Intelligent Detection Method for Car Noises Based on Bionic Auditory Perception Features

Abstract

This invention proposes an intelligent detection method for automotive abnormal noises based on biomimetic auditory perception features, comprising the following steps: Step 1: Converting and framing pre-collected automotive abnormal noise audio database samples to obtain audio signals for each frame in a unified format; Step 2: Calculating the biomimetic auditory perception features of each frame of audio signal, including loudness, sharpness, and total roughness, and weighted summing the loudness, sharpness, and total roughness to obtain a new biomimetic auditory perception feature, denoted as LSR; Step 3: Constructing a multi-model ensemble detection network integrating MLP-Mixer and Gradient Boosting Decision Tree (GBDT), inputting the biomimetic auditory perception feature LSR calculated in Step 2 into the multi-model ensemble detection network to obtain the probability that each frame of audio is an abnormal noise. This invention, by extracting and fusing biomimetic auditory perception features from audio data, avoids the problem of single feature extraction in manual hearing, improves detection efficiency, and reduces reliance on experience.

Description

Technical Field

[0001] This invention relates to the fields of intelligent acoustic detection and automotive NVH technology, specifically to an intelligent detection method for automotive abnormal noises based on biomimetic auditory perception features. Background Technology

[0002] In recent years, the lightweighting of automobiles has significantly increased the probability of abnormal noises. Furthermore, with advancements in NVH (Noise, Vibration and Harness) control technology and the widespread adoption of electric vehicles, conventional interior noise has been significantly reduced, making abnormal noise problems even more prominent. In-vehicle abnormal noise is a key factor affecting ride comfort and consumer satisfaction, and detecting the source of abnormal noise is a crucial prerequisite for controlling it.

[0003] The sources of abnormal noises inside a car are very diverse. They are mainly caused by loosening, friction, vibration, or aging of various structural components, decorative parts, and electrical equipment inside and outside the vehicle due to conditions such as driving, starting, accelerating, decelerating, or traversing bumpy roads. From an acoustic perspective, abnormal noises inside a car can be roughly divided into strong-toned, high-frequency "whistling sounds," such as the airflow sound caused by air leakage from the vents; and non-toned "friction sounds" and "collision sounds," such as the knocking sound caused by gaps between the door panel and the door. Some abnormal noises are also related to the resonance characteristics of vehicle components. For example, at certain specific frequencies, interior parts or motors may couple with the vehicle body structure to produce resonant abnormal noises such as "humming" or "buzzing."

[0004] To eliminate abnormal noises in automobiles, the source of the noise must first be located before a targeted solution can be developed. However, detecting the source of abnormal noises in automobiles is very difficult, mainly for the following reasons: First, the energy of the abnormal noise signal is relatively low, easily submerged in noise signals, and difficult to separate; second, abnormal noise is a highly nonlinear physical phenomenon, with no regularity in its characteristics and duration, making data processing of short-lived, random abnormal noise signals difficult. Currently, in engineering applications, the detection of abnormal noise sources still relies mainly on subjective evaluation, that is, entirely depending on the engineer's human ear to listen and judge the sound source. Subjective evaluation methods have shortcomings such as limited feature extraction, poor environmental adaptability, and over-reliance on human judgment. They require a high level of experience from engineers, involve a large amount of experimental work, and are prone to disputes for abnormal noise problems that are difficult to determine. Developing objective detection and location technology for abnormal noise sources has become an urgent need in the industry. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes an intelligent detection method for automotive abnormal noises based on biomimetic auditory perception features. By extracting and fusing biomimetic auditory perception features from audio data, including loudness, sharpness, and total roughness, a comprehensive characterization of abnormal noise signals is achieved. Simultaneously, the biomimetic auditory perception features are fitted using an MLP-Mixer deep neural network, a LightGBM model, an XGBoost model, and a CatBoost model gradient boosting tree model to obtain the probability that each audio frame is an abnormal noise. Compared to traditional subjective evaluation methods, this patent avoids the problem of single feature extraction in human auditory perception by extracting and fusing biomimetic auditory perception features from audio data. Through deep learning and machine learning, it can automatically process abnormal noise audio data in batches, significantly reducing manual workload. Furthermore, the intelligent detection method has good environmental adaptability and can reduce reliance on human judgment.

