CN116758889A - Robust self-adaptive hybrid active noise reduction method and device and active noise reduction earphone - Google Patents

Abstract

The invention provides a robust self-adaptive hybrid active noise reduction method and device and an active noise reduction earphone, and belongs to the field of active noise control; the method of the invention comprises the following steps: an on-line modeling stage and a control stage, wherein in the on-line modeling stage, a secondary channel modeling signal is designed based on an ear masking curve, and on-line modeling of a secondary channel is completed by utilizing the relation between a playing signal and an error signal; obtaining a real-time primary path through secondary modeling, and further obtaining the frequency response of the transfer function of the current optimal controller; in the control stage, the frequency response of the transfer function of the current optimal controller is input into a pre-established and trained parameter reasoning model of the IIR filter to obtain the optimal controller parameters for noise control, so that noise reduction is realized. The method of the scheme of the invention is more robust in damping variable dominant noise from different directions, and improves the comfort of a user.

Description

Robust self-adaptive hybrid active noise reduction method and device and active noise reduction earphone

Technical Field

The invention belongs to the field of active noise control of headphones, and particularly relates to a robust self-adaptive hybrid active noise reduction method and device and an active noise reduction headphone.

Background

Active control systems employ a secondary sound source or actuator to cancel out primary noise or vibration. Compared with a passive noise vibration control method, the active control system is more effective in low-frequency noise reduction and vibration reduction. Therefore, the active control system has been one of research hotspots in the noise control field at home and abroad for the last 20 years. An active control system generally comprises three parts, namely a secondary sound source, a sensor and a controller, wherein an active control algorithm in the controller is a core part of the active control system and determines the effect of active noise reduction and the adaptability to the environment.

Active noise reduction technology, also called active noise control (Active Noise Control), abbreviated as ANC) technology, is an important research direction of modern noise control, and is widely used due to the characteristics of small circuit size, light weight, convenient control, capability of processing stable broadband noise, etc., so that the active noise reduction technology is greatly developed in the acoustic field and gradually matures in recent years. At present, companies researching the noise reduction earphone are endless, and the noise reduction earphone is widely used in the military and civil fields so as to powerfully protect the health of people. The active noise reduction earphone is mainly divided into the following three types:

(1) Feedforward active noise reduction (FF) headphones, features: the microphone (FF microphone) collects noise signals from the external environment of the earphone, and the earphone speaker plays the inverted signal to cancel the noise. The performance characteristics are as follows: the noise reduction frequency width may be wide, but the noise reduction depth is general.

(2) Feedback type active noise reduction (Feedback ANC) earphone, the characteristics are: the microphone (FB microphone) is positioned in the earphone and human ear coupling cavity, and collects sound (the sound contains normal music signals and environmental noise and residual noise in the coupling cavity) in the earphone and human ear coupling cavity, and the noise reduction purpose is achieved through the noise reduction chip processing. Performance characteristics: the noise reduction bandwidth is difficult to be wide, but the active noise reduction depth can be deep.

(3) Hybrid noise reduction (Hybrid ANC), i.e., feedforward active noise reduction plus feedback active noise reduction (ff+fb). Features are as follows: there is a feed-forward active noise reduction (FF ANC) width, and there is a feedback active noise reduction (FBANC) depth.

The existing active noise reduction earphone technology mostly adopts a fixed parameter mode, namely active control parameters are designed in advance according to earphone characteristics. The designed control parameters are not changed when being used by different people. This approach is also the most mature solution at present, with reduced hardware requirements. However, the active noise reduction effect of the design method is limited mainly because: 1. because of the personalized differences in each human feature, the control parameters adjusted for a particular model are difficult to adapt to each person. Even for the same person, different earphone wearing modes, the control parameters designed in advance cannot achieve the ideal active noise reduction effect. 2. The same control parameter does not guarantee an optimal control effect for all noise environments. Although some active noise reduction headphones in the market today claim to be adaptable to different noise scenarios. But it mostly adopts a semi-adaptive manner, i.e. different control parameters are selected according to different noise scenarios. But for noise environments not in the design scene, the active noise reduction effect is also somewhat reduced.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provides a robust self-adaptive hybrid active noise reduction method and device and an active noise reduction earphone.

