Active noise reduction method, device and system for glass, electronic equipment and computer program product

Abstract

The invention discloses an active noise reduction method, device and system for glass, electronic equipment and a computer program product. The method comprises the steps that an original noise data signal is collected through a feed-forward vibration sensor, and the feed-forward vibration sensor is arranged on the outdoor side of glass; a preset noise prediction model is used for predicting the original noise data signals to obtain anti-vibration signals, the preset noise prediction model is trained with the minimum anti-noise data signals collected by a feedback vibration sensor as the target, and the feedback vibration sensor is arranged on the indoor side of the glass; the vibration controller is controlled to generate anti-vibration signals, the vibration controller is attached to the surface of the glass, and coherence superposition is formed by the anti-vibration signals and noise vibration signals generated by the noise data signals on the glass. The technical problem that in the prior art, an effective mode for resisting noise transmission of glass is lacked is solved.

Description

Technical Field

[0001] The present invention relates to the field of electricity, and in particular to a method, device, system, electronic equipment and computer program product for active noise reduction of glass. Background Art

[0002] In daily office and home environments, we are often disturbed by nearby noise, which affects our normal work and rest.

[0003] However, 90% of indoor noise comes from windows, and the key to window sound insulation is glass. Due to the poor sound insulation effect of glass, the indoor noise is seriously disturbed by external noise. People have developed hollow glass, laminated glass and even vacuum glass for glass sound insulation. The high cost of production and installation makes glass sound insulation not generally applicable to most families. In addition, the cost of replacing glass with high sound insulation material is too high, and there is a lack of effective ways to counteract the transmission of noise through glass.

[0004] With regard to the problem that the prior art lacks an effective way to counteract the propagation of noise by glass, no effective solution has been proposed so far. Summary of the invention

[0005] The embodiments of the present invention provide a method, device, system, electronic device and computer program product for active noise reduction of glass, so as to at least solve the technical problem that the prior art lacks an effective way to counteract the propagation of noise by glass.

[0006] According to one aspect of an embodiment of the present invention, a method for active noise reduction of glass is provided, comprising: collecting original noise data signals through a feedforward vibration sensor, wherein the feedforward vibration sensor is arranged on the outdoor side of the glass; predicting the original noise data signals using a preset noise prediction model to obtain an anti-vibration signal, wherein the preset noise prediction model is pre-trained by machine learning using multiple sets of data, wherein the multiple sets of data are multiple noise sampling segments pre-collected by the feedforward vibration sensor, and the preset noise prediction model is trained with the goal of minimizing the anti-noise data signal collected by the feedback vibration sensor, wherein the feedback vibration sensor is arranged on the indoor side of the glass, and the anti-noise data signal is the noise generated by coherently superimposing a predicted vibration signal predicted by the preset noise prediction model based on the noise sampling segment and a sampled vibration signal generated on the glass by the noise sampling segment; and controlling a vibration controller to generate the anti-vibration signal, wherein the vibration controller is attached to the surface of the glass, and the anti-vibration signal forms a coherent superposition with the noise vibration signal generated by the noise data signal on the glass.

[0007] Optionally, before using the preset noise prediction model to predict the original noise data signal to obtain the anti-vibration signal, the method also includes: using the feedforward vibration sensor, continuously collecting a plurality of the noise sampling segments according to a preset time length, wherein there are overlapping noise segments between adjacent noise sampling segments; using the preset noise prediction model to predict the predicted vibration signal corresponding to each of the noise sampling segments; controlling the vibration controller to generate the corresponding predicted vibration signal within the collection period of each of the noise sampling segments; determining the instantaneous loss function for training the preset noise prediction model based on the anti-noise data signal collected by the feedback vibration sensor within each of the collection period, and adjusting the model parameters in the preset noise prediction model according to the instantaneous loss function.

[0008] Optionally, based on the anti-noise data signal collected by the feedback vibration sensor in each of the collection time periods, determining the instantaneous loss function for training the preset noise prediction model, and adjusting the model parameters in the preset noise prediction model according to the instantaneous loss function includes: in the training stage of the preset noise prediction model, using an adaptive momentum estimation optimizer to adjust the model parameters in the preset noise prediction model of this round of training; after each round of training, detecting whether the anti-noise data signal collected by the feedback vibration sensor is a minimum value; if the anti-noise data signal is not a minimum value, performing the next round of training on the preset noise prediction model; if the anti-noise data signal is a minimum value, determining that the training of the preset noise prediction model is completed.

[0009] Optionally, using a preset noise prediction model to predict the original noise data signal to obtain an anti-vibration signal includes: using the preset noise prediction model to perform a short-time Fourier transform on the original noise data signal to obtain a first signal to be processed; using an encoder in the preset noise prediction model to extract features of the first signal to be processed to obtain a first noise feature; using a long short-term memory model in the preset noise prediction model to eliminate the delay effect in the first noise feature to obtain a second noise feature; using a decoder in the preset noise prediction model to process the first noise feature and the second noise feature to obtain an anti-noise feature; and performing an inverse short-time Fourier transform on the anti-noise feature to obtain the anti-vibration signal.

[0010] Optionally, the encoder includes at least: a first feature convolution layer, a second feature convolution layer and a third feature convolution layer. Using the encoder in the preset noise prediction model, feature extraction is performed on the first signal to be processed to obtain a first noise feature, which includes: using the first feature convolution layer to perform feature extraction on the first signal to be processed to obtain a first noise sub-feature; using the second feature convolution layer to perform feature extraction on the first noise sub-feature to obtain a second noise sub-feature; using the second feature convolution layer to perform feature extraction on the second noise sub-feature to obtain a third noise sub-feature; wherein the first noise feature includes at least: the first noise sub-feature, the second noise sub-feature and the third noise sub-feature.

[0011] Optionally, the long short-term memory model includes at least: a first long short-term memory model, a second long short-term memory model and a third long short-term memory model. The long short-term memory model in the preset noise prediction model is used to eliminate the delay effect in the first noise feature, and obtaining the second noise feature includes: using the first long short-term memory model to process the first noise sub-feature to obtain a fourth noise sub-feature; using the second long short-term memory model to process the second noise sub-feature to obtain a fifth noise sub-feature; using the second long short-term memory model to process the third noise sub-feature to obtain a sixth noise sub-feature; wherein the second noise feature includes at least: the fourth noise sub-feature, the fifth noise sub-feature and the sixth noise sub-feature.

[0012] Optionally, the decoder includes at least: a first feature deconvolution layer, a second feature deconvolution layer and a third feature deconvolution layer, and uses the decoder in the preset noise prediction model to process the first noise feature and the second noise feature to obtain the adversarial noise feature, including: using the first feature deconvolution layer to deconvolve the third noise sub-feature and the sixth noise sub-feature to obtain a first adversarial sub-feature; using the second feature deconvolution layer to deconvolve the first adversarial sub-feature and the fifth noise sub-feature to obtain a second adversarial sub-feature; using the third feature deconvolution layer to deconvolve the second adversarial sub-feature and the fourth noise sub-feature to obtain a third adversarial sub-feature; and determining the third adversarial sub-feature as the adversarial noise feature.

[0013] According to another aspect of an embodiment of the present invention, an active noise reduction device for glass is also provided, comprising: an acquisition module, for acquiring an original noise data signal through a feedforward vibration sensor, wherein the feedforward vibration sensor is arranged on the outdoor side of the glass; a prediction module, for predicting the original noise data signal using a preset noise prediction model to obtain an anti-vibration signal, wherein the preset noise prediction model is pre-trained by machine learning using multiple sets of data, wherein the multiple sets of data are multiple noise sampling segments pre-collected by the feedforward vibration sensor, and the preset noise prediction model is trained with the goal of minimizing the anti-noise data signal acquired by the feedback vibration sensor, wherein the feedback vibration sensor is arranged on the indoor side of the glass, and the anti-noise data signal is the noise generated by coherently superimposing a predicted vibration signal predicted by the preset noise prediction model based on the noise sampling segment and a sampled vibration signal generated by the noise sampling segment on the glass; a noise reduction module, for controlling a vibration controller to generate the anti-vibration signal, wherein the vibration controller is tightly attached to the surface of the glass, and the anti-vibration signal forms a coherent superposition with the noise vibration signal generated by the noise data signal on the glass.

[0014] According to another aspect of an embodiment of the present invention, there is also provided an active noise reduction system for glass, comprising: a feedforward vibration sensor for collecting original noise data signals, wherein the feedforward vibration sensor is arranged on the outdoor side of the glass; a computing device for predicting the original noise data signals using a preset noise prediction model to obtain an anti-vibration signal, wherein the preset noise prediction model is pre-trained by machine learning using multiple sets of data, wherein the multiple sets of data are multiple noise sampling segments pre-collected by the feedforward vibration sensor, and the preset noise prediction model is trained with the goal of minimizing the anti-noise data signal collected by the feedback vibration sensor, wherein the feedback vibration sensor is arranged on the indoor side of the glass, and the anti-noise data signal is the noise generated by coherently superimposing a predicted vibration signal predicted by the preset noise prediction model based on the noise sampling segment and a sampled vibration signal generated on the glass by the noise sampling segment; a vibration controller for generating the anti-vibration signal, wherein the vibration controller is tightly attached to the surface of the glass, and the anti-vibration signal forms a coherent superposition with the noise vibration signal generated by the noise data signal on the glass.

