Data set balancing construction method for NVH identification based on time-frequency domain analysis

Abstract

The application discloses a data set balancing construction method and device for NVH identification based on time-frequency domain analysis, a terminal device and a storage medium, and is applied to NVH analysis and diagnosis in the automobile industry. The method comprises the following steps: collecting a preliminary data set, wherein the preliminary data set is obtained based on real vehicle data under different NVH phenomenon occurring working conditions; performing feature extraction and labeling on the collected preliminary data set to obtain a labeled data set; performing time-frequency domain analysis on the labeled data set by combining short-time Fourier transform and Hilbert-Huang transform to obtain correct time-frequency spectrum, and correcting the labeled data set based on the time-frequency spectrum. The application reduces the error of manual operation, solves the problem of unbalanced number of various samples in the data set, and makes the machine learning model trained based on the data set not produce performance tilt problem and have better classification and identification performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of the automobile industry, and in particular to a data set balancing construction method and device for NVH identification based on time-frequency domain analysis, a terminal device and a storage medium. BACKGROUND

[0002] NVH (Noise, Vibration and Harshness, sound, vibration and sound roughness) has always been a key technology in the automotive industry. Current NVH analysis techniques can be divided into two categories: one is a method based on mechanism and drive train modeling, and the other is a method based on signal analysis. Among them, the former method based on modeling can be used for real-time operation, but consumes more computing power; the latter method based on signal analysis cannot be used for real-time operation diagnosis. The data-driven method based on machine learning algorithm only needs to consume a large amount of computing power during training, and consumes less computing power during real-time operation, which is a new technology in the field of NVH analysis and diagnosis. Among them, constructing a data set for training the model is one of the core technologies of this method.

[0003] The proportion and distribution of the number of various samples in the data set affect the performance of the machine learning model. For a classifier model, it will cause the model performance to be tilted to the class with a larger sample proportion, and the classification and identification performance for the minority sample class will be poor. NVH phenomenon in actual working conditions is frequent but short, so the target NVH phenomenon class only accounts for a small part of the collected data, resulting in poor performance of the trained model in identifying the target NVH phenomenon.

[0004] The data set balancing algorithm, also known as the sampling algorithm, is a common method to solve the above problems. However, this method cannot solve the human error in labeling data. Considering that the method used to label the target NVH phenomenon is to mark points in the time domain, this method is relatively rough for high-frequency acquisition sensor data. The time-frequency domain analysis method can be combined to further refine the distribution of data labels in time and reduce human error.

[0005] Current time-frequency domain analysis methods can be divided into three categories: wavelet transform, short-time Fourier transform, and Hilbert-Huang transform. Among them, the wavelet transform method is commonly used to extract new features of data rather than direct analysis, the short-time Fourier transform method can correctly extract the frequency information of the signal, but cannot balance the time resolution and frequency resolution, and the Hilbert-Huang transform method faces the problems of mode aliasing and IMF (Intrinsic Mode Function) component selection (if the correct frequency information of the target NVH phenomenon is not known in advance, it is impossible to know which IMF component to select, which violates the original intention of NVH analysis).

[0006] From the above, the current data set construction method cannot reduce the human error in data labeling, and the object matching degree of NVH phenomenon is not high. Moreover, the current time-frequency domain analysis method has its inherent defects. The short-time Fourier transform method cannot balance the time resolution and frequency resolution, resulting in limited improvement of data label refinement. The Hilbert-Huang transform method has uncertainty in correctly extracting signal frequency information. SUMMARY

[0007] The main purpose of the present application is to provide a data set balanced construction method, device, terminal equipment and storage medium for NVH identification based on time-frequency domain analysis, aiming to reduce the error of human operation, further solve the problem of unbalanced number of various samples in the data set, and improve the classification and identification effect of the classifier model.

[0008] To achieve the above purpose, the embodiment of the present application provides a data set balanced construction method for NVH identification based on time-frequency domain analysis, which is applied to NVH analysis and diagnosis in the automotive industry. The method comprises the following steps:

[0009] Collecting a preliminary data set, wherein the preliminary data set is obtained based on real vehicle data under different NVH phenomenon occurring working conditions;

[0010] Performing feature extraction and labeling on the collected preliminary data set to obtain a labeled data set;

[0011] Performing time-frequency domain analysis on the labeled data set by combining short-time Fourier transform and Hilbert-Huang transform to obtain correct time-frequency spectrum.

[0012] Optionally, the step of performing time-frequency domain analysis on the labeled data set by combining short-time Fourier transform and Hilbert-Huang transform to obtain correct time-frequency spectrum comprises:

[0013] Selecting a target feature signal from the labeled data set to perform short-time Fourier transform to obtain correct frequency information in each time period;

[0014] Performing empirical mode decomposition on the selected target feature signal to obtain a plurality of intrinsic mode function (IMF) components;

[0015] Combining the correct frequency information in each time period, the correct IMF component is obtained from the plurality of IMF components;

[0016] Performing Hilbert transform on the correct IMF component to obtain correct time-frequency spectrum.

[0017] Optionally, the method further comprises:

[0018] Refine and correct the label data set based on the correct time-frequency spectrum to obtain a corrected data set;

[0019] Balance the corrected data set by a balance algorithm to obtain a new data set.

