A Selectable Active Sound Quality Control System and Testing Method Based on Electroencephalography (EEG)

Abstract

This invention discloses a selective sound quality active control system and testing method based on electroencephalography (EEG). It integrates convolutional neural networks and long short-term memory networks to establish a CNN-LSTM-based classification model. Using EEG, it constructs four sound quality databases suitable for sleep, relaxation, creative thinking, and efficient work. Combined with an active noise control algorithm, it forms an active control system for selectable sound quality within the vehicle. This invention classifies the collected noise according to its corresponding EEG data, and then uses the active noise control system to design the vehicle's sound quality to meet the different needs of passengers.

Description

Technical Field

[0001] This invention relates to the field of active noise control technology, and in particular to a selectable active sound quality control system and testing method based on electroencephalography (EEG). Background Technology

[0002] In the field of automotive sound quality evaluation technology, automotive sound quality evaluation models based on electroencephalogram (EEG) signals can more objectively reflect people's psychological perceptions of car noise. EEG waves can be divided into delta (δ) brain waves, theta (θ) brain waves, alpha (α) brain waves, and beta (β) brain waves. Delta brain waves can promote more restful sleep, theta brain waves are suitable for meditation (such as naps or resting with eyes closed), alpha brain waves are helpful for learning and thinking, and beta brain waves help improve attention and problem-solving abilities. Different sound stimuli can cause different changes in brain waves. By using active noise control systems and adding brainwave music, targeted design can be implemented for automotive noise, creating environmental sound quality suitable for passengers' sleep, relaxation, creative thinking, and high efficiency. Based on the human auditory domain, noise can be divided into 24 Bark domains. By decomposing and reconstructing noise signals based on these 24 Bark domains, the influence of each Bark domain on different brain waves can be identified, and the optimal noise control frequency band can be found to achieve optimal sound quality control.

[0003] In view of this, the present invention provides an active sound quality control system and testing method based on electroencephalogram (EEG) signals. This system can provide passengers with selectable sound quality and design the sound quality of the car through an active noise control system. Summary of the Invention

[0004] The purpose of this invention is to address the technical deficiencies in the existing technology by providing a selective active sound quality control system and testing method based on electroencephalography (EEG).

[0005] The technical solution adopted to achieve the purpose of this invention is:

[0006] A selective sound quality active control system based on electroencephalography (EEG) includes a sound quality selection controller, a sound pressure sensor, a noise-sound quality classification model, a noise active control system, and a secondary loudspeaker, wherein:

[0007] The sound pressure sensor is installed near the passenger seat of the car to receive noise signals and transmit the noise signals to the noise-sound quality classification model. The noise-sound quality classification model outputs the current sound quality type.

[0008] The sound quality selection controller allows passengers to input an ideal sound quality type. If the ideal sound quality type is the same as the current sound quality type, noise is applied. If the ideal sound quality type is different from the current sound quality type, a noise signal is input to the noise active control system. The noise active control system adjusts the secondary loudspeaker to generate corresponding sound waves to cancel the noise signal until the sound quality type output by the noise-sound quality classification model is the same as the ideal sound quality type.

[0009] A selective active control method for sound quality based on electroencephalography (EEG) includes the following steps:

[0010] S1, collect vehicle noise as a noise sample;

[0011] S2. Based on the EEG signals under noise sample stimulation, the noise samples are divided into N types of sound quality, and a noise-sound quality type database is constructed.

[0012] S3, Construct a noise-sound quality classification model with noise signal as input and sound quality type as output;

[0013] S4. The noise signal is decomposed into wavelet based on the 24bark domain to determine the influence factor of each bark domain on different EEG waves.

[0014] S5. Construct an active noise control system based on the FELMS algorithm. Select a certain sound quality type and input the noise signal into the noise-sound quality classification model. If the noise signal meets our desired sound quality requirements, the noise is passed. If the noise signal does not meet our desired sound quality type, the noise is input into the active noise control system. According to the EEG influence factor of each Bark domain of the noise, the controlled frequency bands are increased from large to small to carry out active noise control simulation.