[0015] According to another aspect of an embodiment of the present invention, there is further provided an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to execute the above-mentioned active noise reduction method for glass through the computer program.

[0016] According to another aspect of an embodiment of the present invention, a computer program product is further provided, comprising computer instructions, which, when executed by a processor, implement the steps of the above-mentioned active noise reduction method for glass.

[0017] In an embodiment of the present invention, a feedforward vibration sensor is arranged on the outdoor side of the glass to collect the original noise data signal of the ambient noise, and the original noise data signal is predicted by a preset noise prediction model to determine an anti-vibration signal for offsetting the original noise data signal, and a vibration controller bonded to the glass surface is controlled to generate the anti-vibration signal, and then the anti-vibration signal can be coherently superimposed with the noise vibration signal generated by the noise data signal on the glass to offset the vibration caused by the original noise data signal to the glass, thereby achieving the technical effect of active noise reduction for the glass, and thus solving the technical problem that the prior art lacks an effective way to counteract the propagation of noise by glass. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0019] Figure 1 is a flow chart of an active noise reduction method for glass according to an embodiment of the present invention;

[0020] Figure 2 is a schematic diagram of a device environment on which a dynamic adaptive glass active noise reduction method according to an embodiment of the present invention depends;

[0021] Figure 3 is a schematic diagram of a circuit structure relied upon by a dynamic adaptive glass active noise reduction method according to an embodiment of the present invention;

[0022] Figure 4 is a schematic diagram of a GlassDenoiseNet network model according to an embodiment of the present invention;

[0023] Figure 5 is a schematic diagram of a single LSTM module according to an embodiment of the present invention;

[0024] Figure 6 is a schematic diagram of an active noise reduction device for glass according to an embodiment of the present invention;

[0025] Figure 7 is a schematic diagram of an active noise reduction system for glass according to an embodiment of the present invention;

[0026] Figure 8 It is a structural block diagram of a computer terminal according to an embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0028] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0029] First, some nouns or terms that appear in the description of the embodiments of the present application are subject to the following explanations:

[0030] ANC: Active Noise Cancellation (Control), actively reducing noise (noise control).

[0031] PNC: Passive Noise Cancellation, Passive Noise Cancellation, uses some method to isolate external noise to achieve the purpose of reducing noise (soundproof cotton, soundproof glass, earplugs, soundproof headphones).

[0032] NLANC: Nonlinear Active Noise Control, which uses a nonlinear algorithm compared to linear active noise reduction.

[0033] Damped vibration: belongs to the dissipative system, which refers to any vibration system in which the amplitude gradually decays over time due to friction, medium resistance or other energy consumption of the vibration system. It is also called reduced amplitude vibration or attenuated vibration.

[0034] LSTM: Long Short-Term Memory model is a time recurrent neural network specially designed to solve the long-term dependency problem of general RNN (recurrent neural network). All RNNs have a chain form of repeated neural network modules.

[0035] According to an embodiment of the present invention, an embodiment of an active noise reduction method for glass is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0036] Figure 1 is a flow chart of an active noise reduction method for glass according to an embodiment of the present invention. Figure 1 As shown, the method comprises the following steps:

[0037] Step S102, collecting original noise data signals through a feedforward vibration sensor, wherein the feedforward vibration sensor is arranged on the outdoor side of the glass;

[0038] Step S104, using a preset noise prediction model to predict the original noise data signal to obtain an anti-vibration signal, wherein the preset noise prediction model is pre-trained by machine learning using multiple sets of data, the multiple sets of data are multiple noise sampling segments pre-collected by a feedforward vibration sensor, the preset noise prediction model is trained with the goal of minimizing the anti-noise data signal collected by the feedback vibration sensor, the feedback vibration sensor is arranged on the indoor side of the glass, and the anti-noise data signal is the noise generated by coherently superimposing a predicted vibration signal predicted by the preset noise prediction model based on the noise sampling segment and a sampled vibration signal generated on the glass by the noise sampling segment;

[0039] Step S106, controlling the vibration controller to generate an anti-vibration signal, wherein the vibration controller is attached to the surface of the glass, and the anti-vibration signal and the noise vibration signal generated by the noise data signal on the glass form a coherent superposition.

[0040] In an embodiment of the present invention, a feedforward vibration sensor is arranged on the outdoor side of the glass to collect the original noise data signal of the ambient noise, and the original noise data signal is predicted by a preset noise prediction model to determine an anti-vibration signal for offsetting the original noise data signal, and a vibration controller bonded to the glass surface is controlled to generate the anti-vibration signal, and then the anti-vibration signal can be coherently superimposed with the noise vibration signal generated by the noise data signal on the glass to offset the vibration caused by the original noise data signal to the glass, thereby achieving the technical effect of active noise reduction for the glass, and thus solving the technical problem that the prior art lacks an effective way to counteract the propagation of noise by glass.

[0041] It should be noted that the propagation of sound is a kind of energy propagation in physics, which mainly depends on the vibration of the propagation medium. Sound propagates in the air and is transmitted through air vibration. When it reaches the human ear, the air vibration drives the human eardrum to vibrate, causing the human auditory nerve to feel the sound.

[0042] It should be noted that sound can be transmitted through glass mainly because although glass is a rigid body, it still vibrates with the air, which is difficult to observe with the naked eye. The sound propagation outside causes the air to vibrate, which will drive the glass to vibrate; the vibration of the glass will drive the air in the room to vibrate, so that the sound is transmitted through the glass into the room; therefore, the original noise data signal is transmitted to the glass, which can cause the glass to generate a noise vibration signal, so that the noise outside can be transmitted into the room.

[0043] It should be noted that the three elements of sound are: amplitude, frequency, and timbre. Amplitude indicates the size of the volume. The larger the amplitude, the higher the volume; the smaller the amplitude, the lower the volume; frequency indicates the speed of vibration. The higher the frequency, the higher and sharper the pitch; the lower the frequency, the lower and deeper the pitch; timbre varies from one sound source to another, mainly because the intensity of the sound components of each frequency that make up the sound signal is not the same, such as the voices of different groups of people such as men, women, and children, and the sounds of different instruments such as piano, zither, and guitar.

[0044] It should be noted that the principle of active noise reduction is to collect ambient noise through a microphone, analyze it digitally, and process it with reverse sound waves to achieve cancellation between sound waves and complete noise reduction. Its theoretical basis is mainly based on the coherent superposition of two waves with the same amplitude and opposite directions to form a result of mutual cancellation.

[0045] It should be noted that the vibration of glass due to external noise is a vibration of a plane rigid body. Its vibration is related to factors such as the thickness of the glass, the production process and material, and the installation process. It is more complex than the vibration of a single point sound source. It is impossible to cancel the noise by simply superimposing a single source, nor is it possible to use a fixed active noise reduction model to adapt to all glass in home and office environments. Therefore, it is not possible to simply generate vibrations with the same amplitude and opposite direction as the sound source. It is necessary to consider the collaborative relationship between multiple groups of vibration controllers and the mutual influence relationship between multiple groups of vibration controllers and shock absorbers. Therefore, traditional linear ANC algorithms such as the minimum mean square filtering method are difficult to handle complex nonlinear problems such as sound wave attenuation and amplified pulse distortion of electronic components. In general, a deep learning model based on nonlinear ANC is needed to determine a series of parameters to determine the final noise reduction model coefficients.

[0046] In the above-mentioned embodiments of the present application, the active noise reduction glass achieves the purpose of noise reduction by attaching a vibration controller to the surface of the glass to make the glass vibrate regularly and repel sound waves. However, due to factors such as the thickness of the glass, the production process and material, and the installation process, the regularity of the glass's vibration during sound propagation can be affected, which makes it impossible to use a single fixed deep neural network model to handle the anti-noise reduction process. Therefore, almost every piece of glass requires a separate noise reduction model that meets its characteristics. In addition, the present application provides a technical solution that can train a preset noise prediction model for each piece of glass after the glass is installed, and the glass corresponding to the preset noise prediction model predicts the anti-vibration signal.

[0047] As an optional embodiment, before using a preset noise prediction model to predict the original noise data signal to obtain an anti-vibration signal, the method also includes: using a feedforward vibration sensor to continuously collect multiple noise sampling segments according to a preset time length, wherein there are overlapping noise segments between adjacent noise sampling segments; using a preset noise prediction model to predict a predicted vibration signal corresponding to each noise sampling segment; controlling a vibration controller to generate a corresponding predicted vibration signal within the collection period of each noise sampling segment; determining an instantaneous loss function for training a preset noise prediction model based on the anti-noise data signal collected by the feedback vibration sensor within each collection period, and adjusting model parameters in the preset noise prediction model according to the instantaneous loss function.

