Adaptive active noise control system and method based on deep neural network
By adopting deep neural networks in active noise control and combining the structure of RNN and MLP nodes, the problem of poor nonlinear noise control by the minimum mean square error algorithm is solved, and more efficient noise cancellation and stability improvement is achieved.
Patent Information
- Application Number
- CN202011416763.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-12-07
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2040-12-07
AI Technical Summary
The existing minimum mean square error algorithm is not effective when dealing with nonlinear noise, and the active noise control method based on artificial neural network is complex in calculation and slow in convergence, making it difficult to be applied to nonlinear noise control.
Adaptive active noise control system based on deep neural networks is adopted, and the deep network structure of RNN recurrent neural network and MLP multi-layer perceptron nodes is used to update network parameters through backpropagation algorithm to generate the same phase opposite to the amplitude of the noise signal to be controlled. and convert it into control sound waves to achieve noise cancellation.
It significantly improves the scope of application of active noise control technology to nonlinear noise, reduces the network scale, improves the calculation speed and convergence performance, and enhances the stability and noise reduction effect of the algorithm.
Smart Images

Figure CN112562628B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of active noise control, and in particular to an adaptive active noise control system based on a deep neural network and a method thereof. Background Art
[0002] Noise pollution is an environmental issue that has attracted worldwide attention. General noise interference will affect people's normal life and work, and long-term exposure to high noise environment will cause serious harm to people's psychological and physical health. Therefore, noise control has long been an important task.
[0003] From a strategic point of view, traditional noise control is mainly based on acoustic noise control methods, and technical means include sound absorption treatment, sound insulation treatment, vibration isolation and reduction, etc. The mechanism of these methods is to make the noise sound waves interact with acoustic materials or structures to consume sound energy, thereby achieving the purpose of reducing noise, which is called "passive" noise control. Generally speaking, passive control methods are more effective in reducing mid- and high-frequency noise, but have little effect on low-frequency noise. For this reason, active noise control technology provides a new solution.
[0004] Active noise control technology is also called active noise reduction technology. It controls a secondary sound source to emit a control sound wave with the same amplitude and opposite phase as the original noise through a reasonably designed adaptive filter. The destructive interference of the sound wave is used to make the control sound wave and the original noise overlap and cancel each other at the control point to achieve the purpose of noise elimination. The most commonly used algorithm is the least mean square (LMS) algorithm based on Wiener filtering theory. The algorithm uses the fastest gradient descent method to adjust the filter parameters to achieve adaptive control of unknown noise.
[0005] As a linear control algorithm, the LMS algorithm has the advantages of low computational complexity and good convergence for stationary signals. Since most active noise control systems have good linearity over a large range, this makes them satisfactory to a certain extent. However, for noise with serious nonlinear characteristics, linear control algorithms will find it difficult to achieve ideal control effects. Therefore, how to design a nonlinear control algorithm with good performance is a major difficulty in the development of active noise control.
[0006] One of the solutions is to use artificial neural network (ANN) technology. ANN is a new technology that developed rapidly at the end of the 20th century and is a branch of machine learning. This technology has the characteristics of good nonlinear mapping ability and adaptive learning ability, and provides a new idea for active control of nonlinear noise. However, the existing research on active noise control based on ANN still has major defects: ANN's fitting of nonlinear functions relies on a large hidden layer to perform high-dimensional expansion of the input vector, which undoubtedly greatly increases the computational cost of the algorithm, and its convergence speed and noise reduction effect are also unsatisfactory when the network scale is small. These shortcomings make it difficult for ANN technology to become mainstream in the field of active noise control.
[0007] In 2006, the concept of deep learning was formally proposed based on ANN. Deep learning, also known as deep neural network (DNN), goes beyond the neuroscience view of previous machine learning models and resorts to the more general principle of learning "multi-level combination", which can also be applied to machine learning frameworks that are not inspired by neuroscience. Among the various applications of DNN technology, recurrent neural network (RNN) is a type of neural network used to process sequence data x. (1) , …, x (τ) The special network of the hidden layer is represented by the following formula: (t) ,
[0008] H (t) =f(H (t-1) , x (t) )
[0009] =g (t) (x (t) , x (t-1) , …, x (1) )
[0010] Obviously, the function g (t) The entire past can be entered into x (t) , x (t-1) , …, x (1) Mapped to a fixed length H (t) , and the learned model always has a fixed input size, because the model specifies the transition from one state to another. This feature makes RNN require far fewer training samples than other network structures that are not based on sequences.