In order to achieve the above object, the present invention proposes a robust adaptive hybrid active noise reduction method, the method comprising: an on-line modeling stage and a control stage,

in the online modeling stage, a secondary channel modeling signal is designed based on the human ear masking curve, and online modeling of the secondary channel is completed by utilizing the relation between a playing signal and an error signal; obtaining a real-time primary channel through secondary modeling, and further obtaining the frequency response of the transfer function of the current optimal controller;

in the control stage, the frequency response of the transfer function of the current optimal controller is input into a pre-established and trained parameter reasoning model of the IIR filter to obtain the optimal controller parameters for noise control, so that noise reduction is realized.

As an improvement of the method, the secondary channel modeling signal is designed based on the human ear masking curve, and the secondary channel on-line modeling is completed by utilizing the relation between the playing signal and the error signal; the method specifically comprises the following steps:

the method comprises the steps of obtaining a human ear masking curve through calculation through a playing signal of an active noise reduction earphone, and obtaining distribution of required added modeling noise in each frequency band, namely a weighted value of each frequency band;

according to the playing signal and the error signal acquired by the error sensor of the active noise reduction earphone, updating a secondary channel model by adopting a least mean square algorithm until the algorithm converges, and completing online modeling of the secondary channel; the secondary channel model is a transfer function from a secondary sound source of the active noise reduction earphone to the error sensor.

As an improvement of the method, the real-time primary path is obtained through secondary modeling, and then the frequency response of the transfer function of the current optimal controller is obtained; the method specifically comprises the following steps:

according to signals acquired by a reference microphone and an error sensor of the active noise reduction earphone, estimating a power spectrum by using a weighted overlap average method, and realizing transfer function modeling between the reference microphone and the error sensor to obtain a transfer function P of a primary channel at the moment _c (z) obtaining the transfer function W of the current optimal control filter according to the following formula _o (z) is:

where S (z) is the transfer function of the secondary channel.

As an improvement of the method, the input of the IIR filter parameter reasoning model is frequency response, and the output is estimated filter parameters, and the IIR filter parameter reasoning model comprises a linear connection layer, an activation function layer, a dropout layer, three serially connected full connection modules, a linear mapping layer and an output layer which are connected in sequence; wherein the fully connected module comprises 512 units of fully connected layers, layer normalization and activation functions.

As an improvement of the above method, the cost function J of the IIR filter parameter inference model is:

wherein f _k Is the discrete frequency on the designed target frequency band, f _NR Is the upper band boundary where the main noise reduction is expected, f _stop For E ₂ (f _k ) Defining upper boundary of frequency band corresponding to frequency band, t (f) _k ) Is at a frequency point f _k Calculation of the weight coefficient at E ₁ (f _k ) Represented at f _k The frequency response fitting deviation at the position satisfies the following formula:

wherein H (f) _k ) Representing the iteration variable at f _k Frequency response at W _opt (f _k ) Representing the optimizing target at f _k Is used for the frequency response of the (c), I ₂ Representing a binary norm;

E ₂ (f _k ) Representing the cascade filter at f _k The response condition of the position satisfies the following formula:

E ₃ (f _k ) Represented at f _k The amplitude response gap between the current variable and the optimizing target is satisfied with the following formula:

as an improvement of the method, in the control stage, the frequency response of the transfer function of the current optimal controller is input into a pre-established and trained parameter reasoning model of the IIR filter to obtain the optimal controller parameters for noise control, so as to realize noise reduction; the method specifically comprises the following steps:

transfer function W of the current optimal control filter _o (z) inputting a pre-established and trained parameter reasoning model of the IIR filter to obtain a parameter vector Label:

Label＝[F ₀ ,A ₀ ,Q ₀ ,...,F _i ,A _i ,Q _i ,...F _I-1 ,A _I-1 ,Q _I-1 ,G _Ref ]

wherein F is _i Representing the center frequency of the i-th filter peak or valley; a is that _i Indicating that the amplitude response of the ith filter at its center frequency enhances or attenuates AdB; q (Q) _i Representing the quality factor of the ith filter, determining the bandwidth range of the peak or valley of the filter; i epsilon [0,I-1 ]]I represents the cascade number of second-order IIR filters, the types of the filters are consistent, and the filters are of the peaking/notch type, G _Ref Representing the linear total gain of the cascaded filter;

ith filter parameter subvector F _i ,A _i ,Q _i Sum coefficient [ b ] _i,0 ,b _i,1 ,b _i,2 ,a _i,0 ,a _i,1 ,a _i,2 ]The mapping relation of (2) is expressed as follows:

b _i,1 ＝-2×cos(2πF _i /F _s )，/>

a _i,1 ＝-2×cos(2πF _i /F _s )，/>

wherein F is _s The sampling rate is used for controlling the system to work; the frequency response H (z) of the filter at this time is:

where z represents the frequency domain signal.

In another aspect, the present invention proposes a robust adaptive hybrid active noise reduction device for an active noise reduction earphone, the device comprising: an on-line modeling module and a control module, wherein,

the on-line modeling module is used for designing a secondary channel modeling signal based on the human ear masking curve and completing on-line modeling of the secondary channel by utilizing the relation between the playing signal and the error signal; obtaining a real-time primary path through secondary modeling, and further obtaining the frequency response of the transfer function of the current optimal controller;

the control module is used for inputting the frequency response of the transfer function of the current optimal controller into a pre-established and trained parameter reasoning model of the IIR filter to obtain the optimal controller parameters for noise control, and noise reduction is achieved.

In a third aspect, the present invention provides an active noise reduction earphone, including the robust adaptive hybrid active noise reduction device described above.

Compared with the prior art, the invention has the advantages that:

1. the method is more robust in suppressing variable main noise from different directions, can be adopted in a single-channel or multi-channel adaptive hybrid ANC system, and can update a secondary channel and a primary channel according to a use scene so as to generate proper filter parameters;

2. the cascade IIR filter is used as a controller parameter, so that the calculation load of equipment is reduced;

3. modeling is performed on the non-convex problem of the optimization of the IIR filter coefficient by using a neural network, so that generalization performance is improved, online coefficient generation of the IIR filter is possible, and certain calculation resources can be saved and calculation efficiency is improved unlike the traditional optimization mode;

4. the method is divided into two stages, an online modeling stage and a control stage, wherein the network used in the control stage needs to be trained offline in advance, the online modeling stage models the frequency response of the current optimal controller, and the control stage uses network output parameters for updating the filter according to the modeling result.

Drawings

FIG. 1 is a schematic illustration of online modeling of a secondary channel of the present invention;

FIG. 2 is a block diagram of the system of the present invention when modeling an optimal controller during an online modeling phase;

FIG. 3 is a schematic diagram of the neural network used in the control phase of the present invention;

fig. 4 is a flow chart of the robust adaptive hybrid active noise reduction method of the present invention.

Detailed Description

In order to solve the defects in the prior art, the invention provides a robust self-adaptive hybrid active noise reduction method. In the method, the adaptive robust active noise reduction earphone is researched to automatically estimate the secondary channel model parameters according to different individuals and wearing modes, and the adaptive algorithm is adopted to automatically adjust the control parameters, so that noise fields in different noise environments are offset, and the active noise reduction performance is effectively improved. The secondary channel refers to the transfer function of the secondary sound source to the error sensor in the active noise control system. The method comprises the following steps:

in an online modeling stage, firstly, a secondary channel modeling signal is designed based on an ear masking curve, so that the modeling signal is prevented from being perceived by the ear when being added into a playing signal; and completing the online modeling of the secondary channel by utilizing the relation between the playing signal and the error signal. And then obtaining a real-time primary path through secondary modeling, and finally obtaining the frequency response of the transfer function of the current optimal controller.