[0006] The technical solution of the present invention specifically includes the following steps:

[0007] Step 1: Convert the format and perform frame segmentation on the pre-collected audio database samples of abnormal car noises. The database contains audio signals from different working conditions, different car models, and different parts of the car interior, and obtains audio signals of each frame in a unified format.

[0008] Step 2: Calculate the bionic auditory perception features of each frame of audio signal, including loudness, sharpness, and total roughness. Further, weight the loudness, sharpness, and total roughness and sum them to obtain a new bionic auditory perception feature, denoted as LSR.

[0009] Step 3: Construct a multi-model ensemble detection network that integrates MLP-Mixer and Gradient Boosting Decision Tree (GBDT). Input the bionic auditory perception feature LSR calculated in Step 2 into the multi-model ensemble detection network to obtain the probability that each frame of audio is an abnormal sound.

[0010] The multi-model integrated detection network includes an input layer, a probability prediction layer, and a fusion layer. The input layer receives the bionic auditory perception feature LSR calculated in step 2 and then passes it to the probability prediction layer. The probability prediction layer includes a parallel deep learning path and a machine learning path. The deep learning path includes an MLP-Mixer probability prediction model, and the machine learning path includes three gradient boosting tree probability prediction models: LightGBM, XGBoost, and CatBoost. The probability prediction layer converts the bionic auditory perception feature LSR into the probability that the audio data in the frame is abnormal, and then passes it to the fusion layer. The fusion layer weights and fuses the probabilities of the audio data in the frame being abnormal predicted by multiple probability prediction models to obtain the final probability that each audio frame is abnormal.

[0011] Furthermore, step 1 specifically includes the following sub-steps:

[0012] Step 11: Based on the set quantization precision and sampling frequency f s The database samples are converted into a standardized format, and the multi-channel signal is downmixed to a mono signal to eliminate channel differences.

[0013] Step 12: Based on the set frame length and frame overlap, the audio signal is processed by frame segmentation using the sliding window technique to obtain the audio signal of each frame, and the frequency f of the audio signal of each frame is recorded.

[0014] Furthermore, in step 11, the standardized format is wav, mp4, or mov.

[0015] Furthermore, step 2 specifically includes the following sub-steps:

[0016] Step 21: Perform a nonlinear frequency f-position z-mapping on each frame of the audio signal obtained in Step 1, in order to simulate the spatial decoding characteristics of the cochlea for different frequency components; the mathematical expression for the frequency f-position z-mapping is:

[0017]

[0018] in, This represents the distance from the base of the cochlea along the basilar membrane to a certain point. This indicates the frequency of the audio signal in each frame;

[0019] according to Find the center frequency f c

[0020]

[0021] According to the center frequency f c Find the bandwidth b

[0022]

[0023] Step 22: Based on the calculated center frequency f c With bandwidth b, a Gammatone filter is used to simulate the frequency analysis characteristics of the cochlea. The audio signal of each frame in the uniform format obtained in step 1 is decomposed into different frequency bands, and then the energy of the human ear perception domain is calculated.

[0024] Specifically:

[0025] The time-domain impulse response of the Gammatone filter is:

[0026]

[0027] In the formula, t is time. For amplitude factor, , To determine the filter order, a 4th-order Gammatone filter is used. b is the bandwidth, f c For the center frequency, For the initial phase, ;

[0028] Let the audio signal of each frame in the uniform format obtained in step 1 be the input signal, and let it be... To improve frequency resolution, the input signal is processed through a filter bank, where the i-th filter in the filter bank corresponds to the center frequency. Its bandpass response is

[0029]

[0030] In the formula, ; and It is obtained through the following steps, specifically:

[0031] Take f min =20 Hz, f max =f s / 2, Substitute into the frequency f-position z mapping mathematical expression in step 21 to obtain the corresponding z. min and z max The z corresponding to the i-th filter i for

[0032]

[0033] By z i The corresponding f can be obtained according to step 21. c,i as well as ;

[0034] Then, the output matrix of the filter bank is obtained.

[0035]

[0036] In the formula, M represents the number of filters. ;

[0037] For each filter output Extract its envelope signal e i (t);

[0038]

[0039] () denotes the Hilbert transform, used to obtain analytic signals; these envelope signals are the time response of auditory frequency band energy, reflecting the location and energy concentration of abnormal noise events in the time and frequency domain;

[0040] Finally, logarithmic compression is performed on the extracted envelope signal to obtain the energy of the human ear's perceptual domain for each frame of the audio signal. :

[0041]

[0042] In the formula, ;

[0043] Step 23: Calculate the loudness based on the energy of the human ear's perceptual domain of the audio signal for each frame;

[0044] Specifically:

[0045] First, the instantaneous loudness is calculated using the following formula. :

[0046]

[0047] In the formula, The energy of the human ear's perceptual domain of the audio signal is calculated in step 22 using a 1 kHz, 40 dB pure tone as the input signal.

