KR20230006116A - A method for identifying the path through which noise and vibration generated from the noise source and vibration source of a running vehicle are transmitted to the point of interest, and a system therefor - Google Patents

Abstract

An embodiment provides a system for identifying a path through which noise and vibration generated from a noise source and a vibration source of a running vehicle are transmitted to a point of interest which comprises: a plurality of sensors installed in a vehicle; and a computing device that analyzes a transmission path from a vibration or noise source to a response of interest based on sensing information from the plurality of sensors so as to provide at least one of auditory, physical and visual feedback to a user based on an analysis result of the transmission path.

Description

A method for identifying the path through which noise and vibration generated from the noise source and vibration source of a driving vehicle is transmitted to a point of interest running vehicle are transmitted to the point of interest, and a system therefor}

The present invention relates to a method for determining a main path through which noise/vibration generated from a noise/vibration source of a driving vehicle is transferred to a main point of interest (passenger's ear or major vehicle parts) and a system therefor.

In complex dynamic systems, there are many propagation pathways for the response of interest (ROI) of noise or vibration. The contribution through the ROI path is very important with respect to response control. Therefore, transfer path analysis (TPA) is an important tool in the development of complex mechanical systems such as automobile, airplane and consumer electronics industries. TPA technology was developed as a signal processing method for multi-input and multi-output systems. TPA technology can be classified into three types according to the required measurement data and workflow. 1) classical, component-based and rate-based TPA methods; 2) classical TPA requires an excitation force applied to the relevant path and the frequency response functions (FRFs) of the target subsystem. In conventional TPA methods, measuring FRFs requires separation of the target subsystem from the assembled system, resulting in very high costs. 3) The component-based TPA method can reduce the overall cost in an experimental setup by measuring the FEFs of a target subsystem in an assembled state. In addition, the method introduces equivalent source descriptors, including what are called free velocities and blocked forces, to represent forces from active sources independent of passive structure attachments. Transmissivity-based TPAs use only the measured operational response of the assembled system and do not generally require FRFs of the subsystem of interest. The unique characteristics of the transmittance-based TPA method can provide a fast solution to the TPA process, but inevitably increase the uncertainty of the estimated transmission path due to insufficient information. A least-squares approach, singular value decomposition, and/or additional response points were employed to capture the shortcomings. Thus, the implementation of TPA is the result of certain trade-offs depending on data and time availability.

Recently, artificial neural networks have shown remarkable progress, and networks with various structures such as convolutional neural networks (CNN), RNN (Recurrent Neural Network), VAE (Variational Auto Encoder), VAE (variational auto encoder), and GAN (generative adversarial network) have been developed. has been developed It has been shown that these deep neural networks can detect unknown but important features and reveal input and output relationships in large data sets. Therefore, many studies in the field of noise and vibration research have utilized deep neural networks to predict and/or identify structural damage. However, TPA methodology based on deep neural networks has not been reported yet.

Republic of Korea Patent Registration No. 10-1025163

The embodiment goes through a simple preparation process of continuously measuring a sufficiently large amount of the vibration / noise signal on the path of interest during operation to analyze the operating point propagation path of the vibration / noise path, and constructs a DNN using only the measured data during operation Through this method, it provides a high-accuracy analysis system that can analyze the transmission path of vibration/noise signals.

In addition, compared to the conventional method, the embodiment does not require measurement in a stationary state to identify the characteristics of the structure, so cost and time can be greatly reduced, and the transmission path can be analyzed with high analysis accuracy. Provided is a method and a system for identifying a path through which noise and vibration generated from a noise source and a vibration source are transmitted to a point of interest.

Embodiments include a plurality of sensors installed in a vehicle; and a computing device that analyzes a transmission path from a vibration or noise source to an interest response based on the sensing information from the plurality of sensors, and provides auditory, physical, and visual feedback to the user based on the analysis result of the transmission path. It is possible to provide a system for identifying a path through which noise and vibration generated from a noise source and a vibration source of a driving vehicle that provide at least one of the above are transmitted to a point of interest.

In another aspect, the plurality of sensors may provide a system for identifying a path through which noise and vibration generated from a noise source and a vibration source of a driving vehicle including an acceleration sensor are transmitted to a point of interest.

In another aspect, preparing a training data set using frequency responses of inputs and outputs based on sensing information from sensors of a vehicle body model; Proliferating the training data set to train and verify the deep learning neural network; And analyzing the transmission path during operation using the trained and verified deep learning neural network to analyze the contribution of the transmission path of the vehicle of noise or vibration; A method of determining a path through which vibrations are transmitted to a point of interest may be provided.

In another aspect, the plurality of sensors further include a plurality of touch sensors installed at the plurality of points of interest, and the computing device includes a plurality of touch sensors installed among the plurality of points of interest for a predetermined time from an initial driving time. Detects user's touch information on an area of , ranks a plurality of areas according to the number of touches, displays a vibration value of each of the plurality of ranked areas, and displays the prioritized plurality of areas when the touched area is changed. A system for determining a route through which noise and vibration generated from a noise source and vibration source of a vehicle in motion that induces a change in the user's posture by guiding information on a region having a lower vibration value than a changed touch region among regions of a vehicle to a point of interest can provide

In another aspect, preparing the data set comprises: measuring the input and output frequency responses by defining paths, input and output responses; normalizing the input and output frequency responses; and multiplying phases of the normalized input and output frequency responses; a method for determining a path through which noise and vibration generated from a noise source and a vibration source of a driving vehicle are transmitted to a point of interest may be provided.

In another aspect, in the phase propagation step, noise and vibration generated from a noise source and a vibration source of a driving vehicle, which generate a random phase and move the phases of the input and output frequency responses to propagate the phase, are transmitted to a point of interest. It can provide a way to figure out the route.

In another aspect, the preparing of the data set may further include collecting an entire training data set by multiplying the mutual spectrum of the phase-multiplied input and output frequency responses; It is possible to provide a method for identifying a path through which noise and vibration generated from the noise and vibration are transmitted to a point of interest.

In another aspect, the step of training and verifying the deep learning neural network may include setting parameters of a deep learning neural network model based on the entire collected training data set; predicting an interest response data set by performing deep learning neural network training based on the deep learning model parameters; and verifying the deep learning neural network based on the difference between the interest response and the attention response predicted by the deep learning neural network; noise and vibration generated from the noise and vibration sources of the vehicle being driven are transmitted to the point of interest. It can provide a way to figure out the path to be.

In another aspect, the analyzing of the contribution may include calculating a path contribution by predicting an off response of any one of the transmission paths; and analyzing the path contribution based on the difference between the sum of the contributions of the total input and the interest response predicted by the deep learning neural network. It can provide a way to figure out the path to be.

In another aspect, analyzing the transmission path of the noise or vibration to a plurality of points of interest based on the analysis result of the transmission path of the vehicle; and detecting touch information of the user at the plurality of points of interest for a predetermined time from the initial driving of the vehicle, ranking the plurality of regions of interest according to the number of touches, and generating vibrations of each of the plurality of ranked regions of interest. displaying a value, and inducing a change in posture of the user by guiding information on an area having a lower vibration value than the changed touch area among the plurality of ranked areas of interest when the touched area is changed. A method for determining a path through which noise and vibration generated from a noise source and a vibration source of a vehicle are transmitted to a point of interest may be provided.