[0048] In the above-mentioned embodiment of the present application, before the preset noise prediction model predicts the original noise data signal, the preset noise prediction model corresponding to the glass can be pre-trained, and a feedforward vibration sensor and a feedback vibration sensor are respectively installed on the outdoor side and the indoor side of the glass. The feedforward vibration sensor continuously collects multiple noise sampling segments according to the preset time length, and then uses the preset noise prediction model to predict the predicted vibration signal corresponding to each noise sampling segment, and controls the vibration controller to generate a corresponding predicted vibration signal within the collection period of each noise sampling segment. The predicted vibration signal generated by the vibration controller is coherently superimposed with the sampled vibration signal generated by the noise sampling segment on the glass, thereby eliminating the sampled vibration signal generated by the noise sampling segment on the glass. The sampled vibration signal that is not eliminated can be used as an anti-noise data signal and collected by the feedback vibration sensor installed on the indoor side. Then, in the process of training the preset noise prediction model, the anti-noise data signal collected by the feedback vibration sensor can be used as an instantaneous loss function, and the model parameters in the preset noise prediction model are adjusted with the goal of minimizing the anti-noise data signal, thereby completing the training of the preset noise prediction model.

[0049] Optionally, the instantaneous loss function can be:

[0050]

[0051] Among them, e(t) represents the anti-noise data signal collected by the feedback vibration sensor, and L represents the length of a noise signal data (that is, a noise sampling segment). When the loss is reduced to the minimum, e(t) is the smallest, indicating that the noise elimination effect is good.

[0052] Optionally, each noise sampling segment is sampled by a feedforward vibration sensor at a frequency of 48 kHz, each noise sampling segment x(t) is set to 20 milliseconds, and there is 10 milliseconds of overlapping data between adjacent continuous segments.

[0053] As an optional embodiment, based on the anti-noise data signal collected by the feedback vibration sensor in each collection period, determining the instantaneous loss function for training the preset noise prediction model, and adjusting the model parameters in the preset noise prediction model according to the instantaneous loss function includes: in the training stage of the preset noise prediction model, using an adaptive momentum estimation optimizer to adjust the model parameters in the preset noise prediction model of this round of training; after each round of training, detecting whether the anti-noise data signal collected by the feedback vibration sensor is the minimum value; if the anti-noise data signal is not the minimum value, performing the next round of training on the preset noise prediction model; if the anti-noise data signal is the minimum value, determining that the training of the preset noise prediction model is completed.

[0054] In the above-mentioned embodiment of the present application, when training a preset noise prediction model, an adam optimizer (also known as an adaptive momentum estimation optimizer) is used to calculate the adaptive learning rates of different parameters to reduce the risk of overfitting.

[0055] Optionally, a The dam optimizer (also known as the adaptive momentum estimation optimizer) can effectively calculate the adaptive learning rate of each parameter during the optimization process. In detail, assuming that ut is the parameter to be optimized and gt is the corresponding gradient, the update of θt+1 using adam is as follows:

[0056] α t = r1α t-1 +(1-r1)g t

[0057]

[0058] α t =α t / (1-r1)

[0059] β t =β t / (1-r2)

[0060]

[0061] Among them, α t and β t are the first and second moments of the gradient, η is the learning rate, which is set to 0.001, and the parameters r1, r2, and ∈ are set to 0.9, 0.999, and 0.0000001, respectively.

[0062] It should be noted that the preset noise prediction model is a deep neural network for processing audio, which runs on a computing device trained with a deep neural network. The preset noise prediction model takes the audio data signal (such as the original noise data signal or the noise sampling fragment) collected by the feedforward vibration sensor as input, and outputs the audio data signal (such as an anti-vibration signal for the original noise data signal, or a predicted vibration signal corresponding to the noise sampling fragment) to all vibration controllers.

[0063] As an optional embodiment, using a preset noise prediction model to predict the original noise data signal to obtain an anti-vibration signal includes: using the preset noise prediction model to perform a short-time Fourier transform on the original noise data signal to obtain a first signal to be processed; using an encoder in the preset noise prediction model to extract features of the first signal to be processed to obtain a first noise feature; using a long short-term memory model in the preset noise prediction model to eliminate the delay effect in the first noise feature to obtain a second noise feature; using a decoder in the preset noise prediction model to process the first noise feature and the second noise feature to obtain an anti-noise feature; and performing an inverse short-time Fourier transform on the anti-noise feature to obtain an anti-vibration signal.

[0064] In the above-mentioned embodiment of the present application, after the original noise data signal is input into the preset noise prediction model, the encoder can perform feature extraction, and the long short-term memory model can eliminate the delay effect in the extracted features, and then the decoder can decode the features in the original noise data signal to obtain the anti-vibration signal corresponding to the original noise data signal, thereby achieving the technical effect of determining the anti-vibration signal, and when using the preset noise prediction model to determine the anti-vibration signal corresponding to the original noise data signal, short-time Fourier transform can be used. Since short-time Fourier transform is a frequency domain transformation method commonly used in audio data processing, it makes audio data clearer and easier to process. Therefore, the preset noise prediction model of the present application uses short-time Fourier transform to pre-process the data of the deep neural network, which can effectively improve the model convergence speed.

[0065] As an optional embodiment, the encoder includes at least: a first feature convolution layer, a second feature convolution layer and a third feature convolution layer. The encoder in the preset noise prediction model is used to extract features of the first signal to be processed, and obtaining the first noise feature includes: using the first feature convolution layer to extract features of the first signal to be processed to obtain a first noise sub-feature; using the second feature convolution layer to extract features of the first noise sub-feature to obtain a second noise sub-feature; using the second feature convolution layer to extract features of the second noise sub-feature to obtain a third noise sub-feature; wherein the first noise feature includes at least: the first noise sub-feature, the second noise sub-feature and the third noise sub-feature.

[0066] In the above embodiment of the present application, the noise data x(t) (such as the original noise data signal) sampled by the feedforward vibration sensor at time t is input into three feature convolution layers ConvBlock (such as the first feature convolution layer L1, the second feature convolution layer L2 or the third feature convolution layer L3) of different scales after STFT domain transformation (i.e., short-time Fourier transform). Each layer extracts the 1D convolution feature of the noise (such as the first noise sub-feature, or the second noise sub-feature) and then downsamples it to the next layer, and the third layer (such as the third feature convolution layer L3) jumps the output downsampled feature (such as the first noise sub-feature) to the L3 (i.e., the first feature deconvolution layer L3) corresponding to the decoder; at the same time, the noise feature of the current layer (such as the first noise sub-feature, the second noise sub-feature or the third noise sub-feature) is input into the time-series-related long short-term memory model LSTM, and multiple downsamplings are used to extract noise features at different scales, which are then decoded at multiple scales, so as to better predict the amount of adversarial noise reduction.

[0067] As an optional embodiment, the long short-term memory model includes at least: a first long short-term memory model, a second long short-term memory model and a third long short-term memory model. The long short-term memory model in the preset noise prediction model is used to eliminate the delay effect in the first noise feature, and the second noise feature is obtained, including: using the first long short-term memory model to process the first noise sub-feature to obtain a fourth noise sub-feature; using the second long short-term memory model to process the second noise sub-feature to obtain a fifth noise sub-feature; using the second long short-term memory model to process the third noise sub-feature to obtain a sixth noise sub-feature; wherein the second noise feature includes at least: the fourth noise sub-feature, the fifth noise sub-feature and the sixth noise sub-feature.

[0068] In the above-mentioned embodiments of the present application, the long short-term memory model LSTM of each layer processes the noise characteristics of the current layer (such as the first noise sub-feature, the second noise sub-feature or the third noise sub-feature), and outputs it to the decoder of the current layer; the LSTM model has excellent performance in processing timing-related tasks. In the sound denoising task, errors in the predicted timing fluctuation period are usually caused by propagation delays, acquisition delays, and calculation delays, thereby affecting the final noise reduction result. The introduction of LSTM can effectively predict the fluctuation law of noise characteristics within the time series range, thereby offsetting the problems caused by delays.

[0069] As an optional embodiment, the decoder includes at least: a first feature deconvolution layer, a second feature deconvolution layer and a third feature deconvolution layer. The decoder in the preset noise prediction model is used to process the first noise feature and the second noise feature to obtain the adversarial noise feature, including: using the first feature deconvolution layer to deconvolve the third noise sub-feature and the sixth noise sub-feature to obtain the first adversarial sub-feature; using the second feature deconvolution layer to deconvolve the first adversarial sub-feature and the fifth noise sub-feature to obtain the second adversarial sub-feature; using the third feature deconvolution layer to deconvolve the second adversarial sub-feature and the fourth noise sub-feature to obtain the third adversarial sub-feature; and determining the third adversarial sub-feature as the adversarial noise feature.

