KR-20260067805-A - Method for selecting training data for training an artificial neural network model and device therefor

KR20260067805AKR 20260067805 AKR20260067805 AKR 20260067805AKR-20260067805-A

Abstract

In various embodiments, a method and apparatus for selecting training data for a defect inspection model trained to classify whether vibration data related to a vehicle is defective based on an artificial neural network are disclosed. The method and apparatus include: a step of performing a rule-based inspection to classify whether input vibration data is defective based on predefined rules; a step of inputting the input vibration data into the defect inspection model when the input vibration data is classified as normal or defective by the rule-based inspection; and a step of determining whether to select the input vibration data as training data by comparing the result of the rule-based inspection with the classification information of the defect inspection model.

Inventors

윤재식
우성훈

Assignees

현대모비스 주식회사

Dates

Publication Date: 20260513
Application Date: 20241106

Claims (15)

A method for selecting training data for a defect inspection model trained to classify whether vibration data related to a vehicle is defective based on an artificial neural network, wherein a computing device A step of performing a rule-based inspection to classify whether input vibration data is defective based on predefined rules; If the input vibration data is classified as normal or defective by the above rule-based inspection, the step of inputting the input vibration data into the defect inspection model; and A method comprising the step of determining whether to select the input vibration data as training data by comparing the result of the rule-based inspection and the classification information of the defect inspection model.
In paragraph 1, A method characterized by selecting the vibration data as training data with label information corresponding to the result of the rule-based inspection, based on the fact that the result of the rule-based inspection for the vibration data and the classification information of the defect inspection model are different.
In paragraph 1, A method characterized in that the vibration data is not selected as training data based on the fact that the result of the rule-based inspection of the vibration data matches the classification information of the defect inspection model.
In paragraph 1, The step of performing the above rule-based inspection is, A step of converting the input vibration data into order data based on the number of events occurring per unit rotation; and A method characterized by further including the step of calculating the number of peaks having vibration magnitudes greater than a specific threshold and the total sum of the vibration magnitudes of the peaks in the above order data.
In paragraph 4, The step of converting the above order data further includes the step of correcting the input vibration data by moving a time window having a predetermined size; A method characterized in that the input vibration data is corrected based on the value obtained by subtracting the minimum vibration value from the maximum vibration value within the time window.
In paragraph 4, The above rule-based inspection classifies the input vibration data as normal based on the fact that the number of peaks is less than a first threshold number and the total is less than a first threshold size, and A method characterized by classifying the input vibration data as defective based on the fact that the number of the above peaks exceeds a second threshold number and the total sum of the vibration values of the above pits exceeds a second threshold size.
In paragraph 4, The above rule-based inspection withholds classification of the vibration data based on the fact that the number of peaks is greater than or equal to the first threshold number and less than the second threshold number, and A method characterized in that vibration data for which the above classification is withheld is not selected as the above training data.
In paragraph 4, The step of converting the above order data is a step of correcting the above vibration data into first vibration data for constant velocity motion based on a linear interpolation method based on the vibration data obtained from a rotating body undergoing accelerated motion; and A method characterized by further including the step of converting the first vibration data into the order data.
In paragraph 1, A step of determining a specific time based on a driving pattern for the above vehicle; and A method characterized by further including the step of updating the defect inspection model based on the learning data, based on the measured time information corresponding to the specific time.
In paragraph 1, A step of generating a plurality of augmented vibration data by applying a plurality of augmentation rules to the above input vibration data; A step of inputting the plurality of augmented vibration data into the defect inspection model trained to predict the augmentation rule applied to the augmented vibration data; and The method further includes the step of obtaining an outlier value for the input vibration data calculated based on the prediction accuracy of the plurality of augmentation rules applied to the plurality of augmentation data from the defect inspection model. A method in which classification information of the above defect inspection model is determined based on the above outlier value.
In a computing device for selecting training data for a defect inspection model trained to classify whether vibration data related to a vehicle is defective based on an artificial neural network, processor; and It includes at least one memory that is operablely connected to the above processor and, when executed, causes the at least one processor to perform an operation, and the operation is, A computing device comprising the operation of performing a rule-based inspection to classify whether input vibration data is defective based on predefined rules, inputting the input vibration data into a defect inspection model when the input vibration data is classified as normal or defective by the rule-based inspection, and determining whether to select the input vibration data as training data by comparing the result of the rule-based inspection with the classification information of the defect inspection model.
In Paragraph 11, A computing device characterized by the fact that, based on the fact that the result of the rule-based inspection of the vibration data and the classification information of the defect inspection model are different, the vibration data is selected as training data with label information corresponding to the result of the rule-based inspection set and stored in a database.
In Paragraph 11, A computing device characterized in that the vibration data is not selected as training data based on the fact that the result of the rule-based inspection of the vibration data and the classification information of the defect inspection model are identical.
In Paragraph 11, A computing device characterized by further including, in the above operation, an operation in which at least one processor converts the input vibration data into order data based on the number of events occurring per unit rotation, and calculates the number of peaks having a vibration magnitude greater than a specific threshold in the order data and the total sum of the vibration magnitudes of the peaks.
In Paragraph 14, The above operation further includes the operation of correcting the input vibration data by moving a time window having a predetermined size and converting the corrected input vibration data into the order data. A computing device characterized in that the above input vibration data is corrected based on the value obtained by subtracting the minimum vibration value from the maximum vibration value within the time window.