The embodiment can properly identify the transmission path of the ROI in the car using the introduced DNN model composed of one input, one output and several hidden layers. Further, the embodiment enables real and imaginary parts of an input response at each frequency of an input layer in a trained DNN model to feasibly generate real and imaginary parts of an ROI in an output layer. In addition, in the embodiment, the contribution of each path can be identified by assigning a value of 0 to the corresponding input path or vice versa, and by applying the 1 channel off method, a very simple and efficient method can be provided in terms of computational cost. .

In addition, the embodiment uses only the relative acceleration as an input measured during operation and uses the ROI as an output so that there is no need to measure FRF in a subsystem or a blocked system, reducing the overall time required to perform TPA. there is

In addition, the embodiment provides a method for increasing accuracy by multiplying data even when using a small amount of data when training a deep neural network.

In addition, the phase enhancement utilizing the invariance of the phase shift for the frequency spectrum data set may improve the accuracy of ROI prediction.

1 is a schematic diagram of a system for determining a path through which noise and vibration generated from a noise source and a vibration source of a driving vehicle are transferred to a point of interest according to an embodiment of the present invention.
2A to 2C are flowcharts of a method for determining a path through which noise and vibration generated from a noise source and a vibration source of a driving vehicle are transferred to a point of interest.
3A is a flowchart of providing feedback by analyzing a transmission path during operation using a DNN according to various embodiments of the present invention.
3B is for explaining an example of a method of providing feedback according to a result of analyzing a transmission path.
Figure 4 shows the substructure with the response of interest.
5 shows deep learning composed of fully dense layers.
6 shows a vehicle body introduced to verify the performance of the operational TPA method.
Figure 7 is a graph showing the iteration history of the training process for the largest DNN model.
8 is a graph of the mean square error of the DNN model at force amplitude μ-3σ.
Figure 9 is a graph of the ROI predicted by the DNN model with H = 20 and P = 128 at a force amplitude of μ-3σ compared to the reference value.
10 is the reconstruction error of the X-direction ROI according to the sum of the path contributions using the DNN model with P = 128.
11 is a graph of TPA results during operation for an ROI in the X direction at mean force amplitude.
12a to 12c are graphs showing Operational TPA results for an X-direction ROI in average force amplitude.
Figure 13 shows the individual path contributions at the peak (275 Hz) and the synthesized ROI vector when the force amplitude is averaged.

Since the present invention can apply various transformations and have various embodiments, specific embodiments will be illustrated in the drawings and described in detail in the detailed description. Effects and features of the present invention, and methods for achieving them will become clear with reference to the embodiments described later in detail together with the drawings. However, the present invention is not limited to the embodiments disclosed below and may be implemented in various forms. In the following embodiments, terms such as first and second are used for the purpose of distinguishing one component from another component without limiting meaning. Also, expressions in the singular number include plural expressions unless the context clearly dictates otherwise. In addition, terms such as include or have mean that features or elements described in the specification exist, and do not preclude the possibility that one or more other features or elements may be added. In addition, in the drawings, the size of components may be exaggerated or reduced for convenience of explanation. For example, since the size and thickness of each component shown in the drawings are arbitrarily shown for convenience of description, the present invention is not necessarily limited to the illustrated bar.

- system

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings, and when describing with reference to the drawings, the same or corresponding components are assigned the same reference numerals, and overlapping descriptions thereof will be omitted. .

1 is a schematic diagram of a system for determining a path through which noise and vibration generated from a noise source and a vibration source of a driving vehicle are transferred to a point of interest according to an embodiment of the present invention.

Referring to FIG. 1 , a system 10 for determining a path through which noise and vibration generated from a noise source and a vibration source of a driving vehicle are transmitted to a point of interest according to an embodiment of the present invention is a sensor 100 installed in a vehicle 1 ), and the computing device 200 that analyzes the transmission path of noise or vibration by analyzing the sensing information from the sensor. According to various embodiments of the present invention, the system 10 for identifying a path through which noise and vibration generated from a noise source and a vibration source of a driving vehicle are transmitted to a point of interest according to an embodiment of the present invention displays analysis results and additional information. It may further include device 300 .

The sensor 100 is composed of a plurality of sensors and may include at least one of an acceleration sensor and a microphone sensor.

Computing device 200 is intended to represent various forms of digital computers, such as laptop computers, convertible computers, tablet computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframes, and other suitable computers. Computing device 200 is intended to represent various forms of mobile devices such as personal digital assistants, cellular telephones, smart phones, and other similar computing devices. The components depicted herein, their connections and relationships, and their function are meant to be illustrative only and are not meant to limit implementations of the invention described and/or claimed herein.

In some embodiments, computing device 200 may be a device included in a vehicle.

The computing device 200 may include a processor 210 and a memory 220 . Additionally, computing device 200 may include a high-speed interface coupled to the storage device, memory, and high-speed expansion port, and a low-speed interface coupled to a low-speed bus and the storage device. Each of the components are interconnected using various buses and may be mounted on a common motherboard or in other ways suitable. Processor 210 processes instructions for execution within computing device 200, including instructions stored in memory 220 or a storage device, to display a GUI on an external input/output device such as a display connected to a high-speed interface. graphic information can be displayed. In another implementation, multiple processors and/or multiple buses may be used, as appropriate, along with multiple memories and types of memory. In addition, multiple computing devices 200 may be connected (eg, as a server bank, blade server group, or multi-processor system) with each device providing a portion of a required operation.

Memory 220 stores information within computing device 200 . In one embodiment, memory 220 is a volatile memory unit or units. In another implementation, memory 220 is a non-volatile memory unit or units. Memory 220 may also be another form of computer readable medium, such as a magnetic or optical disk.

The storage device may provide a mass storage device for computing device 200 . In one implementation, the storage device may be an array of devices including a floppy disk device, a hard disk device, an optical disk device or tape device, a flash memory or other similar solid state memory device, or a storage area network or other configuration of devices. there is. A computer program product may be tangibly embodied in an information carrier. The computer program product may also include instructions that, when executed, perform one or more methods as described above. An information carrier is a computer or machine readable medium, such as memory 220, storage device or memory on processor 210.

The high speed controller manages bandwidth intensive operations for computing device 200, while the low speed controller manages low bandwidth intensive operations. The assignment of these functions is exemplary only. In one implementation, the high-speed controller is connected to a high-speed expansion port that can accommodate various expansion cards and memory 220, a display. In various examples, the low speed controller may be coupled to the storage device and the low speed expansion port.

Computing device 200 may be implemented in a number of different forms. For example, it may be implemented as a standard server or multiple times in a group of such servers. It may also be implemented as part of a rack server system. Also, it may be implemented in a personal computer such as a laptop computer.

Computing device 200 may include input/output devices such as processor 210, memory 220, display, communication interface, and transceiver. Computing device 200 may be provided with a storage device such as a micro drive or other device to provide additional storage devices. Each of the components are interconnected using various buses, and some components may be mounted on a common motherboard or in other ways as appropriate.

