is roughly classified into a first process and a second process.

Further, as shown in FIG. 4, when the average value image is input, the model generation unit 23 first receives the weight value ω (see FIG. 5) and the bias b (see FIG. 5) of the classifier before the first processing. ) Is initialized, and the initial value setting process (see FIG. 3) is performed.

The model generation unit 23 may use a predetermined value as the initial value in the initial value setting process, and may determine the initial value by performing pretraining or fine tuning, for example.

The model generation unit 23 performs the first process after the initial value setting process. The first treatment is a treatment in which a set of a convolution layer, an activation function, and a pooling layer is set as one layer, and a plurality of layers (three layers in FIG. 4) are repeated. The details of the first process will be described later with reference to FIG.

Subsequently, the model generation unit 23 performs the second process after the first process. The second process is a process including a fully connected layer, an activation function, a fully connected layer, and a softmax function. The details of the second process will be described later with reference to FIG.

As shown in FIG. 4, the model generation unit 23 first calculates the loss error by performing the first process and the second process in this order by forward propagation. Subsequently, the model generation unit 23 determines the weight value ω that minimizes the loss error by performing the back propagation process based on the loss error calculated by the forward propagation.

Here, the first process and the second process will be specifically described with reference to FIGS. 5 and 6. FIG. 5 is a diagram showing the processing contents of the first processing. FIG. 6 is a diagram showing the processing contents of the second processing.

First, the first process will be described with reference to FIG. Note that FIG. 5 shows the processing contents of the first layer in the first processing. Further, in FIG. 5, the number of pixels of the average value image is 256 × 256. The number of pixels of the average value image shown in FIG. 5, the convolution image described later, and the pooling image is an example, and may be any number of pixels.

As shown in FIG. 5, the model generation unit 23 generates a convolution image in which the number of pixels is compressed by inputting the average value image in the order of the convolution layer and the activation function. Specifically, the model generation unit 23 compresses the average value image using a convolution filter having a predetermined number of pixels (for example, 2 × 2), a weight value ω, and a bias b.

For example, as shown in FIG. 5, the model generation unit 23 extracts the feature amounts of predetermined 4 pixels x1 to x4 in the average value image by the convolution filter, and the weight value ω for each feature amount of each pixel x1 to x4. Is integrated. Subsequently, the model generation unit 23 sums the values of the pixels x1 to x4 in which the weight values ω are integrated, and adds the bias b to the summed values.

The model generation unit 23 substitutes the value to which the bias b is added into the activation function, and sets the value obtained from the activation function as the value of pixel y1 of the convolution image (hereinafter, may be referred to as a convolution value). decide. The activation function can be, for example, a ramp function, but is not limited to this, and for example, a sigmoid function, a max-out function, and a drop-out function can be used.

Subsequently, the model generation unit 23 generates a pooling image in which the number of pixels is further compressed by inputting the generated convolution image into the pooling layer. Specifically, the model generation unit 23 generates a pooling image by scanning the entire convolution image using a pooling filter having a predetermined number of pixels (for example, 2 × 2).

For example, as shown in FIG. 5, the model generation unit 23 extracts predetermined 4 pixels y1 to y4 in the convolution image based on the pooling filter, and among the extracted 4 pixels, the maximum value of the convolution value is the pooling image. It is determined as the value of pixel z1 in.

The model generation unit 23 used the maximum value of the convolution value as the value of pixel z1 in the pooling image among the extracted plurality of pixels (4 pixels in FIG. 5). The average value may be the value of pixel z1.

Thereby, for example, the influence of the deviation of the detected value of the knock sensor 11 and the deviation of the frequency due to noise superimposition can be removed in the pooling image. The model generation unit 23 inputs the generated pooling image to the next layer, the second convolution layer (see FIG. 4).

Next, the second process will be described with reference to FIG. FIG. 6 is a diagram showing a second process. FIG. 6 shows a pooling image generated by the pooling layer of the last layer (third layer in FIG. 4).

As shown in FIG. 6, the model generation unit 23 processes the pooling image in the order of the fully connected layer, the activation function, and the fully connected layer. For the fully connected layer, a linear transformation such as an affine transformation can be used. Further, as the activation function, a ramp function, a sigmoid function, a max-out function, a drop-out function, or the like can be used.

The model generation unit 23 generates a combined image having a predetermined number of pixels (2 × 2 in FIG. 6) by combining the pooling images using the fully connected layer and the activation function. The number of pixels of the combined image can be arbitrarily set according to the softmax function.

