A hybrid active noise reduction superposition method and device based on neural network

Abstract

The present invention discloses a hybrid active noise cancellation superposition method based on a neural network, including: W f W(n) or / and W b (n) is implemented by a linear filter FIR or IIR, and the filter parameters are iterated according to the LMS or RLMS algorithm. Since S(z) is non-linear, S’(z) can still be estimated by a neural network, and the parameters of the neural network are iteratively solved according to the BP algorithm. For the non-linear link existing in the path of the original noise propagation from the A-bit noise source to the B-bit speaker, such as the case where the original noise is too large, resulting in non-linearity of the device, if the feedforward filter can use a non-linear filter, it can better control the noise; for the non-linear link existing in the control channel from the C-bit error microphone to the B-bit speaker, such as the non-linearity of the error microphone or the non-linearity of the amplifier, if the feedback filter can use a non-linear filter, it can better control the noise.

Description

Technical Field

[0001] The present invention belongs to the technical field of hybrid active noise reduction, and more specifically, particularly relates to a hybrid active noise reduction superposition method and device based on a neural network. Background Art

[0002] For digital solutions of hybrid active noise reduction, it generally includes feedforward noise reduction and feedback noise reduction. Figure 1 As shown: Feedforward noise reduction, at position A, the reference microphone picks up the noise signal, and through the feedforward noise reduction controller, a signal opposite to the original noise is generated, which is then output through the speaker, and at position B, a reverse signal with the same amplitude as the original noise signal but opposite phase is generated, so that the original noise signal and the reverse signal cancel each other out. Feedback noise reduction, the error signal microphone at position C picks up the residual signal after the original noise signal and the reverse noise generated by the speaker are superimposed, and then through this residual signal, a reverse signal with the same amplitude as the residual signal and opposite phase is generated through the feedback controller, so as to further eliminate the residual noise.

[0003] General digital noise reduction controllers use FIR (finite impulse response) filters or IIR (infinite impulse response) filters. These two types of filters are linear filters. When there is a nonlinear link in the original noise propagation path from the feedforward control channel, that is, the noise source at position A to the speaker at position B, such as excessive input noise, which causes the device to produce nonlinearity; or there is nonlinearity in the transfer function from the speaker at position B to the error microphone at position C, such as the speaker is saturated, the feedforward and feedback linear filters cannot handle the harmonics and intermodulation distortion caused by nonlinearity, resulting in poor performance of active noise reduction.

[0004] In addition, in conventional hybrid noise reduction controllers, such as Figure 2 As shown, the noise reference microphone signal is x(n), the target signal is d(n), the feedforward filter output is y(n), and the error microphone signal is e(n). P(z) is the transfer function of the original noise channel, W f (n) is the feedforward filter, W b(n) is a feedback filter, and LMS (Least Mean Square) is a minimum mean square error iterative algorithm for iterating filter parameters; S(z) is the secondary channel, that is, the transfer function of the entire path from the control signal output from the speaker, through the speaker to the air, and then to the microphone. In general, the transfer function of the secondary channel cannot be obtained and can only be estimated. In the figure, S'(z) is an estimate of S(Z). In addition, for the feedback controller, the information of the original noise signal cannot be obtained. At this time, the error signal microphone signal e(n) plus the reverse noise signal y(n) output by each control iteration can be used as an estimate d'(n) of the original noise signal d(n); where W f (n), W b (n) can be implemented using FIR or IIR filters.

[0005] General digital feedforward noise reduction controllers use FIR (finite impulse response) filters or IIR (infinite impulse response) filters. These two types of filters are linear filters. When there is a nonlinear link in the path of the original noise propagation from the noise source at position A to the speaker at position B, such as the original noise is too large, causing the device to produce nonlinearity; or there is a nonlinearity in the transfer function from the speaker at position B to the error microphone at position C, such as the speaker is saturated, the linear filter cannot handle the harmonics and intermodulation distortion caused by nonlinearity, thereby reducing the noise reduction performance. Summary of the invention

[0006] The purpose of the present invention is to solve the shortcomings existing in the prior art. For example, if the original noise is too large, causing the device to produce nonlinearity, there is a nonlinear link in the path of the original noise propagation from the A-position noise source to the B-position speaker. If a nonlinear filter can be used as a feedforward filter, the noise can be better controlled; if a nonlinear link exists in the control channel from the C-position error microphone to the B-position speaker, such as the error microphone has nonlinearity, or the amplifier has nonlinearity, if a nonlinear filter can be used as a feedback filter, the noise can be better controlled. A hybrid active noise reduction superposition method and device based on a neural network is proposed.

