Automobile noise active control method and system and storage medium

Abstract

The invention provides an automobile noise active control method and system and a storage medium, relates to the field of automobile noise reduction, and aims to increase available scenes by establishing a local feature library and a local weight library which reflect driving habits and determining the weight of an adaptive filter based on similarity analysis of real-time noise features and noise features in the local feature library. According to the invention, the activation function is introduced, so that the convergence time of the filter can be shortened more widely, and the noise reduction system can achieve the same noise reduction effect more quickly. The invention also provides a real-time updating method and a non-real-time updating method of the local feature library and the local weight library, and provides a judgment scheme, so that the robustness of the local feature library and the local weight library can be ensured. A storage space maintenance scheme is provided, the matching times and the number of days are recorded through a very small amount of extra storage space, the score of each feature can be obtained through simple and effective calculation, and the local feature library and the weight library are kept in a good state.

Description

Technical Field

[0001] The present application relates to the field of automobile noise reduction, and in particular to an automobile noise active control method, system and storage medium. Background Art

[0002] With the rapid development of my country's automobile industry in recent years, automobile manufacturers and consumers have put forward higher and higher quality requirements for automobile driving and riding experience, among which cabin noise control has become one of the focuses. Traditional noise control methods mainly rely on passive noise isolation methods, such as using sound insulation materials and optimizing vehicle structure, but these methods are not only costly, but also have limited noise reduction effects.

[0003] In order to overcome the limitations of traditional noise control methods, active noise control technology has emerged. This technology realizes active control of in-vehicle noise by introducing hardware such as multiple sensors, multiple-stage speakers and multiple error microphones, combined with noise control algorithms. Noise control algorithms have a large demand for computing power. With the rapid development of vehicle-mounted DSP (digital signal processor) chip technology in recent years, computing power issues are gradually being overcome. However, problems such as slow algorithm convergence and insufficient noise reduction in active noise control systems are still key obstacles to commercialization and are currently a hot topic and focus of research. Summary of the invention

[0004] The present application provides a method, system and storage medium for active automobile noise control, which can quickly switch to a more accurate noise reduction path when the user is driving, so as to improve the convergence speed and noise reduction effect of the active noise control system, and provide the user with a better noise reduction experience.

[0005] In one aspect, the present application provides a method for active automobile noise control, the method comprising: Collect noise data, extract noise features and store them to build a local feature library that reflects driving habits; Calculating noise reduction path weights according to the noise data to establish a local weight library, wherein the noise reduction path weights in the local weight library correspond one to one to the noise features in the local feature library; Collecting real-time noise data, extracting real-time noise features, calculating the similarity between the real-time noise features and each noise feature in the local feature library, and rewriting the weight of the adaptive filter according to the similarity, so as to perform active control of noisy vehicle sound; The weight of the adaptive filter is obtained through the noise reduction path weight in the local weight library.

[0006] During implementation, the acceleration sensor is set at the connection point between the suspension and the vehicle body to collect the structural sound caused by the vibration of the vehicle body. The error microphone is set at the headrest of the seat in the car to collect the noise near the ears of the passengers. During the driving process of the car, the acceleration sensor and the error microphone are used to collect noise data under different road conditions in different driving environments and conditions. Based on a large amount of noise data, a local feature library is established so that when performing noise control, more scenarios can use the existing noise reduction path weights in the local weight library. Among them, the noise data collected by the acceleration sensor is recorded as , the noise data collected by the microphone in the car is recorded as After short-time Fourier transform, the time-frequency matrix is ​​obtained, which is recorded as X and D respectively, the timestamp vector is recorded as t, and the input information of the neural network used to extract features is the matrix The obtained feature vector is recorded as , the noise reduction path weight is calculated according to the noise data and recorded as w, wherein s and w have a one-to-one correspondence. In some embodiments, it includes The local "feature and weight library" of a feature is represented as: ; in, , , , Sequentially indicates The feature vector, path weight, number of matches, and number of days of existence of each feature.

[0007] in, It is called "local feature library"; It is called "local weight library".

[0008] In the present application, the method for establishing the local feature library and the local weight library is flexible. In addition to establishing them locally, the feature library and the weight library established by the developer can also be downloaded from the cloud, or the data can be uploaded and the features can be extracted and the weights calculated by the cloud before being downloaded locally.

[0009] When the similarity is between the first activation function threshold and the second activation function threshold, the real-time noise feature partially matches the noise feature in the local feature library, and the weight of the adaptive filter is the weighted sum of the noise reduction path weights corresponding to the partially matched noise features, wherein the weighting coefficient comes from the activation function.

[0010] When the similarity is greater than a second activation function threshold, the real-time noise feature fully matches the noise feature in the local feature library, and the weight of the adaptive filter is the noise reduction path weight corresponding to the fully matched noise feature.

[0011] In a feasible implementation, the activation function is: ; in, , That is the first activation function threshold, is the second activation function threshold, For similarity.

[0012] The similarity is calculated as: ; In the formula, is the feature vector of real-time noise characteristics, is the first l The feature vector of the noise features, is the number of features in the local feature library, express and The cosine similarity of is in the range of [0,1].

[0013] Writing the initial value of weight , where the subscript Indicates the system working time. Indicates As the benchmark, the number of weight updates is zero. The expression is: ; in, is the number of partially matched features, For the The weights corresponding to the partial matching features, is the activation function value for partial matching, is the weight corresponding to the most matching feature, is the subscript of the best matching feature.

[0014] In addition, when the match is sufficient, the method further includes updating the denoising path weights with a small step length to obtain a more refined weight iteration.

[0015] By setting the activation function and the first activation function threshold and the activation function threshold, the use requirements of the noise reduction path weights in the local weight library are reduced, so that partially matched scenes can also obtain the noise reduction path weights of the corresponding activation degree from the local weight library as the weights of the adaptive filter, so that more scenes can call the noise reduction path weights in the local weight library, reduce the occurrence of mismatches, and improve the efficiency of noise control. In addition, since the weighted superposition of the noise reduction path weights of multiple partially matched features is allowed, the solution of this application is scalable, that is, based on the existing features, the weighted combination obtains features close to the current real-time features, thereby improving the coverage of the feature library.