[0015] In the above technical solution, S2 includes four types of sound quality: sleep, relaxation, creative thinking, and efficient work.

[0016] In the above technical solution, in S3, the noise-sound quality classification model is constructed based on the CNN-LSTM model. The noise-sound quality classification model includes a convolutional neural network, a long short-term memory network, two fully connected networks, and a softmax output unit. The convolutional neural network (CNN) takes noise signals as input, and the output of the CNN is fed to the long short-term memory network (LSTM). The output of the LSTM is fed to the fully connected network. The first fully connected network further extracts features from the input data, the second fully connected network is used to reduce the size of the output data, and finally, the sound quality is output through the softmax output unit.

[0017] In the above technical solution, the calculation formula for the convolutional layer of a convolutional neural network (CNN) is as follows:

[0018] h i =σ(w i *x i +b i )

[0019] Where w represents the convolution kernel weight coefficient, x represents the input variable, b represents the bias, σ represents the ReLU activation function, and h represents the output data after computation;

[0020] The formula for calculating the pooling layer in a Convolutional Neural Network (CNN) is:

[0021] h i =max(h i-1 -1)

[0022] Among them, h i-1 h represents the feature data before pooling. i This represents the feature data after pooling.

[0023] In the above technical solution, the calculation formula for the Long Short-Term Memory (LSTM) network is: f t =σ(W f ×[h t-1 x t ])+b f

[0024] i t =σ(W i ×[h t-1 x t ])+b i

[0025] C f =f t ×C t-1 +i t ×tanh(W c ×[h t-1 x t ]+b c )

[0026] O t =σ(W o ×[h t-1 x t ])+b o

[0027] In the formula, h t-1 Indicates the previous sound quality type of output, x t The input, W, represents the characteristics of the current noise signal. f W i W c and W o b represents the weighting coefficient.f b i b c and b o Indicates bias, C t σ represents the sound quality index at the current moment, σ represents the sigmoid activation function, and tanh is the hyperbolic tangent activation function.

[0028] In the above technical solution, the calculation formula for the Softmax output unit is:

[0029]

[0030] In the formula: y i x represents the probability that an in-vehicle noise sample is classified into the i-th class; i and x j This is the input of the Softmax unit, which is the N outputs of the second fully connected unit, where N is 4.

[0031] In the above technical solution, the calculation formula for the influence factor of each bark domain on different EEG waves in S4 is as follows:

[0032] The influencing factor of delta brainwaves is:

[0033]

[0034] The theta brainwave influencing factor is:

[0035]

[0036] The factors influencing alpha brainwaves are:

[0037]

[0038] The influencing factors of β-encephalogram (EEG) are:

[0039]

[0040] In the formula, PSD(δ) represents the delta power spectral density of the subject under the original signal stimulus, PSD(θ) represents the theta power spectral density of the subject under the original signal stimulus, PSD(α) represents the power spectral density of the subject under the original signal stimulus, PSD(β) represents the β power spectral density of the subject under the original signal stimulus, and PSD(δ-Bark) represents the power spectral density of the subject under the original signal stimulus. i ) represents the δ-Bark power spectral density of the reconstructed signal after removing the i-th order Bark domain component, PSD(θ-Bark) i ) represents the θ-Bark power spectral density of the reconstructed signal after removing the i-th order Bark domain component, PSD(α-Bark) i) represents the α-Bark power spectral density of the reconstructed signal after removing the i-th order Bark domain component, PSD(β-Bark) i ) represents the β-EEG power spectral density of the reconstructed signal after removing the i-th order Bark domain component.

[0041] In the above technical solution, in S5, the increase of the control frequency band in the noise active control system is achieved by adjusting the passband of the model residual filter Hw(Z) to control the adaptive filter W(Z) to change its weight coefficient, thereby adjusting the secondary sound source signal f(n) to selectively cancel the noise and reshape the noise spectrum.