[0020] Optionally, the method further comprises:

[0021] Model training is performed based on the preliminary data set and the new data set respectively using the same classifier algorithm, and the new data set is verified by the classification performance indicators of the two trained models.

[0022] Optionally, the step of performing empirical mode decomposition on the selected target feature signal to obtain a plurality of IMF components comprises:

[0023] Local extreme points of the target feature signal are obtained;

[0024] According to the size of the local extreme points, the local maximum points are taken as a first group of data points, and the local minimum points are taken as a second group of data points;

[0025] Curves are fitted by using a cubic spline interpolation method on the two groups of data points respectively to obtain an upper envelope line corresponding to the first group of data points and a lower envelope line corresponding to the second group of data points;

[0026] The mean values of the upper envelope line and the lower envelope line are calculated;

[0027] The target feature signal is subtracted by the mean values of the upper envelope line and the lower envelope line to obtain a sub-signal;

[0028] It is verified whether the sub-signal is an IMF;

[0029] If it is an IMF, an IMF component is obtained, and a new signal obtained by subtracting the IMF component from the target feature signal is taken as an input of a new round of iteration until it cannot be further segmented or the upper limit of the algorithm defined IMF is reached;

[0030] If it is not an IMF, the sub-signal is taken as an input to repeat the above process.

[0031] Optionally, the balance algorithm is a Borderline SMOTE algorithm.

[0032] Optionally, the step of collecting the preliminary data set further comprises:

[0033] The collected data is preprocessed, and the preprocessing comprises one or more of data cleaning and normalization processing.

[0034] The application further provides a device for constructing a balanced dataset for NVH identification based on time-frequency domain analysis, which comprises the following modules:

[0035] A data acquisition module is configured to acquire a preliminary dataset, wherein the preliminary dataset is obtained based on real vehicle data under different NVH phenomenon occurring working conditions;

[0036] A label module is configured to perform feature extraction and labeling on the acquired preliminary dataset to obtain a labeled dataset;

[0037] An analysis module is configured to perform time-frequency domain analysis on the labeled dataset by combining short-time Fourier transform and Hilbert-Huang transform to obtain correct time-frequency spectrum.

[0038] The application further provides a terminal device, which comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor, wherein the computer program is executed by the processor to implement the method for constructing a balanced dataset for NVH identification based on time-frequency domain analysis.

[0039] The application further provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the method for constructing a balanced dataset for NVH identification based on time-frequency domain analysis.

[0040] This invention proposes a data set balancing construction method, apparatus, terminal device, and storage medium based on time-frequency domain analysis for NVH identification. The method involves collecting a preliminary dataset based on real-vehicle data under different NVH conditions; extracting features and labeling the preliminary dataset to obtain a labeled dataset; and then performing time-frequency domain analysis on the labeled dataset using short-time Fourier transform and Hilbert-Huang transform to obtain the correct time-frequency spectrum. This proposed method for NVH identification by combining Hilbert-Huang transform and short-time Fourier transform in time-frequency domain analysis solves the problem of the inability to simultaneously achieve time and frequency resolution in the short-time Fourier transform algorithm, as well as the selection and verification problem of intrinsic mode functions (IMF) components in the Hilbert-Huang transform algorithm. This allows for more precise labeling of target NVH phenomena when collecting sensor signal data, reducing human error. Furthermore, combining this with the Borderline SMOTE machine learning method can address the issue of imbalanced sample sizes across different classes in the dataset. This ensures that machine learning models trained on this dataset do not suffer from performance skew and achieve better classification and recognition performance. Compared to other dataset construction methods, this approach yields a dataset that better matches the NVH (Noise, Vibration, and Harshness) phenomenon, enabling the training of classifier models with superior classification performance. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the functional modules of the terminal device to which the dataset balancing construction device for NVH identification based on time-frequency domain analysis of the present invention belongs;

[0042] Figure 2 This is a flowchart illustrating an exemplary embodiment of the dataset balancing construction method for NVH identification based on time-frequency domain analysis of the present invention.

[0043] Figure 3 This is a flowchart illustrating another exemplary embodiment of the dataset balancing construction method for NVH identification based on time-frequency domain analysis of the present invention.

[0044] Figure 4 This is a flowchart illustrating yet another exemplary embodiment of the dataset balancing construction method for NVH identification based on time-frequency domain analysis of the present invention.

[0045] Figure 5 This is a schematic diagram of the overall process of an embodiment of the dataset balancing construction method for NVH identification based on time-frequency domain analysis of the present invention.

[0046] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0047] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0048] The main solution of this invention is as follows: A preliminary dataset is collected, based on real-vehicle data under different NVH (Noise, Vibration, and Harshness) phenomena. Features are extracted and labeled from the collected preliminary dataset to obtain a labeled dataset. Time-frequency domain analysis is performed on the labeled dataset using short-time Fourier transform and Hilbert-Huang transform to obtain the correct time-frequency spectrum. This invention combines short-time Fourier transform and Hilbert-Huang transform to compensate for their respective shortcomings. The short-time Fourier transform is first used to calculate the correct frequency information for each time period, which is then used to filter and intervene in the IMF component of the Hilbert-Huang transform. Substituting this into the Hilbert transform yields a more accurate and refined time-frequency spectrum, which can be used to correct the labeled data. Compared to datasets obtained using general dataset construction methods, this reduces the impact of human error. Furthermore, by combining time-frequency domain analysis, the target NVH phenomenon can be further precisely located in the time domain, reducing the coarseness of the labeled data caused by the point-based timing method.