[0016] In a feasible implementation, the adaptive filter adopts a cyclic leakage minimum mean square error filter. The function of the weight update process of the cyclic leakage minimum mean square error filter is expressed as: ; in, is the identity matrix, is the adjustable step size factor; Indicates the number of weight updates; For the The weight vector after the update is for right N Take the remainder; is the weight vector The item; is the filter-x signal matrix; is the vector of error microphone signals, The independent variable is and The leakage function of is a The square matrix, only The first The column elements are 1, and the rest are 0. Its function is to perform polling leakage on the weight vector, that is, each time it is updated, only one element in the weight vector is leaked.

[0017] yes and The function is calculated as follows: ; in, ; Compared with the traditional minimum mean square error filter, the cyclic leakage minimum mean square error filter has a small amount of additional calculation and is easy to implement. Since the nonlinear leakage term is similar to the role of the activation function and only one element in the weight vector is leaked each time, it is an unbiased estimate. Therefore, the cyclic leakage minimum mean square error filter takes into account both robustness and estimation performance.

[0018] In a feasible implementation, the method further includes a real-time update method of a local feature library and a local weight library, and the iterative update method includes: When the real-time noise feature does not match or partially matches each noise feature in the local feature library, storing the real-time noise feature in the local feature library to update the local feature library; Based on the real-time noise data, the adaptive filter performs weight iteration to obtain the final weight value. When the real-time noise feature fully matches the noise feature in the local feature library, the real-time noise reduction path weight replaces the noise reduction path weight corresponding to the fully matching noise feature. When the real-time noise feature partially matches or does not match the noise feature in the local feature library, new storage space is allocated to store the real-time noise reduction path weight.

[0019] The update process of the local weight library is as follows: ; in, is the denoising path weight corresponding to the last frame, i.e. -1 frame of data after updating the weights. The cycle for executing the initial weight replacement strategy in the present application scheme has a minimum value of 1, that is, feature and similarity calculations are performed for each frame, and the replacement weights are considered accordingly.

[0020] Local weight library Adding 1 means that the number of matches of the feature corresponding to the weight of the denoising path is increased by 1, which is an important judgment basis for cleaning and updating the local feature library and the local weight library, that is, cleaning the features and corresponding weights that appear too few times. are the subscripts of the features and weights in the local feature library and the local weight library. For the three different matching situations, the situations when writing into the local weight library are different, namely: ; In a feasible implementation, the iterative update method includes a method for maintaining the storage space, and the maintenance method includes: Calculating the noise feature and the weight score of the noise reduction path weight corresponding to the noise feature; Ranking the weight scores in descending order; The noise features with lower weight scores and the noise reduction path weights corresponding to the noise features are discarded.

[0021] When allocating new storage space for storing the real-time noise features and the real-time noise reduction path weights, if the storage space capacity is insufficient, the noise features at the end of the weight score ranking and the noise reduction path weights corresponding to the noise features are discarded.

[0022] The weight score is calculated as follows: ; in, is the cumulative number of matches, There are days, because the calculation cycle of features and similarities is fixed, so the number of times is enough to reflect the duration.

[0023] In a feasible implementation, the method further includes a non-real-time updating method of the local feature library and the local weight library, and the non-real-time updating method includes: Collecting real-time noise data, and storing the real-time noise data and a desired signal of the primary sound field in a temporary storage space; When computing power is idle, the noise features of the real-time noise data are extracted and the noise reduction path weights corresponding to the real-time noise features are calculated to update the local feature library and the weight feature library.

[0024] In a feasible implementation, in the non-real-time updating method, a Wiener filter or a recursive least squares filter is used to calculate the noise reduction path weights.

[0025] The process of calculating the noise reduction path weights using the Wiener filter is expressed as: ; in, is the weight of the Wiener filter, is the filter-x signal matrix, i.e. ; in For the The reference signal of each channel passes through the Secondary sound source to the Secondary path of an error microphone The filtered-x signal vector after ; in ; for microphone signal vector, i.e. ; in It is a signal of expectation for recovery; In addition, R is the filtered-x signal The autocorrelation matrix of p is the filtered-x signal With microphone signal The cross-correlation vector between , the autocorrelation matrix R and the cross-correlation vector p are estimated by the sample data.

[0026] The process of calculating the noise reduction path weights using the recursive least squares filter (RLS) is expressed as: ; ; ; ; in, is the gain vector, is the recursive weight of the recursive least squares. What is needed here is the weight of the final iteration. When the calculation is completed, the number of iterations is hour, for: ; In some embodiments, the effect comparison can be measured using the residual sum of squares, and the data length used is denoted as , taking the Wiener filter as an example, use Filtering, residual sum of squares for: ; The denoising path weight in the local weight library is recorded as , and its corresponding residual sum of squares for: ; Introducing adjustable coefficients , indicating that the performance required by the new weights needs to exceed the degree of the noise reduction path weights in the local weight library, so The situation of writing to the local weight library is described as follows: ; The update of the local feature library and the local weight library ensures the robustness of the local feature library and the local weight library. In the real-time update mode, when the real-time noise feature does not match or partially matches the feature in the local feature library, the real-time noise feature and the corresponding real-time noise reduction path weight are stored in the library, which helps to enrich and improve the local feature library and the local weight library, so that the system can find a matching noise reduction solution more quickly when facing a similar noise environment in the future.

[0027] In a feasible implementation, the desired signal of the primary sound field is obtained according to the residual signal of the error microphone, and the function expression of the desired signal of the primary sound field is: ; in, is the number of secondary sound sources, is the secondary sound source index, ; is the number of error microphones, is the error microphone index, ; For the The residual signal of the error microphone; To deliver to The excitation signal of the secondary sound source is For offline identification Secondary sound source to the The impulse response of an error microphone, for Length; For the recovery The desired signal of the primary sound field at the error microphone.