[0007] To achieve the above object, the present invention provides the following technical solution: a hybrid active noise reduction superposition method based on a neural network, comprising:

[0008] W f (n) or / and W b (n) It is implemented by a linear filter FIR or IIR, and the filter parameters are iterated according to the LMS or RLMS algorithm. Since S(z) is nonlinear, S'(z) can still be estimated by a neural network, and the parameters of the neural network are iterated according to the BP algorithm;

[0009] Two methods are also included:

[0010] Method 1: directly modify the transfer function of the secondary channel;

[0011] S(z) is the secondary channel transfer function;

[0012] (y(n)-e(n)*W b (n))*S=e(n)

[0013] Further solving can yield:

[0014] y(n)*S(Z)-e(n)*W b (n)*S(z)=e(n)

[0015] y(n)*S(Z)=e(n)*(1+W b (n)*S(z))

[0016] Further analysis shows that the actual secondary channel estimate S'(z) is actually the original secondary channel S(z) and the feedback controller W b (n), calculated as follows:

[0017]

[0018] At this time, the feedforward and feedback controllers are superimposed to calculate the feedforward control filter W. f (n) When calculating, the S'(z) used must be a function that includes the original filter and the feedback controller, that is, in the neural network algorithm of the feedforward filter, the input information of the BP algorithm includes the information of the feedback filter, that is,

[0019] W f (n)=f(x(n),S(z),W b (n));

[0020] Method 2: Using online measurement of the secondary channel S'(z);

[0021] After the feedback controller is turned on, an excitation signal is injected into the speaker, and then received at the reference microphone. The overall transfer function of the secondary channel is calculated based on the input and output.

[0022] u(n) is the input excitation signal. Assume that the secondary channel is modeled by a neural network and passes through a feedback controller W. b(n) is output from the speaker to the reference microphone to obtain the error signal e(n). The other way is the secondary channel estimation S'(z) described by the neural network. The output signal is subtracted from e(n) to obtain the difference e'(n) between the two signals. This difference is used as the input value of the back propagation algorithm BP to solve the reverse secondary channel neural network S'(z). By continuously adjusting the parameters of S'(z), e'(n) is approached to zero, so that S'(z) becomes a feedback controller W. b An approximate estimate of the secondary channel transfer function of (n).

[0023] Preferably, when S'(z) is estimated using a neural network, the input signal of S'(z) during BP estimation includes a random signal.

[0024] Preferably, the BP algorithm used for iteratively solving the parameters of the neural network is a back propagation algorithm.

[0025] Preferably, in calculating the feedforward control filter W f (n) When calculating, the S'(z) used must be a function that includes the original filter and the feedback controller. How the function is constructed does not affect the information of the feedback filter required for the calculation of the feedforward filter.

[0026] Preferably, in the superposition of hybrid noise reduction, the filter outputs of the feedforward noise reduction and the feedback noise reduction are directly superimposed, and the calculation of the feedforward noise reduction filter requires the estimated value of the secondary channel; and during the superposition process, the transfer function of the secondary channel will change.

[0027] Preferably, that is, for feedforward noise reduction, the transfer function of the secondary channel without feedback noise reduction and the transfer function with feedback noise reduction are completely different.

[0028] Preferably, in the process of superposition of hybrid noise reduction, the influence of feedback noise reduction on the secondary channel needs to be taken into account. Otherwise, due to inaccurate estimation of the secondary channel transfer function, the superposition of feedforward noise reduction and feedback noise reduction cannot make the noise reduction effect after superposition equal to the effect of feedforward noise reduction plus the effect of feedback noise reduction.

[0029] Preferably, the LMS algorithm is a least mean square algorithm.

[0030] Preferably, in method 2, an excitation signal injected into the speaker includes a random signal and an existing excitation signal of the speaker.