[0049] Technical terms involved in the embodiments of this invention:

[0050] IMF: intrinsic mode function.

[0051] NVH: Noise, Vibration and Harshness.

[0052] The embodiments of the present invention take into account that, in existing related solutions, the current dataset construction methods cannot reduce human error during data labeling and have low matching degree for objects such as NVH phenomena.

[0053] This invention combines time-frequency domain analysis methods to further pinpoint the target NVH phenomenon in the time domain, thereby reducing the coarseness of tag data caused by the dot-matrix timing method.

[0054] Furthermore, current time-frequency domain analysis methods all have their inherent shortcomings. The short-time Fourier transform method cannot simultaneously consider both time resolution and frequency resolution, resulting in limited improvement in the refinement of data labels. The Hilbert-Huang transform method, on the other hand, has uncertainties in accurately extracting signal frequency information.

[0055] The method proposed in this invention combines the two to compensate for their respective shortcomings. First, the correct frequency information for a time period can be calculated using the short-time Fourier transform, which can then be used to filter and intervene in the IMF component in the Hilbert-Huang transform. Substituting this into the Hilbert transform, a more accurate and refined time spectrum can be obtained.

[0056] Specifically, refer to Figure 1 , Figure 1 This diagram illustrates the functional modules of the terminal device to which the time-frequency domain analysis-based NVH identification dataset balancing and construction device of this invention belongs. This time-frequency domain analysis-based NVH identification dataset balancing and construction device can be independent of the terminal device, capable of dynamic networking of streaming media gateways and forwarding of streaming media data. It can be implemented on the terminal device in hardware or software form. The terminal device can be a smartphone, tablet, or other smart mobile terminal, or a network device such as a server.

[0057] In this embodiment, the terminal device to which the time-frequency domain analysis-based NVH identification-oriented dataset balancing construction device belongs includes at least an output module 110, a processor 120, a memory 130, and a communication module 140.

[0058] The memory 130 stores the operating system and a dataset balancing and construction program for NVH identification based on time-frequency domain analysis; the output module 110 can be a display screen, speaker, etc. The communication module 140 can include a WIFI module, a mobile communication module, and a Bluetooth module, etc., and communicates with external devices or servers through the communication module 140.

[0059] In one embodiment, when the time-frequency domain analysis-based dataset balancing construction program for NVH identification in memory 130 is executed by the processor, it performs the following steps:

[0060] A preliminary dataset is collected, which is based on real vehicle data under different NVH (Noise, Vibration, and Harshness) phenomena.

[0061] Feature extraction and labeling are performed on the collected preliminary dataset to obtain a labeled dataset;

[0062] By combining short-time Fourier transform and Hilbert-Huang transform, time-frequency domain analysis was performed on the labeled dataset to obtain the correct time spectrum.

[0063] Furthermore, when the time-frequency domain analysis-based dataset balancing construction program for NVH identification in memory 130 is executed by the processor, it also performs the following steps:

[0064] Target feature signals are selected from the labeled dataset and subjected to short-time Fourier transform to obtain the correct frequency information for each time period;

[0065] Empirical mode decomposition is performed on the selected target feature signal to obtain multiple intrinsic mode function (IMF) components.

[0066] By combining the correct frequency information for each time period, the correct IMF component is obtained from the multiple IMF components;

[0067] The correct IMF components are subjected to Hilbert transform to obtain the correct time spectrum.

[0068] Furthermore, when the time-frequency domain analysis-based dataset balancing construction program for NVH identification in memory 130 is executed by the processor, it also performs the following steps:

[0069] The corrected dataset is balanced using a balancing algorithm to obtain a new dataset.

[0070] Furthermore, when the time-frequency domain analysis-based dataset balancing construction program for NVH identification in memory 130 is executed by the processor, it also performs the following steps:

[0071] The same classifier algorithm was used to train the model on both the initial dataset and the new dataset. The classification performance of the two trained models was then used to validate the new dataset.

[0072] Furthermore, when the time-frequency domain analysis-based dataset balancing construction program for NVH identification in memory 130 is executed by the processor, it also performs the following steps:

[0073] Find the local extrema of the target feature signal;

[0074] Based on the magnitude of the local extreme points, the local maximum points are taken as the first group of data points, and the local minimum points are taken as the second group of data points.

[0075] The curves were fitted using cubic spline interpolation for the two sets of data points to obtain the upper envelope corresponding to the first set of data points and the lower envelope corresponding to the second set of data points.

[0076] Calculate the mean of the upper and lower envelopes;

[0077] Subtract the mean of the upper and lower envelopes from the target feature signal to obtain the sub-signal;

[0078] Verify whether the sub-signal is an IMF;

[0079] If it is an IMF, then an IMF component is obtained. The target feature signal is subtracted from the IMF component to obtain a new signal as the input for a new round of iteration, until it is impossible to continue segmenting or the upper limit of the IMF limit of the algorithm is reached.

[0080] If it is not an IMF, then the above process is repeated with the sub-signal as input.