[0048] To pass The energy of the audio signal in the human ear's perceptual domain, obtained through step 22, is used as the input signal. The set sampling frequency;

[0049] The loudness is obtained by integrating across the entire frequency band:

[0050]

[0051] Step 24: Calculate the sharpness based on the instantaneous loudness using the following formula. :

[0052]

[0053] In the formula, The high-frequency weighting factor is expressed as follows:

[0054]

[0055] Step 25: Analyze the modulation frequency components of the extracted envelope signal in the range of 15–300 Hz, and calculate the specific expressions for the modulation frequency components as follows:

[0056]

[0057] In the formula, is the envelope signal obtained in step 3, F() is the Fourier transform of the envelope signal, and f mod is the modulation frequency;

[0058] Then, based on the calculated modulation frequency components, the total roughness is calculated through the following expression:

[0059]

[0060] In the formula, is the weighting factor. When f ≤ 500 Hz, take 0.3. When 500 < f ≤ 4000 Hz, take 1.0. When 4000 < f ≤ 6000 Hz, take 0.6. When f > 6000 Hz, take 0.3;

[0061] Step 26: Weightedly fuse the three features of loudness, sharpness, and total roughness to obtain a new bionic auditory perception feature, denoted as LSR:

[0062]

[0063] The criterion for selecting the weight coefficient is .

[0064] Furthermore, the MLP-Mixer probability prediction model and the three gradient boosting tree probability prediction models of the LightGBM model, XGBoost model, and CatBoost model in step 3 simultaneously receive the bionic auditory perception feature LSR transmitted by the input layer;

[0065] The MLP-Mixer probability prediction model includes an MLP input mapping layer, a first hidden layer, a second hidden layer, and an MLP output layer; the MLP input mapping layer normalizes the received bionic auditory perception feature vector LSR and performs feature screening to obtain a feature vector adapted to the MLP, and then transmits it to the first hidden layer; the first hidden layer is a fully connected layer, containing 64 neurons, and uses the ReLU activation function to extract features from the input data; a feature representation vector is obtained as the input data for the next hidden layer; the second hidden layer is a fully connected layer, containing 32 neurons, and uses the ReLU activation function to extract features from the input data; a new feature representation vector is obtained as the input data for the MLP output layer; the MLP output layer uses the Sigmoid activation function to perform probability prediction on the feature representation vector data of the second hidden layer to obtain the probability that the audio data of this frame is abnormal sound;

[0066] The LightGBM model, based on a leaf growth strategy, performs feature splitting and node gain calculation on the received bionic auditory perception features (LSR), generates leaf node output values, and accumulates them with weights to obtain the probability that the audio data of that frame is an abnormal sound.

[0067] The XGBoost model is based on the gradient boosting framework. It performs feature splitting and weighted fitting of gradient and second derivative on the received bionic auditory perception feature LSR to obtain the probability that the audio data of the frame is abnormal.

[0068] The CatBoost model is based on a symmetric tree structure. It performs feature splitting, category encoding, and gradient iterative updates on the received bionic auditory perception features (LSR) to obtain the probability that the audio data of the frame is an abnormal sound.

[0069] The fusion layer is a stacking layer. This layer linearly concatenates the predicted probabilities output by the deep learning path with the predicted probabilities of the three gradient boosting tree models in the machine learning path to form a new input vector. This vector is then input into the Logistic Regression meta-learner in the stacking layer for secondary training. This process yields weight coefficients for the probability that the audio data predicted by the MLP-Mixer probabilistic prediction model, the LightGBM model, the XGBoost model, and the CatBoost model is an abnormal sound. Based on these weight coefficients, the probabilities that the audio data predicted by the MLP-Mixer probabilistic prediction model, the LightGBM model, the XGBoost model, and the CatBoost model are an abnormal sound are weighted and summed to obtain the final fused prediction probability.