In the control stage, based on the result obtained in the online modeling stage, the trained network structure is utilized to generate filter parameters in real time, and active noise control is realized at the noise reduction target position.

The method specifically comprises the following steps:

step 1) in the online modeling stage, a secondary channel is first established at the time of initial wear. Modeling a transfer function model of the secondary sound source to the error sensor respectively by using a least mean square (Least mean square, LMS) algorithm; and obtaining the distribution of the required added modeling noise in each frequency band according to the calculated masking curve. And updating a secondary channel model by adopting an LMS algorithm according to the earphone loudspeaker input signal and the error signal, and if the modeling algorithm achieves convergence, further reducing the energy of additional added noise and avoiding being perceived by human ears.

For different individuals or wearing modes, the secondary channel models have large differences, so that for the active noise reduction earphone, the secondary channel online modeling is needed. Conventional on-line modeling of secondary channels typically gives a white noise to the secondary source, and a secondary channel model is derived from the error received signal and the secondary source input signal. However, for active noise reduction headphones this approach is clearly unsuitable, so the method investigates using music played by the headphones themselves or adding unnoticed signals to it for secondary channel modeling to increase the robustness of the active noise reduction headphones. The secondary channel is a transfer function between the earphone loudspeaker and the error sensor, and the accuracy of the model directly influences the noise reduction performance of the adaptive noise reduction earphone. There is a large difference in secondary channels for different individuals and different wearing patterns. Because the function of the adaptive noise reduction earphone is to play voice or music, the secondary channel modeling uses the audio signal played by the earphone loudspeaker as an excitation signal, and is obtained by calculating the response relation between the loudspeaker input and the error point receiving signal. For voice or music, the frequency domain has stronger sparsity, so that certain errors exist in the frequency response of certain frequency points due to insufficient signal-to-noise ratio, and the method researches adding extra noise in the played audio signal to obtain more accurate secondary channel response.

And 2) acquiring signals at the reference point and the error point, estimating a power spectrum by using a weighted overlap average method, and realizing transfer function modeling between the reference point and the error point, namely the primary channel at the moment. Due to the units in the earphoneThe distribution satisfies causality and can obtain the transfer function W of the optimal control filter _o (z) is:

wherein P (z) is the transfer function of the reference microphone to the error microphone, referred to as the primary channel; s (z) is the transfer function of the secondary sound source to the error sensor, i.e. the secondary channel. The control block diagram of the system at this time is shown in fig. 2. When the error signal E (z) is desirably canceled, the interference signal at the error point can be canceled as much as possible by using the filter corresponding to the equation (1).

When the error signal is perfectly cancelled, i.e. E, according to equation (1) _o When (z) =0, then there is the current optimal controller:

P _c (z) is the primary channel transfer function at the moment, in the online modeling stage, the noise reduction system is in a working state, and the frequency response H of the current controller of the system is known _c (z). At this time, error point signal E _c (z) can be expressed as

Thereby, a transfer function model T between the error point and the reference point can be established _c (z)

The primary channel P can be obtained by the conversion of the equation _c (z) can be expressed as:

P _c (z)＝T _c (z)-H _c (z)S(z) (8)

the combination of the two (5) and (8) can be ideal at presentControlling the frequency response W of a filter _o The expression of (z) is:

in summary, the proposed strategy can complete the primary channel measurement while the ANC system is in an online operating state. And the online modeling of the frequency response of the current optimal controller of the system is realized.

Step 3) the result W of formula (1) _opt And (3) inputting the model of the (z) into the trained neural network model to obtain a parameter vector label. The expression form of the parameter vector is as follows:

Label＝[F ₀ ,A ₀ ,Q ₀ ,...,F _i ,A _i ,Q _i ,...F _I-1 ,A _I-1 ,Q _I-1 ,G _Ref ] (10)

wherein F is _i Representing the center frequency of the i-th filter peak or valley; a is that _i Indicating that the amplitude response of the ith filter at its center frequency enhances/attenuates AdB; q (Q) _i Representing the quality factor of the i-th filter, determines the bandwidth range of the peak or valley of the filter.