[0028] The expected signal recovery technology is to subtract the component of the excitation signal after passing through the secondary path from the error microphone signal, and restore the residual signal of the error microphone to the expected signal of the primary sound field. The introduction of this technology makes the preservation of the original data and the real-time noise reduction system compatible and independent of each other. If there is no expected signal recovery technology, the primary sound field will be changed when the real-time noise reduction system works normally. The signal received by the error microphone is the residual signal of the noise reduction system at this time, which cannot be used for non-real-time calculation of weight modules. The real-time noise reduction system needs to be shut down to collect and save the original data. After the introduction of the expected recovery technology, the original data can be saved while performing real-time noise reduction, which improves the system operation efficiency and feasibility.

[0029] On the other hand, the present application provides an active vehicle noise control system, the system comprising: A collection module is used to collect real-time noise data of the car; A control module, wherein the control module comprises a weight calculation module, wherein the weight calculation module obtains the real-time noise data and calculates the real-time noise reduction path weight, and updates a local weight library.

[0030] In a feasible implementation, the weight calculation module includes a real-time weight calculation module and / or a non-real-time weight calculation module, which is used to calculate the noise reduction path weight according to the real-time noise data; wherein the real-time calculation module adopts an adaptive filter, and the non-real-time calculation module adopts a Wiener filter or a recursive least squares filter.

[0031] In a feasible implementation, the system further includes a desired signal recovery module, which is used to restore the residual signal of the error microphone into the desired signal of the primary sound field.

[0032] On the other hand, the present application provides a storage medium, wherein the storage medium stores a vehicle noise active control program, and when the vehicle noise active control program is executed by a processor, the vehicle noise active control method as described above is implemented.

[0033] Compared with the prior art, this application has at least the following beneficial effects: 1. A local feature library and a local weight library that reflect driving habits are established. Based on the similarity analysis of real-time noise features and noise features in the local feature library, the weights of the adaptive filter are replaced when conditions are met, shortening the convergence time of the adaptive filter after the characteristic road section changes. Furthermore, an activation function is introduced. The application of the activation function in the feature matching module not only allows the existing weights in the feature library to be used for fully matching scenarios, but also allows the existing noise reduction path weights in the local weight library to be used for some matching scenarios, thus expanding the scope of application and providing efficient and fast noise reduction solutions for a wider range of scenarios.

[0034] 2. Provides solutions for establishing, updating and maintaining local feature libraries and local weight libraries. The establishment methods of local feature libraries and local weight libraries are flexible. In addition to establishing them locally, you can also download the feature libraries and weight libraries established by developers from the cloud, or you can upload data and then have the cloud extract features and calculate weights before downloading them locally. The update of local feature libraries and local weight libraries provides two methods: real-time and non-real-time update, and provides a judgment scheme to ensure the robustness of local feature libraries and local weight libraries. Since the non-real-time weight calculation module may obtain better weights than the real-time adaptive filter, the data in the original database is used as input, and the weights are obtained by the non-real-time weight calculation module, which are compared with the existing weights to obtain better weights as the weights of the adaptive filter, which can more significantly shorten the convergence time and improve the noise reduction effect.

[0035] 3. This application provides a storage space maintenance solution, which records the number of matches and the number of days they exist by using a very small amount of additional storage space. A simple and effective calculation can be used to obtain the score of each feature, thereby keeping the local feature library and weight library in good condition.

[0036] 4. This application uses the expected signal recovery technology, so there is no need to deliberately collect data from the original database under the premise of turning off the noise reduction system. This method is more convenient and the user is not disturbed, which greatly improves the extensiveness of original data collection. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solution of the present application, the drawings required for use in the embodiments are briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0038] Figure 1 is a flow chart of the active vehicle noise control method provided by an embodiment of the present application; Figure 2 It is a flow chart of a method for real-time updating of a local feature library and a local weight library provided in an embodiment of the present application; Figure 3 It is a flow chart of a method for non-real-time updating of a local feature library and a local weight library provided in an embodiment of the present application; Figure 4 It is a flow chart of a storage space maintenance method of a local feature library and a local weight library provided in an embodiment of the present application; Figure 5 is a feature matching activation function curve diagram provided in an embodiment of the present application; Figure 6 is a leakage term curve diagram of a cyclic leakage adaptive filter provided in an embodiment of the present application; Figure 7 It is a relationship surface diagram of the scores, feature matching times, and feature existence days used in the maintenance process of the local feature library and the local weight library provided in the embodiment of the present application; Figure 8 This is a comparative analysis diagram of the implementation effects of the active noise control method for automobiles provided by the present application and the existing active noise control method. DETAILED DESCRIPTION

[0039] In order to enable those skilled in the art to better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present application.

[0040] The noise inside a car is an important indicator for users to evaluate the ride comfort of the car. Mainstream passive noise reduction technologies, such as sound absorption and sound insulation, are difficult to achieve ideal noise reduction effects. Therefore, active noise control methods are receiving increasing attention. At present, the mainstream active noise control method performs active noise control by designing the sensor layout and combining it with an adaptive filtering algorithm. This method affects the steady-state offset more. After deployment, it often faces the problem of slow algorithm convergence and the noise reduction amount failing to reach the expected level. The present application discloses an active automobile noise control method, system, and storage medium, which can quickly switch to a more accurate noise reduction path when the user is driving, so as to improve the convergence speed and noise reduction effect of the active noise control system and provide users with a better noise reduction experience.

[0041] The following is a detailed description of the active vehicle noise control method, system and storage medium provided by the present application in conjunction with the accompanying drawings.

[0042] Embodiment 1: Reference Figure 1 As shown, Figure 1 This is a flow chart of the active vehicle noise control method provided in the embodiment of the present application. The active vehicle noise control method provided in the present application includes: S100: Collect noise data, extract noise features and store them to establish a local feature library reflecting driving habits.

[0043] During implementation, the acceleration sensor is set at the connection point between the suspension and the vehicle body to collect the structure-borne sound caused by the vibration of the vehicle body, and the error microphone is set at the headrest of the seat in the car to collect the noise near the ears of the passengers. During the driving process of the car, the acceleration sensor and the error microphone are used to collect and store the noise data under different road conditions in different driving environments and conditions. Based on a large amount of noise data, a local feature library is established so that when performing noise control, more scenes can match the existing noise features in the local feature library.