[0070] Beneficial effects:

[0071] This invention provides an intelligent detection method for automotive abnormal noises based on biomimetic auditory perception features. This method extracts and fuses biomimetic auditory perception features from audio data, including loudness, sharpness, and total roughness, to achieve a comprehensive representation of abnormal noise signals. Simultaneously, it fits the biomimetic auditory perception features using an MLP-Mixer deep neural network, a LightGBM model, an XGBoost model, and a CatBoost model gradient boosting tree model to obtain the probability that each audio frame is an abnormal noise. Compared to traditional subjective evaluation methods, this patent avoids the problem of single feature extraction in human hearing by extracting and fusing biomimetic auditory perception features from audio data. Through deep learning and machine learning, it can automatically process abnormal noise audio data in batches, greatly reducing the workload of manual labor. Furthermore, the intelligent detection method has good environmental adaptability and can reduce reliance on human judgment. Attached Figure Description

[0072] Figure 1 This is a flowchart illustrating the implementation of an embodiment of the present invention; Detailed Implementation

[0073] To make the technical problems solved, the technical solutions, and the beneficial effects of this invention clearer and to enable those skilled in the art to better understand the invention, the invention will be further described in detail and in full below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0074] This embodiment takes a pre-collected audio database sample of abnormal car noises as an example and uses the method proposed in this invention to detect abnormal noises in the audio database sample.

[0075] like Figure 1 As shown in the figure, the intelligent detection method for abnormal noises in automobiles based on bionic auditory perception features in this embodiment includes the following steps:

[0076] Step 1: Perform format conversion and frame segmentation on the pre-collected automotive abnormal noise audio database samples. The database contains audio signals from different operating conditions, different car models, and different parts of the vehicle interior, obtaining audio signals of a uniform format for each frame; specifically, it includes the following sub-steps:

[0077] Step 11: Based on the set quantization precision and sampling frequency f s The database samples are converted into a standardized format, and the multi-channel signal is downmixed to a mono signal to eliminate channel differences.

[0078] Step 12: Based on the set frame length and frame overlap, the audio signal is processed by frame segmentation using the sliding window technique to obtain the audio signal of each frame, and the frequency f of the audio signal of each frame is recorded.

[0079] In this embodiment, the database contains 59 abnormal noise audio recordings, each with a duration of 10-12 seconds, a sampling frequency of 44100 Hz or 22050 Hz, and audio formats including wav, mp4, and mov. The sources of abnormal noises cover multiple typical interior components such as car doors, armrests, sunroofs, headrests, center consoles, seat backs, and seat cushions. The samples cover various car models, including sedans and SUVs, and the operating conditions include stationary and driving states.

[0080] In this embodiment, the input audio is uniformly converted to WAV format, the quantization precision is set to 16-bit quantization precision, the sampling frequency is set to 44100 Hz, the multi-channel signal is downmixed to mono, the frame length is set to 30 ms, and the frame overlap is set to 50%.

[0081] Step 2: Calculate the bionic auditory perception features of each frame of audio signal, including loudness, sharpness and total roughness. Further, the loudness, sharpness and total roughness are weighted and summed to obtain new bionic auditory perception features, denoted as LSR.

[0082] Specifically, it includes the following sub-steps:

[0083] Step 21: Perform a nonlinear frequency f-position z-mapping on each frame of the audio signal obtained in Step 1, in order to simulate the spatial decoding characteristics of the cochlea for different frequency components; the mathematical expression for the frequency f-position z-mapping is:

[0084]

[0085] in, This represents the distance from the base of the cochlea along the basilar membrane to a certain point. This indicates the frequency of the audio signal in each frame;

[0086] according to Find the center frequency f c

[0087]

[0088] According to the center frequency f c Find the bandwidth b

[0089]

[0090] Step 22: Based on the calculated center frequency f c With bandwidth b, a Gammatone filter is used to simulate the frequency analysis characteristics of the cochlea. The audio signal of each frame in the uniform format obtained in step 1 is decomposed into different frequency bands, and then the energy of the human ear perception domain is calculated.