[0011] After searching, no application of RNN in the field of active noise control was found. Summary of the invention
[0012] The purpose of the present invention is to overcome the shortcomings of the prior art and propose an adaptive active noise control system and method based on a deep neural network to solve the problem that the mainstream minimum mean square error algorithm is not suitable for nonlinear noise.
[0013] The present invention solves the technical problem by adopting the following technical solutions:
[0014] An adaptive active noise control system based on a deep neural network includes a reference microphone, a controller, an actuator and an error microphone, wherein the controller includes a deep neural network module and a driving circuit;
[0015] The reference microphone is arranged near the noise source and is used to collect the reference signal and input it into the input end of the deep neural network;
[0016] The error microphone is arranged at the control point, and is used to collect the error signal and input it into the income end of the deep neural network;
[0017] The deep neural network module processes the reference signal to generate a control signal with the same amplitude and opposite phase as the noise signal to be controlled, and outputs the generated control signal to the driving circuit; the deep neural network module performs gradient calculation based on the error signal using the back propagation algorithm to update the network parameters;
[0018] The driving circuit outputs the control signal to the actuator;
[0019] The actuator converts the control signal into a control sound wave, which is superimposed on the noise to be controlled at the control point to perform active noise reduction.
[0020] Furthermore, the deep neural network module is a DNN deep neural network structure of an RNN recurrent neural network nested with an MLP multi-layer perceptron node. The deep neural network module is composed of a three-layer deep network consisting of an input layer, a hidden layer and an output layer, wherein the input layer includes 5 input nodes, the hidden layer consists of 10 RNN recurrent neurons and an MLP node, and the output layer includes an output node.
[0021] Furthermore, the controller is an integrated circuit hardware chip, and a deep neural network module and a driving circuit are integrated in the hardware chip.
[0022] A control method for an adaptive active noise control system based on a deep neural network comprises the following steps:
[0023] Step 1: A reference microphone disposed near a noise source collects a reference signal and inputs the reference signal into a deep neural network module; an error microphone disposed at a control point collects an error signal and inputs the error signal into the deep neural network module;
[0024] Step 2: The deep neural network module uses the back propagation algorithm to calculate the gradient and update the network parameters according to the error signal; the deep neural network module processes the reference signal to generate a control signal with the same amplitude and opposite phase as the noise signal to be controlled;
[0025] Step 3: The deep neural network module inputs the generated control signal to the actuator through the driving circuit.
[0026] Step 4: The actuator converts the control signal into a control sound wave, which is superimposed with the noise to be controlled at the control point to perform active noise reduction.
[0027] Furthermore, the specific implementation method of step 2 includes the following steps:
[0028] (1) Use the following formula to calculate the current RNN neuron state H (t) ,
[0029] H (t) =XV+H (t-1) W,
[0030] Among them, H (t-1) is the RNN neuron state at the previous moment; X is the input vector; V and W are weight coefficient matrices;
[0031] (2) Use the following formula to calculate the MLP node state b,
[0032]
[0033] Among them, U is the weight coefficient matrix;
[0034] (3) Use the following formula to calculate the output value of the control signal at the current moment
[0035]
[0036] Among them, α and β are weight coefficient matrices;
[0037] (4) Use the back propagation algorithm to calculate the negative gradient of each weight coefficient matrix. The calculation formula is as follows:
[0038]
[0039] Δβ=eb,
[0040] ΔU i =eβX i ,
[0041] ΔV ij =eα j X i ,
[0042]
[0043] Wherein, e is the sample value of the error signal obtained at the current moment;
[0044] (5) Use the following formula to update the network parameters:
[0045] α=α+ηΔα,
[0046] β=β+ηΔβ,
[0047] U=U+ηΔU,
[0048] V=V+ηΔV,
[0049] W=W+ηΔW
[0050] Among them, η is a preset learning rate.
[0051] The advantages and positive effects of the present invention are:
[0052] 1. As an active noise control method, the present invention fully utilizes the advantages of deep learning technology in fitting nonlinear models, solves the defect that the traditional minimum mean square error algorithm cannot control nonlinear noise, and greatly improves the scope of application of active noise control technology.