[0081] Furthermore, when the time-frequency domain analysis-based dataset balancing construction program for NVH identification in memory 130 is executed by the processor, it also performs the following steps:

[0082] The preliminary dataset is preprocessed, including one or more of data cleaning and normalization processes.

[0083] This embodiment utilizes the above-described scheme, combining Short-Time Fourier Transform (SFT) and Hilbert-Huang Transform (HFT) to perform time-frequency domain analysis on the labeled dataset, obtaining the correct time-frequency spectrum. The proposed time-frequency domain analysis-based dataset balancing construction method for NVH identification combines HFT and SFT in time-frequency domain analysis, solving the shortcomings of the SFT algorithm in achieving both time and frequency resolution, and addressing the selection and verification issues of Intrinsic Mode Functions (IMF) components in the HFT algorithm. This allows for more precise labeling of target NVH phenomena during sensor signal data acquisition, reducing human error. Furthermore, combining the Borderline SMOTE machine learning method solves the problem of imbalanced sample sizes in the dataset, ensuring that the machine learning model trained on this dataset does not suffer from performance skew and achieves better classification and identification performance. Compared to other dataset construction methods, this method yields a dataset that better matches NVH phenomena and can train classifier models with superior classification performance.

[0084] Based on, but not limited to, the terminal device architecture described above, embodiments of the method of the present invention are proposed.

[0085] Reference Figure 2 , Figure 2 This is a flowchart illustrating an exemplary embodiment of the dataset balancing construction method for NVH identification based on time-frequency domain analysis of the present invention.

[0086] This embodiment of the time-frequency domain analysis-based dataset balancing construction method for NVH identification can be applied to NVH analysis and diagnosis in the automotive industry. It constructs a dataset for training models in NVH analysis and diagnosis. The time-frequency domain analysis-based dataset balancing construction method for NVH identification includes:

[0087] Step S101: Collect a preliminary dataset, which is obtained based on real vehicle data under different NVH phenomena.

[0088] The execution subject of the method in this embodiment can be a dataset balancing construction device for NVH identification based on time-frequency domain analysis.

[0089] This embodiment takes into account that current dataset construction methods cannot reduce human error during data labeling, resulting in low matching accuracy for objects such as NVH phenomena. This invention combines time-frequency domain analysis methods to further pinpoint the target NVH phenomenon in the time domain, thereby reducing the coarseness of the label data caused by the point-based timing method.

[0090] Specifically, the first step is to collect a preliminary dataset. This preliminary dataset can be created based on real-vehicle data, which is vehicle data under actual operating conditions. For example, experimental plans and site arrangements can be formulated based on the operating conditions under which different NVH phenomena occur, and experiments can be conducted to collect real-vehicle data under these different NVH phenomena conditions.

[0091] As one implementation method, real vehicle data can be collected using experimental acquisition equipment. This involves collecting vehicle-related voltage signals under different NVH (Noise, Vibration, and Harshness) conditions using sensors, then converting the voltage signals using a pulse generator, and finally transmitting the converted data signals to the experimental acquisition equipment to obtain a preliminary dataset.

[0092] The region where the target NVH phenomenon occurs can be obtained by plotting points in the time domain during the experiment.

[0093] The vehicle data in the preliminary dataset may include various related vehicle data, which is not specifically limited in this embodiment. The preliminary dataset is mainly the initial dataset used to train models for NVH analysis and diagnostics in the automotive industry.

[0094] Furthermore, before or after collecting the initial dataset, the data can be preprocessed, such as parsing, cleaning, and normalizing. This results in a processed initial dataset.

[0095] Step S102: Extract features and label the collected preliminary dataset to obtain a labeled dataset;

[0096] After obtaining the preliminary dataset, features are extracted and labels are applied to the data signals in the preliminary dataset to obtain a labeled dataset. This labeled dataset includes the labeled feature signals.

[0097] Step S103: Combine short-time Fourier transform and Hilbert-Huang transform to perform time-frequency domain analysis on the labeled dataset to obtain the correct time spectrum.

[0098] Specifically, this embodiment also takes into account that current time-frequency domain analysis methods all have their inherent defects. The short-time Fourier transform method cannot take into account both time resolution and frequency resolution, resulting in limited improvement in the refinement of data labels, while the Hilbert-Huang transform method has uncertainty in correctly extracting signal frequency information.

[0099] The method proposed in this embodiment combines the two to make up for their respective shortcomings. First, the correct frequency information for a time period can be calculated using the short-time Fourier transform, which can be used to filter and intervene in the IMF component in the Hilbert-Huang transform. Then, the correct and more refined time spectrum can be obtained by substituting it into the Hilbert transform.

[0100] Specifically, as one implementation, the step of performing time-frequency domain analysis on the labeled dataset using both short-time Fourier transform and Hilbert-Huang transform to obtain the correct time-frequency spectrum may include:

[0101] First, target feature signals are selected from the labeled dataset and subjected to short-time Fourier transform to obtain the correct frequency information for each time period;

[0102] Then, empirical mode decomposition is performed on the selected target feature signal to obtain multiple IMF components;

[0103] Specifically, the step of performing empirical mode decomposition on the selected target feature signal to obtain multiple IMF components includes:

[0104] Find the local extrema of the target feature signal;

[0105] Based on the magnitude of the local extreme points, the local maximum points are taken as the first group of data points, and the local minimum points are taken as the second group of data points.