[0091] Specifically:

[0092] The time-domain impulse response of the Gammatone filter is:

[0093]

[0094] In the formula, t is time. For amplitude factor, , To determine the filter order, a 4th-order Gammatone filter is used. b is the bandwidth, f c For the center frequency, For the initial phase, ;

[0095] Let the audio signal of each frame in the uniform format obtained in step 1 be the input signal, and let it be... To improve frequency resolution, the input signal is processed through a filter bank, where the i-th filter in the filter bank corresponds to the center frequency. Its bandpass response is

[0096]

[0097] In the formula, ; and It is obtained through the following steps, specifically:

[0098] Take f min =20 Hz, f max =f s / 2, in this embodiment, f max =44100 / 2 Hz, substitute into the frequency f-position z mapping mathematical expression in step 21 to obtain the corresponding z. min and z max The z corresponding to the i-th filter i for

[0099]

[0100] By z i The corresponding f can be obtained according to step 21. c,i as well as ;

[0101] Then, the output matrix of the filter bank is obtained.

[0102]

[0103] In the formula, M represents the number of filters. ;

[0104] For each filter output Extract its envelope signal e i (t);

[0105]

[0106] () denotes the Hilbert transform, used to obtain analytic signals; these envelope signals are the time response of auditory frequency band energy, reflecting the location and energy concentration of abnormal noise events in the time and frequency domain;

[0107] Finally, logarithmic compression is performed on the extracted envelope signal to obtain the energy of the human ear's perceptual domain for each frame of the audio signal. :

[0108]

[0109] In the formula, ;

[0110] Step 23: Calculate the loudness based on the energy of the human ear's perceptual domain of the audio signal for each frame;

[0111] Specifically:

[0112] First, the instantaneous loudness is calculated using the following formula. :

[0113]

[0114] In the formula, The energy of the human ear's perceptual domain of the audio signal is calculated in step 22 using a 1 kHz, 40 dB pure tone as the input signal.

[0115] To pass The energy of the audio signal in the human ear's perceptual domain, obtained through step 22, is used as the input signal. The set sampling frequency;

[0116] The loudness is obtained by integrating across the entire frequency band:

[0117]

[0118] Step 24: Calculate the sharpness based on the instantaneous loudness using the following formula. :

[0119]

[0120] In the formula, The high-frequency weighting factor is expressed as follows:

[0121]

[0122] Step 25: Analyze the modulation frequency components of the extracted envelope signal in the range of 15–300 Hz, and calculate the specific expressions for the modulation frequency components as follows:

[0123]

[0124] In the formula, Let f be the envelope signal obtained in step 3, and F() be the Fourier transform of the envelope signal. mod The modulation frequency;

[0125] Then, based on the calculated modulation frequency components, the total roughness is calculated using the following expression:

[0126]

[0127] In the formula, is the weighting factor. When f ≤ 500 Hz, take 0.3. When 500 < f ≤ 4000 Hz, take 1.0. When 4000 < f ≤ 6000 Hz, take 0.6. When f > 6000 Hz, take 0.3;

[0128] Step 26: Weightedly fuse the three features of loudness, sharpness, and total roughness to obtain a new bionic auditory perception feature, denoted as LSR:

[0129]

[0130] The criterion for selecting the weight coefficients is ;

[0131] In this embodiment, the weight coefficients of loudness, sharpness, and total roughness are respectively taken as: , , .

[0132] Step 3: Construct a multi-model integrated detection network that combines MLP-Mixer and Gradient Boosting Decision Tree (GBDT). Input the bionic auditory perception feature LSR calculated in Step 2 into the multi-model integrated detection network to obtain the probability that each frame of audio is an abnormal sound;

[0133] The multi-model integrated detection network includes an input layer, a probability prediction layer, and a fusion layer. The input layer is used to receive the bionic auditory perception feature LSR calculated in Step 2 and then transfer it to the probability prediction layer; the probability prediction layer contains a deep learning path and a machine learning path arranged in parallel. The deep learning path includes an MLP-Mixer probability prediction model, and the machine learning path includes three gradient boosting tree probability prediction models: LightGBM model, XGBoost model, and CatBoost model; the probability prediction layer converts the bionic auditory perception feature LSR into the probability that this frame of audio data is an abnormal sound and then transfers it to the fusion layer; the fusion layer weightedly fuses the probabilities that the audio data of this frame predicted by multiple probability prediction models to obtain the final probability that each frame of audio is an abnormal sound.