[0031] The present invention also proposes a hybrid active noise reduction superposition device based on a neural network, including a hybrid active noise reduction superposition method based on a neural network.

[0032] Technical effects and advantages of the present invention: Compared with the prior art, the hybrid active noise reduction superposition method and device based on neural network provided by the present invention has a nonlinear link in the path of original noise propagation from the noise source at position A to the loudspeaker at position B, such as the original noise is too large, causing the device to produce nonlinearity. If a nonlinear filter can be used as a feedforward filter, the noise can be better controlled; if a nonlinear link exists in the control channel from the error microphone at position C to the loudspeaker at position B, such as the error microphone has nonlinearity, or the amplifier has nonlinearity, if a nonlinear filter can be used as a feedback filter, the noise can be better controlled. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A schematic diagram of hybrid noise reduction in the prior art;

[0034] Figure 2 A schematic diagram of a hybrid noise reduction linear controller and secondary channel identification in the prior art;

[0035] Figure 3 is a schematic diagram of a hybrid noise reduction controller based on a neural network according to the present invention;

[0036] Figure 4 A schematic diagram of a secondary channel estimation based on a neural network according to the present invention;

[0037] Figure 5 A schematic diagram of a secondary channel estimation and hybrid noise reduction controller based on a neural network according to the present invention;

[0038] Figure 6 It is a schematic diagram of the change of the secondary channel after the feedback control superposition of the present invention;

[0039] Figure 7 A schematic diagram of a neural network for online estimation of secondary channels according to the present invention;

[0040] Figure 8 It is a partial schematic diagram of the input layer, hidden layer and output layer and different neurons in each layer in Example 1 of the present invention;

[0041] Fig. 9 A schematic diagram of neuron definition in Example 1 of the present invention;

[0042] Fig.10 Detailed schematic diagram of the input layer, hidden layer, output layer and different neurons in each layer in Example 1 of the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0044] A hybrid active noise reduction superposition method based on a neural network, comprising:

[0045] W f (n) or / and W b (n) It is implemented by a linear filter FIR or IIR, and the filter parameters are iterated according to the LMS or RLMS algorithm. Since S(z) is nonlinear, S'(z) can still be estimated by a neural network, and the parameters of the neural network are iterated according to the BP algorithm;

[0046] General digital feedforward noise reduction controllers use FIR (finite impulse response) filters or IIR (infinite impulse response) filters. These two types of filters are linear filters. When there is a nonlinear link in the path of the original noise propagation from the noise source at position A to the speaker at position B, such as the original noise is too large, causing the device to produce nonlinearity; or there is a nonlinearity in the transfer function from the speaker at position B to the error microphone at position C, such as the speaker is saturated, the linear filter cannot handle the harmonics and intermodulation distortion caused by nonlinearity, thereby reducing the noise reduction performance.

[0047] If there is a nonlinear link in the path of original noise propagation from the noise source at position A to the speaker at position B, for example, the original noise is too large, causing the device to produce nonlinearity, if the feedforward filter can use a nonlinear filter, the noise can be better controlled; if there is a nonlinear link in the control channel from the error microphone at position C to the speaker at position B, such as the nonlinearity of the error microphone or the nonlinearity of the amplifier, if the feedback filter can use a nonlinear filter, the noise can be better controlled; we know that neural networks are a nonlinear controller, so neural networks can be used to implement hybrid noise reduction controllers.

[0048] There is nonlinearity in the transfer function from the speaker at position B to the error microphone at position C. For example, if the speaker is saturated, the linear filter cannot handle the harmonics and intermodulation distortion caused by nonlinearity. If the estimation of the secondary channel adopts nonlinear estimation, the modeling and estimation of the secondary channel will be more accurate.