Subsequently, as shown in FIG. 6, the model generation unit 23 inputs the generated combined image to the softmax function. As a result, the values of the pixels A1 to A4 of the combined image indicate the probability that the pixels A1 to A4 show knocking (or the probability that the pixels do not show knocking), and the sum of the values of the pixels A1 to A4 is the sum. It becomes 1.

Then, the model generation unit 23 calculates the loss error based on the values of the pixels A1 to A4. For example, when a learning image Pa including a knocking feature is input, it is assumed that the knocking feature appears in the pixel A1 shown in FIG.

In such a case, the value of the pixel A1 should be "1", and the value of the other pixels A2 to A4 should be "0". However, as shown in FIG. 6, it is assumed that the value of pixel A1 is "0.64".

In such a case, the model generation unit 23 calculates “0.36”, which is a value obtained by subtracting 0.64 from 1, as a loss error. Then, the model generation unit 23 performs the back propagation process based on the calculated loss error.

Specifically, the model generation unit 23 repeatedly learns the combination of the bias b and the weight value ω by the back propagation process, so that the bias b and the weight value ω that minimize the loss error are finally learned. Determine as a result. As the optimization function of the weight value ω in the back propagation processing, for example, the Adam method, the stochastic gradient descent method (SGD), the Momentum, the Adagrad, or the like can be used.

In this way, by automatically creating a knock model using a convolutional neural network, it is possible to reduce the man-hours of the operator, and it is also possible to capture knocking features (for example, jumping patterns) that cannot be covered by humans. Become.

The model generation unit 23 stores the determined bias b and the weight value ω in the storage unit 3 as the classifier information 32. Here, the classifier information 32 will be described with reference to FIG. 7.

FIG. 7 is an explanatory diagram of the classifier information 32. As shown in FIG. 7, the classifier information 32 includes items such as "engine type", "sensor type", "engine rotation speed", "air load factor (filling efficiency)", "weight value", and "bias". include.

The "engine type" is information on the model of the internal combustion engine 10, and is classified by, for example, a vehicle type. “Sensor type” indicates the type of the knock sensor 11. Specifically, the "knock sensor" is a sensor that outputs the vibration itself of the internal combustion engine 10 as a knock signal. The "microphone" is a sensor that outputs the vibration sound and knocking sound of the internal combustion engine 10 as a knock signal. The "CPS" is a sensor that outputs the combustion pressure inside the internal combustion engine 10 as a knock signal.

"Engine rotation speed" indicates the rotation speed of the internal combustion engine 10. "Air load factor (filling efficiency)" indicates the air filling efficiency of the internal combustion engine 10. The “weight value” indicates the weight value ω of the classifier used by the determination unit 24. “Bias” indicates the bias b of the classifier used by the determination unit 24.

That is, the model generation unit 23 determines the weight value ω and the bias b for each sensor type, engine rotation speed, and air load factor (filling efficiency). As a result, the knocking determination accuracy can be improved by the determination unit 24 described later.

Returning to FIG. 2, the determination unit 24 will be described. The determination unit 24 determines the presence or absence of knocking based on the knock model generated by the model generation unit 23. Specifically, the determination unit 24 performs forward propagation processing on the knock image Pb (determination image Pb) generated by the image generation unit 22 and calculates the probability of showing knocking.

The determination unit 24 calculates the probability of knocking by using the weight value ω and the bias b of the classifier information 32 in the forward propagation process (first process and second process). When the probability of knocking is equal to or greater than a predetermined threshold value, the determination unit 24 determines that knocking has occurred and notifies the control device 100 of the determination result. As a result, the control device 100 can set the ignition timing in consideration of knocking in the next and subsequent cycles.

Next, the processing procedure of the conforming process executed by the knock control device 1 according to the embodiment will be described with reference to FIG. FIG. 8 is a flowchart showing a processing procedure of the conforming process executed by the knock control device 1 according to the embodiment.

As shown in FIG. 8, first, the acquisition unit 21 acquires a knock signal whose waveform changes according to the knocking generated in the internal combustion engine 10 (step S101). Subsequently, the image generation unit 22 generates a knock image Pa (learning image Pa) by converting the knock signal acquired by the acquisition unit 21 into a frequency domain at regular intervals (step S102).

Subsequently, the model generation unit 23 generates a knock model showing the knocking characteristics based on the learning image Pa generated by the image generation unit 22 (step S103).