[0106] The curves were fitted using cubic spline interpolation for the two sets of data points to obtain the upper envelope corresponding to the first set of data points and the lower envelope corresponding to the second set of data points.

[0107] Calculate the mean of the upper and lower envelopes;

[0108] Subtract the mean of the upper and lower envelopes from the target feature signal to obtain the sub-signal;

[0109] Verify whether the sub-signal is an IMF;

[0110] If it is an IMF, then an IMF component is obtained. The target feature signal is subtracted from the IMF component to obtain a new signal as the input for a new round of iteration, until it is impossible to continue segmenting or the upper limit of the IMF limit of the algorithm is reached.

[0111] If it is not an IMF, then the above process is repeated with the sub-signal as input.

[0112] Then, by combining the correct frequency information for each time period, the correct IMF component is obtained from the multiple IMF components;

[0113] Finally, the correct IMF components are subjected to Hilbert transform to obtain the correct time spectrum.

[0114] The time spectrum obtained at this time not only solves the problems of incorrect analytical frequency of Hilbert-Huang transform and selection of IMF components, but also solves the accuracy problem of short-time Fourier transform. It can be used to further refine the occurrence area of ​​target NVH phenomenon and revise the label dataset.

[0115] This embodiment, through the above-described scheme, specifically involves collecting a preliminary dataset based on real-vehicle data under different NVH (Noise, Vibration, and Harshness) conditions; extracting features and labeling the collected preliminary dataset to obtain a labeled dataset; and performing time-frequency domain analysis on the labeled dataset using short-time Fourier transform and Hilbert-Huang transform to obtain the correct time-frequency spectrum. The proposed time-frequency domain analysis-based dataset balancing construction method for NVH identification combines the Hilbert-Huang transform with the short-time Fourier transform in time-frequency domain analysis. This solves the shortcomings of the short-time Fourier algorithm, which cannot simultaneously achieve both time and frequency resolution, and the selection and verification problems of the IMF component in the Hilbert-Huang transform algorithm. It can further refine the labeling time point of the target NVH phenomenon when collecting sensor signal data, reducing human error. Compared with other dataset construction methods, this method can obtain a dataset that better matches the NVH phenomenon and can train a classifier model with better classification performance.

[0116] Reference Figure 3 , Figure 3 This is a flowchart illustrating another exemplary embodiment of the dataset balancing construction method for NVH identification based on time-frequency domain analysis of the present invention.

[0117] Based on the above Figure 2 In the embodiment shown, the method further includes:

[0118] Step S104: Based on the correct time spectrum, refine and correct the label dataset to obtain the corrected dataset;

[0119] Specifically, after obtaining the correct time spectrum, the occurrence area of ​​the target NVH phenomenon can be further refined based on the correct time spectrum, and the label dataset can be revised to obtain the corrected dataset.

[0120] This embodiment combines time-frequency domain analysis methods to further pinpoint the target NVH phenomenon in the time domain, thereby reducing the coarseness of tag data caused by the dot-matrix timing method.

[0121] Step S105: The corrected dataset is balanced using a balancing algorithm to obtain a new dataset.

[0122] After refining and correcting the labeled dataset using the Hilbert-Huang transform, the corrected dataset is then balanced using a balancing algorithm. This resolves the imbalance in the number of samples across different categories, ensuring that the machine learning model trained on this dataset does not suffer from performance bias and achieves better classification and recognition performance.

[0123] Specifically, as one implementation method, the balancing algorithm can be the Borderline SMOTE balancing algorithm, which is used to balance the corrected dataset.

[0124] The Borderline SMOTE algorithm is a new algorithm that takes into account the inherent defects of the SMOTE algorithm and improves upon it. Instead of using all samples of the minority class, it only uses sample points on the boundary within the selected minority class samples to simulate the generation of new samples.

[0125] The Borderline SMOTE algorithm's sampling process involves first dividing minority class samples (in a binary classification context, or each class with a lower distribution proportion in a multi-class problem) into three categories: Safe, Danger, and Noise.

[0126] Safe: Selected minority class sample points, whose surrounding more than half are samples of the same class;

[0127] Danger: Selected minority class sample points, whose surrounding more than half are majority class samples. These are the sample points on the boundary that distinguishes the selected target class from other classes.

[0128] Noise: The target sample is completely surrounded by majority class samples and is considered noise.

[0129] Ultimately, the Borderline SMOTE algorithm generates new samples by oversampling only the Danger class within the minority class. On the surface, it seems that the already limited minority class samples are not being fully utilized, but in reality, it reduces the aliasing of new samples and has a particularly good effect on samples with high information repetition rates.

[0130] As mentioned earlier, the proportion and distribution of samples of each class in the dataset affect the performance of machine learning models. For classifier models, this can lead to a bias in model performance towards the class with a larger sample proportion, resulting in poor classification and identification performance for the minority of sample classes. In actual working conditions, NVH phenomena occur frequently but briefly, thus the target NVH phenomenon class only accounts for a small portion of the collected data, leading to poor identification performance of the trained model for the target NVH phenomenon.

[0131] The solution in this embodiment can obtain a new dataset with consistent proportions of various internal sample types, thus solving the problem of classifier model performance bias towards a certain majority class.