Processor 210 may execute instructions within computing device 200 including instructions stored in memory 220 . The processor may be implemented as a chipset of chips including multiple separate analog and digital processors. The processor may provide coordination of other components of the computing device 200, such as, for example, control of a user interface, applications executed by the computing device 200, and wireless communications by the computing device 200.

In various embodiments, the computing device 200 may be mounted on the vehicle 1 as a component constituting a part of the vehicle 1 .

In various embodiments, the display device 300 may be a display of the computing device 200, but is not limited thereto and may be a component independent of the computing device 200.

Computing device 200 may communicate wirelessly through a communication interface that may include digital signal processing circuitry as needed. The communication interface may provide communication under various modes or protocols such as GSM voice calling, SMS, EMS or MMS messaging, CDMA, TDMA, PDC, WCDMA, CDMA2000 or GPRS. Such communication may occur, for example, via a radio frequency transceiver. Also, short-range communication may occur, such as using Bluetooth, WiFi, or other transceivers. Additionally, the Global Positioning System (GPS) receiver module may provide additional navigation and location-related radio data to the device that can be used as appropriate by applications running on the computing device 200 . Computing device 200 may also communicate aurally using an audio codec that can receive voice information from a user and convert it into usable digital information. An audio codec may likewise produce audible sound for a user, such as through a speaker in a handset of computing device 200, for example. Such sounds may include sounds from voice telephone calls, may include recorded sounds (eg, voice messages, music files, etc.), and may also include sounds generated by applications running on computing device 200. may contain sounds.

- How

2A to 2C are flowcharts of a method for determining a path through which noise and vibration generated from a noise source and a vibration source of a driving vehicle are transferred to a point of interest.

A method for determining a path through which noise and vibration generated from a noise source and a vibration source of a driving vehicle are transmitted to a point of interest according to an embodiment includes 1) preparation of measurement data, 2) training and verification of DNN, and 3) transmission path. It can be divided into analysis steps. In detail, the preparation step of measurement data consists of data measurement and pre-processing. Data measurement consists of frequency response measurements before and after the path of interest of the target structural system, and is a continuous measurement task of various operating states. The data preprocessing process is a process of propagating the measured data, and the propagation consists of the phase propagation of the measured data and the cross-spectrum propagation process between each measured data.

The training and verification phase of the DNN consists of constructing the DNN using the prepared training data, training the artificial neural network to reconstruct the vibration/noise response signal of interest, and verifying it. And, the vibration/noise transmission path analysis step consists of a process of performing transmission path analysis by removing all inputs other than a specific target path or removing only inputs of a specific target path using the verified training DNN.

Hereinafter, the embodiment will be described in detail with reference to FIGS. 2A to 2C.

Referring to FIG. 2A , in the method for determining a path through which noise and vibration generated from a noise source and a vibration source of a vehicle in motion are transferred to a point of interest according to an embodiment, the step of preparing data includes preparing the path, input and output responses. The step of defining (S110), measuring the input and output frequency response (S120), normalizing the input and output frequency response (S130), phase multiplication step (S140), cross spectrum multiplication step (S150) and all It may include collecting a training data set (S160).

Referring to FIG. 2B , in the method for determining a path through which noise and vibration generated from a noise source and a vibration source of a driving vehicle are transmitted to a point of interest according to an embodiment, the DNN training and verification steps include setting a DNN structure. (S210), preparing a data set (S220), setting DNN model parameters (S230), training a DNN (S240), predicting an interest response data set (S250), It may include determining whether the Euclidean norm of the difference between the attention responses predicted by the DNN converges to 0 (S260).

Referring to FIG. 2C , in the method for determining a path through which noise and vibration generated from a noise source and a vibration source of a vehicle in motion are transmitted to a point of interest according to an embodiment, the step of analyzing the transmission path during operation by using a DNN includes: Setting related conditions during operation (S310), measuring input and output responses to the response of interest (S320), predicting a response by turning off one channel (S330), calculating path contribution (S330) S340), determining whether the Euclidean norm of the difference between the sum of the contributions of the interest response predicted by the DNN and the total input converges to 0 (S350), and analyzing path contributions (S360).

을 측정할 수 있다. 그리고, 입력 및 출력 주파수 응답의 정규화를 하는 단계(S130)에서

의 정규화를 수행할 수 있다. 위상 증식 단계(S140)에서 무작위 위상을 생성하고

의 위상을 이동시켜 증식된 응답

을 수집할 수 있다. 그리고, 상호 스펙트럼 증식 단계(S150)에서 입력 및 출력 응답에 기초하여

및

의 증식된 응답을 수집할 수 있다. 그리고, 모든 트레이닝 데이터 세트

를 수집할 수 있다. 그 후, 도 2b의 S220 단계로 진행할 수 있다.More specifically, in FIG. 2A , input and output frequency responses considering perturbations in step S120 of measuring input and output frequency responses

can measure And, in the step of normalizing the input and output frequency responses (S130)

can be normalized. In the phase propagation step (S140), a random phase is generated and

Response multiplied by shifting the phase of

can be collected. And, based on the input and output responses in the cross spectrum multiplication step (S150)

and

Proliferated responses of can be collected. And, all training data sets

can be collected. After that, it may proceed to step S220 of FIG. 2B.

를 설정할 수 있다(S230), 그리고, DNN의 트레이닝

이 되면, 다음 단계로 진행하고, 그렇치 않은 경우 (S230) 단계를 다시 수행할 수 있다(S240). DNN을 트레이닝한 후 관심응답(ROI) 데이터 세트를 예측

하는 테스트 설정이 수행될 수 있다(S250). 그 후, 관심응답과 DNN에 의해 예측된 관심응답의 차이의 유클리드 노름이 0에 수렴

하는지를 판단하여, 수렴하면, 도 2c의 단계로 진행할 수 있다. 그렇치 않은 경우, 도 2b의 S210에서의 DNN 구조를 설정하는 단계가 진행될 수 있다.in Fig. 2b. A data set can be prepared by setting training and setting verification (S220). And, the DNN model parameters

can be set (S230), and training of the DNN

If it is, it proceeds to the next step, and if not, the step (S230) may be performed again (S240). After training the DNN to predict the interest response (ROI) data set

A test setting may be performed (S250). After that, the Euclidean norm of the difference between the attention response and the attention response predicted by the DNN converges to 0.

If it is determined whether or not to converge, the step of FIG. 2c may be performed. If not, the step of setting the DNN structure in S210 of FIG. 2B may proceed.

을 측정하고(S320), 하나의 채널을 오프(off)한 응답

을 예측할 수 있다(S330). 그리고, 경로 기여도

를 계산할 수 있다. 그리고, DNN에 의해 예측된 관심응답과 총 입력의 기여도의 합의 차이의 유클리드 노름이 0에 수렴

하는지를 판단할 수 있다 (S350) 수렴하는 경우 경로 기여도를 분석할 수 있고(S360), 그렇치 않은 경우 도 2b의 S210 단계로 진행하여 DNN 트레이닝 및 검증 과정을 수행할 수 있다.In Figure 2c, the input and output responses to the interest response

Measure (S320), and the response of turning off one channel

can be predicted (S330). and path contribution

can be calculated. And, the Euclidean norm of the difference between the sum of the contributions of the total input and the attention response predicted by the DNN converges to 0.