[0042] In the above technical solution, in step S5, the reconstructed noise is input into the noise-sound quality classification model. If the classification result matches the noise signal in the selected sound quality type, it indicates that the control of the specified sound quality type has been achieved. If it does not match, it is input into the noise active control system for processing. Corresponding brainwave rhythm music is added to the controlled noise signal to optimize the sound quality of the specified type. The noise with added brainwave music is then input into the classification model for testing. If it meets the sound quality requirements, the noise is output.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] This invention classifies the collected noise according to its corresponding EEG data, and then uses an active noise control system to design the car sound quality according to the different needs of passengers, creating environmental sound quality suitable for passengers to sleep, relax, think creatively and work efficiently. Attached Figure Description

[0045] Figure 1 This is a flowchart of the method of the present invention.

[0046] Figure 2 These are four types of brainwave frequency bands and their characteristics.

[0047] Figure 3 It is a structural flowchart.

[0048] Figure 4 It is the Bark domain frequency band division table.

[0049] Figure 5 This is a structural diagram of an active noise control system based on the FELMS algorithm. Detailed Implementation

[0050] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0051] Example 1

[0052] A selective sound quality active control system based on electroencephalography (EEG) includes a sound quality selection controller, a sound pressure sensor, a noise-sound quality classification model, a noise active control system, and a secondary loudspeaker, wherein:

[0053] The sound pressure sensor is installed near the passenger seat of the car to receive noise signals and transmit the noise signals to the noise-sound quality classification model. The noise-sound quality classification model outputs the current sound quality type.

[0054] The sound quality selection controller allows passengers to input an ideal sound quality type. If the ideal sound quality type is the same as the current sound quality type, noise is applied. If the ideal sound quality type is different from the current sound quality type, a noise signal is input to the noise active control system. The noise active control system adjusts the secondary loudspeaker to generate corresponding sound waves to cancel the noise signal until the sound quality type output by the noise-sound quality classification model is the same as the ideal sound quality type.

[0055] Example 2

[0056] A selective active control method for sound quality based on electroencephalography (EEG) includes the following steps:

[0057] Step 1, Vehicle noise data acquisition;

[0058] First, select multiple brands and types of cars, and control the average vehicle speed at four states: 30km / h, 50km / h, 80km / h, and stationary but not turned off. Place multiple sound pressure sensors near each seat of the car to record car noise for 3 minutes. Divide all the recordings into 8-second samples and save them as noise samples.

[0059] Step 2: To create four types of sound quality suitable for sleep, relaxation, creative thinking, and efficient work, four types of brainwave characteristics and frequency bands are as follows: Figure 2 As shown, noise is classified into four categories based on the EEG signals under noise sample stimulation, and a noise-sound quality type database is constructed.

[0060] First, EEG data were collected from 22 subjects under various noise sample stimuli. Subjects wore the EMOTIV wireless EEG acquisition device, and noise samples were played in a quiet space to collect their EEG data. Based on the changes in power spectral density of different EEG rhythms, the noise samples were classified into four sound quality categories: suitable for sleep, rest, creative thinking, and efficient work. After classifying the noise samples, a noise-sound quality type database was constructed by combining the noise samples with their corresponding sound quality labels.

[0061] Step 3: Construct a noise-sound quality classification model;

[0062] The structural flowchart is attached. Figure 3 As shown, a sound quality classification model is trained using noise signals as input and sound quality type as output.

[0063] The noise-sound quality classification model includes a convolutional neural network (CNN), a long short-term memory network (LSTM), two fully connected layers, and a softmax output unit. The CNN takes a noise signal as input, and its output is fed to the LSTM. The LSTM output is fed to the fully connected layers. The first fully connected layer further extracts features from the input data, the second fully connected layer reduces the size of the output data, and finally, the sound quality is output through the softmax output unit.