[0132] This invention proposes a dataset balancing construction method for NVH identification based on time-frequency domain analysis. It combines the Hilbert-Huang Transform and Short-Time Fourier Transform in time-frequency domain analysis, solving the problem of the inability to simultaneously achieve time and frequency resolution in the Short-Time Fourier Transform algorithm, as well as the selection and verification issues of the IMF component in the Hilbert-Huang Transform algorithm. This allows for more precise labeling of target NVH phenomena during sensor signal data acquisition, reducing human error. Furthermore, by incorporating the Borderline SMOTE machine learning method, it addresses the imbalance of sample sizes across different data types, ensuring that machine learning models trained on this dataset do not suffer from performance bias and achieve better classification and identification performance. Compared to other dataset construction methods, this method yields a dataset that better matches NVH phenomena and can train classifier models with superior classification performance. Simultaneously, it reduces the impact of human error on dataset construction.

[0133] Reference Figure 4 , Figure 4 This is a flowchart illustrating yet another exemplary embodiment of the dataset balancing construction method for NVH identification based on time-frequency domain analysis of the present invention.

[0134] Based on the above Figure 3 In the embodiment shown, the method further includes:

[0135] Step S106: Train the model using the same classifier algorithm based on the initial dataset and the new dataset respectively, and verify the new dataset by the classification performance metrics of the two trained models.

[0136] Specifically, different datasets constructed before and after the correction can be trained using the same classifier algorithm, and the superiority of the dataset construction method in this embodiment can be verified by the classification performance indicators of the two models.

[0137] This invention proposes a dataset balancing construction method for NVH identification based on time-frequency domain analysis. It combines the Hilbert-Huang Transform and Short-Time Fourier Transform in time-frequency domain analysis, solving the problem of the inability to simultaneously achieve time and frequency resolution in the Short-Time Fourier Transform algorithm, as well as the selection and verification issues of the IMF component in the Hilbert-Huang Transform algorithm. This allows for more precise labeling of target NVH phenomena during sensor signal data acquisition, reducing human error. Furthermore, by incorporating the Borderline SMOTE machine learning method, it addresses the imbalance of sample sizes across different data types, ensuring that machine learning models trained on this dataset do not suffer from performance bias and exhibit better classification and identification performance. Compared to other dataset construction methods, this approach yields a dataset that better matches NVH phenomena and allows for the training of classifier models with superior classification performance.

[0138] The following combination Figure 5 The overall scheme of the dataset balancing construction method for NVH identification based on time-frequency domain analysis in this invention is described in detail.

[0139] Reference Figure 5 As shown, Figure 5 This is a schematic diagram of the overall process of an embodiment of the dataset balancing construction method for NVH identification based on time-frequency domain analysis of the present invention.

[0140] The technical problem this embodiment aims to solve is to provide a dataset construction method for NVH (Noise, Vibration, and Harshness) phenomena that reduces human error and addresses dataset imbalance. While current data-driven methods are novel for NVH analysis and diagnosis, dataset construction, as a core technology within these methods, remains largely unexplored. Compared to other more common dataset construction methods, the method in this embodiment is better suited to NVH phenomena, achieving dataset balancing while minimizing the impact of human error on dataset construction.

[0141] The overall implementation process of this invention is as follows: Figure 5 As shown, the implementation includes: data processing and acquisition, short-time Fourier transform, Hilbert-Huang transform-based correction of labeled data, balanced dataset, and model validation. The specific implementation process is as follows:

[0142] (1): such as Figure 5 As shown, step 1 requires obtaining real vehicle data to create a preliminary dataset.

[0143] The experimental plan and site arrangement can be formulated according to the different operating conditions in which NVH phenomena occur, and the experiments can be carried out. The experimental data acquisition equipment is Rotec. Since the voltage signal output by the sensor cannot be directly read by the acquisition equipment, it is necessary to convert the data through a pulse generator. The occurrence area of ​​the target NVH phenomenon is obtained by marking points in the time domain when it is sensed during the experiment.

[0144] The collected data needs to be parsed, cleaned, and normalized to complete the data preprocessing and form a preliminary dataset.

[0145] In addition, feature extraction and labeling of the data signals in the preliminary dataset are required to obtain a labeled dataset after feature extraction and labeling.

[0146] (2): The dataset D was obtained after step 1, which serves as the basis for step 2.

[0147] First, the target feature signal is selected for Short-Time Fourier Transform (STFT), and its core formula is as follows:

[0148]

[0149] Where s(τ) is the given time-domain signal, h(t) is the selected window function, h(τ-t) represents the "analysis window" centered at t, and ω is the signal acquisition frequency.

[0150] After short-time Fourier transform, the correct frequency information for each time period can be obtained.

[0151] (3): such as Figure 5 As shown in step 3, this step has two input sources.

[0152] First, Empirical Mode Decomposition (EMD) is performed on the selected target feature signal (original signal). The EMD algorithm can be divided into the following four steps:

[0153] 1) Find the local extreme points of the original signal, group the local maxima into one group and the local minima into another group, and fit the curves using cubic spline interpolation on the two groups of data points respectively. The group of local maxima points corresponds to the upper envelope, and the group of local minima points corresponds to the lower envelope.

