DL-Sig-FxNLMS algorithm-based active noise reduction system and method for dual-feedforward vehicle

Abstract

The invention discloses a DL-Sig-FxNLMS algorithm-based active noise reduction system and method for a dual-feedforward vehicle, and belongs to the technical field of active noise control. The active noise reduction system structurally comprises a reference microphone, an auxiliary microphone, an error microphone, a loudspeaker, a DSP chip and a vehicle-mounted chip. The active noise reduction method comprises the steps that a microphone obtains sound signals, a mode is selected, a sound signal separation module works, a double-feedforward Sig-FxNLMS algorithm is executed, a vehicle-mounted chip executes CNN to update filter coefficients, and a loudspeaker outputs updated reverse sound waves. Through the DL-Sig-FxNLMS algorithm, effective separation and suppression of audio signals in a complex environment are realized. According to the method, the human voice and the environmental noise can be effectively separated and suppressed, so that the accuracy and the practicability of audio signal processing are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of active noise control, and relates to a vehicle active noise reduction technology with an improved noise control algorithm, in particular to a dual feedforward vehicle active noise reduction method based on a DL-Sig-FxNLMS algorithm. Background Art

[0002] In today's society, noise pollution has become a common problem. Especially in high-intensity and long-term noise environments, it will have a negative impact on people's work efficiency, hearing, and physical and mental health. In addition, for drivers, noise can cause physiological reactions such as distraction and headaches, increasing driving safety risks. With the improvement of economic development level, consumers' demand for quiet and comfortable cars is increasing. When a car is driving, it will not only be affected by road noise and wind noise from the road, but the vehicle itself and its components will also generate low-frequency noise due to vibration during driving, which has a more significant impact on the human body. In order to effectively reduce or eliminate the impact of noise, traditional passive noise control (Passive Noise Control, PNC) technology has been developed, which can suppress medium and high-frequency noise in the car by sound absorption, sound insulation, etc., but the noise reduction effect on low-frequency components is minimal. Active noise control (Active Noise Control, ANC) technology based on the principle of destructive interference of sound waves can effectively control noise by generating reverse sound waves. It can design and adjust the control system in a targeted manner according to the characteristics of the noise, making noise control more targeted and targeted, especially for low-frequency noise.

[0003] Among the active noise control methods currently studied by the academic community, the most commonly used is the adaptive filtering theory, which enables the parameters of the adaptive filtering system to be automatically adjusted to achieve the optimal state. On this basis, Morgan, Burgess and others proposed the FxLMS algorithm (Filtered-x Least Mean Square, FxLMS), which provided an important theoretical basis for subsequent active noise control research. However, in practical applications, this algorithm has problems such as slow convergence speed, poor convergence stability, poor convergence adaptability and poor algorithm robustness. On the basis of this algorithm, continuous adjustments and optimization of one or more of the above problems have produced a variety of improved algorithms. With the development of deep learning technology, it has been applied to the field of active noise control. The use of convolutional neural networks (CNNs) can effectively solve the shortcomings of slow convergence speed and poor algorithm robustness. In addition, the selection of a suitable step size formula also has an important influence on the convergence speed and adaptability of the algorithm.

[0004] In the field of active noise control inside cars, there are two types: feedforward active noise control systems and feedback active noise control systems. The feedforward system requires the use of sensors at the noise source to obtain a reference signal, while the feedback system does not require the input of a reference signal. It only constructs a reference signal through the feedback error signal, and the structure is relatively simple. For different noises inside the car, due to their low frequency, it is more suitable to use a feedforward system for noise control. Regardless of the feedforward system or the feedback system, a single system can only reduce the noise of one noise signal or the global noise signal, but in reality, different noises need to be controlled separately. The algorithm of blind signal separation makes it possible to achieve signal separation.

[0005] Therefore, this application improves the FxLMS algorithm and proposes a dual feedforward vehicle active noise reduction method based on the DL-Sig-FxNLMS algorithm. This method applies deep learning to active noise control, proposes a suitable step size μ formula, and designs a dual feedforward active noise reduction system. Through these improvements, the stability, adaptability and robustness of the algorithm convergence can be significantly improved, thereby improving the noise reduction performance. Summary of the invention

[0006] In view of the above problems, the present invention provides a dual feedforward vehicle active noise reduction system and method based on the DL-Sig-FxNLMS algorithm, which modifies the feedforward-feedback system used in the oscillating coil actuator type active hydraulic resistance suspension control and successfully applies it to the field of active noise control technology.