If convergence is reached (S350), the path contribution can be analyzed (S360). If not, the DNN training and verification process can be performed in step S210 of FIG.

을 측정하고(S320), 도 2b의 S250 단계를 진행하여 ROI 데이터 세트를 예측한 후 S260 단계의 조건이 충족되면 S330 단계로 진행할 수 있다.On the other hand, input and output responses to interest responses

After measuring (S320) and predicting the ROI data set through step S250 of FIG. 2B, if the condition of step S260 is satisfied, step S330 may be performed.

In addition, if there is a trained and verified DNN, after setting related in-operation conditions (S310), the transfer path analysis process in operation may be performed using the DNN.

3A is a flowchart of providing feedback by analyzing a transmission path during operation using a DNN according to various embodiments of the present invention. And, FIG. 3B is for explaining an example of a method of providing feedback according to a result of analyzing a transmission path.

In the final step of FIG. 2C in the step of providing feedback by analyzing the transfer path during operation using the DNN according to various embodiments of the present invention, feedback may be provided to the user based on the analysis result of the transfer path (S370). There is, and this feedback can be auditory, physical, or visual feedback. Illustratively, the transmission path information of the vibration source through the display 300 may guide the user to the mounting position of the vibration reduction device. Illustratively referring to FIG. 3B , the display 300 may be a display installed in a vehicle. As shown exemplarily through the display 300, the vehicle frame and main parts are displayed at various angles, and the path through which noise or vibration is transmitted from the displayed vehicle frame and main parts to the point of interest (ia) ( p) can be displayed. In some embodiments, on the display 300, information on a transmission path of noise or vibration to a point of interest analyzed by the computing device 200 at an initial point in time at the time of shipment of the vehicle is compared with information on a transmission path at the current point in time. The analyzed result may be displayed, and current status information of additional transmission paths excluding transmission paths having a similarity to a transmission path analyzed at an initial point or higher than a predetermined value among transmission paths transmitting noise or vibration to the point of interest may be displayed. In addition, the computing device 200 may extract previously set exchangeable vehicle part or structure information from among vehicle parts or structures matched to the additional transmission path. In addition, the computing device 200 may display replaceable vehicle parts or structure information for noise or vibration reduction through the display 300 . Accordingly, by providing information on transmission paths of noise or vibration sources due to deterioration of various parts of the vehicle to the user, it is possible to guide information on parts that are desirable to be replaced in the vehicle. According to various embodiments, a plurality of speakers outputting sound for noise canceling in the vehicle may be independently controlled to output noise canceling sound for noise reduction through a speaker adjacent to a transmission path analyzed as a noise source. there is. For example, based on information on a transmission path of noise or vibration transmitted to a point of interest, a sound for noise canceling may be output from a speaker located in an area adjacent to a transmission path among a plurality of speakers in a vehicle. The computing device 200 may display information on a list of speakers in a vehicle requiring noise canceling execution through the display 300 and, in response to a command signal input from a user, sound for noise canceling from a speaker in the vehicle requiring noise canceling execution. can be output.

According to various embodiments, the computing device 200 displays location information where noise or vibration is analyzed to be the minimum among a plurality of locations of a rear seat of a vehicle based on information on a transmission path of noise or vibration to a plurality of points of interest. By guiding through 300, a convenience function may be provided so that the user sitting in the back seat of the vehicle can sit in a more comfortable position.

According to various embodiments, touch sensors may be provided in a plurality of areas within a vehicle. The computing device 200 installed in the vehicle detects user's touch information on a plurality of areas where the touch sensors are provided by monitoring for a predetermined time from the initial driving of the vehicle based on the sensing information of the plurality of touch sensors, and determines the number of touches. A plurality of regions may be ranked according to the above. For example, in general, the steering wheel area of the vehicle is given first priority, followed by the console box area, shift lever area, door trim armrest area, etc., but these priorities are not limited to the above, and the driver It may vary according to the driver's driving habit or the preference of other occupants' postures while sitting on the seat. Based on the monitoring information, the computing device 200 may display ranking information of a plurality of regions according to the number of touches through the display 300 and display a vibration value that is vibration information of each displayed region. In addition, information on regions from the first to the Nth rank and the vibration value of each region may be displayed. When the user is holding the steering wheel with both hands, the computing device 200 may display the corresponding region if there is a region having a vibration value lower than that of the steering wheel among regions ranked from the first to the Nth order. In addition, when the user touches the console box area with one hand while holding the steering wheel with both hands, and a predetermined time elapses from the point of touch, vibration lower than that of the console box area among the first to Nth ranked areas If a region having values exists, the corresponding region may be displayed. This may be equally applied to a passenger sitting in an assistant seat or another seat other than the driver's seat. In some embodiments, the computing device 200 determines the first through Nth ranks when a user holds the steering wheel with both hands and then touches the console box area with one hand, and a predetermined time elapses from the time of the touch. It is possible to detect area information having a vibration value lower than that of the console box area among the areas of , and display an area having a minimum vibration value among areas adjacent to the console box area among the detected areas. When the user's left hand holds the steering wheel, and the right hand touches the console box after removing the steering wheel, the optimal touch area is selected by considering the vibration value among areas such as the shift lever area within the touchable area with the right hand. is to guide The computing device 200 analyzes the transmission path of noise or vibration to the point of interest in real time while the vehicle is driving, and compares vibration information between areas touched by the occupant in the vehicle so that the occupant feels the least amount of vibration. A change in the posture of the occupant may be induced by inducing a touch to an area that allows the passenger to be seated. In addition, by guiding optimal touch position information among touch areas that are monitored and prioritized for a predetermined period of time rather than all touchable areas, it is possible to respond in a customized way to the preference of the vehicle occupant's posture.

- Transmitting path analysis

1) TPA formula

Figure 4 shows the substructure with the response of interest.

으로 표시된 M개의 입력이 존재할 수 있다. 하부 구조에는

으로 표시된 N개의 ROI도 존재할 수 있다. 하위 구조 시스템이 선형이라고 가정하면 ROI는 각 여기 지점에서 단일 여기로 인한 별도의 응답을 가정하여 결정할 수 있다. 그 후, 주파수 영역의 총 ROI는 수학식 1을 충족할 수 있다.As shown in FIG. 4, a substructure with a response of interest (ROI) having two response sets, i.e. input and output ( X and R ) sets, can be considered. For convenience of description of the invention, it may be assumed that there is no active source in the substructure. Sub-structures are connected to other sub-structures during operation.

There may be M inputs indicated by . in the substructure

N ROIs indicated by may also exist. Assuming that the substructure system is linear, the ROI can be determined assuming a separate response due to a single excitation at each excitation point. Then, the total ROI in the frequency domain may satisfy Equation 1.

[Equation 1]

이고

이고, R은 시스템의 ROI이고, X는 시스템의 입력 응답 벡터이다.in Equation 1

ego

, R is the ROI of the system, and X is the input response vector of the system.

The transfer function matrix T can be a NХM matrix denoted by Tnm for the input at the mth position and the output of the nth ROI.

The contribution ROI of the m-th input (ie, the m-th path) to the n-th input, that is, Cnm, may satisfy Equation 2.

[Equation 2]

Here, Cnm is the contribution of the mth input to the nth ROI.