[0053] 2. The present invention is based on the RNN recurrent neural network and makes full use of its advantages in processing time series signals. Compared with the ordinary neural network denoising method, it reduces the network scale and greatly improves the calculation speed and convergence performance.
[0054] 3. The present invention proposes a new deep network structure by combining the RNN recurrent neural network and the MLP multi-layer perceptron network. This structure greatly improves the stability of the algorithm without affecting the calculation speed by adding an MLP node that does not participate in the cycle in the hidden layer, and the convergence performance is also significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1 is a connection diagram of the adaptive active noise control system of the present invention;
[0056] Figure 2 is a structural diagram of a deep neural network of the present invention;
[0057] Figure 3It is a comparison chart of simulation results of active control of nonlinear noise by the present invention and traditional LMS algorithm and ANN algorithm; wherein, (a) is the time domain waveform of the original noise signal, (b) is the time domain waveform of the error signal controlled by the traditional LMS algorithm, (c) is the time domain waveform of the error signal controlled by the traditional ANN algorithm, and (d) is the time domain waveform of the control using the deep neural network structure proposed in the present invention. DETAILED DESCRIPTION
[0058] The present invention is further described in detail below with reference to the accompanying drawings.
[0059] The design concept of the present invention is:
[0060] The minimum mean square error algorithm is limited by its simple linear structure and cannot effectively control nonlinear noise; while the traditional ANN artificial neural network algorithm has good nonlinear mapping capabilities, but has high requirements for the number of nodes in the hidden layer, which leads to its slow calculation speed and is not conducive to real-time control of noise. With the continuous development of deep learning technology, the idea of using deep neural networks to study nonlinear problems has gradually been extended to different disciplines and fields. Among the many applications of this technology, the most well-known include intelligent speech recognition and natural language processing. Most of this type of research is based on RNN recurrent neural networks, which calculates and analyzes acoustic data based on sequences, which is similar to the data form that needs to be processed in the field of active noise control. The success of these studies has provided new development ideas for the active control of nonlinear noise.
[0061] RNN is a classic deep neural network structure. (1) , …, x (τ) The sequence input, the state of its hidden layer H (t) is represented as
[0062] H (t) =f(H (t-1) , x (t) )
[0063] =g (t) (x (t) , x (t-1) , …, x (1) )
[0064] Obviously, in RNN, the state of the hidden layer at the current moment needs to be represented by the hidden layer at the previous moment. After unrolling the loop, we find that the function g (t) The entire past can be entered into x (t) , x (t-1) , …, x (1) Mapped to a fixed length H (t), and the learned model always has a fixed input size, because the model specifies the transition from one state to another. Obviously, in RNN, the input layer and hidden layer can be small enough without losing all previous input information. This means that the number of samples required to train the model is much smaller than other networks, and the computational cost is also lower, which is very beneficial for the model to be used for real-time control.
[0065] However, RNN also has its drawbacks, namely the vanishing and exploding gradient problem. This is because RNN has to repeatedly apply the same operation at each moment to build a very deep computational graph, and the model parameters are shared. Suppose a computational graph contains a path that repeatedly multiplies the matrix W, then after t steps it is equivalent to multiplying by W. t , if W has an eigenvalue decomposition W = V diag(λ)V -1 , then it is easy to see
[0066] W t =(V diag(λ)V -1 ) t =V diag(λ) t V -1
[0067] When the eigenvalue λ i When it is not near 1, it will "disappear" if it is less than 1 in magnitude, and it will "explode" if it is greater than 1. The gradient on the computational graph will also be affected by diag(λ). t Large changes. The vanishing gradient makes it difficult to know in which direction to move the parameters to improve the cost function, while the exploding gradient makes learning unstable. These changes are particularly obvious when dealing with long time series.
[0068] The present invention adds an MLP multi-layer perceptron node that does not enter the loop outside the RNN recurrent neural network, and uses the node to dynamically balance the gradient changes of the RNN, thereby effectively avoiding the gradient vanishing and gradient exploding problems in the calculation process and enhancing the stability of the entire active noise control system.
[0069] Based on the above design ideas, the present invention proposes an adaptive active noise control system based on deep neural network, such as Figure 1 As shown, it includes at least one set of reference microphones 1 , a controller 4 , an actuator 5 and an error microphone 6 .