[0007] The technical solution adopted by the present invention is as follows:

[0008] A dual feedforward vehicle active noise reduction system based on a DL-Sig-FxNLMS algorithm comprises a reference microphone, an auxiliary microphone, an error microphone, a loudspeaker, a DSP (digital signal processing) chip and an on-board chip. The reference microphone and the auxiliary microphone are used to collect signals generated by a noise source, and the signals are input into a DSP chip. Inside the DSP chip, the audio components in the sound signals collected by the microphones are calculated by a JADE (joint diagonalization of feature matrices) algorithm, and the audio components are respectively input into convolutional neural network calculation units corresponding to the on-board chip. Meanwhile, a dual feedforward adaptive filtering algorithm (Sig-FxNLMS) in the DSP chip calculates and outputs sound signals. During the process of the DSP executing the dual feedforward Sig-FxNLMS algorithm, the on-board chip synchronously executes a CNN algorithm, calculates corresponding filter coefficient values ​​and feeds back for adjustment. The error microphone collects a total error signal, and the loudspeaker outputs a continuously updated reverse sound wave signal for noise processing.

[0009] A dual feedforward vehicle active noise reduction method based on a DL-Sig-FxNLMS algorithm comprises the following steps:

[0010] S1. Microphone acquisition of acoustic signals: Use a reference microphone and an auxiliary microphone to collect audio data x1(n) and x2(n) of one frame length at a sampling rate, where x1(n) and x2(n) are the audio signals collected by the reference microphone and the auxiliary microphone respectively; use an error microphone to collect the total error signal e(n);

[0011] S2, select mode: users can select the system noise reduction mode according to their needs. The available modes are: 1) full noise reduction, that is, S1 and S2 are both closed; 2) S1 noise reduction, that is, S1 is closed and S2 is disconnected; 3) S2 noise reduction, that is, S1 is disconnected and S2 is closed; 4) stop noise reduction, that is, the system stops working;

[0012] S3, the sound signal separation module works: the JADE algorithm is used to calculate the two audio components in the sound signal collected by the microphone, and the two audio components are input to the dual feedforward DL-Sig-FxNLMS system at the sampling rate; at the same time, when the number of sampling points is equal to the size of a frame, the two audio component frames are input to the calculation unit corresponding to the on-board chip;

[0013] S4, dual feedforward Sig-FxNLMS algorithm execution: The first and second feedforward subsystems use the Sig-FxNLMS algorithm to calculate the reverse sound wave y f (n), y b (n), according to the mode selected by the user, the total error signal e(n) is processed by different algorithm processes, and the error signals e of the first and second feedforward subsystems are separated from the total error signal e(n) by the error separation subsystem. f (n) and e b (n), and send the error signal to the corresponding system;

[0014] S5, the on-board chip executes CNN to update the filter coefficient: while S4 is being executed, the on-board chip executes CNN to calculate the filter coefficient values ​​in the first and second feedforward subsystems according to the audio component frames sent in S3. When the calculated filter coefficient value is different from the last calculation result of CNN, it is updated to the fixed filter coefficient value of the corresponding number;

[0015] S6. The speaker outputs the updated reverse sound wave.

[0016] Preferably, the specific steps of the acoustic signal separation module in step S3 include:

[0017] S31. Using the JADE algorithm in ICA (independent component analysis), two audio components in the audio data of one frame length are calculated based on the acoustic signals x1(n) and x2(n) collected by the two microphones, and amplitude normalization is performed to keep the amplitude of the audio component unchanged from that before separation; the audio components of x1(n) and x2(n) are composed as follows:

[0018]

[0019] In the above formula, x1(n) is the acoustic signal collected by the reference microphone, x2(n) is the acoustic signal collected by the auxiliary microphone, and x d (n) and x v (n) are the two audio components of the acoustic signal at the reference microphone, a and b are the coefficients of the two audio components of the reference microphone in the auxiliary microphone, both of which are constants;

[0020] S32, collect audio components x respectively d (n) and x v (n) The last sampling point data, x d The sampling points of (n) are stored in x d (n) In the audio component frame, x v The sampling points of (n) are stored in x v (n) audio component frame, and input the last sampling point data of the current two audio components to the dual feedforward Sig-FxNLMS system;

[0021] S33. When the two audio component frames are fully stored (ie, the number of sampling points is equal to the size of one frame), the two audio component frames are respectively input into the CNN computing units corresponding to the on-board chip.