[0070] In the above embodiment of the present application, in the decoder, there are three scales of deconvolution blocks (such as the first feature deconvolution layer, the second feature deconvolution layer and the third feature deconvolution layer). In each layer of deconvolution blocks, the low-scale features transmitted from the next layer are first deconvolved, and then the result is concatenated with the output of LSTM to obtain the decoding features of the current layer (such as deconvolving the third noise sub-feature and connecting it in series with the sixth noise sub-feature, or deconvolving the first adversarial sub-feature and connecting it in series with the fifth noise sub-feature, or deconvolving the second adversarial sub-feature and connecting it in series with the fourth noise sub-feature); then the decoding features of the current layer are passed to the deconvolution block of the previous layer. The decoder combines the noise features at multiple scales output by the encoder and the empirical prediction of the noise features in time series, thereby re-encoding and restoring the required adversarial noise reduction amount (i.e., adversarial noise features).

[0071] Optionally, the result output by the decoder L1 layer (such as the anti-noise feature) is restored to electronic signal data y(t) that can be used by the vibration controller after short-time inverse Fourier transform.

[0072] The present invention also provides a preferred embodiment, which provides a dynamic and adaptive glass active noise reduction method, using a noise reduction algorithm model GlassDenoiseNet and an overall device that can learn and adapt to the actual environment, combined with mature deep learning technology, and can effectively learn and train deep neural network noise reduction models under different glass thicknesses, production processes and materials, and installation processes, so that good active noise reduction countermeasure signals can be generated in different actual environments. Combined with a vibration controller, the glass can produce coherent vibrations, forming coherent superposition locally to offset the vibrations propagated by the sound source. At the same time, in the coherent superposition enhancement area of ​​the vibration, a vibration energy diversion shock-absorbing damper is used to release the superimposed energy into the natural environment. Combined together, a complete set of glass noise reduction devices is formed to exert the effect of noise reduction in different actual environments.

[0073] Because the thickness, production process, material, and installation process of glass in different office and home environments are different, it is impossible to use a single fixed deep neural network model to handle the adversarial noise reduction process. Almost every piece of glass requires a separate noise reduction model that meets its characteristics.

[0074] The noise reduction technical solution for a specific piece of glass is mainly divided into four steps: building a noise reduction training equipment environment, designing the GlassDenoiseNet deep neural network for glass noise reduction, training the network model in the constructed training environment (model learning adaptive noise reduction), and daily noise reduction. The specific steps are as follows:

[0075] Step S1, the equipment assembly stage, builds the equipment environment for noise reduction training.

[0076] Figure 2 is a schematic diagram of a device environment on which a dynamic adaptive glass active noise reduction method according to an embodiment of the present invention depends, such as Figure 2 As shown, the device environment includes at least: a feedforward vibration sensor, a computing device, a vibration controller, a feedback vibration sensor and a shock absorbing damper.

[0077] Optionally, a feedforward vibration sensor is arranged outside the glass window (not bonded to the glass, with a spacing of 10-20 cm) to collect vibration information of the noise source and convert it into an electronic signal.

[0078] Optionally, a computing device for deep neural network training is used to train the model to adapt the optimal sound source confrontation coefficient.

[0079] Optionally, a vibration controller is distributed on the glass surface and closely attached to the glass, with a fixed position (related to the parameters of the model) for coherently superimposing the vibration generated by the input electrical signal with the vibration caused by the external noise.

[0080] Optionally, a feedback vibration sensor is arranged in the glass window (not bonded to the glass, and 10-20 cm away from the glass) to collect vibration information after the noise passes through the glass and convert it into an electronic signal.

[0081] Optionally, a shock-absorbing damper forms a damped vibration, is distributed on the glass surface (spaced apart from the vibration controller), is closely attached to the glass, and has a fixed position (related to the parameters of the model) for channeling the vibration energy of the glass.

[0082] Optionally, a vibration sensor (i.e., a feedforward vibration sensor and a feedback vibration sensor) is used to collect external vibration information and convert it into an electrical signal. Among them, vibration sensors are divided into resistive, capacitive, eddy current, dynamic, dynamic magnetic, strain, etc. according to their implementation principles. Common microphones are only one type of vibration sensor, and their effect on collecting low-frequency noise is average. However, in actual environments, although high-frequency noise is harsh and sharp, it often decays rapidly with distance and obstacles. Low-frequency noise can be transmitted over long distances and has strong penetration. Different vibration sensors can be used according to different noise environments.

[0083] Optionally, a vibration controller is used to convert electrical signals into physical vibrations. Common speakers, in-ear headphones, and bone conduction headphones are a small part of the applications of vibration controllers.

[0084] It should be noted that most of the existing active noise reduction methods use speakers to play sound. This application uses a vibration controller attached to the glass surface. Its principle is roughly similar to that of bone conduction headphones. Its core components include acoustic electronic signal amplifiers, electromagnetic vibration modules, etc., which are mainly used to control the glass to produce corresponding low-frequency vibrations, which is different from the existing ANC methods.

[0085] Optionally, in Figure 2 In the hardware environment shown, four vibration controllers and one vibration damper are configured.

[0086] Optionally, the actual configuration of the vibration controller and the shock absorbing damper can be matched in quantity according to the actual size of the glass.

[0087] Optionally, a vibration controller is usually set every 60-80cm, and a shock-absorbing damper is set at the center equidistant from the vibration controller. When the vibration controller generates antagonistic vibration, the local noise vibration forms a coherent cancellation. However, the vibration energy will not disappear in vain, and will form a superposition of peaks at the far point of resonance. Therefore, a shock-absorbing damper is set in the middle of the vibration controller to divert the vibration energy of the resonance peak. The shock-absorbing damper needs to use a thin flexible steel wire so that it can better resonate with the glass at the same frequency and consume residual energy in the gradually attenuated damping movement.

[0088] Figure 3 is a schematic diagram of a circuit structure relied upon by a dynamic adaptive glass active noise reduction method according to an embodiment of the present invention, such as Figure 3 As shown in the figure, it shows how the adaptive noise reduction control algorithm can play the role of glass noise reduction in the whole system, where d(t) represents the vibration data value after the original noise passes through the glass, a(t) represents the anti-vibration value transmitted to the glass by the vibration controller, and e(t) represents the noise signal received by the feedback vibration sensor.

[0089] Step S2, constructing the GlassDenoiseNet network model.

[0090] Figure 4 is a schematic diagram of a GlassDenoiseNet network model according to an embodiment of the present invention, such as Figure 4 As shown in the figure, the GlassDenoiseNet network model adopts the GlassDenoiseNet network structure, where ConvBlock represents the convolution downsampling block, DeconvBlock represents the deconvolution block, LSTM represents the long short-term memory model, STFT represents the short-time Fourier transform, and ISTFT represents the inverse short-time Fourier transform.

[0091] Optionally, short-time Fourier transform is a frequency domain transformation method commonly used in audio data processing. It makes the audio data clearer and easier to process. As data preprocessing for deep neural networks, it can effectively improve the model convergence speed.

[0092] As an optional embodiment, Figure 2 As shown, GlassDenoiseNet is a deep neural network for processing audio. It runs on a computing device trained with a deep neural network. It takes the audio data signal collected by the feedforward vibration sensor as input and outputs the audio data signal to all vibration controllers. The specific steps are as follows:

[0093] Step S2-1, constructing an encoder for noise segment data.

[0094] Optionally, the encoder includes: multiple convolution downsampling blocks ConvBlock, which are used to take the noise data x(t) sampled from the feedforward vibration sensor at time t as input, and after STFT domain transformation, input into three feature convolution layers ConvBlock L1 / L2 / L3 of different scales. Each layer extracts the 1D convolution features of the noise and then downsamples and passes them to the next layer, and at the same time, the noise features of the current layer are input into the time-related long short-term memory model LSTM. In the third layer ConvBlockL3, the output downsampled features are jump-connected to the L3 corresponding to the decoder. Multiple downsampling can extract noise features at different scales, and then multi-scale decoding can be performed, so as to better predict the amount of adversarial noise reduction.

[0095] Step S2-2, construct a time-series-associated long short-term memory model LSTM module array. Each layer of LSTM processes the noise characteristics of the current layer and outputs it to the decoder of the current layer. The LSTM model has excellent performance in processing time-series-related tasks. In the sound denoising task, there are usually errors in the predicted time series fluctuation period due to propagation delay, collection delay, and calculation delay, which affects the final noise reduction result. The introduction of LSTM can effectively predict the fluctuation law of noise characteristics within the time series range, thereby offsetting the problems caused by delay.

[0096] Figure 5 is a schematic diagram of a single LSTM module according to an embodiment of the present invention, such as Figure 5 As shown, the state transfer equation represented by a single LSTM module is:

[0097] f t =σ(W f [h t-1 , x t ]+b f )

[0098] i t =σ(W i [h t-1 , x t ]+b i )

[0099]

[0100] o t =σ(W o [h t-1 , x t ]+b o )

[0101] h t =o t *tanh(C t ),

[0102] Among them, W f W i W C W o is the weight, b f b i b C b o is the bias, σ(*) and tanh(*) represent the sigmoid and tanh activation functions, h t i t f t o t They represent hidden state, gated input, gated forget, and gated output respectively. and C t It is the memory state update unit and the memory output unit.