[0134] In this embodiment, the MLP-Mixer probability prediction model and the three gradient boosting tree probability prediction models of LightGBM model, XGBoost model, and CatBoost model simultaneously receive the bionic auditory perception feature LSR transferred by the input layer;

[0135] The MLP-Mixer probabilistic prediction model includes an MLP input mapping layer, a first hidden layer, a second hidden layer, and an MLP output layer. The MLP input mapping layer standardizes and filters the received bionic auditory perception feature vector (LSR) to obtain a feature vector adapted to the MLP, which is then passed to the first hidden layer. The first hidden layer is a fully connected layer containing 64 neurons, and uses the ReLU activation function to extract features from the input data, resulting in a feature representation vector, which serves as the input data for the next hidden layer. The second hidden layer is a fully connected layer containing 32 neurons, and uses the ReLU activation function to extract features from the input data, resulting in a new feature representation vector, which serves as the input data for the MLP output layer. The MLP output layer uses the Sigmoid activation function to perform probabilistic prediction on the feature representation vector data from the second hidden layer, obtaining the probability that the audio data in that frame is an abnormal sound.

[0136] The LightGBM model, based on a leaf growth strategy, performs feature splitting and node gain calculation on the received bionic auditory perception features (LSR), generates leaf node output values, and accumulates them with weights to obtain the probability that the audio data of that frame is an abnormal sound.

[0137] The XGBoost model is based on the gradient boosting framework. It performs feature splitting and weighted fitting of gradient and second derivative on the received bionic auditory perception feature LSR to obtain the probability that the audio data of the frame is abnormal.

[0138] The CatBoost model is based on a symmetric tree structure. It performs feature splitting, category encoding, and gradient iterative updates on the received bionic auditory perception features (LSR) to obtain the probability that the audio data of the frame is an abnormal sound.

[0139] The fusion layer is a stacking layer. This layer linearly concatenates the predicted probabilities output by the deep learning path with the predicted probabilities of the three gradient boosting tree models in the machine learning path to form a new input vector. This vector is then input into the Logistic Regression meta-learner in the stacking layer for secondary training. This process yields weight coefficients for the probability that the audio data predicted by the MLP-Mixer probabilistic prediction model, the LightGBM model, the XGBoost model, and the CatBoost model is an abnormal sound. Based on these weight coefficients, the probabilities that the audio data predicted by the MLP-Mixer probabilistic prediction model, the LightGBM model, the XGBoost model, and the CatBoost model are an abnormal sound are weighted and summed to obtain the final fused prediction probability.

[0140] In this embodiment, the car noise audio database samples from step 1 are randomly divided into training and test sets at an 8:2 ratio. The LSR features constructed based on the training set are simultaneously input into the MLP-Mixer probabilistic prediction model, LightGBM model, XGBoost model, and CatBoost model for training. Each probabilistic prediction model outputs the noise prediction probability corresponding to each frame of audio. The prediction outputs of the MLP-Mixer probabilistic prediction model, LightGBM model, XGBoost model, and CatBoost model are linearly concatenated into a new input vector, which is used as the input feature of the fusion layer and input into the LogisticRegression meta-learner for secondary training. The weight coefficients of the probability that the audio data predicted by the MLP-Mixer probabilistic prediction model, LightGBM model, XGBoost model, and CatBoost model is a noise are obtained. Then, based on the obtained weight coefficients, the probabilities that the audio data predicted by the MLP-Mixer probabilistic prediction model, LightGBM model, XGBoost model, and CatBoost model is a noise are weighted and summed to obtain the final fused prediction probability. At this point, the training of the four probabilistic prediction models is complete.

[0141] In this embodiment, to verify the effectiveness of the intelligent detection method for abnormal noises in automobiles based on biomimetic auditory perception features proposed in this patent, the LSR features constructed based on the test set were input into the trained probabilistic prediction model to predict abnormal noises. The prediction results were then compared with those of the widely used Mel-Frequency Cepstral Coefficients (MFCC) feature, which is based on the human auditory perception mechanism. The performance comparison results of the two features under the multi-model integrated detection network are shown in Table 1.

[0142] Table 1 Comparison of results for different features

[0143]

[0144] In addition, using LSR as input features, the detection results of the multi-model ensemble detection network and several classic networks were compared, and the results are shown in Table 2.

[0145] Table 2 Comparison of detection results from different networks

[0146]

[0147] The results in Table 1 show that the extracted LSR features significantly outperform traditional MFCC features in detection performance, achieving higher values ​​in multiple evaluation metrics such as accuracy, precision, recall, F1 score, and AUC, thus verifying the effectiveness of the features in the abnormal sound detection task. The results in Table 2 further show that the integrated detection network proposed in this patent outperforms commonly used algorithms such as ResNet18, KNN, and SVM in overall performance, demonstrating superior detection capabilities in all core metrics, reflecting the advanced nature and applicability of the proposed method.