[0070] The reference microphone 1 is arranged near the noise source to collect the reference signal, and the error microphone 6 is arranged at the control point to collect the error signal; the reference signal and the error signal are converted into electrical signals and input to the input end of the controller 4.
[0071] The controller 4 processes the reference signal and the error signal to generate a control signal with the same amplitude and opposite phase as the noise signal to be controlled, and inputs the control signal to the input end of the actuator 5 .
[0072] In this embodiment, the controller 4 includes a deep neural network module 2 and a driving circuit 3.
[0073] The deep neural network module 2 processes the reference signal, generates a control signal with the same amplitude and opposite phase as the noise signal to be controlled, and outputs the generated control signal to the driving circuit 3; at the same time, the deep neural network module 2 performs gradient calculation based on the error signal using the backpropagation (BP) algorithm and updates the network parameters.
[0074] The drive circuit 3 outputs a control signal to the actuator 5 .
[0075] The actuator 5 converts the control signal into a control sound wave, which is superimposed on the noise to be controlled at the control point, thereby achieving an active noise reduction effect.
[0076] In practical applications, the controller 4 can be integrated into a hardware chip, and the deep neural network module and the driving circuit are integrated in the hardware chip.
[0077] In the present invention, the deep neural network module is a novel DNN deep neural network structure in which RNN recurrent neural network is nested with MLP multi-layer perceptron nodes, such as Figure 2 As shown in the figure, it is a three-layer deep network consisting of 1 input layer, 1 hidden layer and 1 output layer, where the input layer includes 5 input nodes, the hidden layer consists of 10 RNN recurrent neurons and one MLP node, and the output layer includes one output node.
[0078] Based on the above-mentioned adaptive active noise control system based on deep neural network, the present invention also proposes an adaptive active noise control method based on deep neural network, comprising the following steps:
[0079] Step 1: A reference microphone placed near a noise source collects a reference signal and inputs it into a deep neural network module; an error microphone placed at a control point collects an error signal and inputs it into a deep neural network module.
[0080] Step 2: The deep neural network module uses the back propagation algorithm to calculate the gradient and update the network parameters according to the error signal; the deep neural network module processes the reference signal to generate a control signal with the same amplitude and opposite phase as the noise signal to be controlled.
[0081] The specific implementation method of this step includes the following steps:
[0082] (1) Use the following formula to calculate the current RNN neuron state H (t) ,
[0083] H (t) =XV+H (t-1) W,
[0084] Right now
[0085] Among them, H (t-1) is the RNN neuron state at the previous moment; X is the input vector; V and W are weight coefficient matrices;
[0086] (2) Use the following formula to calculate the MLP node state b:
[0087]
[0088] Among them, U is the weight coefficient matrix;
[0089] (3) Use the following formula to calculate the output value of the control signal at the current moment.
[0090]
[0091] Among them, α and β are weight coefficient matrices;
[0092] (4) Use the back propagation algorithm to calculate the negative gradient of each weight coefficient matrix. The calculation formula is as follows:
[0093]
[0094] Δβ=eb,
[0095] ΔU i =eβX i ,
[0096] ΔV ij =eα j X i ,
[0097]
[0098] Wherein, e is the sample value of the error signal obtained at the current moment;
[0099] (5) Use the following formula to update the network parameters:
[0100] α=α+ηΔα,
[0101] β=β+ηΔβ,
[0102] U=U+ηΔU,
[0103] V=V+ηΔV,
[0104] W=W+ηΔW
[0105] Among them, η is a preset learning rate.
[0106] Step 3: The deep neural network module inputs the generated control signal to the actuator through the driving circuit.
[0107] Step 4: The actuator converts the control signal into a control sound wave, which is superimposed with the noise to be controlled at the control point to perform active noise reduction.
[0108] The adaptive active noise control system proposed in the present invention is simulated using the Simulink module of MATLAB, and a comparative example of the active control of nonlinear noise by the present invention, the LMS minimum mean square error algorithm, and the traditional ANN artificial neural network algorithm is given.