Thus, if we can use the transfer function and the amount of input on the path, we can estimate the contribution.

In classical transfer path analysis (TPA), the matrix T can be measured by applying a force to a structural path or a volume velocity to an air path. The force in the path (where χm stands for force) can be identified, for example, from joint stiffness and motion displacement measurements.

In the Operational TPA method, only the operational response of the connection part and the ROI can be used to identify the path contribution of Equation 2 without measuring the transfer function T.

In Equation 2, if the response χm (where χm is not a force but, for example, acceleration) during motion on the connection is measured, the transfer function due to the measured connection response can provide the path contribution. Thus, an Operational TPA method can essentially be a process for estimating transfer functions in a multi-input multi-output environment.

To estimate T , if the complex conjugate of the input is multiplied on both sides of Equation 1, the result satisfies Equation 3, and when the complex conjugate of the output is multiplied on both sides of Equation 1, the result satisfies Equation 4 can do.

[Equation 3]

[Equation 4]

where S is the covariance matrix and T is the transfer function matrix for the subsystem. And, in the above equation, the matrix S is the matrix of the cross spectral density function between the two responses.

As can be seen from Equations 3 and 4, S ^XX Alternatively, a transfer matrix can be identified if an inverse matrix of S ^XR exists, which means that the response can be transformed into M uncorrelated components. However, in general, the measured in-motion response in the path of interest is correlated because it is a simultaneous signal with a limited number of excitation sources. Operational TPA methods can usually address the lack of rank problem by utilizing a least-squares approach, singular value decomposition (SVD), and additional ROIs.

2) Operational TPA based on deep neural networks

5 shows deep learning composed of fully dense layers.

After collecting the motion responses X and R in Equation 1, a deep neural network (DNN) can be constructed to express the matrix T. In the present invention, as shown in FIG. 5, fully dense layers for constructing a DNN are configured.

The physical input X and output R vectors composed of complex numbers at all frequencies can be encoded as vectors of the actual input U and output Y of the DNN, respectively.

, 하나의 출력 벡터

그리고 H 히든 레이어로 구성될 수 있다.DNNs have one input vector

, one output vector

And it may be composed of an H hidden layer.

The lengths of the U and Y vectors (eg, M ^* and N ^* ) are 2MХn _f and 2NХn _f , where each n _f is a frequency.

Each hidden layer has a P node (i.e. a hidden variable). In DNN, the hidden and output variables for a given input and output data set ( U , Y ) can be calculated as follows.

[Equation 5]

는 현재 모델 매개 변수가 있는 예측된 출력 벡터이다. 즉, 각각의 Z ^h, W ^h 및 B ^h는 h번째 히든 변수 벡터와 h 번째 가중치 행렬 및 벡터이다. 그리고, H는 히든 레이어 수 +1 이다.Here, W is the weight matrix multiplied by the variables of the nodes in the DNN, U is the input vector of the DNN, B is the vector of constants added to the nodes of the DNN, and Z is the vector of hidden variables of the DNN. And,

is the predicted output vector with the current model parameters. That is, each of Z ^h , W ^h , and B ^h is an h-th hidden variable vector and an h-th weight matrix and vector. And, H is the number of hidden layers +1.

는 트레이닝 데이터 세트를 사용하여 네트워크를 트레이닝하여 결정할 수 있다(S230). 데이터 세트는 다음 수학식 6과 같이 입력 및 출력 쌍의 모음이다.Model parameters of DNN

may be determined by training the network using the training data set (S230). A data set is a collection of input and output pairs as shown in Equation 6 below.

[Equation 6]

는 트레이닝 데이터 세트이고,

는 M^*차원의 실수이다. 데이터 세트의 실제 출력 데이터(

: DNN의 출력 벡터)와 신경망에서 예측되어 계산된 출력(

: 현재 모델 매개 변수가 있는 예측된 출력 벡터) 간의 차이를 최소화하면 수학식 7을 통해 모델 매개 변수를 결정할 수 있다.here,

is the training data set,

is a real number of dimension M ^* . The actual output data of the data set (

: DNN's output vector) and predicted and calculated output from the neural network (

: predicted output vector with current model parameters), the model parameters can be determined through Equation 7.

That is, parameters satisfying Equation 7 below may be determined according to DNN training in order to determine trained model parameters (S240).

[Equation 7]

는 트레이닝된 모델 매개 변수이고,

는 유클리드 노름을 나타내며

는 벡터의 노름이다.here,

is the trained model parameter,

denotes the Euclidean norm and

is the norm of the vector.

If the trained DNN is constructed from the prepared data set, the ROI for which the path contribution is requested for the in-operation data set ( X , R ) can be predicted as shown in Equation 8 (S250).

[Equation 8]

은 DNN에 의해 예측된 ROI이다. f(.)는 수학식 5의 함수이다.where R is the ROI of the system,

is the ROI predicted by the DNN. f(.) is a function of Equation 5.

차이의 유클리드 노름이 0에 수렴하는지 판단하고, 수렴하는 경우 DNN의 트레이닝 및 검증이 완료된 것으로 판단할 수 있다(S260).R and

It is determined whether the Euclidean norm of the difference converges to 0, and if it converges, it can be determined that training and verification of the DNN are completed (S260).

In addition, the trained DNN model can predict the response when the mth input is virtually removed or only the mth input is present using the equation below.

[Equation 8]

m은 m번째 입력을 0으로 설정하여 예측된 ROI이며,

m은 m번째 입력을 제외한 입력에 모두 0을 넣어 ROI을 예측한 것이고,

m은 m번째 입력 값이 0인 입력 벡터이며,

m은 m번째 입력을 제외한 나머지 입력에 모두 0이 있는 입력 벡터이다. here,

m is the ROI predicted by setting the mth input to 0,

m is the prediction of the ROI by putting 0 in all inputs except the mth input,

m is an input vector in which the mth input value is 0,

m is an input vector in which all inputs except for the mth input have 0.

을 예측할 수 있다(S330). That is, the response of turning off one channel according to Equation 8

can be predicted (S330).

The contribution of the mth path can be calculated as follows (S340).

[Equation 9-1]

[Equation 9-2]

은 DNN에 의한 예측 ROI이고,

m은 m번째 입력을 0으로 설정하여 예측된 ROI이며,

m은 m번째 입력을 제외한 입력에 모두 0을 넣어 ROI을 예측한 것이다.where Cm is the contribution of the mth input to the ROI,

is the predicted ROI by DNN,

m is the ROI predicted by setting the mth input to 0,

m is the ROI prediction by putting 0 in all inputs except for the mth input.

Therefore, the mth contribution vector c ^m (m=1,...,M) in Equation 9-1 or Equation 9-2 may form the mth column of the path contribution matrix c in Equation 2. The sum of the contribution vectors should recreate the predicted response as shown in Equation 10 below.

[Equation 10]

은 DNN에 의해 예측된 ROI이다.here,

is the ROI predicted by the DNN.

과 1부터 M까지의 기여 벡터의 총합의 차이의 유클리드 노름의 0으로의 수렴 여부를 판다(S350)하고, 수렴하는 경우 경로 기여도를 분석(S360)할 수 있다.in other words,

It is determined whether the Euclidean norm of the difference between the sum of the contribution vectors from 1 to M converges to 0 (S350), and in case of convergence, the path contribution may be analyzed (S360).