[0148] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A method for intelligent detection of abnormal noises in automobiles based on biomimetic auditory perception features, characterized in that: Includes the following steps: Step 1: Convert the format and perform frame segmentation on the pre-collected audio database samples of abnormal car noises. The database contains audio signals from different working conditions, different car models, and different parts of the car interior, and obtains audio signals of each frame in a unified format. Step 2: Calculate the bionic auditory perception features of each frame of audio signal, including loudness, sharpness and total roughness. Then, the loudness, sharpness and total roughness are weighted and summed to obtain new bionic auditory perception features, denoted as LSR. Step 3: Construct a multi-model ensemble detection network that integrates MLP-Mixer and Gradient Boosting Decision Tree (GBDT). Input the bionic auditory perception feature LSR calculated in Step 2 into the multi-model ensemble detection network to obtain the probability that each frame of audio is an abnormal sound. The multi-model integrated detection network includes an input layer, a probability prediction layer, and a fusion layer. The input layer receives the bionic auditory perception feature LSR calculated in step 2 and then passes it to the probability prediction layer. The probability prediction layer includes a parallel deep learning path and a machine learning path. The deep learning path includes an MLP-Mixer probability prediction model, and the machine learning path includes three gradient boosting tree probability prediction models: LightGBM, XGBoost, and CatBoost. The probability prediction layer converts the bionic auditory perception feature LSR into the probability that the audio data in the frame is abnormal, and then passes it to the fusion layer. The fusion layer weights and fuses the probabilities of the audio data in the frame being abnormal predicted by the four probability prediction models to obtain the final probability that each audio frame is abnormal.

2. The intelligent detection method for abnormal noises in automobiles based on bionic auditory perception features according to claim 1, characterized in that: Step 1 specifically includes the following sub-steps: Step 11: Based on the set quantization precision and sampling frequency f s The database samples are converted into a standardized format, and the multi-channel signal is downmixed to a mono signal to eliminate channel differences. Step 12: Based on the set frame length and frame overlap, the audio signal is processed by frame segmentation using the sliding window technique to obtain the audio signal of each frame, and the frequency f of the audio signal of each frame is recorded.

3. The intelligent detection method for abnormal noises in automobiles based on bionic auditory perception features according to claim 2, characterized in that: In step 11, the standardized format is wav, mp4, or mov.

4. The intelligent detection method for abnormal noises in automobiles based on bionic auditory perception features according to claim 1, characterized in that: Step 2 specifically includes the following sub-steps: Step 21: Perform a nonlinear frequency f-position z-mapping on each frame of the audio signal obtained in Step 1, in order to simulate the spatial decoding characteristics of the cochlea for different frequency components; the mathematical expression for the frequency f-position z-mapping is: in, This represents the distance from the base of the cochlea along the basilar membrane to a certain point. This represents the frequency of the audio signal in each frame; according to Find the center frequency f c According to the center frequency f c Find the bandwidth b Step 22: Based on the calculated center frequency f c With bandwidth b, a Gammatone filter is used to simulate the frequency analysis characteristics of the cochlea. The audio signal of each frame in the uniform format obtained in step 1 is decomposed into different frequency bands, and then the energy of the human ear perception domain is calculated. Specifically: The time-domain impulse response of the Gammatone filter is: In the formula, t represents time. For amplitude factor, , To determine the filter order, a 4th-order Gammatone filter is used. b is the bandwidth, f c For the center frequency, For the initial phase, ; Let the audio signal of each frame in the uniform format obtained in step 1 be the input signal, and let it be... To improve frequency resolution, the input signal is processed through a filter bank, where the i-th filter in the filter bank corresponds to the center frequency. Its bandpass response is In the formula, ; and It is obtained through the following steps, specifically: Take f min =20 Hz, f max =f s / 2, Substitute into the frequency f-position z mapping mathematical expression in step 21 to obtain the corresponding z. min and z max The z corresponding to the i-th filter i for By z i The corresponding f can be obtained according to step 21. c,i as well as ; Then, the output matrix of the filter bank is obtained. In the formula, M represents the number of filters. ; For each filter output Extract its envelope signal e i (t); () denotes the Hilbert transform, used to obtain analytic signals; these envelope signals are the time response of auditory frequency band energy, reflecting the location and energy concentration of abnormal noise events in the time and frequency domain; Finally, logarithmic compression is performed on the extracted envelope signal to obtain the energy of the human ear's perceptual domain for each frame of the audio signal. : In the formula, ; Step 23: Calculate the loudness based on the energy of the human ear's perceptual domain of the audio signal for each frame; Specifically: First, the instantaneous loudness is calculated using the following formula. : In the formula, The energy of the human ear's perceptual domain of the audio signal is calculated in step 22 using a 1 kHz, 40 dB pure tone as the input signal. To pass The energy of the audio signal in the human ear's perceptual domain, obtained through step 22, is used as the input signal. The set sampling frequency; The loudness is obtained by integrating across the entire frequency band: Step 24: Calculate the sharpness based on the instantaneous loudness using the following formula. : In the formula, This is the high-frequency weighting factor, and its specific expression is as follows: Step 25: Analyze the modulation frequency components of the extracted envelope signal in the range of 15–300 Hz, and calculate the specific expressions for the modulation frequency components as follows: In the formula, Let f be the envelope signal obtained in step 3, and F() be the Fourier transform of the envelope signal. mod The modulation frequency; Then, based on the calculated modulation frequency components, the total roughness is calculated using the following expression: In the formula, is the weighting factor. When f ≤ 500 Hz, take 0.