Subsequently, the model generation unit 23 updates the weight value of the classifier used by the determination unit 24, which is the weight value included in the classifier information 32 stored in the storage unit 3 based on the generated knock model (). Step S104), the process is terminated.

Next, the processing procedure of the determination process executed by the knock control device 1 according to the embodiment will be described with reference to FIG. FIG. 9 is a flowchart showing a processing procedure of the determination process executed by the knock control device 1 according to the embodiment.

As shown in FIG. 9, first, the acquisition unit 21 acquires a knock signal whose waveform changes according to the knocking generated in the internal combustion engine 10 (step S201). Subsequently, the image generation unit 22 generates a knock image Pb (determination image Pb) by converting the knock signal acquired by the acquisition unit 21 into a frequency domain at regular intervals (step S202).

Subsequently, the determination unit 24 determines the presence or absence of knocking based on the knock model generated by the model generation unit 23 (step S203). Subsequently, the determination unit 24 notifies the control device 100 of the determination result (step S204), and ends the process.

As described above, the knock control device 1 according to the embodiment includes an acquisition unit 21, an image generation unit 22, a model generation unit 23, and a determination unit 24. The acquisition unit 21 acquires a knock signal whose waveform changes according to the knocking generated in the internal combustion engine 10. The image generation unit 22 generates knock images Pa and Pb by converting the knock signal acquired by the acquisition unit 21 into a frequency domain at regular intervals. The model generation unit 23 generates a knock model showing knocking characteristics based on the knock images Pa and Pb generated by the image generation unit 22. The determination unit 24 determines the presence or absence of knocking based on the knock model generated by the model generation unit 23. As a result, the matching accuracy of the knock model can be improved, so that the knocking detection accuracy can be improved.

In the above-described embodiment, it is determined whether or not knocking has occurred in the internal combustion engine 10, but for example, it may be determined in which cylinder of the internal combustion engine 10 the knocking has occurred. This point will be described with reference to FIGS. 10 and 11.

FIG. 10 is a block diagram showing the configuration of the knock control device 1 according to the modified example of the embodiment. FIG. 11 is an explanatory diagram of the classifier information 32 according to the modified example. In the following, the same reference numerals will be given to the configurations common to the above embodiments, and the description thereof will be omitted.

As shown in FIG. 10, four CPSs 11a to 11d are provided in each cylinder of the internal combustion engine 10 (for example, four cylinders) and measure the combustion pressure in the cylinders. The acquisition unit 21 acquires a plurality of knock signals (four in FIG. 10) for each cylinder based on the detected values of the CPS sensors 11a to 11d.

The image generation unit 22 generates a plurality of knock images Pa and Pb (4 images in FIG. 10) by converting each of the plurality of knock signals into a frequency domain at regular intervals. The model generation unit 23 generates a plurality of knock models based on the plurality of knock images Pa and Pb, and stores them as the classifier information 32.

As shown in FIG. 11, the classifier information 32 includes "engine type", "cylinder ID", "sensor ID", "engine rotation speed", "air load factor (filling efficiency)", "weight value", and " Includes items such as "bias". The "cylinder ID" is information that identifies each cylinder. The "sensor ID" is information for identifying the CPS sensors 11a to 11d provided in each cylinder. Items other than the "cylinder ID" and the "sensor ID" are described above and will be omitted.

That is, the determination unit 24 determines the presence or absence of knocking for each cylinder of the internal combustion engine 10 based on the plurality of knock models. Thereby, it is possible to detect in which cylinder of the internal combustion engine 10 the knocking has occurred. Further, the control device 100 controls the ignition timing for each cylinder, so that the fuel consumption can be improved.

Further, in the above-described embodiment, the image generation unit 22 generates knock images Pa and Pb by using the noise signal as it is. For example, the noise signal is passed through the bandpass filter and knocked by using the signal after passing. Images Pa and Pb may be generated. This point will be described with reference to FIG.

FIG. 12 is a block diagram showing the configuration of the knock control device 1 according to the modified example of the embodiment. In the following, the same reference numerals will be given to the configurations common to the above embodiments, and the description thereof will be omitted.

As shown in FIG. 12, the control unit 2 of the knock control device 1 according to the modified example further includes a bandpass filter 25 (BPF25). The BPF 25 passes only a specific knock frequency band among the knock signals input from the acquisition unit 21, and outputs the signal after passing to the image generation unit 22.

Specifically, the BPF 25 is set to a knock frequency band (for example, 5 to 20 kHz) that attenuates only the normal vibration component of the internal combustion engine 10 (see FIG. 2) among the knock signals.