[0103] Optionally, in the iterative training, the LSTM module inputs the data features at time t into the input layer and outputs the result through the activation function. t , the hidden layer output h at time t-1 t-1 and the unit storage C at time t-1 t-1 The information is input into the nodes of the LSTM structure, and the data is output to the next hidden layer h through the processing of gated input, gated output, gated forgetting and memory units. t Or output layer C t , output the result o of the LSTM structure node t To the output layer neurons, calculate the back propagation error, and update each weight.

[0104] Step S2-3, construct a noise feature decoder.

[0105] Optionally, in the decoder, there are three scale deconvolution blocks. In each deconvolution block, the low-scale features passed from the next layer are first deconvolved, and then the result is concatenated with the output of LSTM. e , obtain the decoded features of the current layer. Then pass the decoded features of the current layer to the deconvolution block of the previous layer. The decoder combines the noise features of the encoder output at multiple scales and the empirical prediction of the noise features in time series to re-encode and restore the required adversarial noise reduction. The result of the decoder L1 layer output is restored to the electronic signal data y(t) that can be used by the vibration controller through the short-time inverse Fourier transform.

[0106] Step S3, using the equipment environment of step S1 and the network structure constructed in step S2, train the network model (perform adaptive noise reduction training of the model).

[0107] Optionally, the training model is fed with a noise sampling segment each time. The feedforward vibration sensor is sampled at a frequency of 48kHz. Each sampling segment x(t) is set to 20 milliseconds, and there is 10 milliseconds of overlapping data between adjacent consecutive segments. The model training uses the adam optimizer, which can effectively calculate the adaptive learning rate of each parameter during the optimization process. In detail, assuming that ut is the parameter to be optimized and gt is the corresponding gradient, the adam update θt+1 is as follows:

[0108] α t = r1α t-1 +((1-r1)g t

[0109]

[0110] α t =α t / (1-r1)

[0111] β t =β t / (1-r2)

[0112]

[0113] Among them, α t and β t are the first and second moments of the gradient, η is the learning rate, set to 0.001, and the parameters r1, r2, and ∈ are set to 0.9, 0.999, and 0.0000001, respectively. Overfitting is a common problem in learning-based systems, and the adam optimizer is used here, which can calculate adaptive learning rates for different parameters to reduce the risk of overfitting.

[0114] Optionally, the training model is to offset the noise received by the feedback vibration sensor, and the final noise is expressed as e(t) = d(t) - a(t), where d(t) represents the vibration data value of the original noise after passing through the glass, and a(t) represents the counteracting vibration value transmitted to the glass by the vibration controller. Therefore, the noise signal e(t) received by the feedback vibration sensor is used as the main basis for the model loss, and the instantaneous loss function at time t is expressed as:

[0115]

[0116] Among them, e(t) is the vibration data collected by the feedback vibration sensor, and L is the length of a noise signal data. When the loss is reduced to the minimum, e(t) is the smallest, indicating that the noise elimination effect is good.

[0117] Step S4, using the trained network model obtained in step S3 to perform glass noise reduction.

[0118] Optionally, the feedback vibration sensor is removed, and the feedforward vibration sensor, the computing device for deep neural network training, the shock absorber and a set of vibration controllers attached to the glass surface are retained. The noise data signal is collected in the feedforward vibration sensor, and after being predicted by the deep neural network noise reduction model, the output anti-vibration information is transmitted to the vibration controller, and the vibration controller generates corresponding vibration. Finally, the glass forms a vibration superposition in the vibration controller to cancel each other out and eliminate the noise, thereby achieving the purpose of noise reduction.

[0119] Compared with the traditional glass noise reduction solution, the above-mentioned embodiments of the present application have dynamic adaptive capabilities and can effectively play a noise reduction role under the influence of comprehensive factors such as the thickness of different glass, production process and material, and installation process.

[0120] The technical solution provided in this application focuses on using a deep learning algorithm to control the vibration of the glass, thereby offsetting the vibration of the noise on the glass during its propagation path. This is very different from most traditional methods of offsetting the noise through the superposition of sound waves from speakers. The former implements noise reduction on the glass medium, and the latter implements noise reduction on the air medium.

[0121] The technical solution provided by this application constructs a GlassDenoiseNet deep neural network structure and performs an adaptive noise reduction learning process. The network uses the data signal of the feedforward vibration sensor as input and the data signal of the real-time feedback vibration sensor as the instant loss of network training, which enables the network model to effectively adapt to the thickness of the glass in the installation environment, the production process and material, the installation process, and the relative position of the vibration controller and the shock absorber, so as to obtain the best model parameters during the model training convergence process.

[0122] The technical solution provided in this application integrates a multi-element vibration damper, fully considers the energy conservation law of the closed system, and releases the residual noise energy under the consideration of the physical vibration and wave propagation laws, thereby further enhancing the noise reduction performance of the glass.

[0123] According to an embodiment of the present invention, an embodiment of an active noise reduction device for glass is also provided. It should be noted that the active noise reduction device for glass can be used to execute the active noise reduction method for glass in the embodiment of the present invention, and the active noise reduction method for glass in the embodiment of the present invention can be executed in the active noise reduction device for glass.

[0124] Figure 6 is a schematic diagram of an active noise reduction device for glass according to an embodiment of the present invention, such as Figure 6As shown, the device may include: an acquisition module 62, which is used to acquire the original noise data signal through a feedforward vibration sensor, wherein the feedforward vibration sensor is arranged on the outdoor side of the glass; a prediction module 64, which is used to predict the original noise data signal using a preset noise prediction model to obtain an anti-vibration signal, wherein the preset noise prediction model is pre-trained by machine learning using multiple sets of data, and the multiple sets of data are multiple noise sampling segments pre-collected by the feedforward vibration sensor, and the preset noise prediction model is trained with the goal of minimizing the anti-noise data signal acquired by the feedback vibration sensor, and the feedback vibration sensor is arranged on the indoor side of the glass, and the anti-noise data signal is the noise generated by the coherent superposition of the predicted vibration signal predicted by the preset noise prediction model based on the noise sampling segment and the sampled vibration signal generated on the glass by the noise sampling segment; a noise reduction module 66, which is used to control the vibration controller to generate an anti-vibration signal, wherein the vibration controller is tightly attached to the surface of the glass, and the anti-vibration signal and the noise vibration signal generated on the glass by the noise data signal form a coherent superposition.

[0125] It should be noted that the acquisition module 62 in this embodiment can be used to execute step S102 in the embodiment of the present application, the prediction module 64 in this embodiment can be used to execute step S104 in the embodiment of the present application, and the noise reduction module 66 in this embodiment can be used to execute step S106 in the embodiment of the present application. The examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the contents disclosed in the above embodiments.

[0126] In an embodiment of the present invention, a feedforward vibration sensor is arranged on the outdoor side of the glass to collect the original noise data signal of the ambient noise, and the original noise data signal is predicted by a preset noise prediction model to determine an anti-vibration signal for offsetting the original noise data signal, and a vibration controller bonded to the glass surface is controlled to generate the anti-vibration signal, and then the anti-vibration signal can be coherently superimposed with the noise vibration signal generated by the noise data signal on the glass to offset the vibration caused by the original noise data signal to the glass, thereby achieving the technical effect of active noise reduction for the glass, and thus solving the technical problem that the prior art lacks an effective way to counteract the propagation of noise by glass.

[0127] As an optional embodiment, the method also includes: a collection submodule, which is used to predict the original noise data signal using a preset noise prediction model to obtain an anti-vibration signal, and continuously collect multiple noise sampling segments according to a preset time length through a feedforward vibration sensor, wherein there are overlapping noise segments between adjacent noise sampling segments; a prediction submodule, which is used to predict the predicted vibration signal corresponding to each noise sampling segment using a preset noise prediction model; a generation submodule, which is used to control the vibration controller to generate a corresponding predicted vibration signal within the collection period of each noise sampling segment; and an adjustment submodule, which is used to determine the instantaneous loss function for training the preset noise prediction model based on the anti-noise data signal collected by the feedback vibration sensor within each collection period, and adjust the model parameters in the preset noise prediction model according to the instantaneous loss function.

[0128] As an optional embodiment, the adjustment submodule includes: an adjustment subunit, which is used to use an adaptive momentum estimation optimizer to adjust the model parameters in the preset noise prediction model of this round of training during the training stage of the preset noise prediction model; a detection subunit, which is used to detect whether the anti-noise data signal collected by the feedback vibration sensor is a minimum value after each round of training; a training subunit, which is used to perform the next round of training on the preset noise prediction model when the anti-noise data signal is not a minimum value; and a determination subunit, which is used to determine that the training of the preset noise prediction model is completed when the anti-noise data signal is a minimum value.