[0022] Preferably, the specific steps of executing the dual feedforward DL-Sig-FxNLMS algorithm in step S4 include:

[0023] When S1 and S2 are both closed:

[0024] S41, storing the audio component sampling point data separated in S3 into the matrix x d (n) and x v (n), and respectively d (n) and x v (n) is transmitted to the first and second feedforward subsystems, and the first and second feedforward subsystems respectively calculate the acoustic signal y to be output f (n) and y b (n), y f (n) and y b The calculation formula of (n) is as follows:

[0025]

[0026] in, are the coefficient values ​​of the first and second feedforward subsystem filters, respectively, d (n), x v (n) are two audio components obtained by the acoustic signal separation module;

[0027] S42. Calculate the total sound wave y(n). The calculation formula of y(n) is as follows:

[0028] y(n)=y f (n)+y b (n)

[0029] S43, the loudspeaker outputs the canceling reverse sound wave y′(n), where y′(n) is the canceling reverse sound wave generated by the total sound wave y(n) calculated by the algorithm after passing through the secondary channel S(z). The calculation formula of y′(n) is as follows:

[0030] y′(n)=s(n)*y(n)

[0031] S44, the error microphone collects the total error signal e(n) after the reverse sound wave in the car is superimposed and noise-reduced. e(n) can be expressed by the following mathematical formula:

[0032] e(n)=d(n)+v(n)-y′(n)

[0033] Among them, d(n) and v(n) are x d (n) and x d (n) The signal generated after passing through the primary channel P(z);

[0034] S45, inputting the error total error separation subsystem signal e(n) to adder 3;

[0035] S46, the filter of the error separation subsystem adopts the NLMS algorithm, and its output is e f (n), and input to the first feedforward subsystem, e f (n) The calculation formula is as follows:

[0036]

[0037] Among them, w e (n) is the weight vector of the error separation subsystem filter;

[0038] S47, calculate and obtain e in adder 3 b (n), e b (n) The calculation formula is as follows:

[0039] e b (n) = e(n) - e f(n)

[0040] S48, e f (n) is input to the first feedforward subsystem, e b (n) are input to the error separation subsystem and the second feedforward subsystem respectively;

[0041] S49, respectively updating the filter coefficients w of the first and second feedforward subsystems f (n+1), w b (n+1) and the filter coefficients w of the error separation subsystem e (n+1);

[0042] The weight update formula of the first feedforward subsystem can be expressed as:

[0043]

[0044] The weight update formula of the second feedforward subsystem can be expressed as:

[0045]

[0046] Among them, w f (n), w b (n) are the filter coefficients of the first and second feedforward subsystems, μ1(n) and μ2(n) are the update step sizes of the filters of the first and second feedforward subsystems, respectively. f (n), e b (n) are the error signals of the first and second feedforward subsystems, respectively, and α and β are the introduced parameter values;

[0047] w e (n+1) is calculated as follows:

[0048] w e (n+1)=w e (n)+2μ e e b (n)x d (n)

[0049] Among them, μ e is the update step size of the filter of the error separation subsystem;

[0050] If S1 is closed and S2 is open:

[0051] At this time, compared with the situation when S1 and S2 are both closed, only y f (n) Output, y b (n) is not output, so the calculation formula of y(n) in S42 is changed to:

[0052] y(n)=y f (n)

[0053] And at this time, w in S49 must be e (n+1) The calculation formula is changed to:

[0054] w e (n+1)=w e (n)

[0055] After correction, the filter coefficients w of the error separation subsystem are e (n) will not be affected by e b (n) and x d (n) is controlled to ensure that in the weight update formula of the first feedforward subsystem in S49, the output e of the error separation subsystem f The calculation process of (n) is not affected by e b (n) and x d (n) influence, that is, to ensure the error signal e of the second feedforward subsystem in S47 b (n) the calculation is normal, so that the first feedforward subsystem is not affected by the second feedforward subsystem having no output;

[0056] If S2 is closed and S1 is open:

[0057] At this time, compared with the situation when S1 and S2 are both closed, only y b (n) Output, y f (n) is not output, so the calculation formula of y(n) in S42 is changed to:

[0058] y(n)=y b (n)

[0059] In addition, in S49 w e (n+1) e in the calculation formula b (n) will remain unchanged, then e in S46 f (n) In the calculation formula, the output of the error separation subsystem e f The calculation process of (n) will only be affected by x d (n) control, and e b (n) is irrelevant, then e in S47 b (n) In the calculation formula, e(n) minus x d (n) related e f (n), the remaining component e b (n) and x d (n) irrelevant, only related to x v (n), that is, at this time, the first and second feedforward subsystems are independent of each other, and the second feedforward subsystem will not be affected by the fact that the first feedforward subsystem has no output.