[0044] As an example, the noise data collected by the acceleration sensor is recorded as , the noise data collected by the microphone in the car is recorded as After short-time Fourier transform, the time-frequency matrix is ​​obtained, which is recorded as X and D respectively, the timestamp vector is recorded as t, and the input information of the neural network used to extract features is the matrix The obtained feature vector is recorded as , the local feature library is .

[0045] The establishment of a local feature library can be achieved in two ways: Approach 1: Developers collect vehicle noise data under different driving environments and conditions, and use a pre-trained neural network model to extract features, store the extracted noise features, and obtain a local feature library.

[0046] When this method is used to construct a local feature library, the driving environment and conditions that may be encountered in the actual environment should be covered as much as possible, so as to increase the probability of fully matching similar noise features in noise control.

[0047] Approach 2: While the car is driving, real-time noise data is collected and uploaded to the cloud server through the on-board communication module. On the cloud server, a pre-trained neural network model is used to extract features from the uploaded real-time noise data, and the noise features are downloaded from the cloud to the local server to form a local feature library.

[0048] Among them, the neural network model used to establish the local feature library can be a BP neural network or an Elman neural network, etc. This application does not limit the neural network model used.

[0049] S200: Calculating noise reduction path weights according to the noise data to establish a local weight library, wherein the noise reduction path weights in the local weight library and the noise features in the local feature library correspond one to one.

[0050] The noise reduction path weight refers to the system response from the reference signal to the excitation signal when active noise control is performed. The noise reduction path weight is stored in the local weight library. In the subsequent active noise control process, the noise reduction path weight can be used as the weight of the adaptive filter.

[0051] In some embodiments, the noise reduction path weights in the local weight library can be calculated and stored centrally by non-real-time weight calculation to calculate the noise reduction path weights corresponding to the noise features in the local feature library, thereby constructing a weight feature library. As an example, the non-real-time weight calculation can use a Wiener filter or a recursive least squares filter. In some other embodiments, the noise reduction path weights in the local weight library can be calculated and stored in real time. As an example, real-time noise data is collected and the noise reduction path weights are calculated by an adaptive filter. In some other embodiments, the noise reduction path weights in the local features are calculated in a non-real-time calculation method, and during the use of the car, the local feature library is continuously optimized by real-time noise reduction path weight calculation to obtain a local weight library that is more in line with the user's driving habits, so as to better complete noise control.

[0052] The noise reduction path weights in the local weight library should correspond to the noise features in the local feature library. During driving, real-time noise data is collected and feature extraction is performed to obtain real-time noise features, which should correspond to the real-time noise reduction path weights calculated based on the real-time noise data. The local "feature and weight library" of a feature is represented as: ; in, , , , Sequentially indicates The feature vector, path weight, number of matches, and number of days of existence of each feature.

[0053] in, It is called "local feature library"; It is called "local weight library".

[0054] S300: Collect real-time noise data, extract real-time noise features, calculate the similarity between the real-time noise features and each noise feature in the local feature library, and rewrite the weight of the adaptive filter according to the similarity to perform active control of the noise. The weight of the adaptive filter is obtained through the noise reduction path weight in the local weight library. The following is an introduction to the method for determining the weight of the adaptive filter, as follows: Reference Figure 2 As shown, Figure 2 This is a flow chart of a real-time processing module in active automobile noise control using an adaptive filter provided in an embodiment of the present application. The weight of the adaptive filter is determined by similarity and an activation function. Specifically, the activation function is provided with a first activation function threshold and a second activation function threshold. The similarity is compared with the first activation function and the second activation function respectively to determine the matching degree between the real-time noise feature and the noise feature in the local feature library. The matching degree can be divided into three different situations: "mismatch", "partial match" and "sufficient match", specifically: Case 1: When the similarity is greater than or equal to the second activation function threshold, it is in a "sufficient match" state.

[0055] Case 2: when the similarity is between the first activation function threshold and the second activation function threshold, it belongs to the "partial match" state.

[0056] Case 3: When the similarity is less than or equal to the activation function parameter 1, it is in a "mismatch" state.

[0057] In some embodiments, the similarity is calculated as follows: ; In the formula, is the feature vector of real-time noise characteristics, is the first l The feature vector of the noise features, is the number of features in the local feature library, express and The cosine similarity of is in the range of [0,1].

[0058] The activation function is: ; ； in, is the first activation function threshold, is the second activation function threshold, both of which are preset fixed values. , 85 hours, refer to Figure 5 As shown, the feature matching activation function curve provided in the embodiment of the present application.

[0059] When the similarity is greater than the second activation function threshold, the real-time noise feature fully matches the noise feature in the local feature library, and the weight of the adaptive filter is the noise reduction path weight corresponding to the fully matched noise feature. At this time, the noise reduction path weight corresponding to the "most fully matched" noise feature in the local feature library is selected, where "most fully matched" means that the similarity value is the largest.

[0060] When the similarity is between the first activation function threshold and the second activation function threshold, the real-time noise feature partially matches the noise feature in the local feature library, and the weight of the adaptive filter is the weighted sum of the noise reduction path weights corresponding to the partially matched noise features, wherein the weight coefficient comes from the activation function, K = , where is the activation function value for partial matching, B is the number of noise features that are partially matched.

[0061] The advantage of the partial matching scheme is that it reduces the use requirements of the existing noise reduction path weights in the local weight library, so that partially matched scenes can also obtain weights of corresponding activation levels from the local weight library, so that more scenes can use the noise reduction path weights in the local weight library, and reduce the probability of noise control failure caused by mismatches. In addition, since the weighted superposition of noise reduction path weights of multiple partially matched features is allowed, the scheme of the present application is scalable, that is, based on the existing features, the weighted combination obtains noise features close to the current real-time noise features, thereby improving the coverage of the local feature library. Compared with only matching and non-matching schemes, the partial matching scheme has a small amount of additional computation and is easy to implement.