In the method, the type of the second-order IIR filter used is consistent, and the second-order IIR filter is a filter of the peak/notch type. Ith filter parameter subvector F _i ,A _i ,Q _i Sum coefficient [ b ] _i,0 ,b _i,1 ,b _i,2 ,a _i,0 ,a _i,1 ,a _i,2 ]The mapping of (2) may be expressed as follows:

b _i,1 ＝-2×cos(2πF _i /F _s )，/>

wherein F is _s To control the sampling rate at which the system operates. The frequency response of the filter at this time can be expressed as:

wherein G is _Ref The linear total gain of the cascaded second-order IIR filter is represented, and I represents the number of filter cascades.

Fig. 1 and 2 show a structural block diagram of secondary channel modeling and a structural block diagram of optimal controller frequency response modeling in an online modeling stage, which are used for frequency response modeling of a transfer function of an optimal solution of a current minimum phase system.

The network architecture used in connection with the present method of network training is shown in fig. 3. The model comprises an IIR filter parameter reasoning network, after the model is input into the network, the characteristic extraction is carried out through a linear connection layer, and the function layer is activated to play a role in avoiding overfitting through a dropout layer. The main components of the network are three fully connected modules in series. Each module consists of 512 units of fully connected layers, layer normalization and ReLU activation functions. This part of the network architecture is similar to a multi-layer perceptron network. The final output of the network is 3*K, which is divided into 3 vectors of size K, corresponding to the center frequencies, gains and quality factors of the different IIR filters. K is the number of cascaded second-order IIR filters.

One notable point in such a network design is that the output parameters must be constrained, and adding constraints directly to the network output layer by activating functions can eliminate gradients during model training, easily leading to extremum saturation. In the method, normalization processing is carried out on data input, so that unified constraint is carried out on output layers through a sigmoid activation function, so that output has physical significance, and meanwhile, the output is in a desired range through constraint clipping. Because the frequency response of the cascade second-order IIR filter can be calculated by using the formula (4), the conversion of the parameter-IIR filter coefficient and the calculation of the corresponding frequency response are simultaneously carried out on the parameters output by the sample and the model when the cost function is calculated, so that the frequency response of the sample and the model is ensured to be evaluated in the same calculation environment, the calculation of the cost function is further advanced, and the output of the system training model is formed. The cost function of the model is defined as follows:

wherein f _k (k=1,., K) is the discrete frequency on the design target frequency band, t (f _k ) Is at a frequency point f _k Calculation of the weight coefficient at E ₁ (f _k ) Represented at f _k Fitting deviation of frequency response at E ₁ (f _k ) Expressed as follows

H(f _k ) Representing the iteration variable at f _k Frequency response at W _opt (f _k ) Representing the optimizing target at f _k Is a frequency response of (a) to (b). The calculation comprises phase information and amplitude information of iteration variables and optimization targets so as to achieve better fitting effect. E (E) ₂ (f _k ) Representing the cascade filter at f _k Response conditions at E ₂ (f _k ) The formula is expressed as follows

E ₃ (f _k ) Represented at f _k The amplitude response gap between the current variable and the optimizing target is expressed as follows:

to sum up, weThe objective function is divided into four parts here. First part E ₁ Noise reduction amount corresponding to the target frequency band. In order to maximize the amount of noise reduction, the gap between the fitting result and the target is as small as possible, and the design objective of this part should be minimized. The weighting coefficient of the frequency band is relatively maximum among the three parts. f (f) _NR Is the upper band bound where major noise reduction is desired; second part E ₂ Corresponding to the frequency response of the defined frequency band. Although the noise reduction is not of major concern in this frequency band, the main noise should not be enhanced, i.e. it is desirable that the smaller the amplitude response of the controller in this frequency band is, the better. f (f) _stop Is the upper bound of the frequency band corresponding to the part. Third section E ₃ Corresponding to the target frequency band at f _stop The difference between the amplitude response mode and the design target mode of the above controller. At f _stop The phase response changes quickly and the fitting difficulty is high, so that only the amplitude response is limited. Similar to the first part, the design goals of this part should also be minimized, with the corresponding weight coefficients being the smallest of the first three parts.