[0154] 2) Calculate the mean of the upper and lower envelopes;

[0155] 3) Subtract the mean of the upper and lower envelopes from the original signal to obtain the sub-signal, and verify whether it is an IMF.

[0156] 4) If it is an IMF, an IMF component is obtained. The new signal obtained by subtracting this component from the original signal is used as the input for the new round of iteration (that is, the new signal is used as the original signal and the above steps 1)-4) are repeated until it is impossible to continue to divide or the upper limit of the IMF limit of the algorithm is reached. The remaining part after division is called the residual component, which is used to describe the gradual change trend of the signal. The IMF is used to describe the oscillation changes in different frequency bands.

[0157] If it is not IMF, then repeat the above process for the sub-signal obtained in 3). The difference from the previous branch is that the input for the new iteration is different. One is the original signal minus the sub-signal that meets the IMF condition, and the other is the sub-signal that does not meet the IMF condition. This is a binary recursive process.

[0158] After obtaining multiple IMF components through empirical mode decomposition, the correct IMF component can be found by combining the frequency information of each time period obtained in step 2, and then substituted into the Hilbert transform to obtain the correct time spectrum.

[0159] The core formula of the Hilbert transform is as follows:

[0160]

[0161]

[0162] Here, x(t) is a given time-domain signal, and the Hilbert transform can be viewed as the convolution of x(t) and the function h(t); h(t) is the selected window function, and h(τ-t) represents the "analysis window" centered at t.

[0163] At this point, the obtained time spectrum not only solves the problems of incorrect analytical frequency of Hilbert-Huang transform and selection of IMF components, but also solves the accuracy problem of short-time Fourier transform. It can be used to further refine the occurrence area of ​​target NVH phenomenon and revise the label dataset.

[0164] (4): Figure 5 The corrected dataset obtained in step 3 can be balanced using the Borderline SMOTE method.

[0165] The Borderline SMOTE algorithm is a new algorithm that takes into account the inherent defects of the SMOTE algorithm and improves upon it. Instead of using all samples of the minority class, it only uses sample points on the boundary within the selected minority class samples to simulate the generation of new samples.

[0166] The sampling process of the Borderline SMOTE algorithm is as follows: First, the minority class samples (in a binary classification context, or each class with a lower distribution proportion in a multi-class problem) are divided into three classes: Safe, Danger, and Noise.

[0167] Safe: Selected minority class sample points, whose surrounding more than half are samples of the same class;

[0168] Danger: Selected minority class sample points, whose surrounding more than half are majority class samples. These are the sample points on the boundary that distinguishes the selected target class from other classes.

[0169] Noise: The target sample is completely surrounded by majority class samples and is considered noise.

[0170] Ultimately, the Borderline SMOTE algorithm uses only the Danger class within the minority class samples to oversample and generate new samples. On the surface, it seems that the already limited number of minority class samples are not fully utilized, but in reality, it reduces the aliasing of new samples and has a particularly good effect on samples with high information repetition rates.

[0171] This step yields a new dataset with consistent proportions of samples from all classes, thus resolving the issue of the classifier model's performance being biased towards a particular majority class.

[0172] (5): Figure 5 In step 5, different datasets constructed based on the old and new methods are used to train the models using the same classifier algorithm. The superiority of the new method is verified by the classification performance metrics of the two models.

[0173] As shown above, the method in this embodiment combines the Hilbert-Huang transform and the short-time Fourier transform in time-frequency domain analysis, solving the problem of the inability to simultaneously achieve time and frequency resolution in the short-time Fourier algorithm, as well as the selection and verification issues of the IMF component in the Hilbert-Huang transform algorithm. This allows for more precise labeling of the target NVH phenomenon when acquiring sensor signal data, reducing errors from human operation. Furthermore, by combining this with the Borderline SMOTE machine learning method, the problem of imbalanced sample sizes in the dataset can be addressed, ensuring that the machine learning model trained on this dataset does not suffer from performance bias and has better classification and identification performance. Compared to other dataset construction methods, this method yields a dataset that better matches NVH phenomena and can train classifier models with better classification performance.

[0174] Furthermore, embodiments of the present invention also propose a dataset balancing construction device for NVH identification based on time-frequency domain analysis, the device comprising:

[0175] The data acquisition module is used to collect a preliminary dataset, which is obtained based on real vehicle data under different NVH (Noise, Vibration, and Harshness) conditions.

[0176] The labeling module is used to extract features and label the initial collected dataset to obtain a labeled dataset.

[0177] The analysis module is used to perform time-frequency domain analysis on the labeled dataset by combining short-time Fourier transform and Hilbert-Huang transform to obtain the correct time spectrum.

[0178] The principle and implementation process of the dataset balancing construction for NVH identification based on time-frequency domain analysis in this embodiment are described in the above embodiments and will not be repeated here.

[0179] Furthermore, this embodiment of the invention also proposes a terminal device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the dataset balancing construction method for NVH identification based on time-frequency domain analysis as described in any of the above embodiments.

[0180] Since the dataset balancing construction program for NVH identification based on time-frequency domain analysis adopts all the technical solutions of all the aforementioned embodiments when executed by the processor, it has at least all the beneficial effects brought by all the technical solutions of all the aforementioned embodiments, which will not be repeated here.