[0062] When the similarity is less than the first activation function threshold, the real-time noise feature does not match the noise feature in the local feature library, and the weight of the adaptive filter uses the weight updated by the adaptive filter algorithm at the last moment. At the initial moment, the weight of the adaptive filter is zero.

[0063] Based on the above, the initial value of the weight is written as , where the subscript Indicates the system working time. Indicates As the benchmark, the number of weight updates is zero. The expression is: ; in, is the number of partially matched features, For the The weights corresponding to the partial matching features, is the activation function value for partial matching, is the weight corresponding to the most matching feature, is the subscript of the best matching feature.

[0064] In addition, when the match is sufficient, the method further includes updating the denoising path weights with a small step length to obtain a more refined weight iteration.

[0065] In some embodiments, the adaptive filter may use a cyclic leakage minimum mean square error filter. The cyclic leakage filter will only perform leakage processing if the weight exceeds the threshold, and the degree of leakage is related to the degree of exceeding the threshold, and the leakage judgment period of each weight is N Iteration cycle, the cyclic leakage minimum mean square error filter weight update process is expressed as: in represents the identity matrix, is the adjustable step size factor; Indicates the number of weight updates; Indicates The weight vector after the update has a weight length of ; express right N Take the remainder; Represents the weight vector The item; is the filter-x signal matrix; is a vector of error microphone signals, i.e. ; is a The square matrix, only The first The column elements are 1, and the rest are 0. Its function is to perform polling leakage on the weight vector, that is, each time it is updated, only one element in the weight vector is leaked.

[0066] yes and The function is calculated as ; in, ; when , , When, refer to Figure 6 As shown, Figure 6 It is a leakage term curve diagram of the cyclic leakage adaptive filter provided in an embodiment of the present application.

[0067] Compared with the traditional minimum mean square error filter, the cyclic leakage minimum mean square error filter has better robustness due to the leakage mechanism. In addition, since the nonlinear leakage term is similar to the role of the activation function and only one element in the weight vector is leaked each time, it is an unbiased estimate. In summary, the cyclic leakage minimum mean square error filter takes into account both robustness and estimation performance. The cyclic leakage minimum mean square error filter has a small amount of additional calculations and is easy to implement.

[0068] The active automobile noise control method provided by the present application can accurately set the weights of the adaptive filter by analyzing the similarity between the real-time noise features and the noise features in the local feature library, thereby avoiding the tedious process of the filter converging from zero and significantly shortening its convergence time. Furthermore, an activation function is introduced, which can not only make full use of the existing weights in the feature library in fully matched scenarios, but also effectively utilize the existing noise reduction path weights in the local weight library in the case of partial matching. This strategy greatly expands the scope of applicable scenarios, so that the noise reduction process in more scenarios does not need to converge from zero, and the same level of noise reduction effect can be achieved more quickly.

[0069] Reference Figure 8 As shown, it is a comparative analysis diagram of the implementation effect of the active noise control method for automobiles provided by the present application and the existing active noise control method provided by the embodiment of the present application. The core of the present application is to inherit the weights, and the algorithm of the present application is referred to as the "scheme of the present application". The existing active noise control is referred to as the traditional scheme algorithm. By comparing the effects of the two active noise control methods for automobiles in different time periods of two types of noise signals, the implementation effect of the active noise control method for automobiles provided by the present application is explained, as follows: In this embodiment, the active noise control effect of two types of noise in different time periods is described. Among them, the two types of noise are Class A noise and Class B noise, and Class A noise and Class B noise are noises in different scenarios. As an example, A represents a high-speed section and B represents a low-speed section. In Class A noise, A1, A2, and A3 respectively represent the noise data collected by Class A noise in different sections or at different times. In Class B noise, B1, B2, and B3 respectively represent the noise data collected by Class B noise in different sections or at different times. In this embodiment, the noise environment changes are shown in the following table: Figure 8 middle: 0 to 90 seconds, noise environment is A1, local feature library has no match, reference Figure 2 and Figure 5 The present application scheme has the same effect as the traditional scheme.

[0070] 90 to 180 seconds, the noise environment is A2, there is no match in the local feature library, refer to Figure 2 and Figure 5 The present application scheme has the same effect as the traditional scheme.

[0071] At 180 seconds, the noise environment changes to A2. At this time, the similarity between A2 and A features (features obtained from A1) is higher than the second activation function threshold. The application scheme switches to the weight corresponding to the A1 segment data (the weight is calculated based on the 0-90 second data). Figure 3The algorithm of the present application scheme has a maximum improvement of about 3.8dBA in noise reduction effect compared with the traditional algorithm.

[0072] At 270 seconds, the noise environment changes to B2. At this time, the similarity between B2 and B features (features obtained from B1) is higher than the second activation function threshold. The present application solution switches to the weight corresponding to the B1 data (the weight is calculated based on the data from 90 to 180 seconds). Figure 3 The algorithm of the present application scheme has a maximum improvement of about 3.2dBA in noise reduction effect compared with the traditional algorithm.

[0073] At 360 seconds, the noise environment changes to A3. At this time, the similarity between the features of A3 and A (the features obtained from A1 and A2) is higher than the threshold. The present application scheme switches to the weight corresponding to the feature of A (the weight is the weight at the end of 270 seconds). Figure 3 The noise reduction effect of the algorithm of the present application is improved by about 3.3dBA at most compared with the traditional algorithm.

[0074] At 450 seconds, the noise environment changes to B3. At this time, the similarity between the features of B3 and B (the features obtained from B1 and B2) is higher than the second activation function threshold. The present application scheme switches to the weight corresponding to the feature of B (the weight is the weight at the end of 360 seconds). Figure 3 The noise reduction effect of the algorithm of the present application is improved by about 2.7dB at most compared with the traditional algorithm.

[0075] From 180 to 270 seconds, from 270 to 360 seconds, from 360 to 450 seconds, and from 450 to 540 seconds, the gap between the present application solution and the traditional solution is gradually narrowing, because the present application solution has converged to a steady state, and the traditional solution is gradually converging.