[0060] Preferably, in step S5, the vehicle-mounted chip executes CNN to update the filter coefficients, and the specific steps include:

[0061] S51, several sets of fixed filter coefficient values ​​obtained after training are pre-stored in DSP; when the deep learning module receives a frame of audio data (x d (n) or x v When (n) audio component frame is obtained, the frame data is sent to the one-dimensional CNN operation. The initial input data is initially extracted by the convolution layer-batch normalization layer and then introduced into the nonlinear transformation by the ReLU activation layer;

[0062] S52, further extracting information features from the intermediate data generated in S51 through a hierarchical structure of maximum pooling-batch normalization-maximum pooling;

[0063] S53, after a residual module with a convolutional layer, a batch normalization layer and a ReLU activation layer, the residual module can extract the intrinsic features of the speech signal more deeply, making the network layer deeper without performance degradation;

[0064] S54, after the model passes through the global pooling layer and the fully connected layer, the probability of each pre-trained fixed filter coefficient is obtained and stored in a one-dimensional probability matrix;

[0065] S55, comparing and obtaining the maximum element in the one-dimensional probability matrix obtained in S4, and obtaining the number of the corresponding fixed filter coefficient;

[0066] S56. If the result obtained in S55 is different from the calculation result of the audio component of the previous frame, the on-board chip transmits the calculation result to the DSP chip via the serial port line, and w(n) in the Sig-FxNLMS algorithm is updated to the fixed filter coefficient value of the corresponding number. Otherwise, the filter coefficient value w(n) in the DSP chip will not be updated (here, if x is received in S51 d (n) audio component frame, then update w f (n); if x is received in S51 v (n) audio component frame, then update w b (n)).

[0067] Beneficial effects:

[0068] 1. The present invention realizes the effective separation and suppression of audio signals in complex environments (such as in a car) through the DL-Sig-FxNLMS algorithm. This method can effectively separate and suppress human voice and environmental noise, thereby significantly improving the accuracy and practicality of audio signal processing.

[0069] 2. The present invention adopts an iterative step size adjustment based on the Sigmoid function, which makes the filter weight update more reasonable, improves the convergence speed and stability of the system, and ensures reliable performance in different application scenarios.

[0070] 3. The present invention achieves selective suppression of specific parts of the audio signal by selectively controlling the output of the reverse sound wave, thereby enhancing the flexibility and applicability of the system, enabling it to accurately suppress specific audio components according to actual needs, and expanding its scope of application. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is a hardware connection diagram of the present invention.

[0072] Figure 2 It is the overall flow chart of the present invention.

[0073] Figure 3 It is the operation flow chart of the DSP chip of the present invention.

[0074] Figure 4 It is the calculation flow chart of the vehicle-mounted chip of the present invention.

[0075] Figure 5 It is a block diagram of the DSP chip algorithm execution of the present invention.

[0076] Figure 6 This is a structural model diagram of the deep learning module of the present invention.

[0077] Figure 7 This is a frequency band division diagram of the pre-trained filter of the present invention.

[0078] Figure 8 This is a comparison chart of the noise reduction results of Example 2 of the present invention. DETAILED DESCRIPTION

[0079] Example 1

[0080] The hardware components used in the present invention include a reference microphone, an auxiliary microphone, an error microphone, a speaker, a digital signal processing (DSP) chip and an on-board chip, and the corresponding hardware connection structure is as follows: Figure 1 The reference microphone and the auxiliary microphone collect the signals generated by the noise source, denoted as x1(n) and x2(n), and input these signals into the DSP chip. Inside the DSP chip, the two audio components x in the acoustic signal collected by the microphone are calculated through the Joint Diagonalization of Feature Matrix (JADE) algorithm. d (n) and x v (n), and are input into the convolutional neural network computing unit corresponding to the vehicle chip. At the same time, the dual feedforward adaptive filtering algorithm (Sig-FxNLMS) in the DSP chip calculates the output sound signal y f(n) and y b (n). While the DSP is executing the dual-feedforward Sig-FxNLMS algorithm, the onboard chip is synchronously executing the CNN algorithm, calculating the corresponding filter coefficient values ​​and providing feedback for adjustment. The error microphone collects the total error signal e(n) and outputs a continuously updated reverse sound wave signal in the speaker.