[0049] like Figure 5As shown in FIG. 1 , when there are nonlinear links in the channel from the noise source at position A to the loudspeaker at position B, in the control channel from the error microphone at position C to the loudspeaker at position B, and in the channel from the loudspeaker at position B to the error microphone at position C, the feedforward and feedback controllers and the secondary channel estimation can be realized by using a neural network;

[0050] In the superposition of the above mixed noise reduction, the filter outputs of the feedforward noise reduction and the feedback noise reduction are directly superimposed, and the calculation of the feedforward noise reduction filter requires the estimated value of the secondary channel; and in the superposition process, the transfer function of the secondary channel will change. That is, for feedforward noise reduction, the transfer function of the secondary channel without feedback noise reduction and the transfer function with feedback noise reduction are completely different. Therefore, in the above superposition process, the impact of feedback noise reduction on the secondary channel needs to be taken into account, otherwise, due to the inaccurate estimation of the secondary channel transfer function, the superposition of feedforward noise reduction and feedback noise reduction cannot make the noise reduction effect after superposition equal to the effect of feedforward noise reduction plus the effect of feedback noise reduction.

[0051] There are two ways to solve this superposition problem. One way is to directly modify the transfer function of the secondary channel. The other way to solve the superposition of the feedforward and feedback controllers is to use the online measurement of the secondary channel S'(z)

[0052] like Figure 2-7 As shown, two methods are also included:

[0053] Method 1: directly modify the transfer function of the secondary channel;

[0054] S(z) is the secondary channel transfer function;

[0055] (y(n)-e(n)*W b (n))*S=e(n)

[0056] Further solving can yield:

[0057] y(n)*S(Z)-e(n)*W b (n)*S(z)=e(n)

[0058] y(n)*S(Z)=e(n)*(1+W b (n)*S(z))

[0059] Further analysis shows that the actual secondary channel estimate S'(z) is actually the original secondary channel S(z) and the feedback controller W b (n), calculated as follows:

[0060]

[0061] At this time, the feedforward and feedback controllers are superimposed to calculate the feedforward control filter W.f (n) When calculating, the S'(z) used must be a function that includes the original filter and the feedback controller, that is, in the neural network algorithm of the feedforward filter, the input information of the BP algorithm includes the information of the feedback filter, that is,

[0062] W f (n)=f(x(n),S(z),W b (n));

[0063] Method 2: Using online measurement of the secondary channel S'(z);

[0064] After the feedback controller is turned on, an excitation signal is injected into the speaker, and then received at the reference microphone. The overall transfer function of the secondary channel is calculated based on the input and output.

[0065] u(n) is the input excitation signal. Assume that the secondary channel is modeled by a neural network and passes through a feedback controller W. b (n) is output from the speaker to the reference microphone to obtain the error signal e(n). The other way is the secondary channel estimation S'(z) described by the neural network. The output signal is subtracted from e(n) to obtain the difference e'(n) between the two signals. This difference is used as the input value of the back propagation algorithm BP to solve the reverse secondary channel neural network S'(z). By continuously adjusting the parameters of S'(z), e'(n) is approached to zero, so that S'(z) becomes a feedback controller W. b An approximate estimate of the secondary channel transfer function of (n).

[0066] When S'(z) is estimated using a neural network, the input signal of S'(z) during BP estimation includes a random signal.

[0067] The BP algorithm used for iterative solution of the neural network parameters is the back propagation algorithm.

[0068] In calculating the feedforward control filter W f (n) When calculating, the S'(z) used must be a function that includes the original filter and the feedback controller. How the function is constructed does not affect the information of the feedback filter required for the calculation of the feedforward filter.

[0069] In the superposition of hybrid noise reduction, the outputs of the filter of the feedforward noise reduction and the filter of the feedback noise reduction are directly superimposed, and the calculation of the feedforward noise reduction filter requires the estimated value of the secondary channel; and in the superposition process, the transfer function of the secondary channel will change.

[0070] That is, for feedforward noise reduction, the secondary channel transfer function without feedback noise reduction and the transfer function with feedback noise reduction are completely different.

[0071] In the process of superposition of hybrid noise reduction, the influence of feedback noise reduction on the secondary channel needs to be taken into account. Otherwise, due to inaccurate estimation of the secondary channel transfer function, the superposition of feedforward noise reduction and feedback noise reduction cannot make the noise reduction effect after superposition equal to the effect of feedforward noise reduction plus the effect of feedback noise reduction.

[0072] The LMS algorithm is a least mean square algorithm.

[0073] In the second method, an excitation signal injected into the speaker includes a random signal and an existing excitation signal of the speaker.