[0076] Embodiment 2: This embodiment only describes the parts that are different from Embodiment 1, and the same parts are not repeated again.

[0077] Reference Figure 2 As shown, Figure 2 : is a flow chart of a method for real-time updating of a local feature library and a local weight library provided in an embodiment of the present application, and the real-time updating method includes: S111: When the real-time noise feature does not match or partially matches each noise feature in the local feature library, the real-time noise feature is stored in the local feature library to update the local feature library.

[0078] S112: Based on the real-time noise data, the adaptive filter performs weight iteration to obtain a final weight value, and stores the final weight value as a real-time noise reduction path weight in a local weight library to update the local weight library.

[0079] When the car is driving, real-time noise data is collected, and the noise reduction path weight is iterated through the adaptive filter. At the end of the characteristic road section, the final value of the weight is stored in the local weight library as the real-time noise reduction path weight to update the local weight library. If the real-time noise feature fully matches the noise feature in the local feature library, the real-time noise reduction path weight replaces the original noise reduction path weight. If the real-time noise feature does not match or partially matches the noise feature in the local feature library, a new storage space is allocated to store the real-time noise reduction path weight and the corresponding real-time noise feature.

[0080] The update process of the local weight library is as follows: ; in, is the denoising path weight corresponding to the last frame, i.e. n -1 frame of data after updating the weight. Adding 1 means that the number of matches of the feature corresponding to the weight of the denoising path is increased by 1, which is an important judgment basis for cleaning and updating the local feature library and the local weight library, that is, cleaning the features and corresponding weights that appear too few times. are the subscripts of the features and weights in the feature library and weight library. For the three different matching situations, the situations when writing into the local weight library are different, namely: ; Reference Figure 3 As shown, Figure 3 A flowchart of a method for non-real-time updating of a local feature library and a local weight library provided in an embodiment of the present application. The non-real-time updating method includes: S121: collecting real-time noise data, and storing the real-time noise data and a desired signal of the primary sound field in a temporary storage space; S122: When computing power is idle, extract the noise features of the real-time noise data and calculate the noise reduction path weights corresponding to the real-time noise features to update the local feature library and the weight feature library.

[0081] In some embodiments, the non-real-time updating method uses a Wiener filter or a recursive least squares filter to calculate the noise reduction path weights.

[0082] The process of calculating the noise reduction path weights using the Wiener filter is expressed as: ; in, is the weight of the Wiener filter, is the filter-x signal matrix, i.e. ; in For the The reference signal of each channel passes through the Secondary sound source to the Secondary path of an error microphone The filtered-x signal vector after ; in ; for microphone signal vector, i.e. ; in It is a signal of expectation for recovery; In addition, R is the filtered-x signal The autocorrelation matrix of p is the filtered-x signal With microphone signal The cross-correlation vector between , the autocorrelation matrix R and the cross-correlation vector p are estimated by the sample data.

[0083] The process of calculating the noise reduction path weights using the recursive least squares filter (RLS) is expressed as: ; ; ; ; in, is the gain vector, is the recursive weight of the recursive least squares. What is needed here is the weight of the final iteration. When the calculation is completed, the number of iterations is hour, for: ; In some embodiments, the effect comparison can be measured using the residual sum of squares, and the data length used is denoted as , taking the Wiener filter as an example, use Filtering, residual sum of squares for: ; The denoising path weight in the local weight library is recorded as , and its corresponding residual sum of squares for: ; Introducing adjustable coefficients , indicating that the performance required by the new weights needs to exceed the degree of the noise reduction path weights in the local weight library, so The situation of writing to the local weight library is described as follows: ; The update of the local feature library and the local weight library ensures the robustness of the local feature library and the local weight library. In the real-time update mode, when the real-time noise feature does not match or partially matches the feature in the local feature library, the real-time noise feature and the corresponding real-time noise reduction path weight are stored in the library, which helps to enrich and improve the local feature library and the local weight library, so that the system can find a matching noise reduction solution more quickly when facing a similar noise environment in the future.

[0084] When performing non-real-time weight calculations, the collected noise data is calculated using idle computing power. Performance judgment is required before writing. When the new non-real-time calculated weights are determined to be better than the existing weights in the weight library, they can be replaced. The non-real-time weight calculation module cannot be performed in real time due to the large amount of calculation or the need for statistical information of the data. However, the non-real-time weight calculation module can obtain better noise reduction path weights as the weights of the adaptive filter, which can significantly shorten the convergence time and improve the noise reduction effect.

[0085] Embodiment 3: This embodiment only describes the parts that are different from Embodiment 1 and Embodiment 2, and the same parts are not repeated again.

[0086] Reference Figure 4 As shown, Figure 4 Flowchart of a method for maintaining storage space of a local feature library and a local weight library provided in an embodiment of the present application. The present application provides a method for maintaining storage space of a local feature library and a local weight library, the method comprising: S131: Calculate the weight scores of the noise features and the noise reduction path weights corresponding to the noise features; and rank the weight scores in descending order.

[0087] In some embodiments, the weight score calculation scheme is: ; in, is the cumulative number of matches, is the number of days of existence, where . Because the calculation cycle of features and similarities is fixed, the number of times is sufficient to reflect the duration. Figure 7 As shown, Figure 7 It is a surface diagram showing the relationship between the score, the number of feature matches, and the number of days the feature exists, which are used in the maintenance process of the local feature library and the local weight library provided in the embodiment of the present application.

[0088] S132: Discarding the noise features with a preset proportion of weight scores that are ranked lower and the noise reduction path weights corresponding to the noise features.