[0081] Example 2

[0082] like Figure 2 As shown, the present invention comprises the following steps:

[0083] S1. Microphone acquisition of acoustic signals: using a reference microphone and an auxiliary microphone, respectively collecting audio data x1(n) and x2(n) with a frame length of 10800 sampling points at a sampling rate of 10800 sampling points / second, where x1(n) and x2(n) are the audio signals collected by the reference microphone and the auxiliary microphone, respectively; using an error microphone, collecting the total error signal e(n);

[0084] S2, select mode: Select the mode as full noise reduction (S1 and S2 are both closed);

[0085] S3, the sound signal separation module works: the JADE algorithm is used to calculate the two audio components in the sound signal collected by the microphone, and the two audio components are input to the dual feedforward DL-Sig-FxNLMS system at the sampling rate; at the same time, when the number of sampling points is equal to the size of a frame, the two audio component frames are input to the calculation unit corresponding to the on-board chip;

[0086] S4, double feedforward Sig-FxNLMS algorithm execution: DSP chip operation process is as follows Figure 3 As shown, the specific algorithm operation is as follows: the first and second feedforward subsystems use the Sig-FxNLMS algorithm to calculate the reverse sound wave y f (n), y b (n), the total error signal e(n) is processed by different algorithm processes according to the mode selected by the user. The error separation subsystem separates the error signals e of the first and second feedforward subsystems from the total error signal e(n). f (n) and e b (n), and send the error signal to the corresponding system;

[0087] S5, the on-board chip executes CNN to update the filter coefficients: Figure 4 As shown, while S4 is being executed, the on-board chip executes CNN, and calculates the filter coefficient values ​​in the first and second feedforward subsystems according to the audio component frames sent in S3. When the calculated filter coefficient value is different from the last calculation result of CNN, it is updated to the fixed filter coefficient value of the corresponding number;

[0088] S6, the speaker continuously outputs updated reverse sound waves;

[0089] Step S3 specifically includes the following steps:

[0090] S31. Using the JADE algorithm in ICA, two audio components in the audio data of one frame length are calculated based on the acoustic signals x1(n) and x2(n) collected by the two microphones, and the amplitude is normalized so that the amplitude of the audio component remains unchanged compared with that before separation. The audio components of x1(n) and x2(n) are composed as follows:

[0091]

[0092] In formula (1), x1(n) is the acoustic signal collected by the reference microphone, x2(n) is the acoustic signal collected by the auxiliary microphone, and x d (n) and x v (n) are the two audio components of the acoustic signal at the reference microphone, a and b are the coefficients of the two audio components of the reference microphone in the auxiliary microphone, both of which are constants;

[0093] S32, collect audio components x respectively d (n) and x v (n) The last sampling point data, x d The sampling points of (n) are stored in x d (n) In the audio component frame, x v The sampling points of (n) are stored in x v (n) audio component frame, and input the last sampling point data of the current two audio components to the dual feedforward Sig-FxNLMS system;

[0094] S33. When the two audio component frames are fully stored (ie, the number of sampling points is equal to the size of one frame), the two audio component frames are respectively input into the CNN computing units corresponding to the on-board chip.

[0095] Step S4 specifically includes the following steps, and the algorithm flow chart is as follows: Figure 5 As shown:

[0096] S41, storing the audio component sampling point data separated in S3 into the matrix x d (n) and x v (n), and respectively d (n) and x v (n) is transmitted to the first and second feedforward subsystems, which respectively calculate the acoustic signal y to be output f (n) and y b (n), y f (n) and y bThe calculation formula of (n) is as follows:

[0097]

[0098] in, are the coefficient values ​​of the first and second feedforward subsystem filters, respectively, d (n), x v (n) are two audio components obtained by the acoustic signal separation module;

[0099] S42. Calculate the total sound wave y(n). The calculation formula of y(n) is as follows:

[0100] y(n)=y f (n)+y b (n) (4)

[0101] S43, the loudspeaker outputs the canceling reverse sound wave y′(n), where y′(n) is the canceling reverse sound wave generated by the total sound wave y(n) calculated by the algorithm after passing through the secondary channel S(z). The calculation formula of y′(n) is as follows:

[0102] y′(n)=s(n)*y(n) (5)

[0103] S44, the error microphone collects the total error signal e(n) after the reverse sound wave in the car is superimposed and noise-reduced. e(n) can be expressed by the following mathematical formula:

[0104] e(n)=d(n)+v(n)-y′(n) (6)

[0105] Among them, d(n) and v(n) are x d (n) and x d (n) The signal generated after passing through the primary channel P(z);

[0106] S45, inputting the total error signal e(n) to adder 3;

[0107] S46, the filter of the error separation subsystem adopts the NLMS algorithm, and its output is e f (n), and is used as input to the first feedforward subsystem, e f (n) The calculation formula is as follows:

[0108]

[0109] Among them, w e (n) is the weight vector of the error separation subsystem filter;

[0110] S47, calculate and obtain e in adder 3 b (n), e b (n) The calculation formula is as follows:

[0111] e b (n) = e(n) - e f (n) (8)