That is, the BPF 25 removes the normal vibration component of the internal combustion engine 10 (see FIG. 2) as noise from the noise signal, and passes only the knock frequency band of the component corresponding to knocking.

As a result, the image generation unit 22 can generate knock images Pa and Pb from a signal containing only the component corresponding to knocking. That is, it is possible to realize the generation of a highly accurate knock model by the model generation unit 23 in the subsequent stage and the improvement of the knocking determination accuracy by the determination unit 24.

In the above-described embodiment, the determination unit 24 only performs the knocking determination using the discriminator of the convolutional neural network, but for example, the weight value ω is re-updated based on the loss error calculated at the time of the knocking determination. May be good. That is, the determination unit 24 may continue to learn the classifier repeatedly (sequential learning or online learning), so that the determination process may always be performed using the latest classifier.

Further effects and variations can be easily derived by those skilled in the art. For this reason, the broader aspects of the invention are not limited to the particular details and representative embodiments expressed and described as described above. Therefore, various modifications can be made without departing from the spirit or scope of the general concept of the invention as defined by the appended claims and their equivalents.

1 Knock control device 2 Control unit 3 Storage unit 10 Internal combustion engine 11 Knock sensor 21 Acquisition unit 22 Image generation unit 23 Model generation unit 24 Judgment unit 31 Image information 32 Discriminator information 100 Control device Pa, Pb Knock image

Claims (7)

An acquisition unit that acquires a knock signal whose waveform changes according to knocking generated in an internal combustion engine,
An image generation unit that generates a knock image by converting the knock signal acquired by the acquisition unit into a frequency domain at regular intervals, and an image generation unit.
A model generation unit that generates a knock model showing the knocking characteristics based on the knock image generated by the image generation unit, and a model generation unit.
A determination unit that determines the presence or absence of knocking based on the knock model generated by the model generation unit .
Equipped with a,
The model generator
The weight value of the classifier using the convolutional neural network is generated as the knock model, and
The determination unit
A knock control device for determining the presence or absence of knocking based on the classifier using the weight value. The acquisition unit
The knock signal is acquired for each cylinder of the internal combustion engine, and the knock signal is acquired.
The determination unit
Knock control apparatus according to claim 1, wherein the determining the presence or absence of the knocking for each of the cylinders. The acquisition unit
The knock control device according to claim 1 or 2 , wherein the knock sound generated by the knocking is acquired as the knock signal. A bandpass filter that selectively passes a signal in the knock frequency band corresponding to the knocking among the knock signals acquired by the acquisition unit is further provided.
The image generation unit
The knock control device according to any one of claims 1 to 3 , wherein the knock image is generated by converting a signal in the knock frequency band that has passed through the bandpass filter. The acquisition process of acquiring a knock signal whose waveform changes according to the knocking generated in the internal combustion engine, and
An image generation step of generating a knock image by converting the knock signal acquired by the acquisition step into a frequency domain at regular intervals, and an image generation step.
A model generation step of generating a knock model showing the knocking characteristics based on the knock image generated by the image generation step, and a model generation step.
A knock matching method comprising an update step of updating a weight value of a classifier used for determining knocking based on the knock model generated by the model generation step. An acquisition procedure for acquiring a knock signal whose waveform changes according to knocking generated in an internal combustion engine, and
An image generation procedure for generating a knock image by converting the knock signal acquired by the acquisition procedure into a frequency domain at regular intervals, and an image generation procedure.
A model generation procedure for generating a knock model showing the knocking characteristics based on the knock image generated by the image generation procedure, and a model generation procedure.
A knock matching program comprising causing a computer to perform an update procedure for updating a weight value of a classifier used for determining knocking based on the knock model generated by the model generation procedure. The acquisition process of acquiring a knock signal whose waveform changes according to the knocking generated in the internal combustion engine, and
An image generation step of generating a knock image by converting the knock signal acquired by the acquisition step into a frequency domain at regular intervals, and an image generation step.
A model generation step of generating a knock model showing the knocking characteristics based on the knock image generated by the image generation step, and a model generation step.
A determination step of determining the presence or absence of knocking based on the knock model generated by the model generation step, and a determination step of determining the presence or absence of knocking.
Including
The model generation step is
The weight value of the classifier using the convolutional neural network is generated as the knock model, and
The determination step is
Determining the presence or absence of knocking based on the classifier using the weight value.
A knock control method characterized by.

Images