[0109] The simulation results are as follows Figure 3 As shown, the whole figure is divided into 4 parts: (a) is the time domain waveform of the original noise signal; (b) is the time domain waveform of the error signal controlled by the traditional LMS algorithm; (c) is the time domain waveform of the error signal controlled by the traditional ANN algorithm; (d) is the time domain waveform of the deep neural network structure proposed by the present invention. In this example, the learning rate is 8e-3. In actual use, the value can be adjusted according to the specific noise environment or in combination with other adaptive algorithms. The noise to be controlled is set as a variable frequency signal with a frequency of 200Hz to 400Hz, the system sampling rate is 2000Hz, and the control time is 5 seconds; the horizontal axis of the coordinate is the time axis, and the vertical axis is the amplitude of the error signal. The smaller the vibration amplitude, the more obvious the control effect. As can be seen from the figure, compared with the traditional algorithm, the new algorithm has obvious advantages in convergence speed and control accuracy, and is more suitable for research and application in nonlinear acoustic environments.
[0110] Any matters not described in the present invention are applicable to the prior art.
[0111] It should be emphasized that the embodiments described in the present invention are illustrative rather than restrictive. Therefore, the present invention includes but is not limited to the embodiments described in the specific implementation manner. Any other implementation manners derived by those skilled in the art based on the technical solution of the present invention also fall within the scope of protection of the present invention.
Claims
1. An adaptive active noise control system based on deep neural network, characterized in that: It includes a reference microphone, a controller, an actuator and an error microphone, wherein the controller includes a deep neural network module and a driving circuit; The reference microphone is arranged near the noise source and is used to collect the reference signal and input it into the input end of the deep neural network; The error microphone is arranged at the control point, and is used to collect the error signal and input it into the income end of the deep neural network; The deep neural network module processes the reference signal to generate a control signal with the same amplitude and opposite phase as the noise signal to be controlled, and outputs the generated control signal to the driving circuit; the deep neural network module performs gradient calculation based on the error signal using the back propagation algorithm to update the network parameters; The driving circuit outputs the control signal to the actuator; The actuator converts the control signal into a control sound wave, which is superimposed on the noise to be controlled at the control point to perform active noise reduction; The deep neural network module is a DNN deep neural network structure in which an RNN recurrent neural network nests an MLP multi-layer perceptron node, and the MLP multi-layer perceptron node is used to dynamically balance the gradient changes of the RNN recurrent neural network; the deep neural network module is composed of a three-layer deep network consisting of an input layer, a hidden layer and an output layer, wherein the input layer includes 5 input nodes, the hidden layer consists of 10 RNN recurrent neurons and an MLP node, and the output layer includes an output node.
2. The adaptive active noise control system based on deep neural network according to claim 1, characterized in that: The controller is an integrated circuit hardware chip, in which a deep neural network module and a driving circuit are integrated.
3. A control method for an adaptive active noise control system based on a deep neural network as claimed in claim 1 or 2, characterized in that: The following steps are involved: Step 1: A reference microphone disposed near a noise source collects a reference signal and inputs the reference signal into a deep neural network module; an error microphone disposed at a control point collects an error signal and inputs the error signal into the deep neural network module; Step 2: The deep neural network module uses the back propagation algorithm to calculate the gradient and update the network parameters according to the error signal; the deep neural network module processes the reference signal to generate a control signal with the same amplitude and opposite phase as the noise signal to be controlled; Step 3: The deep neural network module inputs the generated control signal to the actuator through the driving circuit; Step 4: The actuator converts the control signal into a control sound wave, which is superimposed with the noise to be controlled at the control point to perform active noise elimination; The specific implementation method of step 2 includes the following steps: (1) Use the following formula to calculate the current RNN neuron state H (t) , H (t) =XV+H (t-1) W, Among them, H (t-1) is the RNN neuron state at the previous moment; X is the input vector; V and W are weight coefficient matrices; (2) Use the following formula to calculate the MLP node state b, Among them, U is the weight coefficient matrix; (3) Use the following formula to calculate the output value of the control signal at the current moment Among them, α and β are weight coefficient matrices; (4) Use the back propagation algorithm to calculate the negative gradient of each weight coefficient matrix. The calculation formula is as follows: Δβ=eb, ΔU i =eβX i , ΔV ij =eα j X i , Wherein, e is the sample value of the error signal obtained at the current moment; (5) Use the following formula to update the network parameters: α=α+ηΔα, β=β+ηΔβ, U=U+ηΔU, V=V+ηΔV, W=W+ηΔW Among them, η is a preset learning rate.