[0064] The noise-sound quality classification model is built on the CNN-LSTM model. The CNN-LSTM model effectively combines the advantages of LSTM in extracting temporal features and extracting deep features. It establishes a fusion feature extraction layer and uses a combination of fully connected and Softmax output units to construct the classifier module.

[0065] Convolutional Neural Networks (CNNs) primarily rely on convolutional kernels in convolutional layers for feature extraction. These kernels scan the input data and apply a non-linear mapping using the ReLU activation function to extract the spatial features of the data. The calculation formula is as follows:

[0066] h i =σ(w i *x i +b i )

[0067] Where w represents the convolution kernel weight coefficient, x represents the input variable (noise signal), b represents the bias, σ represents the ReLU activation function, and h represents the output data after processing.

[0068] Pooling layers primarily reduce computational complexity by lowering the dimensionality of data features, typically employing either max pooling or average pooling. This study uses max pooling, and the calculation formula for the pooling layer is as follows.

[0069] h i =max(h i-1 -1)

[0070] Among them, h i-1 h represents the feature data before pooling. i This represents the feature data after pooling.

[0071] Long Short-Term Memory (LSTM) networks are a variant of recurrent neural networks. This network effectively solves the gradient vanishing problem during traditional model training and is suitable for analyzing and predicting long-term sequences. Compared to standard RNN networks, LSTM adds a memory state unit to the hidden layer neurons to store past information and uses three gating structures (input gate, forget gate, and output gate) to control the forgetting and updating of historical information. Its calculation formula is as follows:

[0072] f t =σ(W f ×[h t-1 x t ])+b f

[0073] i t =σ(W i ×[h t-1 x t ])+b i

[0074] C f =f t ×C t-1 +i t ×tanh(W c ×[h t-1 x t ]+b c )

[0075] O t =σ(W o ×[h t-1 x t ])+b o

[0076] In the formula, h t-1 Indicates the previous sound quality type of output, x t The input, W, represents the characteristics of the current noise signal. f W i W c and W o b represents the weighting coefficient. f b i b c and b o Indicates bias, C t σ represents the sound quality index at the current moment, σ represents the sigmoid activation function, and tanh is the hyperbolic tangent activation function.

[0077] The classifier of the noise-sound quality classification model consists of two fully connected units and one Softmax output unit. The final output of the model is the sound quality type corresponding to the noise sample.

[0078] The classifier in a noise-sound quality classification model typically uses two fully connected layers. The first fully connected layer is used to further extract features from the input data, while the second fully connected layer is used to reduce the final output size. The number of neurons in the second fully connected layer is equal to the number of classes in the model.

[0079] The Softmax function is a commonly used output unit for multi-class classification tasks. It represents the probability distribution of N different classes, and its calculation formula is as follows:

[0080]

[0081] In the formula: y i x represents the probability that an in-vehicle noise sample is classified into the i-th class; i and x j This is the input to the Softmax unit, which is the N outputs of the second fully connected layer. Since the in-vehicle noise is divided into 4 states as described above, the value of N is 4.

[0082] Step 4: Determine the influencing factors of each bark domain on different EEG waves;

[0083] The noise signal was decomposed using wavelet decomposition in the 24-bark domain. The bark domain frequency band division table is attached. Figure 4 As shown, the influence factor of each bark domain on different EEG waves is determined, and the calculation formula is as follows:

[0084] The influencing factor of delta brainwaves is:

[0085]

[0086] The theta brainwave influencing factor is:

[0087]

[0088] The factors influencing alpha brainwaves are:

[0089]

[0090] The influencing factors of β-encephalogram (EEG) are:

[0091]