Step 4) repeatedly calling the step 3) until the preset iteration times are reached, screening the most suitable filter parameters according to the proposed cost function, using the corresponding filter coefficients for updating the actual control filter, and driving the secondary source to send out signals for canceling interference at the error point; and then initializing, and continuing to calculate the next time interval by using the method.

The technical scheme of the invention is described in detail below with reference to the accompanying drawings and examples.

Example 1

The embodiment 1 of the invention provides a robust self-adaptive hybrid active noise reduction method for an active noise reduction earphone. Comprising the following steps: an online modeling stage and a control stage.

Specifically, the implementation process of online modeling is as follows:

1. modeling the secondary channel online, and modeling transfer function models of the secondary sound sources to the error sensors respectively by using a least mean square (Least mean square, LMS) algorithm; and obtaining the distribution of the required added modeling noise in each frequency band according to the calculated masking curve. And the secondary channel model is updated by adopting an LMS algorithm according to the input signal and the error signal of the earphone loudspeaker, and if the modeling algorithm achieves convergence, the energy of additional added noise is further reduced, and the noise is prevented from being perceived by human ears.

2. When the sampling time accumulation reaches a preset time interval t, taking 1/4 section of signals after the sampling interval, calculating the power spectrum density between the reference point and the error point at the moment according to a formula (7), and substituting the power spectrum density, the frequency response and a secondary channel of the controller at the moment of the system into the formula (9), so that the frequency response of the theoretical optimal controller of the system at the moment can be obtained. This calculation is performed once every predetermined time interval to save calculation power.

3. And (5) training a network.

Data is a key component of the training and evaluation of deep learning algorithms. In this method, we based on the conventional design experience of the verification platform, set the sampling rate to fs=384 kHz, the number of filters k=4, the number of fft points n=1024 by cascading the randomly generated amplitude-frequency response of the IIR filters, and configure the network to match this configuration. This provides a sample pair of the desired amplitude-frequency response and corresponding cascaded IIR filter parameters. The distribution for random sampling of IIR filter bank parameters depends on the band type. The training set and the validation set have equal distributions. In the method, the IIR filters used for training and design are of peak type, and the sampling range of the center frequency is 50-500Hz according to the reference design experience of a verification platform used for experiments, so that uniform distribution is obeyed. The peak filter frequency bands are arranged in ascending frequency order to minimize confusion of parameter loss terms to the arrangement and increase the parameter symmetry. The gain of the sample set parameters [ -45,30] obeys the beta distribution (2, 5). The distribution of the data set generation relies on more traditional design experience, which also means that the model can have better fit while training is also more challenging. The quality factor Q is concentrated in a higher range, its distribution obeys the beta distribution (5, 1). The model evaluation was performed using the same distribution of validation sets as the training set, with a data volume of 1/8 of that of the training set. Here, we normalize the parameters to (0, 1) according to the respective upper and lower limits, considering the magnitude and distribution range of the center frequency gain and quality factor, preventing the output from being over saturated.

The network architecture used in the present method is shown in fig. 3. The model comprises an IIR filter parameter reasoning network, after the model is input into the network, the characteristic extraction is carried out through a linear connection layer, and the function layer is activated to play a role in avoiding overfitting through a dropout layer. The main components of the network are three fully connected modules in series. Each module consists of 512 units of fully connected layers, layer normalization and ReLU activation functions. This part of the network architecture is similar to a multi-layer perceptron network. The final output of the network is 3*K, which is divided into 3 vectors of size K, corresponding to the center frequencies, gains and quality factors of the different IIR filters.