[0129] As an optional embodiment, the prediction module includes: a Fourier transform unit, which is used to use a preset noise prediction model to perform a short-time Fourier transform on the original noise data signal to obtain a first signal to be processed; an encoding unit, which is used to use an encoder in the preset noise prediction model to extract features from the first signal to be processed to obtain a first noise feature; a processing unit, which is used to use a long short-term memory model in the preset noise prediction model to eliminate the delay effect in the first noise feature to obtain a second noise feature; a decoding unit, which is used to use a decoder in the preset noise prediction model to process the first noise feature and the second noise feature to obtain an anti-noise feature; and an inverse Fourier transform unit, which is used to perform a short-time inverse Fourier transform on the anti-noise feature to obtain an anti-vibration signal.

[0130] As an optional embodiment, the encoder includes at least: a first feature convolution layer, a second feature convolution layer and a third feature convolution layer, and the encoding unit includes: a first encoding subunit, used to use the first feature convolution layer to perform feature extraction on the first signal to be processed to obtain a first noise sub-feature; a second encoding subunit, used to use the second feature convolution layer to perform feature extraction on the first noise sub-feature to obtain a second noise sub-feature; a third encoding subunit, used to use the second feature convolution layer to perform feature extraction on the second noise sub-feature to obtain a third noise sub-feature; wherein the first noise feature includes at least: a first noise sub-feature, a second noise sub-feature and a third noise sub-feature.

[0131] As an optional embodiment, the long short-term memory model includes at least: a first long short-term memory model, a second long short-term memory model and a third long short-term memory model, and the processing unit includes: a first processing sub-unit, used to use the first long short-term memory model to process the first noise sub-feature to obtain a fourth noise sub-feature; a second processing sub-unit, used to use the second long short-term memory model to process the second noise sub-feature to obtain a fifth noise sub-feature; a third processing sub-unit, used to use the second long short-term memory model to process the third noise sub-feature to obtain a sixth noise sub-feature; wherein the second noise feature includes at least: a fourth noise sub-feature, a fifth noise sub-feature and a sixth noise sub-feature.

[0132] As an optional embodiment, the decoder includes at least: a first feature deconvolution layer, a second feature deconvolution layer and a third feature deconvolution layer, and the decoding unit includes: a first decoding sub-unit, used to use the first feature deconvolution layer to deconvolve the third noise sub-feature and the sixth noise sub-feature to obtain a first adversarial sub-feature; a second decoding sub-unit, used to use the second feature deconvolution layer to deconvolve the first adversarial sub-feature and the fifth noise sub-feature to obtain a second adversarial sub-feature; a third decoding sub-unit, used to use the third feature deconvolution layer to deconvolve the second adversarial sub-feature and the fourth noise sub-feature to obtain a third adversarial sub-feature; a decoding determination sub-unit, used to determine the third adversarial sub-feature as an adversarial noise feature.

[0133] Figure 7 is a schematic diagram of an active noise reduction system for glass according to an embodiment of the present invention, such as Figure 7As shown, the system includes: a feedforward vibration sensor 72 for collecting original noise data signals, wherein the feedforward vibration sensor is arranged on the outdoor side of the glass 71; a computing device 74 for predicting the original noise data signal using a preset noise prediction model to obtain an anti-vibration signal, wherein the preset noise prediction model is pre-trained by machine learning using multiple sets of data, and the multiple sets of data are multiple noise sampling segments pre-collected by the feedforward vibration sensor, and the preset noise prediction model is trained with the goal of minimizing the anti-noise data signal collected by the feedback vibration sensor 75, and the feedback vibration sensor is arranged on the indoor side of the glass, and the anti-noise data signal is the noise generated by the coherent superposition of the predicted vibration signal predicted by the preset noise prediction model based on the noise sampling segment and the sampled vibration signal generated on the glass by the noise sampling segment; a vibration controller 76 for generating an anti-vibration signal, wherein the vibration controller is tightly attached to the surface of the glass, and the anti-vibration signal and the noise vibration signal generated on the glass by the noise data signal form a coherent superposition.

[0134] In an embodiment of the present invention, a feedforward vibration sensor is arranged on the outdoor side of the glass to collect the original noise data signal of the ambient noise, and the original noise data signal is predicted by a preset noise prediction model to determine an anti-vibration signal for offsetting the original noise data signal, and a vibration controller bonded to the glass surface is controlled to generate the anti-vibration signal, and then the anti-vibration signal can be coherently superimposed with the noise vibration signal generated by the noise data signal on the glass to offset the vibration caused by the original noise data signal to the glass, thereby achieving the technical effect of active noise reduction for the glass, and thus solving the technical problem that the prior art lacks an effective way to counteract the propagation of noise by glass.

[0135] As an optional embodiment, the feedforward vibration sensor is spaced apart from the outdoor side surface of the glass by a first preset distance.

[0136] Optionally, the first preset distance may be 10-20 cm from the glass.

[0137] As an optional embodiment, the feedback vibration sensor is spaced apart from the indoor side surface of the glass by a second preset distance.

[0138] Optionally, the second preset distance may be 10-20 cm from the glass.

[0139] As an optional embodiment, when there are multiple vibration controllers, the multiple vibration controllers are arranged according to a third preset distance.

[0140] Optionally, the third preset distance may be 60-80 cm.

[0141] As an optional embodiment, Figure 7As shown, the system further includes: shock absorbing dampers 78, which are arranged at equidistant centers of the plurality of vibration controllers, wherein the shock absorbing dampers are made of flexible steel wires.

[0142] An embodiment of the present invention may provide an electronic device, which may be a computer terminal, and the computer terminal may be any computer terminal device in a computer terminal group. Optionally, in this embodiment, the computer terminal may also be replaced by a terminal device such as a mobile terminal.

[0143] Optionally, in this embodiment, the computer terminal may be located in at least one network device among a plurality of network devices of the computer network.

[0144] In this embodiment, the above-mentioned computer terminal can execute the program code of the following steps in the active noise reduction method of glass: collecting original noise data signals through a feedforward vibration sensor, wherein the feedforward vibration sensor is arranged on the outdoor side of the glass; using a preset noise prediction model to predict the original noise data signal to obtain an anti-vibration signal, wherein the preset noise prediction model is pre-trained by machine learning using multiple sets of data, and the multiple sets of data are multiple noise sampling segments pre-collected by the feedforward vibration sensor, and the preset noise prediction model is trained with the goal of minimizing the anti-noise data signal collected by the feedback vibration sensor, and the feedback vibration sensor is arranged on the indoor side of the glass, and the anti-noise data signal is the noise generated by the coherent superposition of the predicted vibration signal predicted by the preset noise prediction model based on the noise sampling segment and the sampled vibration signal generated on the glass by the noise sampling segment; controlling the vibration controller to generate an anti-vibration signal, wherein the vibration controller is attached to the surface of the glass, and the anti-vibration signal and the noise vibration signal generated on the glass by the noise data signal form a coherent superposition.

[0145] Figure 8 is a structural block diagram of a computer terminal according to an embodiment of the present invention. Figure 8 As shown, the computer terminal 80 may include: one or more (only one is shown in the figure) processors 82 , and a memory 84 .

[0146] Among them, the memory can be used to store software programs and modules, such as program instructions / modules corresponding to the active noise reduction method and device for glass in the embodiment of the present invention. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, that is, the active noise reduction method for glass mentioned above is realized. The memory may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory may further include a memory remotely arranged relative to the processor, and these remote memories can be connected to the terminal 80 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0147] The processor can call the information and application stored in the memory through the transmission device to perform the following steps: collect the original noise data signal through the feedforward vibration sensor, wherein the feedforward vibration sensor is arranged on the outdoor side of the glass; use the preset noise prediction model to predict the original noise data signal to obtain an anti-vibration signal, wherein the preset noise prediction model is pre-trained by machine learning using multiple sets of data, and the multiple sets of data are multiple noise sampling segments pre-collected by the feedforward vibration sensor, and the preset noise prediction model is trained with the goal of minimizing the anti-noise data signal collected by the feedback vibration sensor, and the feedback vibration sensor is arranged on the indoor side of the glass, and the anti-noise data signal is the noise generated by the coherent superposition of the predicted vibration signal predicted by the preset noise prediction model based on the noise sampling segment and the sampled vibration signal generated on the glass by the noise sampling segment; control the vibration controller to generate the anti-vibration signal, wherein the vibration controller is attached to the surface of the glass, and the anti-vibration signal and the noise vibration signal generated on the glass by the noise data signal form a coherent superposition.