[0092] In the formula, PSD(δ) represents the delta-level EEG power spectral density of the subject under the original signal stimulus, PSD(θ) represents the theta-level EEG power spectral density of the subject under the original signal stimulus, PSD(α) represents the EEG power spectral density of the subject under the original signal stimulus, and PSD(β) represents the β-level EEG power spectral density of the subject under the original signal stimulus. PSD(δ-Bark) i ) represents the δ-Bark power spectral density of the reconstructed signal after removing the i-th order Bark domain component. PSD(θ-Bark)i ) represents the θ-Bark power spectral density of the reconstructed signal after removing the i-th order Bark domain component. PSD(α-Bark) i ) represents the α-Bark power spectral density of the reconstructed signal after removing the i-th order Bark domain component. PSD(β-Bark) i ) represents the β-EEG power spectral density of the reconstructed signal after removing the i-th order Bark domain component.

[0093] Step 5: Construct a noise active control system module based on the FELMS algorithm. The structure diagram of the noise active control system based on the FELMS algorithm is attached. Figure 5 As shown, this allows for the design of sound quality for a specified type;

[0094] By selecting a specific sound quality type through the sound quality selection controller, noise samples are input into the sound quality classification model. If the noise meets our desired sound quality requirements, the noise is passed through; otherwise, the noise is input into the noise active control system.

[0095] A noise active control model based on the FELMS algorithm was built. According to the EEG influence factor of each Bark domain of noise, the controlled frequency bands were increased sequentially from large to small to carry out noise active control simulation. The noise samples after each simulation were predicted using the noise-sound quality classification model. If the predicted sound quality matches the selected sound quality type when the frequency band is increased to a certain control band, the increase was stopped.

[0096] The increase in the control frequency band is achieved by adjusting the passband of the residual filter Hw(Z) of the model to control the adaptive filter W(Z) and change its weight coefficients, thereby adjusting the secondary sound source signal f(n) to selectively cancel noise and reshape the noise spectrum. The reconstructed noise is input into the noise-sound quality classification model. If the classification result matches the noise signal in the selected sound quality type, it indicates that the control of the specified sound quality type has been achieved. If not, it is input again into the noise active control system for processing. Corresponding brainwave rhythm music is added to the controlled noise signal to optimize the sound quality of the specified type. The noise with added brainwave music is then input into the classification model for testing. If it meets the sound quality requirements, the noise is output.

[0097] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A selective active sound quality control method based on electroencephalography (EEG), characterized in that, Includes the following steps: S1, collect vehicle noise as a noise sample; S2. Based on the EEG signals under noise sample stimulation, the noise samples are divided into N types of sound quality, and a noise-sound quality type database is constructed. S3, Construct a noise-sound quality classification model with noise signal as input and sound quality type as output; S4. The noise signal is decomposed into wavelet based on the 24bark domain to determine the influence factor of each bark domain on different EEG waves. S5. Construct an active noise control system based on the FELMS algorithm. Select a certain sound quality type and input the noise signal into the noise-sound quality classification model. If the noise signal meets our desired sound quality requirements, the noise is passed. If the noise signal does not meet our desired sound quality type, the noise is input into the active noise control system. According to the EEG influence factor of each Bark domain of the noise, the controlled frequency bands are increased from large to small to carry out active noise control simulation.

2. The selective active sound quality control method based on electroencephalography (EEG) as described in claim 1, characterized in that, The S2 includes four sound qualities: sleep, relaxation, creative thinking, and efficient work.

3. The selective active sound quality control method based on electroencephalography (EEG) as described in claim 1, characterized in that, In S3, the noise-sound quality classification model is constructed based on the CNN-LSTM model. The noise-sound quality classification model includes a convolutional neural network, a long short-term memory network, two fully connected networks, and a softmax output unit. The convolutional neural network (CNN) takes noise signals as input, and the output of the CNN is fed to the long short-term memory network (LSTM). The output of the LSTM is fed to the fully connected network. The first fully connected network further extracts features from the input data, and the second fully connected network is used to reduce the size of the output data. Finally, the sound quality is output through the softmax output unit.