And (3) a control stage: and inputting the frequency response of the transfer function of the current optimal controller into a built and trained parameter reasoning model of the IIR filter to obtain the parameters of the optimal controller for noise control, thereby realizing noise reduction.

As shown in fig. 4, a flow chart of the robust adaptive hybrid active noise reduction method of the present invention is shown. The above fully shows that the invention provides a robust self-adaptive hybrid active noise reduction method.

Example 2

Embodiment 2 of the present invention proposes a robust adaptive hybrid active noise reduction device for active noise reduction headphones, implemented based on the method of embodiment 1, the device comprising: an on-line modeling module and a control module, wherein,

the on-line modeling module is used for designing a secondary channel modeling signal based on the human ear masking curve and completing on-line modeling of the secondary channel by utilizing the relation between the playing signal and the error signal; obtaining a real-time primary path through secondary modeling, and further obtaining the frequency response of the transfer function of the current optimal controller;

the control module is used for inputting the frequency response of the transfer function of the current optimal controller into a pre-established and trained parameter reasoning model of the IIR filter to obtain the optimal controller parameters for noise control, and noise reduction is achieved.

It should be noted that the above-described system can be implemented in a variety of ways, such as software, hardware, or a combination of hardware and software. The hardware platform may be a central processing unit (Central processing unit, CPU), a field programmable gate array (Field programmable gate array, FPGA), a programmable logic device (Programmable logic device, PLD) or other application specific integrated circuit (Application specific integrated circuit, ASIC). The software platform includes a digital signal processor (Digital signal processing, DSP), ARM, or other microprocessor. The combination of software and hardware, for example, part of the modules are implemented in DSP software and part of the modules are implemented in hardware accelerators.

Example 3

Embodiment 3 of the present invention proposes an active noise reduction earphone, comprising the robust adaptive hybrid active noise reduction apparatus of embodiment 2.

Finally, it should be noted that the above embodiments are only for illustrating the technical solution of the present invention and are not limiting. Although the present invention has been described in detail with reference to the embodiments, it should be understood by those skilled in the art that modifications and equivalents may be made thereto without departing from the spirit and scope of the present invention, which is intended to be covered by the appended claims.

Claims (8)

1. A robust adaptive hybrid active noise reduction method, the method comprising: an on-line modeling stage and a control stage,

in the online modeling stage, a secondary channel modeling signal is designed based on the human ear masking curve, and online modeling of the secondary channel is completed by utilizing the relation between a playing signal and an error signal; obtaining a real-time primary channel through secondary modeling, and further obtaining the frequency response of the transfer function of the current optimal controller;

in the control stage, the frequency response of the transfer function of the current optimal controller is input into a pre-established and trained parameter reasoning model of the IIR filter to obtain the optimal controller parameters for noise control, so that noise reduction is realized.

2. The robust adaptive hybrid active noise reduction method of claim 1, wherein the designing a secondary channel modeling signal based on the human ear masking curve, and completing the secondary channel on-line modeling using a relationship between a play signal and an error signal; the method specifically comprises the following steps:

the method comprises the steps of obtaining a human ear masking curve through calculation through a playing signal of an active noise reduction earphone, and obtaining distribution of required added modeling noise in each frequency band, namely a weighted value of each frequency band;

according to the playing signal and the error signal acquired by the error sensor of the active noise reduction earphone, updating a secondary channel model by adopting a least mean square algorithm until the algorithm converges, and completing online modeling of the secondary channel; the secondary channel model is a transfer function from a secondary sound source of the active noise reduction earphone to the error sensor.

3. The robust adaptive hybrid active noise reduction method of claim 2, wherein the secondary modeling is performed to obtain a real-time primary path, thereby obtaining a frequency response of a transfer function of the current optimal controller; the method specifically comprises the following steps:

according to signals acquired by a reference microphone and an error sensor of the active noise reduction earphone, estimating a power spectrum by using a weighted overlap average method, and realizing transfer function modeling between the reference microphone and the error sensor to obtain a transfer function P of a primary channel at the moment _c (z) obtaining the transfer function W of the current optimal control filter according to the following formula _o (z) is:

where S (z) is the transfer function of the secondary channel.