3. When 500 < f ≤ 4000 Hz, take 1.

0. When 4000 < f ≤ 6000 Hz, take 0.

6. When f > 6000 Hz, take 0.3; Step 26: Weighted fusion of loudness, sharpness, and total roughness features yields a new bionic auditory perception feature, denoted by LSR: The criteria for selecting weighting coefficients are as follows .

5. The intelligent detection method for abnormal noises in automobiles based on bionic auditory perception features according to claim 1, characterized in that: In step 3, the MLP-Mixer probabilistic prediction model and the three gradient boosting tree probabilistic prediction models, LightGBM, XGBoost, and CatBoost, simultaneously receive the bionic auditory perception features (LSR) transmitted from the input layer. The MLP-Mixer probabilistic prediction model includes an MLP input mapping layer, a first hidden layer, a second hidden layer, and an MLP output layer. The MLP input mapping layer standardizes and filters the received bionic auditory perception feature vector (LSR) to obtain a feature vector adapted to the MLP, which is then passed to the first hidden layer. The first hidden layer is a fully connected layer containing 64 neurons, and uses the ReLU activation function to extract features from the input data, resulting in a feature representation vector, which serves as the input data for the next hidden layer. The second hidden layer is a fully connected layer containing 32 neurons, and uses the ReLU activation function to extract features from the input data, resulting in a new feature representation vector, which serves as the input data for the MLP output layer. The MLP output layer uses the Sigmoid activation function to perform probabilistic prediction on the feature representation vector data from the second hidden layer, obtaining the probability that the audio data in that frame is an abnormal sound. The LightGBM model, based on a leaf growth strategy, performs feature splitting and node gain calculation on the received bionic auditory perception features (LSR), generates leaf node output values, and accumulates them with weights to obtain the probability that the audio data of that frame is an abnormal sound. The XGBoost model is based on the gradient boosting framework. It performs feature splitting and weighted fitting of gradient and second derivative on the received bionic auditory perception feature LSR to obtain the probability that the audio data of the frame is abnormal. The CatBoost model is based on a symmetric tree structure. It performs feature splitting, category encoding, and gradient iterative updates on the received bionic auditory perception features (LSR) to obtain the probability that the audio data of the frame is an abnormal sound. The fusion layer is a stacking layer. This layer linearly concatenates the predicted probabilities output by the deep learning path with the predicted probabilities of the three gradient boosting tree models in the machine learning path to form a new input vector. This vector is then input into the Logistic Regression meta-learner in the stacking layer for secondary training. This process yields weight coefficients for the probability that the audio data predicted by the MLP-Mixer probabilistic prediction model, the LightGBM model, the XGBoost model, and the CatBoost model is an abnormal sound. Based on these weight coefficients, the probabilities that the audio data predicted by the MLP-Mixer probabilistic prediction model, the LightGBM model, the XGBoost model, and the CatBoost model are an abnormal sound are weighted and summed to obtain the final fused prediction probability.

Images