[0181] Furthermore, embodiments of the present invention also propose a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the dataset balancing construction method for NVH identification based on time-frequency domain analysis as described in any of the above embodiments.

[0182] Since the dataset balancing construction program for NVH identification based on time-frequency domain analysis adopts all the technical solutions of all the aforementioned embodiments when executed by the processor, it has at least all the beneficial effects brought by all the technical solutions of all the aforementioned embodiments, which will not be repeated here.

[0183] Compared to existing technologies, the dataset balancing construction method for NVH identification based on time-frequency domain analysis proposed in this invention combines the short-time Fourier transform method and the Hilbert-Huang transform method, eliminating the defects of both and obtaining a more refined and accurate time-frequency spectrum, which can be used to correct label data. Compared with the datasets obtained by general dataset construction methods, the datasets reduce the impact of human error. On the basis of using the same dataset balancing method, the classifier model trained with this dataset has better classification and identification performance, with at least a 10% improvement in classification performance.

[0184] Moreover, compared to other dataset construction methods, this method does not rely solely on dataset balancing algorithms. Instead, it combines time-frequency analysis to obtain time-spectrum information before balancing, further refining and revising the label data to obtain a dataset that better matches NVH phenomena. This invention is applicable to fields where data-driven methods are used for the analysis and identification of NVH phenomena.

[0185] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0186] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0187] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, controlled terminal, or network device, etc.) to execute the methods of each embodiment of the present invention.

[0188] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A data set balancing construction method for NVH identification based on time-frequency domain analysis, characterized in that, The method for NVH analysis and diagnosis applied in the automotive industry includes the following steps: A preliminary dataset is collected, which is based on real vehicle data under different NVH (Noise, Vibration, and Harshness) phenomena. Feature extraction and labeling are performed on the collected preliminary dataset to obtain a labeled dataset; By combining short-time Fourier transform and Hilbert-Huang transform, time-frequency domain analysis was performed on the labeled dataset to obtain the correct time spectrum; The steps for performing time-frequency domain analysis on the labeled dataset using short-time Fourier transform and Hilbert-Huang transform to obtain the correct time-frequency spectrum include: Target feature signals are selected from the labeled dataset and subjected to short-time Fourier transform to obtain the correct frequency information for each time period; Empirical mode decomposition is performed on the selected target feature signal to obtain multiple intrinsic mode function (IMF) components. By combining the correct frequency information for each time period, the correct IMF component is obtained from the multiple IMF components; The correct IMF components are subjected to Hilbert transform to obtain the correct time spectrum.

2. The method according to claim 1, characterized in that, The method further includes: Based on the correct time spectrum, the labeled dataset is refined and corrected to obtain the corrected dataset; The corrected dataset is balanced using a balancing algorithm to obtain a new dataset.

3. The method according to claim 2, characterized in that, The method further includes: The same classifier algorithm was used to train the model on both the initial dataset and the new dataset. The classification performance of the two trained models was then used to validate the new dataset.

4. The method according to any one of claims 2-3, characterized in that, The step of performing empirical mode decomposition on the selected target feature signal to obtain multiple IMF components includes: Find the local extrema of the target feature signal; Based on the magnitude of the local extreme points, the local maximum points are taken as the first group of data points, and the local minimum points are taken as the second group of data points. The curves were fitted using cubic spline interpolation for the two sets of data points to obtain the upper envelope corresponding to the first set of data points and the lower envelope corresponding to the second set of data points. Calculate the mean of the upper and lower envelopes; Subtract the mean of the upper and lower envelopes from the target feature signal to obtain the sub-signal; Verify whether the sub-signal is an IMF; If it is an IMF, then an IMF component is obtained. The target feature signal is subtracted from the IMF component to obtain a new signal as the input for a new round of iteration, until it is impossible to continue segmenting or the upper limit of the IMF limit of the algorithm is reached. If it is not an IMF, then the above process is repeated with the sub-signal as input.

5. The method according to claim 2, characterized in that, The balancing algorithm is the Borderline SMOTE algorithm.

6. The method according to any one of claims 1-3, characterized in that, The step of collecting the initial dataset also includes: The collected data is preprocessed, including one or more of data cleaning and normalization processes.

7. A dataset balancing and construction device for NVH identification based on time-frequency domain analysis, characterized in that, The device includes: The data acquisition module is used to collect a preliminary dataset, which is obtained based on real vehicle data under different NVH (Noise, Vibration, and Harshness) conditions. The labeling module is used to extract features and label the initial collected dataset to obtain a labeled dataset. The analysis module is used to perform time-frequency domain analysis on the labeled dataset by combining short-time Fourier transform and Hilbert-Huang transform to obtain the correct time spectrum, specifically including: Target feature signals are selected from the labeled dataset and subjected to short-time Fourier transform to obtain the correct frequency information for each time period; Empirical mode decomposition is performed on the selected target feature signal to obtain multiple intrinsic mode function (IMF) components. By combining the correct frequency information for each time period, the correct IMF component is obtained from the multiple IMF components; The correct IMF components are subjected to Hilbert transform to obtain the correct time spectrum.

8. A terminal device, characterized in that, The terminal device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the dataset balancing construction method for NVH identification based on time-frequency domain analysis as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the dataset balancing construction method for NVH identification based on time-frequency domain analysis as described in any one of claims 1-6.

Images