[0112] S48, e f (n) is input to the first feedforward subsystem, e b (n) are input to the error separation subsystem and the second feedforward subsystem respectively;

[0113] S49, respectively updating the filter coefficients w of the first and second feedforward subsystems f (n+1), w b (n+1) and the filter coefficients w of the error separation subsystem e (n+1);

[0114] The weight update formula of the first feedforward subsystem can be expressed as:

[0115]

[0116] The weight update formula of the second feedforward subsystem can be expressed as:

[0117]

[0118] Among them, w f (n), w b (n) are the filter coefficients of the first and second feedforward subsystems, μ1(n) and μ2(n) are the update step sizes of the filters of the first and second feedforward subsystems, respectively. f (n), e b (n) are the error signals of the first and second feedforward subsystems, respectively; α and β are the introduced parameter values, where α is set to 2 and β is set to 0.005;

[0119] w e (n+1) is calculated as follows:

[0120] w e (n+1)=w e (n)+2μ e e b (n)x d (n) (11)

[0121] Among them, μ e is the update step size of the filter of the error separation subsystem, which is set to 0.001 in this example.

[0122] Step S5 specifically includes the following steps: the deep learning module structure model is as follows: Figure 6 As shown:

[0123] S51, the 13 sets of fixed filter coefficient values ​​obtained after training are pre-stored in the DSP; when the deep learning module receives a frame of audio data (x d (n) or x v When (n) audio component frame is generated, the frame data is sent to the one-dimensional CNN operation. The initial input data is initially extracted by the convolution layer-batch normalization layer and then introduced into the nonlinear transformation by the ReLU activation layer. Among them, the 13 groups of fixed filter coefficients are divided according to the frequency band during pre-training. The division rules are as follows: Figure 7 shown.

[0124] S52, further extracting information features from the intermediate data generated in S51 through a hierarchical structure of maximum pooling-batch normalization-maximum pooling;

[0125] S53, after a residual module with convolutional layer, batch normalization layer and ReLU activation layer. The residual module can extract the intrinsic features of the speech signal more deeply, making the network layer deeper without performance degradation;

[0126] S54, after the model passes through the global pooling layer and the fully connected layer, the probability of each pre-trained fixed filter coefficient is obtained and stored in a one-dimensional probability matrix;

[0127] S55, comparing and obtaining the maximum element in the one-dimensional probability matrix obtained in S4, and obtaining the number of the corresponding fixed filter coefficient;

[0128] S56. If the result obtained in S55 is different from the calculation result of the audio component of the previous frame, the on-board chip transmits the calculation result to the DSP chip via the serial port line, and w(n) in the Sig-FxNLMS algorithm is updated to the fixed filter coefficient value of the corresponding number. Otherwise, the filter coefficient value w(n) in the DSP chip will not be updated (here, if x is received in S51 d (n) audio component frame, then update w f (n); if x is received in S51 v (n) audio component frame, then update w b (n)). The results are as follows Figure 8 shown.

[0129] Figure 8 (a) shows the waveform of the human voice signal. Figure 8 (b) shows the waveform of the ambient noise. The two original sound signals are mixed to form the noise reference signal of the reference microphone. Different output results are obtained after outputting in different noise reduction modes. When the algorithm indiscriminately filters out the human voice signal and the ambient noise, the effect is as follows Figure 8(c) shows that the red area represents the mixed noise reference signal, and the blue area represents the total error signal after noise reduction. It can be observed from the figure that the noise signal amplitude is significantly reduced after noise reduction, which shows that the noise reduction algorithm has a good noise reduction effect on both acoustic signal components. When the algorithm only chooses to filter out the ambient noise, the effect is as follows Figure 8 As shown in (d), the red area represents the mixed noise reference signal, and the blue area represents the error signal after filtering out the ambient noise. Figure 8 (a) The contours are basically consistent, which shows that the denoising algorithm can filter out part of the noise separately in this mode and has a good effect.

Claims

1. A dual feedforward active noise reduction system for vehicles based on the DL-Sig-FxNLMS algorithm, the structure of which includes a reference microphone, an auxiliary microphone, an error microphone, a speaker, a DSP chip and an on-board chip; the reference microphone and the auxiliary microphone are used to collect signals generated by the noise source, and input these signals into the DSP chip and the on-board chip for calculation and processing, the error microphone is used to collect the total error signal, and the speaker is used to output the continuously updated reverse sound wave signal for noise processing.