4. The robust adaptive hybrid active noise reduction method of claim 1, wherein the input of the IIR filter parameter inference model is a frequency response and the output is an estimated filter parameter, and the IIR filter parameter inference model comprises a linear connection layer, an activation function layer, a dropout layer, three serially connected full connection modules, a linear mapping and output layer, which are sequentially connected; wherein the fully connected module comprises 512 units of fully connected layers, layer normalization and activation functions.

5. The robust adaptive hybrid active noise reduction method of claim 4, wherein the cost function J of the IIR filter parametric inference model is:

wherein f _k Is the discrete frequency on the designed target frequency band, f _NR Is the upper band boundary where the main noise reduction is expected, f _stop For E ₂ (f _k ) Defining upper boundary of frequency band corresponding to frequency band, t (f) _k ) Is at a frequency point f _k Calculation of the weight coefficient at E ₁ (f _k ) Represented at f _k The frequency response fitting deviation at the position satisfies the following formula:

wherein H (f) _k ) Representing the iteration variable at f _k Frequency response at W _opt (f _k ) Representing the optimizing target at f _k Is used for the frequency response of the (c), I ₂ Representing a binary norm;

E ₂ (f _k ) Representing the cascade filter at f _k The response condition of the position satisfies the following formula:

E ₃ (f _k ) Represented at f _k The amplitude response gap between the current variable and the optimizing target is satisfied with the following formula:

6. the robust adaptive hybrid active noise reduction method according to claim 4, wherein in the control stage, the frequency response of the current optimal controller transfer function is input into a pre-established and trained IIR filter parameter inference model to obtain optimal controller parameters for noise control, so as to realize noise reduction; the method specifically comprises the following steps:

transfer function W of the current optimal control filter _o (z) inputting a pre-established and trained parameter reasoning model of the IIR filter to obtain a parameter vector Label:

Label＝[F ₀ ,A ₀ ,Q ₀ ,...,F _i ,A _i ,Q _i ,...F _I-1 ,A _I-1 ,Q _I-1 ,G _Ref ]

wherein F is _i Representing the center frequency of the i-th filter peak or valley; a is that _i Indicating that the amplitude response of the ith filter at its center frequency enhances or attenuates AdB; q (Q) _i Representing the quality factor of the ith filter, determining the bandwidth range of the peak or valley of the filter; i epsilon [0,I-1 ]]I represents the cascade number of second-order IIR filters, the types of the filters are consistent, and the filters are of the peaking/notch type, G _Ref Representing the linear total gain of the cascaded second-order IIR filter;

ith filter parameter subvector F _i ,A _i ,Q _i Sum coefficient [ b ] _i,0 ,b _i,1 ,b _i,2 ,a _i,0 ,a _i,1 ,a _i,2 ]The mapping relation of (2) is expressed as follows:

b _i,1 ＝-2×cos(2πF _i /F _s )，/>

a _i,1 ＝-2×cos(2πF _i /F _s )，/>

wherein F is _s The sampling rate is used for controlling the system to work; the frequency response H (z) of the filter at this time is:

where z represents the frequency domain signal.

7. A robust adaptive hybrid active noise reduction device for an active noise reduction earphone, the device comprising: an on-line modeling module and a control module, wherein,

the on-line modeling module is used for designing a secondary channel modeling signal based on the human ear masking curve and completing on-line modeling of the secondary channel by utilizing the relation between the playing signal and the error signal; obtaining a real-time primary path through secondary modeling, and further obtaining the frequency response of the transfer function of the current optimal controller;

the control module is used for inputting the frequency response of the transfer function of the current optimal controller into a pre-established and trained parameter reasoning model of the IIR filter to obtain the optimal controller parameters for noise control, and noise reduction is achieved.

8. An active noise reduction earphone, characterized in that the earphone comprises a robust adaptive hybrid active noise reduction device according to claim 7.