3) Operational data preparation

A DNN-based operational TPA may require two types of data sets: training and associated ROI. A training data set may consist of in-motion responses used to determine DNN model parameters. The ROI data set is an in-motion response requiring path contribution, for example, ( X , R ) in Equation 1.

로 사용하면 트레이닝 데이터 세트는 수학식 11을 충족한다.For the training of a DNN, we need to collect many pairs of input and output data instances of the substructure of interest. A data set may be provided through measurement of the response during operation at the input and output locations (S120). These measurements can be performed on the assembled system during operation. During measurement, the magnitude of the active source can be disturbed and kept near the relevant operating conditions. Simultaneously measured input and output

When used as , the training data set satisfies Equation 11.

[Equation 11]

는 측정된 입력 응답이고,

은 측정된 ROI이며, J₀는 측정된

쌍의 수로써, 측정된 트레이닝 데이터 세트수를 나타낸다.where ( U , Y ) is an element of the training data set,

is the measured input response,

is the measured ROI, and J ₀ is the measured

As the number of pairs, it indicates the number of training data sets measured.

4) Data augmentations

A larger number of training data sets can reduce the risk of overfitting when determining model parameters in the training phase.

In addition, in order to identify the transfer function matrix T of Equation 1, it is necessary to accurately consider the path contribution as well as the response during operation, and the training data set includes characteristic information (ie, response and path contribution) unique to the data itself required for estimation. Should be.

In the present invention, the measured data set of Equation 11 can be augmented such that the trained DNN can obtain both the randomness of the reference phase of the frequency spectrum and the correlation of the measured response at the input and output pairs.

의 측정을 재검토하여, 기준 위상은 임의로 선택되고 주파수 응답은 기준 위상에 상대적인 의미에서 결정될 수 있다.　따라서,

쌍의 주파수 스펙트럼의 위상에 상수값 ø(응답의 위상)를 추가하여, 원래 응답을 위상 증식 쌍

으로 복제할 수 있다.of Equation 11

By reviewing the measurement of , the reference phase can be chosen arbitrarily and the frequency response determined in a sense relative to the reference phase. thus,

By adding a constant value ø (the phase of the response) to the phase of the pair's frequency spectrum, the original response can be obtained from the phase multiplying pair.

can be replicated with

에서

를 생성하면 수학식 12와 같이

의 i번째 요소에 대해 단계적으로 증식된 데이터 세트를 수집할 수 있다.randomly generate J ₁ step,

at

, as shown in Equation 12

A step-wise multiplied data set can be collected for the i-th element of .

[Equation 12]

는 단계적으로 증식된 트레이닝 데이터 세트이다(S140).here,

Is a training data set multiplied step by step (S140).

및

에 대한 추가 증식은 수학식 3, 4에 기초한다.

and

Further multiplication for is based on equations (3) and (4).

Comparing Equations 3 and 4 with Equation 1, all columns of the cross-spectrum matrices S ^XX and S ^XR can be considered as inputs of the DNN, the purpose of which is to identify the transfer matrix T.

Similarly, columns of cross-spectrum matrices S ^RX and S ^RR can be used as corresponding output vectors during DNN training. Therefore, cross spectrum multiplication can be performed as shown in Equation 13 (S150).

[Equation 13]

는 입력에 의한 상호 스펙트럼 증식 트레이닝 데이터 세트이고,

는 출력 별 상호 스펙트럼 증식 트레이닝 데이터 세트이며, 위첨자 *는 인수의 켤레를 나타낸다.here

is the cross-spectral multiplication training data set by input,

is the output-by-output cross-spectral multiplication training data set, and the superscript * denotes the conjugate of the argument.

It should also be noted that the order of phase and cross-spectral multiplication can be reversed.

In addition, cross-spectral broadening during DNN model training is essential to identify transmission paths of ROIs using DNN models.

는 다음과 같이 단계 및 상호 스펙트럼에 의해 증식된 데이터 세트의 모든 조합이 된다(S160).full data set

becomes all combinations of data sets multiplied by step and reciprocal spectrum as follows (S160).

[Equation 14]

The total number of multiplied data sets is J ₀ Х(J ₁ +1)Х(M+N), which is significantly expanded compared to the amount of data initially measured.

를　나눈다．정규화된　데이터　세트의　평균을　０으로　이동하는 평균 이동이　응답의　위상　정보를 왜곡한다는 점을 유의해야 한다.Prior to the propagation step, the initial data set should be normalized (S130) to minimize the numerical instability associated with the difference in magnitude of the input and output variables. Here, for normalization, the standard deviation of the responses can be used for each response individually. The standard deviation can be calculated for the response of the training data set over the entire frequency range. The standard deviation is

Divide . It should be noted that a mean shift of the normalized data set to zero distorts the phase information of the response.

- Numerical verification

1) body body

6 shows a vehicle body introduced to verify the performance of the operational TPA method.

In order to verify the performance of the operational TPA method according to an embodiment of the present invention, a simplified body as shown in FIG. 6 was introduced. The simplified version itself has two substructures. Substructure A may consist of 70mm x 70mm rectangular steel beams with a thickness of 10mm, and substructure B may consist of 50mm x 50mm rectangular steel beams with a thickness of 5mm. Three-way translational linear springs and viscous dampers connect the two substructures at four joints. A three-dimensional force is applied to substructure B. The 3D forces can act randomly on substructure B. In addition, the point located on the lower structure A due to this force is the acceleration ROI in three directions. Path contributions can be identified through this experiment. That is, to identify the path contributions of connections to the ROI of substructure A, which has 12 paths from input to output (4 connections and 3 degrees of freedom per connection).

In performance verification, a finite element analysis model can be used to create a training data set. We can focus on the verification of the proposed method because it is assumed that the data set numerically generated from the finite element (FE) model is noise-free and accurate. The self-structured FE model has 154 nodes with beam and spring elements in floating boundary conditions. Using a finite element analysis model, the frequency response of the acceleration at the junction and ROI location is calculated under harmonic excitation with randomly generated amplitude values. The log magnitude of the external force is normally distributed with means and coefficients of variance log ₁₀ (200), log ₁₀ (200), and log ₁₀ (300) N (Newtons), and 0.14, 0.12, and 0.10 in the X, Y, and Z directions. have The force signature for each direction is also randomly determined.

A total of 1000 data sets are prepared for training using the software MSC/NASTRAN used.

Each data set contains the accelerations of two substructures at junction points in three directions and the acceleration of substructure A in three directions at ROI locations. The resolution and bandwidth of the frequency response are 1Hz and 400Hz, respectively.

2) DNN's structure and training

A DNN model capable of predicting ROI and identifying path contributions is constructed for the simplified sub-order problem. In the simplified car body problem, substructure A has 12 inputs through connections. The ROI is a function of the relative acceleration of the joint and the (unknown) dynamic stiffness of the joint spring. Thus, the input is set to the relative acceleration of the three-way junction. Also, the frequency response is a complex number with a real part and an imaginary part. Therefore, when considering the frequency band of the frequency response, there are 9600 nodes (M) for the input layer of the DNN. The output of the DNN is the 3-way acceleration at the ROI point. Thus, similar to the input layer, there are 2400 nodes (N) for the output layer.