2. A dual feedforward vehicle active noise reduction method based on DL-Sig-FxNLMS algorithm, comprising the following steps: S1. Microphone acquisition of acoustic signals: Use a reference microphone and an auxiliary microphone to collect audio data x1(n) and x2(n) of one frame length at a sampling rate, where x1(n) and x2(n) are the audio signals collected by the reference microphone and the auxiliary microphone respectively; use an error microphone to collect the total error signal e(n); S2, select mode: users can select the system noise reduction mode according to their needs. The available modes are: 1) full noise reduction, that is, S1 and S2 are both closed; 2) S1 noise reduction, that is, S1 is closed and S2 is disconnected; 3) S2 noise reduction, that is, S1 is disconnected and S2 is closed; 4) stop noise reduction, that is, the system stops working; S3, the sound signal separation module works: the JADE algorithm is used to calculate the two audio components in the sound signal collected by the microphone, and the two audio components are input to the dual feedforward DL-Sig-FxNLMS system at the sampling rate; at the same time, when the number of sampling points is equal to the size of a frame, the two audio component frames are input to the calculation unit corresponding to the on-board chip; S4, dual feedforward Sig-FxNLMS algorithm execution: The first and second feedforward subsystems use the Sig-FxNLMS algorithm to calculate the reverse sound wave y f (n), y b (n), according to the mode selected by the user, the total error signal e(n) is processed by different algorithm processes, and the error signals e of the first and second feedforward subsystems are separated from the total error signal e(n) by the error separation subsystem. f (n) and e b (n), and send the error signal to the corresponding system; S5, the on-board chip executes CNN to update the filter coefficient: while S4 is being executed, the on-board chip executes CNN to calculate the filter coefficient values ​​in the first and second feedforward subsystems according to the audio component frames sent in S3. When the calculated filter coefficient value is different from the last calculation result of CNN, it is updated to the fixed filter coefficient value of the corresponding number; S6. The speaker outputs the updated reverse sound wave.

3. The dual feedforward vehicle active noise reduction method based on the DL-Sig-FxNLMS algorithm according to claim 2 is characterized in that: The specific steps of the acoustic signal separation module in step S3 include: S31. Using the JADE algorithm in ICA, two audio components in the audio data of one frame length are calculated based on the acoustic signals x1(n) and x2(n) collected by the two microphones, and the amplitude is normalized so that the amplitude of the audio component remains unchanged compared with that before separation. The audio components of x1(n) and x2(n) are composed as follows: In the above formula, x1(n) is the acoustic signal collected by the reference microphone, x2(n) is the acoustic signal collected by the auxiliary microphone, and x d (n) and x v (n) are the two audio components of the acoustic signal at the reference microphone, a and b are the coefficients of the two audio components of the reference microphone in the auxiliary microphone, both of which are constants; S32, collect audio components x respectively d (n) and x v (n) The last sampling point data, x d The sampling points of (n) are stored in x d (n) In the audio component frame, x v The sampling points of (n) are stored in x v (n) audio component frame, and input the last sampling point data of the current two audio components to the dual feedforward Sig-FxNLMS system; S33. When the two audio component frames are fully stored (ie, the number of sampling points is equal to the size of one frame), the two audio component frames are respectively input into the CNN computing units corresponding to the on-board chip.