[0089] In terms of the storage space maintenance solution for the local feature library and the local weight library, this application provides an efficient and easy-to-implement maintenance strategy to keep the local feature library and the local weight library effective and lightweight. Take the following scenario as an example: 20 reference signals, 4 secondary sound sources, the filter weight is an int32 of 256 points, that is, 4 bytes, and the noise feature vector is an int32 of 2048 length. The number of bytes occupied by each set of noise feature features and the corresponding noise reduction path weights is: 4MB of space can store 11 sets of noise features and noise reduction path weights, calculated as follows: =44 In this embodiment, when the used storage space exceeds 80%, that is, the 35 sets of features and weights of the above scenario, the last 30% sets of noise features and the noise reduction path weights corresponding to the noise features are discarded according to the weight scores of the features, that is, the 13 sets of features and weights of the above scenario. The discarding principle is: for the same number of days, the more matches, the higher the weight score; for the same number of matches, the longer the number of days, the lower the weight score. In some other embodiments, the percentage of storage space and the percentage of discarded noise features and noise reduction path weights can be set according to actual conditions.

[0090] In addition, when the updated denoising path weights need to be written into the local weight library, if the noise feature at this time is a newly added noise feature (that is, the noise feature does not match or partially matches the noise feature in the local feature library), when the storage space is sufficient, new storage space is directly allocated to store the noise feature and the corresponding denoising path weights; when the storage space is insufficient, the noise feature and the corresponding denoising path weights that were ranked last in the local feature library and the local weight library are directly discarded.

[0091] The above maintenance and update scheme of the local feature library and the local weight library records the number of matches and the number of days of existence through a very small amount of additional storage space, and obtains the score of each feature through simple and effective calculation, so that the local feature library and the local weight library are kept in good condition. By reserving margins, it is ensured that there is enough space to use before the next maintenance; in extreme cases, that is, when the reserved space is insufficient, the system directly discards the last-ranked features and scores based on the scores at the time of the last maintenance.

[0092] Embodiment 4: This embodiment only describes the parts that are different from Embodiment 1, Embodiment 2 and Embodiment 3, and the same parts are not repeated again.

[0093] The present application provides a method for iterating and maintaining noise data. The method for collecting noise data has been described in Example 1 and will not be repeated here. In order to store the collected noise data, the present application allocates a temporary storage space in the storage space as a local original database for storing noise data.

[0094] When in use, the real-time noise data and the recovered expected signal are stored in the local original database, waiting for non-real-time weight calculation to optimize the existing weights in the weight library. Since the original noise data occupies a large storage space, the non-real-time weight calculation needs to be performed as quickly as possible. After the non-real-time weight calculation is completed, the data in the database needs to be cleaned up in time.

[0095] Take the following scenario as an example: 20 reference signals, 4 error microphones, 32-bit PCM data format, and storage at a 4kHz sampling rate. The number of bytes of temporary storage space required to store 1 minute of noise data is: 4GB of space can store 173 minutes of noise data, which is enough for the non-real-time weight calculation module. =173 Based on this, the non-real-time weight calculation and original data cleaning strategy is: when a certain noise data exceeds 5 minutes, the non-real-time weight calculation module can be started, and the noise data will be cleaned up after the calculation is completed. The 173-minute cache is sufficient for use.

[0096] In this embodiment, the desired signal recovery technology refers to subtracting the component of the excitation signal after passing through the secondary path from the error microphone signal, and restoring the residual signal of the microphone signal to the desired signal of the primary sound field, specifically: ; in, is the number of secondary sound sources, is the secondary sound source index, ; is the number of error microphones, is the error microphone index, ; For the The signal of an error microphone; To deliver to The excitation signal of the secondary sound source is For offline identification Secondary sound source to the The impulse response of an error microphone, for Length; For the recovery The desired signal of the primary sound field at the error microphone.

[0097] By adopting the expected signal recovery technology, the storage of original data and the real-time noise reduction system can be compatible and do not affect each other. If there is no expected signal recovery technology, the primary sound field will be changed when the real-time noise reduction system works normally. The signal received by the error microphone is the residual signal of the noise reduction system at this time, which cannot be used for the non-real-time calculation weight module. The real-time noise reduction system needs to be turned off to collect and save original data. After the introduction of the expected signal recovery technology, the original data can be saved while the real-time noise reduction system is working, which improves the operation efficiency and feasibility of the system.

[0098] Embodiment 5: On the other hand, the present application provides an active vehicle noise control system, which adopts the active vehicle noise control method provided in the above-mentioned embodiment 1, embodiment 2, embodiment 3 and embodiment 4.

[0099] The automobile active control system includes an acquisition module and a control module, wherein the acquisition module is used to collect noise data of the automobile, and the noise data can be used in the process of establishing a local feature library, and can also be used in the process of updating the local feature library, and is used to collect real-time noise data during the operation of the automobile. The control module of the system includes a weight calculation module, which obtains noise data and calculates the weight of the noise reduction path, and is used to build or update the local weight library.

[0100] The weight calculation module includes a real-time weight calculation module and / or a non-real-time weight calculation module, which is used to calculate the noise reduction path weight according to the noise data collected by the acquisition module. Here, "and / or" means that the composition of the weight module includes three situations, namely: the first situation is that only the real-time weight calculation module is included. In this case, the construction and update of the local weight library in the system are carried out in real time. At this time, the real-time weight calculation module can use an adaptive filter. The second situation is that only the non-real-time weight calculation module is included. In this case, the local weight library in the system cannot be updated in real time. The local weight library adopts a non-real-time update method, that is, the noise data collected by the acquisition module is stored and updated when the computing power is idle. At this time, the non-real-time weight calculation module can use a Wiener filter or a recursive least squares filter. The third situation is that the weight calculation module includes both a real-time calculation module and a non-real-time calculation module. At this time, the system also includes an expected signal recovery module for restoring the residual signal of the error microphone to the expected signal of the primary sound field.

[0101] If there is no expected signal recovery module, the real-time weight calculation module will change the primary sound field when working normally, thereby affecting the normal operation of the non-real-time calculation module. After setting up the expected signal recovery module, the noise data can be saved while the noise reduction system is working. When the computing power is idle, the non-real-time weight calculation module is used to complete the optimization of the local weight library, thereby improving the operating efficiency and feasibility of the system.