[0148] Optionally, the processor may also execute the program code of the following steps: continuously collecting multiple noise sampling segments according to a preset time length through a feedforward vibration sensor, wherein there are overlapping noise segments between adjacent noise sampling segments; using a preset noise prediction model to predict a predicted vibration signal corresponding to each noise sampling segment; controlling the vibration controller to generate a corresponding predicted vibration signal within the collection period of each noise sampling segment; determining an instantaneous loss function for training a preset noise prediction model based on the anti-noise data signal collected by the feedback vibration sensor within each collection period, and adjusting the model parameters in the preset noise prediction model according to the instantaneous loss function.

[0149] Optionally, the processor may also execute the program code of the following steps: in the training stage of the preset noise prediction model, using an adaptive momentum estimation optimizer to adjust the model parameters in the preset noise prediction model of this round of training; after each round of training, detecting whether the anti-noise data signal collected by the feedback vibration sensor is a minimum value; if the anti-noise data signal is not a minimum value, performing the next round of training on the preset noise prediction model; if the anti-noise data signal is a minimum value, determining that the training of the preset noise prediction model is completed.

[0150] Optionally, the processor may also execute the following program code: using a preset noise prediction model to perform a short-time Fourier transform on the original noise data signal to obtain a first signal to be processed; using an encoder in the preset noise prediction model to perform feature extraction on the first signal to be processed to obtain a first noise feature; using a long short-term memory model in the preset noise prediction model to eliminate the delay effect in the first noise feature to obtain a second noise feature; using a decoder in the preset noise prediction model to process the first noise feature and the second noise feature to obtain an anti-noise feature; performing an inverse short-time Fourier transform on the anti-noise feature to obtain an anti-vibration signal.

[0151] Optionally, the encoder includes at least: a first feature convolution layer, a second feature convolution layer and a third feature convolution layer, and the processor may also execute the following program code: using the first feature convolution layer to perform feature extraction on the first signal to be processed to obtain a first noise sub-feature; using the second feature convolution layer to perform feature extraction on the first noise sub-feature to obtain a second noise sub-feature; using the second feature convolution layer to perform feature extraction on the second noise sub-feature to obtain a third noise sub-feature; wherein the first noise feature includes at least: a first noise sub-feature, a second noise sub-feature and a third noise sub-feature.

[0152] Optionally, the long short-term memory model includes at least: a first long short-term memory model, a second long short-term memory model and a third long short-term memory model, and the processor may also execute the following program code: using the first long short-term memory model to process the first noise sub-feature to obtain a fourth noise sub-feature; using the second long short-term memory model to process the second noise sub-feature to obtain a fifth noise sub-feature; using the second long short-term memory model to process the third noise sub-feature to obtain a sixth noise sub-feature; wherein the second noise feature includes at least: a fourth noise sub-feature, a fifth noise sub-feature and a sixth noise sub-feature.

[0153] Optionally, the decoder includes at least: a first feature deconvolution layer, a second feature deconvolution layer and a third feature deconvolution layer, and the processor may also execute the following steps of program code: using the first feature deconvolution layer to deconvolve the third noise sub-feature and the sixth noise sub-feature to obtain a first antagonistic sub-feature; using the second feature deconvolution layer to deconvolve the first antagonistic sub-feature and the fifth noise sub-feature to obtain a second antagonistic sub-feature; using the third feature deconvolution layer to deconvolve the second antagonistic sub-feature and the fourth noise sub-feature to obtain a third antagonistic sub-feature; and determining the third antagonistic sub-feature as an antagonistic noise feature.

[0154] In an embodiment of the present invention, a feedforward vibration sensor is arranged on the outdoor side of the glass to collect the original noise data signal of the ambient noise, and the original noise data signal is predicted by a preset noise prediction model to determine an anti-vibration signal for offsetting the original noise data signal, and a vibration controller bonded to the glass surface is controlled to generate the anti-vibration signal, and then the anti-vibration signal can be coherently superimposed with the noise vibration signal generated by the noise data signal on the glass to offset the vibration caused by the original noise data signal to the glass, thereby achieving the technical effect of active noise reduction for the glass, and thus solving the technical problem that the prior art lacks an effective way to counteract the propagation of noise by glass.

[0155] It can be understood by those skilled in the art that Figure 8 The structure shown is for illustration only, and the computer terminal may also be a smart phone (such as an Android phone, an iOS phone, etc.), a tablet computer, a handheld computer, a mobile Internet device (Mobile Internet Devices, MID), a PAD, or other terminal devices. Figure 8 The structure of the electronic device is not limited. For example, the computer terminal 80 may also include Figure 8 More or fewer components (such as network interfaces, display devices, etc.) shown in, or having Figure 8 Different configurations shown.

[0156] A person of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the hardware related to the terminal device through a computer program. The computer program can be stored in a non-volatile medium. The non-volatile storage medium may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0157] The embodiment of the present invention further provides a non-volatile storage medium. Optionally, in this embodiment, the non-volatile storage medium can be used to store the program code executed by the active noise reduction method for glass provided in the above embodiment.

[0158] Optionally, in this embodiment, the non-volatile storage medium may be located in any computer terminal in a computer terminal group in a computer network, or in any mobile terminal in a mobile terminal group.

[0159] Optionally, in this embodiment, the non-volatile storage medium is configured to store program codes for executing the following steps: collecting original noise data signals through a feedforward vibration sensor, wherein the feedforward vibration sensor is arranged on the outdoor side of the glass; using a preset noise prediction model to predict the original noise data signal to obtain an anti-vibration signal, wherein the preset noise prediction model is pre-trained through machine learning using multiple sets of data, the multiple sets of data are multiple noise sampling segments pre-collected by the feedforward vibration sensor, the preset noise prediction model is trained with the goal of minimizing the anti-noise data signal collected by the feedback vibration sensor, the feedback vibration sensor is arranged on the indoor side of the glass, the anti-noise data signal is the noise generated by the coherent superposition of the predicted vibration signal predicted by the preset noise prediction model based on the noise sampling segment and the sampled vibration signal generated on the glass by the noise sampling segment; controlling the vibration controller to generate an anti-vibration signal, wherein the vibration controller is attached to the surface of the glass, and the anti-vibration signal and the noise vibration signal generated on the glass by the noise data signal form a coherent superposition.

[0160] Optionally, in this embodiment, the non-volatile storage medium is configured to store program codes for executing the following steps: continuously collecting multiple noise sampling segments according to a preset time length through a feedforward vibration sensor, wherein there are overlapping noise segments between adjacent noise sampling segments; using a preset noise prediction model to predict a predicted vibration signal corresponding to each noise sampling segment; controlling the vibration controller to generate a corresponding predicted vibration signal within the collection period of each noise sampling segment; determining an instantaneous loss function for training a preset noise prediction model based on the anti-noise data signal collected by the feedback vibration sensor within each collection period, and adjusting the model parameters in the preset noise prediction model according to the instantaneous loss function.

[0161] Optionally, in this embodiment, the non-volatile storage medium is configured to store program code for executing the following steps: in the training phase of the preset noise prediction model, using an adaptive momentum estimation optimizer to adjust the model parameters in the preset noise prediction model of this round of training; after each round of training, detecting whether the anti-noise data signal collected by the feedback vibration sensor is a minimum value; if the anti-noise data signal is not a minimum value, performing the next round of training on the preset noise prediction model; if the anti-noise data signal is a minimum value, determining that the training of the preset noise prediction model is completed.

[0162] Optionally, in this embodiment, the non-volatile storage medium is configured to store program codes for executing the following steps: using a preset noise prediction model to perform a short-time Fourier transform on the original noise data signal to obtain a first signal to be processed; using an encoder in the preset noise prediction model to extract features from the first signal to be processed to obtain a first noise feature; using a long short-term memory model in the preset noise prediction model to eliminate the delay effect in the first noise feature to obtain a second noise feature; using a decoder in the preset noise prediction model to process the first noise feature and the second noise feature to obtain an anti-noise feature; performing an inverse short-time Fourier transform on the anti-noise feature to obtain an anti-vibration signal.

[0163] Optionally, in this embodiment, the encoder includes at least: a first feature convolution layer, a second feature convolution layer and a third feature convolution layer, and the non-volatile storage medium is configured to store program codes for executing the following steps: using the first feature convolution layer to perform feature extraction on the first signal to be processed to obtain a first noise sub-feature; using the second feature convolution layer to perform feature extraction on the first noise sub-feature to obtain a second noise sub-feature; using the second feature convolution layer to perform feature extraction on the second noise sub-feature to obtain a third noise sub-feature; wherein the first noise feature includes at least: a first noise sub-feature, a second noise sub-feature and a third noise sub-feature.

[0164] Optionally, in this embodiment, the long short-term memory model includes at least: a first long short-term memory model, a second long short-term memory model and a third long short-term memory model, and the non-volatile storage medium is configured to store program codes for executing the following steps: using the first long short-term memory model to process the first noise sub-feature to obtain a fourth noise sub-feature; using the second long short-term memory model to process the second noise sub-feature to obtain a fifth noise sub-feature; using the second long short-term memory model to process the third noise sub-feature to obtain a sixth noise sub-feature; wherein the second noise feature includes at least: a fourth noise sub-feature, a fifth noise sub-feature and a sixth noise sub-feature.