4. The dual feedforward vehicle active noise reduction method based on the DL-Sig-FxNLMS algorithm according to claim 2 is characterized in that: The specific steps of the dual feedforward DL-Sig-FxNLMS algorithm in step S4 include: When S1 and S2 are both closed: S41, storing the audio component sampling point data separated in S3 into the matrix x b (n) and x v (n), and respectively d (n) and x v (n) is transmitted to the first and second feedforward subsystems, and the first and second feedforward subsystems respectively calculate the acoustic signal y to be output f (n) and y b (n), y f (n) and y b The calculation formula of (n) is as follows: in, are the coefficient values ​​of the first and second feedforward subsystem filters, respectively, d (n), x v (n) are two audio components obtained by the acoustic signal separation module; S42. Calculate the total sound wave y(n). The calculation formula of y(n) is as follows: y(n)=y f (n)+y b (n) S43, the loudspeaker outputs the canceling reverse sound wave y′(n), where y′(n) is the canceling reverse sound wave generated by the total sound wave y(n) calculated by the algorithm after passing through the secondary channel S(z). The calculation formula of y′(n) is as follows: y′(n)=s(n)*y(n) S44, the error microphone collects the total error signal e(n) after the reverse sound wave in the car is superimposed and noise-reduced. e(n) can be expressed by the following mathematical formula: e(n)=d(n)+v(n)-y′(n) Among them, d(n) and v(n) are x d (n) and x d (n) The signal generated after passing through the primary channel P(z); S45, inputting the error total error separation subsystem signal e(n) to adder 3; S46, the filter of the error separation subsystem adopts the NLMS algorithm, and its output is e f (n), and input to the first feedforward subsystem, e f (n) The calculation formula is as follows: Among them, w e (n) is the weight vector of the error separation subsystem filter; S47, calculate and obtain e in adder 3 b (n), e b (n) The calculation formula is as follows: e b (n)=e(n)-e f (n) S48, e f (n) is input to the first feedforward subsystem, e b (n) are input to the error separation subsystem and the second feedforward subsystem respectively; S49, respectively updating the filter coefficients w of the first and second feedforward subsystems f (n+1), w b (n+1) and the filter coefficients w of the error separation subsystem e (n+1); The weight update formula of the first feedforward subsystem can be expressed as: The weight update formula of the second feedforward subsystem can be expressed as: Among them, w f (n), w b (n) are the filter coefficients of the first and second feedforward subsystems, μ1(n) and μ2(n) are the update step sizes of the filters of the first and second feedforward subsystems, respectively. f (n), e b (n) are the error signals of the first and second feedforward subsystems, respectively, and α and β are the introduced parameter values; w e (n+1) is calculated as follows: In e (n+1)=in e (n)+2μ e and b (n)x d (n) Among them, μ e is the update step size of the filter of the error separation subsystem; If S1 is closed and S2 is open: At this time, compared with the situation when S1 and S2 are both closed, only y f (n) Output, y b (n) is not output, so the calculation formula of y(n) in S42 is changed to: y(n)=y f (n) And at this time, w in S49 should be e (n+1) The calculation formula is changed to: w e (n+1)=w e (n) After correction, the filter coefficients w of the error separation subsystem are e (n) will not be affected by e b (n) and x d (n) is controlled to ensure that in the weight update formula of the first feedforward subsystem in S49, the output e of the error separation subsystem f The calculation process of (n) is not affected by e b (n) and x d (n) influence, that is, to ensure the error signal e of the second feedforward subsystem in S47 b (n) the calculation is normal, so that the first feedforward subsystem is not affected by the second feedforward subsystem having no output; If S2 is closed and S1 is open: At this time, compared with the situation when S1 and S2 are both closed, only y b (n) Output, y f (n) is not output, so the calculation formula of y(n) in S42 is changed to: y(n)=y b (n) In addition, in S49 w e (n+1) e in the calculation formula b (n) will remain unchanged, then e in S46 f (n) In the calculation formula, the output of the error separation subsystem e f The calculation process of (n) will only be affected by x d (n) control, and e b (n) is irrelevant, then e in S47 b (n) In the calculation formula, e(n) minus x d (n) related e f (n), the remaining component e b (n) and x d (n) irrelevant, only related to x v (n), that is, at this time, the first and second feedforward subsystems are independent of each other, and the second feedforward subsystem will not be affected by the fact that the first feedforward subsystem has no output.

5. The dual feedforward vehicle active noise reduction method based on the DL-Sig-FxNLMS algorithm according to claim 2, characterized in that: Step S5: The vehicle-mounted chip executes CNN to update the filter coefficients. The specific steps include: S51, several sets of fixed filter coefficient values ​​obtained after training are pre-stored in DSP; when the deep learning module receives a frame of audio data (x d (n) or x v When (n) audio component frame is obtained, the frame data is sent to the one-dimensional CNN operation. The initial input data is initially extracted by the convolution layer-batch normalization layer and then introduced into the nonlinear transformation by the ReLU activation layer; S52, further extracting information features from the intermediate data generated in S51 through a hierarchical structure of maximum pooling-batch normalization-maximum pooling; S53, after a residual module with a convolutional layer, a batch normalization layer and a ReLU activation layer, the residual module can extract the intrinsic features of the speech signal more deeply, making the network layer deeper without performance degradation; S54, after the model passes through the global pooling layer and the fully connected layer, the probability of each pre-trained fixed filter coefficient is obtained and stored in a one-dimensional probability matrix; S55, comparing and obtaining the maximum element in the one-dimensional probability matrix obtained in S4, and obtaining the number of the corresponding fixed filter coefficient; S56. If the result obtained in S55 is different from the calculation result of the previous frame audio component, the on-board chip transmits the calculation result to the DSP chip via the serial port line, and w(n) in the Sig-FxNLMS algorithm is updated to the fixed filter coefficient value of the corresponding number. Otherwise, the filter coefficient value w(n) in the DSP chip will not be updated. If x is received in S51, d (n) audio component frame, then update w f (n); if x is received in S51 v (n) audio component frame, then update w b (n).

Images