[Table 1]

(number of model parameters and GPU time (sec.) for the DNN model in the body problem)

A plurality of hidden layers (H) and the number of related nodes (P) are very important hyperparameters that affect the performance of a DNN model. Construct a combinatorial DNN model to determine hyperparameters and examine the performance results. Table 1 lists selected hyperparameters and their corresponding numbers of model parameters to be determined through deep learning. The hidden layer does not use an activation function, but uses a linear activation function.

The data set prepared before training of the DNN model was finalized through a phase and cross-spectral multiplication process. 20 random phases are created for phase propagation. The cross spectral multiplication process uses 12 input and 3 output responses. Thus, there are a total of 315,000 training data sets, which are calculated by 1000×(20+1)×(12+3). It can be seen that this is expanded by a factor of 315 compared to the original data set. It should be noted that 80% of the data set was used for training and the remaining 20% was used for validation. Standard deviation calculations for normalization used only responses with the training data set.

For the development of the DNN model, Python including the Keras (ver. 2.2.4) library was used in the training and data processing stages. In the training process, the batch size, learning rate, and convergence criterion were set to 64, 0.0001, and 0.0001, respectively. In order to minimize the objective function according to Equation 7, RMSprop was selected as an optimization algorithm. Data processing was performed on an HP840 PC workstation with two CPUs and 196 GB of memory. Training of the DNN model was performed on a Titan V GPU with 12 GB of memory. Table 1 also lists the typical GPU time required to repeat one epoch on the GPU.

In general, the training process converges well with respect to the minimization parameters specified for the DNN model. Figure 7 is a graph showing the iteration history of the training process for the largest DNN model. As in Fig. 7, it shows the iteration history of the largest DNN model, showing 20 hidden layers and 128 nodes per layer.

3) Verification result

We investigate the ability of a DNN model to predict the transmission path of an ROI. The ROI and path contribution to the ROI were predicted using the trained DNN model and compared with reference values. Here, the reference ROI corresponds to the measured ROI. In the case of TPA, the Classical TPA formula in Equation 2 provides a reference value in case the transfer function and dynamic stiffness of the connection are available. Therefore, the transfer function calculated by the aforementioned finite element analysis model and the known dynamic stiffness of the connection are used. Obtain the reference TPA result and compare it with the path contribution obtained from the DNN model.

Seven ROI data sets are prepared for Operational TPA. The ROI data set is a data set containing the ROI and all corresponding input responses for which path contributions are required. The seven ROI data sets correspond to responses when the excitation force amplitudes are μ, μ±σ, μ±2σ, and μ±3σ, where μ and σ are the mean and standard deviation of the force amplitude, respectively. And, an excitation force amplitude far from the average value indicates a low probability that the DNN model has learned a response under this excitation level in the training data set.

8 is a graph of the mean square error of the DNN model at force amplitude μ-3σ. And, Fig. 9 is a graph of the ROI predicted by the DNN model with H = 20 and P = 128 at a force amplitude of μ-3σ compared to the reference value.

The DNN model of FIG. 8 shows the mean square error of the DNN model according to the hidden layer (H) and the number of nodes per hidden layer (P). In Fig. 8, the result is displayed only when the force amplitude is μ-3σ, because the accuracy is the lowest in all cases. As shown in FIG. 8, the number of nodes per hidden layer has a greater effect on the accuracy of the DNN model than the network depth. Because the total number of model parameters in this problem increases rapidly with the number of nodes. This characteristic is due to the fact that the number of nodes in the input and output layers greatly exceeds the corresponding value in the hidden layer. Also, FIG. 8 shows that too many hidden layers can degrade ROI prediction performance. Figure 9 shows that the two responses are almost indistinguishable by comparing the ROI predicted by the DNN model to the reference value.

It can be seen that a few hidden layers, such as 1 or 3, are sufficient to predict the ROI using the DNN model. For this problem, one or three hidden layers with 128 nodes per hidden layer accurately predict the ROI. The phase propagation in the ROI prediction process is also significantly improved.

10 is the reconstruction error of the X-direction ROI according to the sum of the path contributions using the DNN model with P = 128.

Next, the path contribution to the ROI is calculated using the trained DNN model along with the first equation of Equation 9. The reconstruction error of the ROI is estimated based on Equation 10. Figure 10 shows the expected error for the ROI in the X direction. Reconstruction errors of ROIs for other orientations show similar trends. The DNN model according to the embodiment agrees well with the superposition principle of the linear system and shows the linearity of the developed DNN model. Similar to ROI prediction, the reconstruction error grows with network depth (H), while it gets smaller with greater power.

11 is a graph of TPA results during operation for an ROI in the X direction at mean force amplitude.

The exact path contribution was calculated using Equation 2 by the DNN model and used as a reference value. For accurate calculation, accurate values of the transfer function from the input location to the ROI location and the joint stiffness are also required. The transfer function is obtained through finite element analysis of the substructure. 11 shows the Operational TPA result of the DNN model (H=20. P=128) compared separately to the reference Classical TPA for the ROI in the X direction in the entire frequency band. The reference TPA result in FIG. 11(b) shows that in the case of the X-direction ROI, there is no contribution from the Z-direction input at the junction, whereas contributions in the X and Y directions vary with frequency. The reference TPA results in Fig. 11(b) were obtained using a significant amount of additional information, in which case the frequency response function (FRF) and joint stiffness are added to the response during operation. 11(a) shows that the proposed Operational TPA direction can recover the path contribution using only the response during operation. This feature of the DNN model is mainly related to cross-spectral multiplication. It can be seen that the ability to recover the path contribution disappears and the ability to predict the response is maintained by performing the training of the DNN model only through phase propagation.

As shown in FIG. 11 (a), although the level of contribution in the Z direction by Operational TPA is much lower than other levels, it can be seen that the level of contribution in the Z direction by Operational TPA is maintained compared to the case where the reference TPA does not contribute. . Therefore, the Z contribution level in FIG. 11(a) can be a metric representing the error amount of the DNN model in TPA. Comparing the TPA results of the DNN model, it can be seen that the noise level tends to decrease as the network depth increases.

12a to 12c are graphs showing Operational TPA results for an X-direction ROI in average force amplitude.

To investigate the accuracy of individual input contributions, each directional contribution to the same directional ROI at the joint was compared with the reference TPA result. 12a to 12c show comparison results for connection part 1 of FIG. 6 . In Figure 12, only four DNN models are shown for brevity. The DNN model reconstructs the frequency-dependent nature of the path contribution fairly well over the entire frequency band, although the accuracy decreases at higher frequencies in the X and Z directions. It should be emphasized that even large in-motion data sets, i.e. in-motion inputs and outputs, may not contain enough information to accurately recover all path contributions, unlike the data set used in the reference TPA. TPA, as described above, always involves a trade off process between accuracy and time. The proposed operational TPA direction uses the response during operation only at the junction and ROI location.