[0074] Example 1

[0075] like Figure 8 As shown, the neural network implements W f (n), W b (n). S'(z) is implemented using FIR or IIR;

[0076] Assume W f (n) or W b (n) is implemented by the following 3*N*1 forward neural network, that is, the input layer is 3 neurons, the hidden layer is N neurons, and the output layer is 1 neuron; where w ij,h represents the weight value of the i-th input to the j-th neuron in the hidden layer, w j,o Represents the weight value from the neurons in the j hidden layers to the output;

[0077] Obviously, different numbers of layers and different numbers of neurons in each layer are extensions that are easy to deduce and understand, and also fall within the scope of protection of this patent;

[0078] like Fig. 9 As shown, the figure shows the definition of neurons;

[0079] Assume there are several input signals x i ,i=1,2,…,N;w i represents the weight coefficient corresponding to the i-th input, θ represents the threshold of the neuron, which can also be understood as another constant input value; ∑ represents accumulation, Net represents the accumulated value after each input value is multiplied by the weight value, and y is the output value; f is a nonlinear function, such as the Sigmoid function, or the hyperbolic tangent function as follows;

[0080]

[0081] Assume that the cost function of the feedforward active noise reduction control is the error mean square function:

[0082] Cost function = e 2 =(yd) 2

[0083] like Fig.10 As shown, according to the BP (Back propagation) back propagation algorithm, the iterative calculation formula of each weight of the output layer can be calculated as follows, where f' o (Net o ) is f o (Net o ) is the derivative of:

[0084] w j,o (n+1)=w j,o (n)+2e f o ′(Net o )x j,h (n)

[0085] The iterative calculation formula for each weight of the hidden layer is as follows:

[0086] w ij,h (n+1)=w ij,h (n)+2e f o ′(Net o ) j,o f h ′(Net j,h )x i (n)

[0087] Or write it uniformly

[0088] w node,o (n+1)=w node,o (n)+λ1δ node (n)x node,h (n)

[0089] Where λ1 is the convergence coefficient, δ node (n) is calculated as follows:

[0090]

[0091] At the same time, assuming that the secondary channel S'(z) is implemented by a FIR filter or an IIR filter, for example, a FIR filter, the iterative calculation formula of the weight coefficient of each order of the FIR filter is as follows, where l = 1, 2, ..., L

[0092] w l (n+1)=w l (n)+λ2e x(nl)

[0093] Example 2

[0094] FIR or IIR implementation of W f (n),W b(n), S'(z) is implemented using a feed-forward neural network;

[0095] Assume that FIR realizes W f (n),W b (n). At this time W f (n),W b The weight coefficient of (n) is iteratively calculated using the following formula according to the LMS least mean square algorithm:

[0096] w l (n+1)=w l (n)+λ2e x(nl)

[0097] S'(z) is implemented using a feedforward neural network, and its weight coefficient is iteratively calculated by the following formula:

[0098] w node,o (n+1)=w node,o (n)+λ1δ node (n)x node,h (n)

[0099]

[0100] Example 3

[0101] W f (n),W b Both (n) and S'(z) are implemented using neural networks;

[0102] W f (n),W b (n) and S'(z) are implemented using a feedforward neural network, and their weight coefficients are iteratively calculated using the following formula.

[0103] w node,o (n+1)=w node,o (n)+λ1δ node (n)x node,h (n)

[0104]

[0105] Of course, W f (n),W b The number of neural network layers and the number of neuron nodes in each layer used by S'(n) and S'(z) can be different, but the iterative process and formula of their weight coefficients are similar.

[0106] Example 4

[0107] The secondary channel estimate is iteratively calculated using the original secondary channel estimate and a feedback controller function.

[0108] As mentioned above, the secondary channel estimate S'(z) is actually the original secondary channel S(z) and the feedback controller W b (n), for example, the calculation is as follows:

[0109]

[0110] Example 5

[0111] The secondary channel estimation is obtained using the feedforward neural network BP algorithm, where the feedback controller is used as the input of the BP algorithm.