4. The selective active sound quality control method based on electroencephalography (EEG) as described in claim 3, characterized in that, The formula for calculating the convolutional layer of a Convolutional Neural Network (CNN) is: h i =σ(w i *x i +b i ) Where w represents the convolution kernel weight coefficient, x represents the input variable, b represents the bias, σ represents the ReLU activation function, and h represents the output data after computation; The formula for calculating the pooling layer in a Convolutional Neural Network (CNN) is: h i =max(h i-1 -1) Among them, h i-1 h represents the feature data before pooling. i This represents the feature data after pooling.

5. The selective active sound quality control method based on electroencephalography (EEG) as described in claim 3, characterized in that, The calculation formula for the Long Short-Term Memory (LSTM) network is: f t =σ(W f ×[h t-1 x t ])+b f i t =σ(W i ×[h t-1 ,x t ])+b i C f =f t ×C t-1 +i t ×tanh(W c ×[h t-1 ,x t ]+b c ) The t =σ(W o ×[h t-1 ,x t ])+b o In the formula, h t-1 This indicates the previous output quality type, x t The input W represents the characteristics of the current noise signal. f W i W c and W o b represents the weighting coefficient. f b i b c and b o Indicates bias, C t-1 σ represents the sound quality index at the previous moment, σ represents the sigmoid activation function, and tanh is the hyperbolic tangent activation function.

6. The selective active sound quality control method based on electroencephalography (EEG) as described in claim 3, characterized in that, The formula for calculating the Softmax output unit is: In the formula: y i x represents the probability that an in-vehicle noise sample is classified into the i-th class; i and x j This is the input of the Softmax unit, which is the N outputs of the second fully connected unit, where N is 4.

7. The selective active sound quality control method based on electroencephalography (EEG) as described in claim 1, characterized in that, In S4, the formula for calculating the influence factor of each bark domain on different EEG waves is as follows: The influencing factor of delta brainwaves is: The theta brainwave influencing factor is: The factors influencing alpha brainwaves are: The influencing factor of β-encephalogram (EEG) is: In the formula, PSD(δ) represents the delta power spectral density of the subject under the original signal stimulus, PSD(θ) represents the theta power spectral density of the subject under the original signal stimulus, PSD(α) represents the α power spectral density of the subject under the original signal stimulus, PSD(β) represents the β power spectral density of the subject under the original signal stimulus, and PSD(δ-Bark) represents the β power spectral density of the subject under the original signal stimulus. i ) represents the δ-Bark power spectral density of the reconstructed signal after removing the i-th order Bark domain component, PSD(θ-Bark) i ) represents the θ-Bark power spectral density of the reconstructed signal after removing the i-th order Bark domain component, PSD(α-Bark) i ) represents the α-Bark power spectral density of the reconstructed signal after removing the i-th order Bark domain component, PSD(β-Bark) i ) represents the β-EEG power spectral density of the reconstructed signal after removing the i-th order Bark domain component.

8. The selective active sound quality control method based on electroencephalography (EEG) as described in claim 1, characterized in that, In the S5, the increase in the control frequency band in the noise active control system is achieved by adjusting the passband of the model residual filter Hw(Z) to control the adaptive filter W(Z) to change its weight coefficients, thereby adjusting the secondary sound source signal f(n) to selectively cancel the noise and reshape the noise spectrum.

9. The selective active sound quality control method based on electroencephalography (EEG) as described in claim 1, characterized in that, In step S5, the reconstructed noise is input into the noise-sound quality classification model. If the classification result matches the noise signal in the selected sound quality type, it indicates that the control of the specified sound quality type has been achieved. If it does not match, it is input into the noise active control system for processing. Corresponding brainwave rhythm music is added to the controlled noise signal to optimize the sound quality of the specified type. The noise with added brainwave music is then input into the classification model for testing. If it meets the sound quality requirements, the noise is output.

Images