The Operational TPA result in Fig. 12 is surprising given that the rank of the covariance matrix in Equation 3 with dimensions 12Х12 for this body problem is almost 1. In contrast, the lack of information in the training data set could explain the large variance seen in the high frequencies in Fig. 12c. The actionable performance of Operational TPA based on a DNN model represents the ability of a deep learning process to gather information and synthesize it into the overall features of a data set. In addition, it can be noted that the proposed Operational TA result is not very sensitive to the network depth or the number of nodes.

The response of a linear system is the sum of each path contribution along with magnitude and phase values (i.e. complex numbers). The accuracy of operational TPA is investigated by considering vector summation at a specific frequency. Figure 13 shows the individual path contributions at the peak (275 Hz) and the synthesized ROI vector when the force amplitude is averaged. In FIG. 13(a), thick arrows are ROIs synthesized by the proposed TPA and the reference TPA. The line segments for the synthesized ROI are the corresponding contributions of each path, and the same color represents the same directional component along the junction sequence, while the same line style represents the same path. In FIG. 13(b), the magnitude of the path contribution is also compared. As shown in FIG. 13(a), the two ROIs synthesized by the proposed TPA and the reference TPA have the same size and phase, indicating high accuracy of the proposed method. However, the contribution steps of the individual pathways are not comparable because the input quantities of the two methods are different. In Fig. 13(b), the proposed method according to the embodiment shows that there are four significant paths to the ROI, while the rest have negligible contributions similar to the reference TPA results, but the relative size accuracy is slightly blurry. It should also be noted that TPA aims to identify the relative importance among possible pathways. Therefore, it can be concluded that the proposed TPA method estimates the path contribution fairly well only in motion responses.

In summary, Operational TPA based on the DNN model can well retrieve the path contribution using not only the ROI itself but also the in-motion response of the path. In the numerical verification process, the DNN model training includes the extraction process of transmission path information included in the original 1000 data sets. The number of these initial data sets depends on the resolution and frequency band involved, but in real measurements it takes less than an hour to record. With a smaller data set, phase augmentation can make up for any insufficiency with finer additional phase generation, which greatly improves ROI prediction performance and increases the amount of cross-spectral augmentation.

Embodiments according to the present invention described above may be implemented in the form of program instructions that can be executed through various computer components and recorded on a computer-readable recording medium. The computer readable recording medium may include program instructions, data files, data structures, etc. alone or in combination. Program instructions recorded on the computer-readable recording medium may be specially designed and configured for the present invention, or may be known and usable to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and magneto-optical media such as floptical disks. medium), and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include high-level language codes that can be executed by a computer using an interpreter or the like as well as machine language codes generated by a compiler. A hardware device may be modified with one or more software modules to perform processing according to the present invention and vice versa.

Specific implementations described in the present invention are examples and do not limit the scope of the present invention in any way. For brevity of the specification, description of conventional electronic components, control systems, software, and other functional aspects of the systems may be omitted. In addition, the connection of lines or connecting members between the components shown in the drawings are examples of functional connections and / or physical or circuit connections, which can be replaced in actual devices or additional various functional connections, physical connection, or circuit connections. In addition, if there is no specific reference such as "essential" or "important", it may not necessarily be a component necessary for the application of the present invention.

In addition, the detailed description of the present invention described has been described with reference to preferred embodiments of the present invention, but those skilled in the art or those having ordinary knowledge in the art will find the spirit of the present invention described in the claims to be described later. And it will be understood that the present invention can be variously modified and changed without departing from the technical scope. Therefore, the technical scope of the present invention is not limited to the contents described in the detailed description of the specification, but should be defined by the claims.

Claims (10)

A plurality of sensors installed in the vehicle; and
A computing device that analyzes a transmission path from a vibration or noise source to a plurality of points of interest based on the sensing information from the plurality of sensors;
Providing at least one of auditory, physical and visual feedback to the user based on the analysis result of the transmission path
A system that identifies the path through which noise and vibration generated from the noise and vibration sources of a driving vehicle are transmitted to a point of interest. According to claim 1,
The plurality of sensors include an acceleration sensor
A system that identifies the path through which noise and vibration generated from the noise and vibration sources of a driving vehicle are transmitted to a point of interest. According to claim 2,
The plurality of sensors further include a plurality of touch sensors installed at the plurality of points of interest, and the computing device detects a user on a plurality of areas where the touch sensors are provided among the plurality of points of interest for a predetermined time from an initial driving time. Detects touch information, ranks a plurality of regions according to the number of touches, displays vibration values of each of the plurality of ranked regions, and when the touched region is changed, changes among the plurality of ranked regions Inducing a change in posture of the user by guiding area information having a vibration value lower than that of the touch area
A system that identifies the path through which noise and vibration generated from the noise and vibration sources of a driving vehicle are transmitted to a point of interest. preparing a training data set using input and output frequency responses based on sensing information from a sensor of a vehicle body model;
Proliferating the training data set to train and verify the deep learning neural network; and
Using a trained and verified deep learning neural network to analyze the transmission path during operation to analyze the contribution of the transmission path of the vehicle of noise or vibration; including
A method for determining the path through which noise and vibration generated from the noise and vibration sources of a driving vehicle are transmitted to a point of interest. According to claim 4,
To prepare the data set,
measuring the input and output frequency responses by defining paths, input and output responses;
normalizing the input and output frequency responses; and
multiplying the phases of the normalized input and output frequency responses;
A method for determining the path through which noise and vibration generated from the noise and vibration sources of a driving vehicle are transmitted to a point of interest. According to claim 5,
In the step of multiplying the phase, generating a random phase and shifting the phase of the input and output frequency responses to multiply the phase
A method for determining the path through which noise and vibration generated from the noise and vibration sources of a driving vehicle are transmitted to a point of interest. According to claim 6,
To prepare the data set,
Collecting an entire training data set by multiplying the cross spectrum of the phase-multiplied input and output frequency responses.
A method for determining the path through which noise and vibration generated from the noise and vibration sources of a driving vehicle are transmitted to a point of interest. According to claim 7,
The step of training and verifying the deep learning neural network,
setting deep learning neural network model parameters based on the entire collected training data set;
predicting an interest response data set by performing deep learning neural network training based on the deep learning model parameters; and
Verifying the deep learning neural network based on the difference between the interest response and the interest response predicted by the deep learning neural network;
A method for determining the path through which noise and vibration generated from the noise and vibration sources of a driving vehicle are transmitted to a point of interest. According to claim 8,
Analyzing the contribution,
calculating a path contribution by predicting an off response of any one of the transmission paths; and
Analyzing the path contribution based on the difference between the sum of the contributions of the total input and the interest response predicted by the deep learning neural network;
A method for determining the path through which noise and vibration generated from the noise and vibration sources of a driving vehicle are transmitted to a point of interest. According to claim 4,
analyzing transmission paths of the noise or vibration to a plurality of points of interest based on analysis results of transmission paths of the vehicle; and
Touch information of the user at the plurality of points of interest is detected for a predetermined time from the initial driving of the vehicle, the plurality of regions of interest are ranked according to the number of touches, and the vibration values of each of the plurality of ranked regions of interest are determined. displaying, and inducing a change in posture of the user by guiding information on an area having a lower vibration value than the changed touch area among the plurality of ranked areas of interest when the touched area is changed.
A method for determining the path through which noise and vibration generated from the noise and vibration sources of a driving vehicle are transmitted to a point of interest.

Images