Description

Method for selecting training data for training an artificial neural network model and device therefor This invention relates to a method for selecting training data from vibration data generated in a vehicle to train an AI model, and a device for this purpose. NVH (Noise, Vibration and Harshness) for vehicles is a term referring to the various phenomena of vibration and noise in automobiles. Noise refers to loud sounds expressed in decibels (dB) that are unpleasant to human emotions, and is broadly classified into interior noise generated by vehicle parts and external noise generated from outside the vehicle. Vibration refers to the phenomenon in which a vehicle body shakes at regular intervals, and is caused by the repeated transfer of kinetic and potential energy. Vibration can broadly include internal vibrations caused by the operation of internal components such as the engine, and external vibrations transmitted to the vehicle through the body, tires, and suspension from friction with the road surface, wind, etc. Harshness refers to noise and vibration caused by irregular impacts, such as when passing over speed bumps on roads or railway tracks. NVH is an important factor in determining the emotional quality of a vehicle. While past NVH research aimed simply to reduce sound and vibration to create a quiet car, recent research has evolved in the direction of creating sound frequencies or waveforms that provide pleasure to the driver by producing emotional and high-quality sound. Conventional NVH tests for noise detection use a rule-based methodology that identifies and classifies specific frequency components according to predefined rules or conditions. The drawings attached to this specification are intended to provide an understanding of the present invention, to illustrate various embodiments of the invention, and to explain the principles of the invention together with the description in the specification. Figures 1 and 2 are diagrams illustrating a method for analyzing NVH for vibration data occurring in a vehicle. Figures 3 and 4 are diagrams illustrating a method for actively training an AI model based on an artificial neural network. Figures 5 to 8 are diagrams illustrating a method for performing rule-based inspection based on order analysis. Figures 9 to 11 are drawings illustrating an improvement method for improving order analysis. FIG. 12 is a block diagram briefly illustrating a computing device according to one embodiment. Figures 13 and 14 are diagrams illustrating a method for pre-processing vibration data input to a defect inspection model based on an artificial neural network. Figure 15 is a diagram illustrating a method for training a defect inspection model. Figures 16 and 17 are diagrams illustrating a method for obtaining outlier values by inputting vibration data into a defect inspection model. The following detailed description of the present invention refers to the accompanying drawings, which illustrate specific embodiments in which the present invention can be practiced in order to clarify the objects, technical solutions, and advantages of the present invention. These embodiments are described in sufficient detail to enable a person skilled in the art to practice the present invention. Throughout the detailed description and claims of this specification, "learning" or "training" refers to the performance of machine learning through procedural computing and is not intended to refer to mental processes such as human educational activities; furthermore, "training" is used in the generally accepted sense regarding machine learning. For example, "deep learning" refers to machine learning using deep artificial neural networks. A deep neural network is a machine learning model that automatically learns the characteristics of each data point by training a large amount of data within a structure composed of multi-layered artificial neural networks, and proceeds with learning in a manner that minimizes the error of the objective/loss function, i.e., classification accuracy. It is capable of extracting and classifying features at various levels, ranging from low-level features such as points, lines, and surfaces to complex and meaningful high-level features. And throughout the detailed description and claims of this specification, the word “comprising” and its variations are not intended to exclude other technical features, additions, components, or steps. Also, “one” or “one” is used in the sense of one or more, and “another” is limited to at least a second. Other objects, advantages, and characteristics of the present invention will become apparent to a person skilled in the art, in part from this specification and in part from the practice of the present invention. The following examples and drawings are provided as examples and are not intended to limit the invention. Accordingly, details disclosed in this specification regarding specific structures or functions should not be inter