[0165] Optionally, in this embodiment, the decoder includes at least: a first feature deconvolution layer, a second feature deconvolution layer and a third feature deconvolution layer, and the non-volatile storage medium is configured to store program codes for executing the following steps: using the first feature deconvolution layer to deconvolve the third noise sub-feature and the sixth noise sub-feature to obtain a first antagonistic sub-feature; using the second feature deconvolution layer to deconvolve the first antagonistic sub-feature and the fifth noise sub-feature to obtain a second antagonistic sub-feature; using the third feature deconvolution layer to deconvolve the second antagonistic sub-feature and the fourth noise sub-feature to obtain a third antagonistic sub-feature; and determining the third antagonistic sub-feature as an antagonistic noise feature.

[0166] The embodiment of the present invention further provides a computer program product, including a computer program. Optionally, in this embodiment, when the computer program is executed by a processor, the steps of the active noise reduction method for glass provided in the above embodiment are implemented.

[0167] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0168] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0169] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0170] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0171] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0172] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a non-volatile storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a non-volatile storage medium, including several instructions for a computer device (which can be a personal computer, server or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned non-volatile storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

[0173] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for active noise reduction of glass, characterized in that: include: Collecting the original noise data signal through a feedforward vibration sensor, wherein the feedforward vibration sensor is arranged on the outdoor side of the glass; The original noise data signal is predicted using a preset noise prediction model to obtain an anti-vibration signal, wherein the preset noise prediction model is pre-trained by machine learning using multiple sets of data, the multiple sets of data are multiple noise sampling segments pre-collected by a feedforward vibration sensor, the preset noise prediction model is trained with the goal of minimizing the anti-noise data signal collected by the feedback vibration sensor, the feedback vibration sensor is arranged on the indoor side of the glass, and the anti-noise data signal is the noise generated by coherently superimposing a predicted vibration signal predicted by the preset noise prediction model based on the noise sampling segment and a sampled vibration signal generated by the noise sampling segment on the glass; The vibration controller is controlled to generate the anti-vibration signal, wherein the vibration controller is attached to the surface of the glass, and the anti-vibration signal and the noise vibration signal generated by the noise data signal on the glass form a coherent superposition.

2. The method according to claim 1, characterized in that Before using a preset noise prediction model to predict the original noise data signal to obtain an anti-vibration signal, the method further includes: By means of the feedforward vibration sensor, a plurality of the noise sampling segments are continuously collected according to a preset time length, wherein there are overlapping noise segments between adjacent noise sampling segments; Using the preset noise prediction model to predict the predicted vibration signal corresponding to each of the noise sampling segments; Controlling the vibration controller to generate the corresponding predicted vibration signal within the collection period of each noise sampling segment; Based on the anti-noise data signal collected by the feedback vibration sensor in each collection period, an instantaneous loss function for training the preset noise prediction model is determined, and model parameters in the preset noise prediction model are adjusted according to the instantaneous loss function.

3. The method according to claim 2, characterized in that Determining an instantaneous loss function for training the preset noise prediction model based on the anti-noise data signal collected by the feedback vibration sensor in each collection period, and adjusting the model parameters in the preset noise prediction model according to the instantaneous loss function includes: During the training phase of the preset noise prediction model, using an adaptive momentum estimation optimizer to adjust the model parameters in the preset noise prediction model of this round of training; After each round of training, detecting whether the anti-noise data signal collected by the feedback vibration sensor is a minimum value; When the anti-noise data signal is not a minimum value, performing a next round of training on the preset noise prediction model; When the anti-noise data signal is at a minimum value, it is determined that the training of the preset noise prediction model is completed.

4. The method according to claim 1, characterized in that: Predicting the original noise data signal using a preset noise prediction model to obtain an anti-vibration signal includes: Using the preset noise prediction model, performing a short-time Fourier transform on the original noise data signal to obtain a first signal to be processed; Using the encoder in the preset noise prediction model, extracting features from the first signal to be processed to obtain a first noise feature; Using the long short-term memory model in the preset noise prediction model, eliminating the delay effect in the first noise feature, and obtaining a second noise feature; Using a decoder in the preset noise prediction model to process the first noise feature and the second noise feature to obtain an adversarial noise feature; The anti-noise feature is subjected to short-time inverse Fourier transform to obtain the anti-vibration signal.

5. The method according to claim 4, characterized in that The encoder includes at least: a first feature convolution layer, a second feature convolution layer and a third feature convolution layer. The encoder in the preset noise prediction model is used to extract features of the first signal to be processed, and the first noise feature obtained includes: Using the first feature convolution layer to perform feature extraction on the first signal to be processed to obtain a first noise sub-feature; Using the second feature convolution layer to perform feature extraction on the first noise sub-feature to obtain a second noise sub-feature; Using the second feature convolution layer to perform feature extraction on the second noise sub-feature to obtain a third noise sub-feature; The first noise feature at least includes: the first noise sub-feature, the second noise sub-feature and the third noise sub-feature.

6. The method according to claim 5, characterized in that The long short-term memory model at least includes: a first long short-term memory model, a second long short-term memory model and a third long short-term memory model. The long short-term memory model in the preset noise prediction model is used to eliminate the delay effect in the first noise feature, and the second noise feature is obtained including: Processing the first noise sub-feature using a first long short-term memory model to obtain a fourth noise sub-feature; Processing the second noise sub-feature using a second long short-term memory model to obtain a fifth noise sub-feature; Processing the third noise sub-feature using a second long short-term memory model to obtain a sixth noise sub-feature; The second noise feature at least includes: the fourth noise sub-feature, the fifth noise sub-feature and the sixth noise sub-feature.

7. The method according to claim 6, characterized in that The decoder includes at least: a first feature deconvolution layer, a second feature deconvolution layer and a third feature deconvolution layer, and uses the decoder in the preset noise prediction model to process the first noise feature and the second noise feature to obtain the anti-noise feature, including: Deconvolving the third noise sub-feature and the sixth noise sub-feature using a first feature deconvolution layer to obtain a first adversarial sub-feature; Deconvolving the first adversarial sub-feature and the fifth noise sub-feature using a second feature deconvolution layer to obtain a second adversarial sub-feature; Deconvolving the second adversarial sub-feature and the fourth noise sub-feature using a third feature deconvolution layer to obtain a third adversarial sub-feature; The third antagonistic sub-feature is determined as the antagonistic noise feature.

8. An active noise reduction device for glass, characterized in that: include: A collection module, used for collecting the original noise data signal through a feedforward vibration sensor, wherein the feedforward vibration sensor is arranged on the outdoor side of the glass; A prediction module, configured to predict the original noise data signal using a preset noise prediction model to obtain an anti-vibration signal, wherein the preset noise prediction model is pre-trained by machine learning using multiple sets of data, the multiple sets of data are multiple noise sampling segments pre-collected by a feedforward vibration sensor, the preset noise prediction model is trained with the goal of minimizing the anti-noise data signal collected by a feedback vibration sensor, the feedback vibration sensor is arranged on the indoor side of the glass, and the anti-noise data signal is the noise generated by coherently superimposing a predicted vibration signal predicted by the preset noise prediction model based on the noise sampling segment and a sampled vibration signal generated by the noise sampling segment on the glass; The noise reduction module is used to control the vibration controller to generate the anti-vibration signal, wherein the vibration controller is tightly attached to the surface of the glass, and the anti-vibration signal and the noise vibration signal generated by the noise data signal on the glass form a coherent superposition.

9. An active noise reduction system for glass, characterized in that: include: A feedforward vibration sensor, used for collecting raw noise data signals, wherein the feedforward vibration sensor is arranged on the outdoor side of the glass; A computing device, configured to predict the original noise data signal using a preset noise prediction model to obtain an anti-vibration signal, wherein the preset noise prediction model is pre-trained by machine learning using multiple sets of data, the multiple sets of data are multiple noise sampling segments pre-collected by a feedforward vibration sensor, the preset noise prediction model is trained with the goal of minimizing the anti-noise data signal collected by a feedback vibration sensor, the feedback vibration sensor is arranged on the indoor side of the glass, and the anti-noise data signal is the noise generated by coherently superimposing a predicted vibration signal predicted by the preset noise prediction model based on the noise sampling segment and a sampled vibration signal generated by the noise sampling segment on the glass; A vibration controller is used to generate the anti-vibration signal, wherein the vibration controller is tightly attached to the surface of the glass, and the anti-vibration signal and the noise vibration signal generated by the noise data signal on the glass form a coherent superposition.

10. An electronic device comprising a memory and a processor, characterized in that: The memory stores a computer program, and the processor is configured to execute the active noise reduction method for the glass according to any one of claims 1 to 7 through the computer program.

11. A computer program product comprising computer instructions, characterized in that: When the computer instructions are executed by a processor, the steps of the active noise reduction method for glass according to any one of claims 1 to 7 are implemented.

Images