[0102] The automotive noise control system can flexibly respond to different noise processing requirements by integrating real-time weight calculation modules and non-real-time weight calculation modules (or using one of them alone). Real-time weight calculation ensures that the system can respond quickly to rapidly changing noise environments and provide immediate noise reduction effects; while non-real-time weight calculation allows the system to perform more in-depth noise feature analysis and weight optimization when computing power is idle, improving the accuracy and stability of noise reduction processing.

[0103] On the other hand, the present application provides a storage medium storing a vehicle noise active control program. When the vehicle noise active control program is executed by a processor, the above-mentioned vehicle noise active control method is implemented.

[0104] It is easy to understand that those skilled in the art can combine, split, reorganize, etc. the embodiments of the present application to obtain other embodiments based on the several embodiments provided in the present application, and these embodiments do not exceed the protection scope of the present application.

[0105] The above specific implementation methods further explain in detail the purpose, technical solutions and beneficial effects of the embodiments of the present application. It should be understood that the above are only specific implementation methods of the embodiments of the present application and are not used to limit the protection scope of the embodiments of the present application. Any modifications, equivalent substitutions, improvements, etc. made on the basis of the technical solutions of the embodiments of the present application should be included in the protection scope of the embodiments of the present application.

Claims

1. A method for active automobile noise control, characterized in that: The method comprises: Collect noise data, extract noise features and store them to build a local feature library that reflects driving habits; Calculating noise reduction path weights according to the noise data to establish a local weight library, wherein the noise reduction path weights in the local weight library correspond one to one to the noise features in the local feature library; Collecting real-time noise data, extracting real-time noise features, calculating the similarity between the real-time noise features and each noise feature in the local feature library, and rewriting the weight of the adaptive filter according to the similarity to perform active control of noisy vehicle sound; The weight of the adaptive filter is obtained through the noise reduction path weight in the local weight library.

2. The active vehicle noise control method according to claim 1, characterized in that: When the similarity is between the first activation function threshold and the second activation function threshold, the real-time noise feature partially matches the noise feature in the local feature library, and the weight of the adaptive filter is the weighted sum of the noise reduction path weights corresponding to the partially matched noise features, wherein the weight coefficient comes from the activation function; When the similarity is greater than a second activation function threshold, the real-time noise feature fully matches the noise feature in the local feature library, and the weight of the adaptive filter is the noise reduction path weight corresponding to the fully matched noise feature.

3. The active vehicle noise control method according to claim 2, characterized in that: The activation function is: ； in, , That is the first activation function threshold, is the second activation function threshold, For similarity.

4. The active vehicle noise control method according to claim 1, characterized in that: The adaptive filter adopts a cyclic leakage minimum mean square error filter; the function of the cyclic leakage minimum mean square error filter weight update process is expressed as: ； in, is the identity matrix, is the adjustable step size factor; Indicates the number of weight updates; For the The weight vector after the update is for right N Take the remainder; is the weight vector The item; is the filter-x signal matrix; is the vector of error microphone signals, The independent variable is and The leakage function of is a The square, The first The column elements are 1 and the rest are 0.

5. The method for active automobile noise control according to claim 1, characterized in that: The method also includes a real-time update method for a local feature library and a local weight library, and the real-time update method includes: When the real-time noise feature does not match or partially matches each noise feature in the local feature library, storing the real-time noise feature in the local feature library to update the local feature library; Based on the real-time noise data, the adaptive filter performs weight iteration to obtain the final weight value. When the real-time noise feature fully matches the noise feature in the local feature library, the real-time noise reduction path weight replaces the noise reduction path weight corresponding to the fully matching noise feature. When the real-time noise feature partially matches or does not match the noise feature in the local feature library, new storage space is allocated to store the real-time noise reduction path weight.

6. The active vehicle noise control method according to claim 5, characterized in that: The iterative update method includes a maintenance method for the storage space, and the maintenance method includes: Calculating the noise feature and the weight score of the noise reduction path weight corresponding to the noise feature; Ranking the weight scores in descending order; A preset proportion of the noise features ranked lower in weight scores and the noise reduction path weights corresponding to the noise features are discarded; wherein, when allocating new storage space for storing the real-time noise features and the real-time noise reduction path weights, if the storage space capacity is insufficient, the noise features ranked lower in weight scores and the noise reduction path weights corresponding to the noise features are discarded; The weight score is calculated as follows: ； in, is the cumulative number of matches, The number of days it exists.

7. The active vehicle noise control method according to claim 1, characterized in that: The method further includes a non-real-time updating method for a local feature library and a local weight library, wherein the non-real-time updating method includes: Collecting real-time noise data, and storing the real-time noise data and a desired signal of the primary sound field in a temporary storage space; When computing power is idle, the noise features of the real-time noise data are extracted and the noise reduction path weights corresponding to the real-time noise features are calculated to update the local feature library and the weight feature library.

8. The method for active automobile noise control according to claim 7, characterized in that: The desired signal of the primary sound field is obtained according to the residual signal of the error microphone, and the function expression of the desired signal of the primary sound field is: ； in, is the number of secondary sound sources, is the secondary sound source index, ; is the number of error microphones, is the error microphone index, ; For the The residual signal of the error microphone; To deliver to The excitation signal of the secondary sound source is For offline identification Secondary sound source to the The impulse response of an error microphone, for Length; For the recovery The desired signal of the primary sound field at the error microphone.

9. An active automobile noise control system, characterized in that: The system comprises: A collection module is used to collect real-time noise data of the car; A control module, wherein the control module comprises a weight calculation module, wherein the weight calculation module obtains the real-time noise data and calculates the real-time noise reduction path weight, and updates a local weight library.

10. The active vehicle noise control system according to claim 9, characterized in that: The weight calculation module includes a real-time weight calculation module and / or a non-real-time weight calculation module, which is used to calculate the noise reduction path weight according to the noise data; wherein the real-time calculation module adopts an adaptive filter, and the non-real-time calculation module adopts a Wiener filter or a recursive least squares filter.

11. A storage medium, characterized in that: The storage medium stores an active vehicle noise control program. When the active vehicle noise control program is executed by a processor, the active vehicle noise control method according to any one of claims 1 to 8 is implemented.

Images