[0112] In summary: Compared with the prior art, the hybrid active noise reduction superposition method and device based on neural network provided by the present invention has a nonlinear link in the path of original noise propagation from the noise source at position A to the speaker at position B, such as the original noise is too large, causing the device to produce nonlinearity. If the feedforward filter can use a nonlinear filter, the noise can be better controlled. There is a nonlinear link in the control channel from the error microphone at position C to the speaker at position B, such as the nonlinearity of the error microphone or the nonlinearity of the amplifier. If the feedback filter can use a nonlinear filter, the noise can be better controlled. We know that neural network is a nonlinear controller, so neural network can be used to implement a hybrid noise reduction controller.

[0113] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A hybrid active noise reduction superposition method based on neural network, characterized in that: include: Wf(n) and / or Wb(n) are implemented by linear filters FIR or IIR, and the filter parameters are iterated according to the LMS or RLMS algorithm. Since S(z) is nonlinear, S'(z) can still be estimated by a neural network, and the parameters of the neural network are iteratively solved according to the BP algorithm. Two methods are also included: Method 1, directly modifying the transfer function of the secondary channel; S(z) is the secondary channel transfer function; (y(n)-e(n)*Wb(n))*S=e(n) Further solution can yield: y(n)*S(Z)-e(n)*Wb(n)*S(z)=e(n)y(n)*S(Z)=e(n)*(1+Wb(n)*S(z)) Further solution can yield that the actual secondary channel estimate S'(z) is actually a function of the original secondary channel S(z) and the feedback controller Wb(n), calculated as follows: At this time, the feedforward and feedback controllers are superimposed. When calculating the feedforward control filter Wf(n), the S'(z) used must be a function that includes the original filter and the feedback controller, that is, in the neural network algorithm of the feedforward filter, the input information of the BP algorithm includes the information of the feedback filter, that is, Wf(n) = f(x(n), S(z), Wb(n)); Method 2, using online measurement of the secondary channel S'(z); After the feedback controller is turned on, an excitation signal is injected into the speaker, and then received at the reference microphone, and the overall transfer function of the secondary channel is calculated based on the input and output; u(n) is the input The excitation signal is assumed to be a secondary channel modeled by a neural network. One path is output from the speaker to the reference microphone through a feedback controller Wb(n) to obtain an error signal e(n). The other path is a secondary channel estimate S'(z) described by a neural network. The output signal is subtracted from e(n) to obtain the difference e'(n) between the two signals. This difference is used as the input value of the back propagation algorithm BP for solving the reverse secondary channel neural network S'(z). By continuously adjusting the parameters of S'(z), e'(n) is made close to zero, so that S'(z) becomes an approximate estimate of the secondary channel transfer function with feedback controller Wb(n). In the superposition of hybrid noise reduction, the outputs of the filter of the feedforward noise reduction and the filter of the feedback noise reduction are directly superimposed, and the calculation of the feedforward noise reduction filter requires the estimated value of the secondary channel; and in the superposition process, the transfer function of the secondary channel will change; That is, for feedforward noise reduction, the transfer function of the secondary channel without feedback noise reduction and the transfer function with feedback noise reduction are completely different; In the process of superposition of hybrid noise reduction, the influence of feedback noise reduction on the secondary channel needs to be taken into account. Otherwise, due to inaccurate estimation of the secondary channel transfer function, the superposition of feedforward noise reduction and feedback noise reduction cannot make the noise reduction effect after superposition equal to the effect of feedforward noise reduction plus the effect of feedback noise reduction.

2. The hybrid active noise reduction superposition method based on neural network according to claim 1, characterized in that: When S'(z) is estimated using a neural network, the input signal of S'(z) during BP estimation includes a random signal.

3. The hybrid active noise reduction superposition method based on neural network according to claim 1, characterized in that: The BP algorithm used for iterative solution of the neural network parameters is the back propagation algorithm.

4. The hybrid active noise reduction superposition method based on neural network according to claim 1, characterized in that: When calculating the feedforward control filter Wf(n), the S'(z) used must be a function that includes the original filter and the feedback controller. How the function is constructed does not affect the information of the feedback filter required for the calculation of the feedforward filter.

5. The hybrid active noise reduction superposition method based on neural network according to claim 1, characterized in that: The LMS algorithm is a least mean square algorithm.

6. The hybrid active noise reduction superposition method based on neural network according to claim 1, characterized in that: In the second method, an excitation signal injected into the speaker includes a random signal and an existing excitation signal of the speaker.

Images